JADT’ 18
PROCEEDINGS OF THE
14TH INTERNATIONAL CONFERENCE
ON STATISTICAL ANALYSIS OF TEXTUAL DATA

JADT’ 18
PROCEEDINGS OF THE
14TH INTERNATIONAL CONFERENCE
ON STATISTICAL ANALYSIS OF TEXTUAL DATA

(Rome, 12-15 June 2018)

Vol. I

UniversItalia
2018

PROPRIETÀ LETTERARIA RISERVATA
Copyright 2018 - UniversItalia - Roma
ISBN 978-88-3293-137-2
A norma della legge sul diritto d’autore e del codice civile è vietata la riproduzione di
questo libro o di parte di esso con qualsiasi mezzo, elettronico, meccanico, per mezzo
di fotocopie, microfilm, registra-tori o altro. Le fotocopie per uso personale del lettore
possono tuttavia essere effettuate, ma solo nei limiti del 15% del volume e dietro
pagamento alla SIAE del compenso previsto dall’art. 68, commi 4 e 5 della legge 22
aprile 1941 n. 633. Ogni riproduzione per finalità diverse da quelle per uso personale
deve essere autorizzata specificatamente dagli autori o dall’editore.

Program Committee
Ramón Álvarez Esteban: Univ. of León, E
Valérie Beaudouin: Telecom ParisTech, F
Mónica Bécue: Poly. Univ. of Catalunya, E
Sergio Bolasco: Sapienza Univ. of Rome, I
Isabella Chiari: Sapienza Univ. of Rome, I
François Daoust, UQÀM, Montreal, CDN
Anne Dister, FUSL, Bruxelles / UCL, Louvain, B
Jules Duchastel: UQÀM, Montreal, CDN
Serge Fleury: Univ. Paris 3, F
Cédrick Fairon: UCL, Louvain, B
Luca Giuliano: Sapienza Univ. of Rome, I
Serge Heiden, ENS, Lyon, F
Domenica Fioredistella Iezzi, Univ. of Tor Vergata, I
Margareta Kastberg, Univ. of Franche Comté, F
Ludovic Lebart: CNRS / ENST, Paris, F
Jean-Marc Leblanc: Univ. of Créteil, F

Alain Lelu: Univ. of Franche Comté, F
Dominique Longrée, Univ. of Liège, B
Véronique Magri: Univ. of Nice Sophia-Antipolis, F
Pascal Marchand: Univ. of Toulouse, F
William Martinez: Univ. of Lisboa, P
Damon Mayaffre: CNRS, Nice, F
Sylvie Mellet: CNRS, Nice, F
Michelangelo Misuraca: Univ. of Calabria, I
Denis Monière: Univ. of Montréal, CDN
Bénédicte Pincemin: CNRS, Lyon, F
Céline Poudat: Univ. of Nice Sophia-Antipolis, F
Pierre Retinaud: Univ. of Tolouse, F
André Salem: Univ. Paris 3, F
Monique Slodzian: Inalco, F
Arjuna Tuzzi: Univ. of Padua, I
Mathieu Valette: Inalco, F

Organising Committee
Domenica Fioredistella Iezzi: Univ. of Tor Vergata, I
Sergio Bolasco: Sapienza Univ. of Rome, I
Livia Celardo: Sapienza Univ. of Rome, I
Isabella Chiari: Sapienza Univ. of Rome, I
Francesca della Ratta: ISTAT, I
Fiorenza Deriu: Sapienza Univ. of Rome, I
Francesca Dolcetti: Sapienza Univ. of Rome, I

Andrea Fronzetti Colladon: Univ. of Tor Vergata, I
Francesca Greco: Sapienza Univ. of Rome, I
Isabella Mingo: Sapienza Univ. of Rome, I
Michelangelo Misuraca: Univ. of Calabria, I
Arjuna Tuzzi: Univ. of Padua, I
Maurizio Vichi: Sapienza Univ. of Rome, I
Francesco Zarelli: ISTAT, I

Local Organisation
Francesco Alò, Giulia Giacco,
Paolo Meoli, Vittorio Palermo, Viola Talucci

Table of contents

Introduction ............................................................................................................... XVII
Acknowledgements ....................................................................................................XIX
Invited Speakers
GERMAN KRUSZEWSKI
Memorize or generalize? Searching for a compositional RNN in a haystack
Adam Liška ......................................................................................................... XXIII
BING LIU
Scaling-up Sentiment Analysis through Continuous Learning .................. XXIV
PASCAL MARCHAND
La textométrie comme outil d’expertise :
application à la négociation de crise. ................................................................ XXV
GEORGE K. MIKROS
Author Identification Combining Various Author Profiles. Towards a Blended
Authorship Attribution Methodology ............................................................. XXVI
ROBERTO NAVIGLI
From text to concepts and back: going multilingual
with BabelNet in a step or two ....................................................................... XXVII
Contributors
MOTASEM ALRAHABI1, CHIARA MAINARDI1
Identification automatique de l’ironie et des formes apparentées dans un
corpus de controverses théâtrales ........................................................................... 1
MOHAMMAD ALSADHAN, SASCHA DIWERSY,
AGATA JACKIEWICZ, GIANCARLO LUXARDO
Migrants et réfugiés : dynamique de la nomination de l'étranger ................... 10
R. ALVAREZ-ESTEBAN, M. BÉCUE-BERTAUT, B. KOSTOV, F. HUSSON, J-A
SÁNCHEZ-ESPIGARES
Xplortext, a R package. Multidimensional statistics for textual data science . 19
ELENA, AMBROSETTI, ELEONORA MUSSINO, VALENTINA TALUCCI
L'evoluzione delle norme: analisi testuale delle politiche sull'immigrazione in
Italia ........................................................................................................................... 26

VIII

JADT’ 18

MASSIMO ARIA, CORRADO CUCCURULLO
A bibliometric meta-review of performance measurement, appraisal,
management research ............................................................................................. 35
LAURA ASCONE
Textual Analysis of Extremist Propaganda and Counter-Narrative: a quantiquali investigation ................................................................................................... 44
LAURA ASCONE, LUCIE GIANOLA
Analyse de données textuelles appliquée à des problématiques de sécurité et
d'enquête judiciaire ................................................................................................. 52
SIMONA BALBI, MICHELANGELO MISURACA, MARIA SPANO
A two-step strategy for improving categorisation of short texts ..................... 60
CHRISTINE BARATS, ANNE DISTER, PHILIPPE GAMBETTE, JEAN-MARC
LEBLANC, MARIE PERES
Appeler à signer une pétition en ligne : caractéristiques linguistiques des
appels ........................................................................................................................ 68
MANUEL BARBERA, CARLA MARELLO
Newsgroup e lessicografia: dai NUNC al VoDIM .............................................. 76
IGNAZIA BARTHOLINI
Techniques for detecting the normalized violence in the perception of refugee
/ asylum seekers between lexical analysis and factorial analysis...................... 83
PATRIZIA BERTINI MALGARINI, MARCO BIFFI, UGO VIGNUZZI
Dal corpus al dizionario: prime riflessioni lessicografiche sul Vocabolario
storico della cucina italiana postunitaria (VoSCIP) ............................................ 90
MARCO BIFFI
Strumenti informatico-linguistici per la realizzazione di un dizionario
dell’italiano postunitario ........................................................................................ 99
ANNICK FARINA, RICCARDO BILLERO
Comparaison de corpus de langue « naturelle » et de langue « de traduction »
: les bases de données textuelles LBC, un outil essentiel pour la création de
fiches lexicographiques bilingues........................................................................ 108
FELICE BISOGNI, STEFANO PIRROTTA
Il rapporto tra famiglie di anziani non autosufficienti e servizi territoriali:
un'analisi dei dati esploratoria con l'Analisi Emozionale del Testo (AET) .... 117
ANTONELLA BITETTO, LUIGI BOLLANI
Esperienza di analisi testuale di documentazione clinica e di flussi informativi
sanitari, di utilità nella ricerca epidemiologica e per indagare la qualità
dell'assistenza......................................................................................................... 126
GUIDO BONINO, DAVIDE PULIZZOTTO, PAOLO TRIPODI
Exploring the history of American philosophy in a computer-assisted
framework .............................................................................................................. 134

JADT’ 18

IX

MARC-ANDRE BOUCHARD, SYLVIA KASPARIAN
La classification hiérarchique descendante pour l’analyse des représentations
sociales dans une pétition antibilinguisme au Nouveau-Brunswick,
Canada .................................................................................................................... 142
LIVIA CELARDO, RITA VALLEROTONDA, DANIELE DE SANTIS,CLAUDIO
SCARICI, ANTONIO LEVA
Analysing occupational safety culture through mass media monitoring..... 150
BARBARA CORDELLA, FRANCESCA GRECO, PAOLO MEOLI,VITTORIO
PALERMO, MASSIMO GRASSO
Is the educational culture in Italian Universities effective? A case study ...... 157
MICHELE A. CORTELAZZO, GEORGE K. MIKROS, ARJUNA TUZZI
Profiling Elena Ferrante: a Look Beyond Novels .............................................. 165
FABRIZIO DE FAUSTI, MASSIMO DE CUBELLIS, DIEGO ZARDETTO1
Word Embeddings: a Powerful Tool for Innovative Statistics at Istat .......... 174
Gibbons A. (1985). Algorithmic Graph Theory. Cambridge University Press. . 182
VIVIANA DE GIORGI, CHIARA GNESI
Analisi di dati d’impresa disponibili online: un esempio di data science tratto
dalla realtà economica dei siti di e-commerce ................................................... 183
ALESSANDRO CAPEZZUOLI, FRANCESCA DELLA RATTA,
STEFANIA MACCHIA,MANUELA MURGIA, MONICA SCANNAPIECO,
DIEGO ZARDETTO
The use of textual sources in Istat: an overview ................................................ 192
FRANCESCA DELLA RATTA, GABRIELLA FAZZI, MARIA ELENA
PONTECORVO, CARLO VACCARI, ANTONINO VIRGILLITO
Twitter e la statistica ufficiale: il dibattito sul mercato del lavoro ................. 200
SAMI DIAF
Gauging An Author’s Mood Using Hidden Markov Chains ......................... 209
MARC DOUGUET
Les hémistiches répétés ........................................................................................ 215
FRANCESCA DRAGOTTO, SONIA MELCHIORRE
«Mangiata dall’orco e tradita dalle donne». Vecchi e nuovi media raccontano
la vicenda di Asia Argento, tra storytelling e Speech Hate ............................. 223
CRISTIANO FELACO, ANNA PAROLA
Il cosa e il come del processo narrativo. L’uso combinato della Text Analysis e
Network Text Analysis al servizio della precarietà lavorativa ....................... 233
ANA NORA FELDMAN
Hablando de crisis: las comunicaciones del Fondo Monetario Internacional 242
VALERIA FIASCO
Brexit in the Italian and the British press:
a bilingual corpus-driven analysis ...................................................................... 250
VIVIANA FINI, GIUSEPPE LUCIO GAETA, SERGIO SALVATORE
Textual analysis to promote innovation within public policy evaluation .... 259

X

JADT’ 18

ALESSIA FORCINITI, SIMONA BALBI
A proposal for Cross-Language Analysis:
violence against women and the Web ................................................................ 268
BEATRICE FRACCHIOLLA, OLINKA SOLENE DE ROGER
La verbalisation des émotions ............................................................................. 276
LUISA FRANCHINA, FRANCESCA GRECO, ANDREA LUCARIELLO,
ANGELO SOCAL, LAURA TEODONNO
Improving Collection Process for Social Media Intelligence: A Case Study . 285
ANDREA FRONZETTI COLLADON, JOHANNE SAINT-CHARLES, PIERRE
MONGEAU
The impact of language homophily and similarity of social position on
employees’ digital communication ..................................................................... 293
MATTEO GERLI
Looking Through the Lens of Social Sciences: The European Union in the EUFunded Research Projects Reporting .................................................................. 300
LUCIE GIANOLA, MATHIEU VALETTE
Spécialisation générique et discursive d’une unité lexical L’exemple de
joggeuse dans la presse quotidienne régionale ................................................... 312
PETER A. GLOOR, JOAO MARCOS DE OLIVEIRA, DETLEF SCHODER
The Transparency Engine – A Better Way to Deal with Fake News .............. 319
FRANCESCA GRECO, LEONARDO ALAIMO, LIVIA CELARDO
Brexit and Twitter: The voice of people.............................................................. 327
FRANCESCA GRECO, GIULIO DE FELICE, OMAR GELO
A text mining on clinical transcripts of good and poor outcome
psychotherapies ..................................................................................................... 335
FRANCESCA GRECO, DARIO MASCHIETTI, ALESSANDRO POLLI
DOMINIO: A Modular and Scalable Tool for the Open Source Intelligence 343
LEONIE GRÖN, ANN BERTELS, KRIS HEYLEN
Is training worth the trouble? A PoS tagging experiment with Dutch clinical
records..................................................................................................................... 351
FRANCE GUERIN-PACE, ELODIE BARIL
Les outils de la statistique textuelle pour analyser
les corpus de données d’enquêtes de la statistique publique .......................... 359
SERGE HEIDEN
Annotation-based Digital Text Corpora Analysis within the TXM Platform 367
DANIEL HENKEL
Quantifying Translation : an analysis of the conditional perfect in EnglishFrench comparable-parallel corpus..................................................................... 375
DANIEL DEVATMAN HROMADA
Extraction of lexical repetitive expressions from complete works of William
Shakespeare ............................................................................................................ 384

JADT’ 18

XI

OLIVIER KRAIF, JULIE SORBA
Spécificités des expressions spatiales et temporelles dans quatre sous-genres
romanesques (policier, science-fiction, historique et littérature générale) .... 392
CYRIL LABBE, DOMINIQUE LABBE
Les phrases de Marcel Proust .............................................................................. 400
LUDOVICA LANINI, MARÍA CARLOTA NICOLÁS MARTÍNEZ
Verso un dizionario corpus-based del lessico dei beni culturali: procedure di
estrazione del lemmario ....................................................................................... 411
DANIELA LARICCHIUTA, FRANCESCA GRECO, FABRIZIO PIRAS, BARBARA
CORDELLA, DEBORA CUTULI, ELEONORA PICERNI, FRANCESCA
ASSOGNA, CARLO LAI, GIANFRANCO SPALLETTA, LAURA PETROSINI
“The grief that doesn’t speak”: Text Mining and Brain Structure 419
GEVISA LA ROCCA, CIRUS RINALDI
Icone gay: tra processi di normalizzazione e di resistenza. Ricostruire la
semantica degli hashtag........................................................................................ 428
LUDOVIC LEBART
Looking for topics: a brief review......................................................................... 436
GAËL LEJEUNE, LICHAO ZHU
Analyse Diachronique de Corpus : le cas du poker .......................................... 444
JULIEN LONGHI, ANDRE SALEM
Approche textométrique des variations du sens ............................................... 452
LAURENT VANNI1, DAMON MAYAFFRE, DOMINIQUE LONGREE
ADT et deep learning, regards croisés. Phrases-clefs, motifs et nouveaux
observables ............................................................................................................. 459
LUCIE LOUBERE
Déconstruction et reconstruction de corpus... À la recherche de la pertinence
et du contexte ......................................................................................................... 467
HEBA METWALLY
L’apport du corpus-maquette à la mise en évidence des niveaux descriptifs de
la chronologie du sens. Essai sur une Série Textuelle Chronologique du Monde
diplomatique (1990-2008). ....................................................................................... 474
JUN MIAO, ANDRE SALEM
Séries textuelles homogènes................................................................................. 491
SILVIO MIGLIORI, ANDREA QUINTILIANI, DANIELA ALDERUCCIO,
FIORENZO AMBROSINO, ANTONIO COLAVINCENZO, MARIALUISA
MONGELLI, SAMUELE PIERATTINI, GIOVANNI PONTI SERGIO BOLASCO,
FRANCESCO BAIOCCHI, GIOVANNI DE GASPERIS
TaLTaC in ENEAGRID Infrastructure................................................................ 501
ISABELLA MINGO, MARIELLA NOCENZI
The dimensions of Gender in the International Review of Sociology. A
lexicometric approach to the analysis of the publications in the last twenty
years ........................................................................................................................ 509

XII

JADT’ 18

ADIEL MITTMANN, ALCKMAR LUIZ DOS SANTOS
The Rhythm of Epic Verse in Portuguese From the 16th to the 21st Century514
DENIS MONIERE, DOMINIQUE LABBE
Le vocabulaire des campagnes électorales ......................................................... 522
CYRIELLE MONTRICHARD
Faire émerger les traces d’une pratique imitative dans la presse de tranchées à
l’aide des outils textométriques ........................................................................... 532
ALBERT MORALES MORENO
Evolución diacrónica de la terminología y la fraseología jurídicoadministrativa en los Estatutos de autonomía de Catalunya de 1932, 1979 y
2006 .......................................................................................................................... 541
CEDRIC MOREAU
Comment penser la recherche d’un signe pour une plateforme multilingue et
multimodale français écrit / langue des signes française ? .............................. 556
JEAN MOSCAROLA, BORIS MOSCAROLA
Conclusion ADT et visualisation, pour une nouvelle lecture des corpus Les
débats de 2ème tour des Présidentielles (1974-2017) ........................................ 563
MAURIZIO NALDI
A conversation analysis of interactions in personal finance forums ............. 571
STEFANO NOBILE
Analisi testuale, rumore semantico e peculiarità morfosintattiche:
problemi e strategie di pretrattamento di corpora speciali.............................. 578
DANIEL PELISSIER
L’individu dans le(s) groupe(s) : focus group et partitionnement
du corpus ................................................................................................................ 586
BENEDICTE PINCEMIN, CELINE GUILLOT-BARBANCE, ALEXEI
LAVRENTIEV
Using the First Axis of a Correspondence Analysis as an Analytical Tool.
Application to Establish and Define an Orality Gradient for Genres of
Medieval French Texts .......................................................................................... 594
CELINE POUDAT
Explorer les désaccords dans les fils de discussion du Wikipédia francophone
.................................................................................................................................. 602
MATTHIEU QUIGNARD, SERGE HEIDEN, FREDERIC LANDRAGIN,
MATTHIEU DECORDE
Textometric Exploitation of Coreference-annotated Corpora with TXM:
Methodological Choices and First Outcomes .................................................... 610
PIERRE RATINAUD
Amélioration de la précision et de la vitesse de l’algorithme de classification
de la méthode Reinert dans IRaMuTeQ ............................................................. 616

JADT’ 18

XIII

LUISA REVELLI
Il parametro della frequenza tra paradossi e antinomie:
il caso dell’italiano scolastico .................................................................................. 626
PIERGIORGIO RICCI
How Twitter emotional sentiments mirror on the Bitcoin
transaction network .............................................................................................. 635
CHANTAL RICHARD, SYLVIA KASPARIAN
Analyse de contenu versus méthode Reinert : l’analyse comparée d’un corpus
bilingue de discours acadiens et loyalistes du N.-B., Canada ......................... 643
VALENTINA RIZZOLI, ARJUNA TUZZI
Bridge over the ocean: Histories of social psychology in Europe and North
America. An analysis of chronological corpora ................................................ 651
LOUIS ROMPRE, ISMAÏL BISKRI
Les « itemsets fréquents » comme descripteurs de documents textuels ....... 659
CORINNE ROSSARI, LJILJANA DOLAMIC, ANNALENA HÜTSCH, CLAUDIA
RICCI, DENNIS WANDEL
Discursive Functions of French Epistemic Adverbs: What can Correspondence
Analysis tell us about Genre and Diachronic Variation? ................................. 668
VANESSA RUSSO, MARA MARETTI, LARA FONTANELLA, ALICE
TONTODIMAMMA
Misleading information in online propaganda networks ............................... 676
ELIANA SANANDRES, CAMILO MADARIAGA, RAIMUNDO ABELLO
Topic modeling of Twitter conversations .......................................................... 684
FRANCESCO SANTELLI, GIANCARLO RAGOZINI, MARCO MUSELLA
What volunteers do? A textual analysis of voluntary activities in the Italian
context ..................................................................................................................... 692
S. SANTILLI, S. SBALCHIERO, L. NOTA, S. SORESI
A longitudinal textual analysis of abstract presented at Italian Association for
Vocational guidance and Career Counseling’
Conferences from 2002 to 2017 ............................................................................ 700
JACQUES SAVOY
A la poursuite d’Elena Ferrante........................................................................... 707
JACQUES SAVOY
Regroupement d’auteurs dans la littérature du XIXe siècle ........................... 716
STEFANO SBALCHIERO, ARJUNA TUZZI
What’s Old and New? Discovering Topics in the American Journal of
Sociology................................................................................................................. 724
NILS SCHAETTI, JACQUES SAVOY
Comparison of Neural Models for Gender Profiling ........................................ 733
LIONEL SHEN
Segments répétés appliqués à l'extraction de connaissances trilingues ......... 740

XIV

JADT’ 18

SANDRO STANCAMPIANO
Misurare, Monitorare e Governare le città con i Big Data .............................. 748
FADILA TALEB, MARYVONNE HOLZEM
Exploration textométrique d’un corpus de motifs juridiques dans le droit
international des transports ................................................................................. 755
JAMES M. TEASDALE
The Framing of the Migrant: Re-imagining a Fractured Methodology in the
Context of the British Media. ............................................................................... 763
MARJORIE TENDERO1, CECILE BAZART
Results from two complementary textual analysis software (Iramuteq and
Tropes) to analyze social representation of contaminated brownfields ........ 771
MATTEO TESTI, ANDREA MERCURI, FRANCESCO PUGLIESE
Multilingual Sentiment Analysis ......................................................................... 780
JUAN MARTÍNEZ TORVISCO
A linguistic analysis of the image of immigrants’ gender in Spanish
newspapers............................................................................................................. 788
FRANCESCO URZÌ
Lo strano caso delle frequenze zero nei testi legislativi euroistituzionali...... 796
SYLVIE VANDAELE
Les traductions françaises de The Origin of Species : pistes lexicométriques . 805
PIERRE WAVRESKY, MATTHIEU DUBOYS DE LABARRE, JEAN-LOUP
LECOEUR
Circuits courts en agriculture : utilisation de la textométrie dans le traitement
d’une enquête sur 2 marchés ............................................................................... 814
MARIA ZIMINA, NICOLAS BALLIER
On the phraseology of spoken French: initial salience, prominence and
lexicogrammatical recurrence in a prosodic-syntactic treebank Rhapsodie .... 822

Abstracts
FILIPPO CHIARELLO, GUALTIERO FANTONI, ANDREA BONACCORSI,
SILVIA FARERI
What kind of contributions does research provides? Mapping issue based
statements in research abstracts .......................................................................... 833
FILIPPO CHIARELLO, GIACOMO OSSOLA, GUALTIERO FANTONI,ANDREA
BONACCORSI, ANDREA CIMINO, FELICE DELL’ORLETTA
Technical sentiment analysis: predicting the success of new products using
social media ............................................................................................................ 835

JADT’ 18

XV

FIORENZA DERIU, DOMENICA FIOREDISTELLA IEZZI
Citizens and neighbourhood life: mapping population sentiment in Italian
cities ......................................................................................................................... 837
FRANCESCA DI CARLO, ROSY INNARELLA, BRIZIO LEONARDO TOMMASI
Vax network: profiling influential nodes with social network analysis on
twitter ...................................................................................................................... 838
DAVIDE DONNA
Alteryx .................................................................................................................... 840
VALERIO FICCADENTI, ROY CERQUETI, MARCEL AUSLOOS
Complexity of US President Speeches ................................................................ 841
PETER A. GLOOR
Measuring the Dynamics of Social Networks with Condor ........................... 842
IOLANDA MAGGIO, DOMENICA FIOREDISTELLA IEZZI, MATTEO
FATIGHENTI
“BIG DATA” Words Trend Analysis using the multidimensional analysis of
texts ......................................................................................................................... 844
MARIO MASTRANGELO
Itinerari turistici, network analysis e text mining ............................................. 845
MARIA FRANCESCA ROMANO, GUIDO REY, ANTONELLA BALDASSARINI
PASQUALE PAVONE
Text Mining per l’analisi qualitativa e quantitativa dei dati amministrativi
utilizzati dalla Pubblica Amministrazione......................................................... 847
ALESSANDRO CESARE ROSA
Taglio cesareo e Vbac in Italia al tempo dei Big Data: una proposta di ulteriore
contributo informativo.......................................................................................... 849

Introduction

The International Conference on the Statistical Analysis of Textual Data (JADT,
Journées d’Analyse Statistique des Données Textuelles) has been at its 14th
edition. It was held for the third time in Rome, from 12 to 15 June 2018,
organized by the DII - Department of Enterprise Engineering “Mario
Lucertini” at Tor Vergata University of Rome and the DSS - Department of
Statistical Sciences at Sapienza University of Rome. This biennial conference
has continuously gained importance since its first occurrence in Barcelone
(1992), and with the editions of Montpellier (1994), Rome (1996), Nice (1998),
Lausanne (2000), Saint-Malo (2002), Louvain-la Neuve (2004), Besançon
(2006), Lyon (2008), Rome (2010), Liège (2012), Paris (2014), Nice (2016).
Every two years, the JADT conference presented the state of the art
concerning theories, problems, methods, algorithms, software and
applications in several domains, sharing a quantitative approach to the study
of lexical, textual, pragmatic or discursive features of information expressed
in natural language.
The proceedings of the 2018 Conference collect 113 contributions by 243
scholars from 15 countries spread all over the world. These papers include
contributions open to all scholars and researchers working in the field of
textual data analysis, ranging from lexicography to the analysis of political
discourse, from information retrieval to marketing research, from
computational linguistics to sociolinguistics, from text mining to content
analysis. The invited speakers focused on the central topics of the conference,
discussing open and new themes, e.g. machine learning algorithms to
profiling users of social media, new multilingual approaches, textometry,
and authorship. The proceedings follow an alphabetical order by the
surname of the first author of the contributions.
In this edition, several innovations have been introduced with respect to the
past. In a roundtable, we discussed the past, present and future of Statistical
Analysis of Textual Data and Text Mining methods, by examining the point
of view of Universities and enterprises. The papers, which followed a review
process carried out with two and sometimes three reviewers, are maximum
of 6 pages. The idea is that the papers were not yet in their final version, and
the exchange with other scholars during the conference led to an
improvement. For the first time, a selection of extended papers presented at

XVIII

JADT’ 18

the JADT conference will be published, after another reviewing process, in a
book published by Springer and in several special issues of acknowledged
Journals (Advances in Data Analysis and Classification, International Review
of Sociology, Italian Journal of Applied Statistics, Social Indicators Research,
RPC Rivista di Psicologia Clinica). The perspective of enhancing the papers
discussed during JADT conference will allow the scholar community to keep
the network of active contacts and lively exchanges.

D. Fioredistella Iezzi, Livia Celardo, Michelangelo Misuraca

Acknowledgements

We express our gratitude to the 56 reviewers who offered their assistance in
selecting and anonymously reviewing the papers of this volume: Massimo
Aria, Barbara Baldazzi, Nadia Battisti, Valérie Beaudouin, Sergio Bolasco,
Etienne Brunet, Mónica Bécue, Isabella Chiari, Livia Celardo, Michele
Cortelazzo, Pasquale Del Vecchio, Francesca Della Ratta, Fiorenza Deriu,
Anne Dister, Francesca Dolcetti, Annick Farina, Serge Fleury, Andrea
Fronzetti, Luca Giuliano, Peter Gloor, Francesca Greco, Francesca Grippa,
Serge Heiden, D. Fioredistella Iezzi, Antonio Iovanella, Sylvia Kasparian,
Margareta Kastberg, Dominique Labbé, Ludovica Lanini, Alexei Lavrentev,
Ludovic Lebart, Jean-Marc Leblanc, Alain Lelu, Dominique Longrée,
Véronique Magri, Pascal Marchand, Damon Mayaffre, Sylvie Mellet, Silvia
Micheli, Michelangelo Misuraca, Denis Monière, Gianluca Murgia, Pa-squale
Pavone, Bénédicte Pincemin, Céline Poudat, Pierre Ratinaud, Piergiorgio
Ricci, Maria Francesca Romano, Johanne Saint-Charles, André Salem,
Massimiliano Schiraldi, Max Silberztein, Maria Spano, Arjuna Tuzzi, Mathieu
Valette, Ramón Álvarez Esteban.
JADT2018 was held under the patronage of ISTAT (Istituto Nazionale di
Statistica - National Institute of Statistics). We are also very grateful to the
following sponsors: ISTAT, Le Sphinx, The Information Lab, Master in Data
Science at Tor Vergata University, Prisma.
As regards the organisation of the conference, we would like to thank all the
members of the local organising team: Francesco Alò, Silvia Castellan, Giulia
Giacco, Paolo Meoli, Vittorio Palermo, Viola Talucci.
Special thanks go to Livia Celardo, Isabella Chiari, Andrea Fronzetti
Colladon, Francesca Della Ratta, Fiorenza Deriu, Francesca Dolcetti,
Francesca Greco, for the organisation of special tracks concerning Official
Statistics, Linguistics, Applications on social and psychological domains,
Social Network and Semantic Analysis.

Invited Speakers

JADT’ 18

XXIII

Memorize or generalize? Searching for a
compositional RNN in a haystack Adam Liška
German Kruszewski
Facebook- germank@fb.com

Abstract
Machine learning systems have made rapid progress in the past few years, as
evidenced by the remarkable feats they have accomplished on fields as
diverse as computer vision or reinforcement learning. Yet, as impressive as
these achievements are, they rely on learning algorithms that require orders
of magnitude more data than a human learner would. This disparity could be
rooted in many different factors. In this talk, we will draw on the hypothesis
that compositional learning — that is, the ability to recombine previously
acquired skills and knowledge to solve new problems– could be one
important element of fast and efficient learning (Lake et al, 2017). In this
direction, we will discuss our ongoing efforts towards building systems that
can learn in compositional ways. Concretely, we will present a simple
benchmark based on function composition to measure the compositionality
of learning systems and use it to draw insights for whether current learning
systems learn or can learn in a compositional manner.

XXIV

JADT’ 18

Scaling-up Sentiment
Analysis through Continuous Learning
Bing Liu
University of Illinois at Chicago - liub@uic.edu

Abstract
Sentiment analysis (SA) or opinion mining is the computational study of
people’s opinions, sentiments, emotions, and evaluations. Due to numerous
research challenges and almost unlimited applications, SA has been a very
active research area in natural language processing and text mining. In this
talk, I will first give a brief introduction to SA, and then move on to discuss
some major difficulties with the current technologies when one wants to
perform sentiment analysis in a large number of domains, e.g., all products
sold in a large retail store. To tackle the scaling-up problem, I will describe
our recent work on lifelong machine learning (LML) (or lifelong learning)
that tries to enable the machine learn like humans, i.e., learning continuously,
retaining or accumulating the knowledge learned in the past, and using the
knowledge to help future learning and problem solving. This paradigm is
quite suitable for SA and can help scale up SA to a large number of domains
with little manual involvement.

JADT’ 18

XXV

La textométrie comme outil d’expertise :
application à la négociation de crise.
Pascal Marchand
Université de Toulouse – pascal.marchand@iut-tlse3.fr

Résumé
Pour aborder la pertinence de la pratique textométrique dans des
problématiques de terrains et comme outil d’expertise, on étudiera les
échanges réels impliquant les négociateurs des Forces d’intervention de
Police, dans des contextes de barricades, prises d’otages, terrorisme ou
intention suicidaire à haut niveau de dangerosité.
Nous envisagerons donc la négociation au travers des dynamiques de choix
lexical et nous chercherons à cartographier le lexique, classer des segments
de textes et comparer des profils de locuteurs et de situations.
On se propose ainsi de répondre aux questions suivantes :
 Y a-t-il des thèmes récurrents dans les crises ?
 Y a-t-il une chronologie lexicale de la crise ?
 Comment se gèrent les émotions ?
 Quelles sont les spécificités des situations « radicalisées » ?
L’objectivation des échanges et la mise en évidence des séquences formelles
peut alors fournir une aide au diagnostic, dans le but de tirer des éléments
concrets pour des objectifs de retour d'expérience et de formalisation des
pratiques des professionnels de la négociation.

XXVI

JADT’ 18

Author Identification Combining Various Author
Profiles. Towards a Blended Authorship Attribution
Methodology
George K. Mikros
National and Kapodistrian University of Athens – gmikros@gmail.com

Abstract
The aim of this presentation is to describe a new method of attributing texts
to their real authors using combined author profiles, modern computational
stylistic methods based on shallow text features (n-grams) and machine
learning algorithms. Until recently, authorship attribution and author
profiling were considered similar methods (nearly identical feature sets and
classification algorithms), but with different aims, i.e. in the former to
identify the author’s identity and in the latter to detect author’s
characteristics such as gender, age, psychological profile etc. Both of these
methods have been used independently aiming at different research aims
and in different real-life tasks. However, in this talk we will present a unified
methodological framework where standard authorship attribution
methodology and author profiling are combined so that we can approach
more effectively open or semi-open authorship attribution problems, a
category known as authorship verification which is particularly difficult to
tackle with present computational stylistic methods. More specifically, we
will present preliminary research results from the application of this blended
methodology to a real semi-open authorship problem, the Ferrante’s
authorship case. Using a corpus of 40 modern Italian literary authors
compiled by Arjuna Tuzzi and Michele Cortelazzo from the University of
Padua (Tuzzi & Cortelazzo, under review), we will explore the dynamics of
author profiling in gender, age and region and various methods we can
combine the extracted profiles so that we can entail the identity of the real
author behind Ferrante’s books. Moreover, we will extend this methodology
and validate its usefulness in social media texts using the English Blog
Corpus (Argamon, Koppel, Pennebaker, & Schler, 2007). Using, simulated
scenarios of authorship attribution cases (the real author to be included in the
training data and the real author to be missing from the training corpus) we
will further evaluate the usefulness of the proposed blended methodology
which can lead to some exciting new possibilities for investigating author
identities in both closed and open authorship attribution tasks.

JADT’ 18

XXVII

From text to concepts and back: going multilingual
with BabelNet in a step or two
Roberto Navigli
Sapienza University of Rome – roberto.navigli@uniroma1.it

Abstract
Multilinguality is a key feature of today’s Web, and it is this feature that we
leverage and exploit in our research work at the Sapienza University of
Rome’s Linguistic Computing Laboratory, which I am going to overview and
showcase in this talk. I will describe the most recent developments of the
BabelNet technology. I will introduce BabelNet live – the largest,
continuously-updated multilingual encyclopedic dictionary – and then
discuss a range of cutting-edge industrial use cases implemented by
Babelscape, our Sapienza startup company, including: multilingual
interpretation of terms; multilingual concept and entity extraction from text;
cross-lingual text similarity.

Contributors

JADT’ 18

1

Identification automatique de l’ironie et des formes
apparentées dans un corpus de controverses
théâtrales
Motasem Alrahabi1, Chiara Mainardi2
1

Université Paris-Sorbonne Abu Dhabi – motasem.alrahabi@gmail.com
2 Université Sorbonne Nouvelle – chiara.mainardi@univ-paris3.fr

Abstract
This paper presents the results of an automatic analysis on a corpus of French
texts about theatre debates (16th –19th centuries). The purpose of this study
is to highlight the important role of different forms of irony in the theatre
controversy and to reveal the stand point of authors and established
authorities towards theatre performances. Despite the difficulty of this task,
our research shows encouraging results. This unprecedent comparison of
these kind of texts, in which authors condemn the theatre or approve it,
enables to a broader understanding of the authors’ positions, arguments and
rhetorical strategies relating to theatre controversies.
Résumé
Cet article présente les résultats de notre analyse automatique d’un corpus de
débats sur le théâtre (16ème – 19ème siècle). L’objectif de cette étude est
d’illustrer le rôle important que jouent les différentes formes de l’ironie dans
la polémique autour du théâtre et de mettre en évidence la position des
auteurs ou des autorités antiques citées vis-à-vis des spectacles. Les résultats
obtenus sont encourageants malgré la difficulté de la tâche et ils nous
permettent de comparer d’une façon inédite les textes des auteurs défenseurs
avec ceux des auteurs pourfendeurs du théâtre et d’avoir une meilleure
compréhension de certains arguments et stratégies d’auteurs dans le champ
de la controverse.
Keywords: Ironie, théâtre, marqueurs linguistiques, annotation sémantique,
système à base de règles.
1. Introduction
Nous proposons une analyse automatique d’un corpus en français qui
rassemble des débats sur le théâtre depuis le milieu du 16e siècle jusque dans
les années 1840. Notre objectif est d’illustrer le rôle important que jouent les
expressions de l’ironie dans la polémique autour du théâtre et de mettre en
évidence la position des auteurs ou des autorités antiques citées vis-à-vis des

2

JADT’ 18

spectacles. Nous présentons d’abord les ressources linguistiques
développées, l’outil d’annotation utilisé et le corpus ; ensuite, nous
commentons les résultats d’analyse automatique et, avant de conclure, nous
explorons les perspectives de ce projet en cours.
2. Prémisses sur l’ironie
L’ironie est un fait de langue utilisé afin de transmettre un message
directement ou indirectement opposé à ce qui est dit littéralement.
Largement étudiée en philosophie, en rhétorique ou en linguistique
(Berrendonner, Sperber et Wilson, Kerbrat-Orecchioni, Ducrot, Grice…),
l’ironie représente un concept hétérogène extrêmement difficile à définir du
fait de ses nombreuses formes et de la complexité des phénomènes qui sont
en jeu. L’ironie fonctionne à l’aide d’indices laissés par le locuteur à
l’interlocuteur pour lui faire comprendre ses intentions par des jeux de
parallélismes, de contradictions, d’exagérations et d’hyperboles plus ou
moins marqués. Ces indices – souvent pragmatiques ou extralinguistiques –
sont plus ou moins évidents, d’où l’importance de la prise en compte du
contexte (référentiel, locuteur, interlocuteur…), des connaissances partagées
et des normes sociales et culturelles. La présente étude constitue la première
étape pour une détection automatique du champ de l’ironie au sein de notre
corpus. Conscients de la difficulté de la tâche et de l’absence de ressources
linguistiques adaptées à notre corpus et à nos objectifs, nous nous sommes
tournés vers une approche symbolique en nous basant sur un travail
précédent autour de l’annotation automatique des modalités énonciatives
(Riguet et Alrahabi, 2017). Employés dans les stratégies argumentatives, ces
marqueurs observables aident à exprimer ou à rapporter l’ironie ou d’autres
cas qui s’y apparentent (sarcasme, raillerie, satire, moquerie…). Exemple :
De sorte qu'on ne peut mieux définir la Comédie, qu'une « assemblée de
railleurs où personne ne se connait, et où chacun rit des défauts qui les
rendent tous également coupables et ridicules ». [Lelevel, 1694]
Les marqueurs utilisés sont principalement des verbes comme se moquer,
ironiser, parodier… Ensuite, par l’observation d’une partie du corpus, nous
avons enrichi ces ressources par des substantifs, des adjectifs et des adverbes.
Nous avons ensuite classé ces marqueurs dans des sous-catégories selon
différentes nuances sémantiques : 1) ironie, dérision, se moquer, sarcastique,
parodier… ; 2) chicaner, taquiner, narguer… ; 3) faire rire, comique, pitre,
grotesque, idiot… ; 4) mordant, piquant, pinçant, aigre… ; 5) mépriser,
dénigrer, sous-estimer, vilipender… ; 6) calomnier, hypocrisie, ruse,
malice… ; etc. En tout, nous avons collecté autour de 70 marqueurs

JADT’ 18

3

linguistiques.
3. Méthodologie et choix techniques
La détection automatique de l’ironie est une tâche difficile, notamment à
cause de la multitude des moyens linguistiques qui expriment, souvent de
manières subtiles, l’ironie ou les autres formes apparentées. Différents
travaux computationnels s’intéressent à la détection automatique de ces
phénomènes linguistiques (Joshi et al., 2016): approches à base de règles,
approches statistiques et approches d’apprentissage profond. Dans le présent
projet, nous avons utilisé Excom2 (Alrahabi, 2010), un outil d’annotation à
base de règles qui nous a permis d’avoir le contrôle sur le processus
d’annotation et d’améliorer progressivement la pertinence des ressources
linguistiques exploitées. Pour le système, la présence dans une phrase d’un
marqueur de l’ironie déclenche les règles associées qui explorent le contexte
et vérifient la présence ou l’absence de marqueurs complémentaires.
Dans la phrase qui suit, la présence de l’adverbe moqueusement dans le
contexte d’un marqueur de parole permet à Excom2 d’attribuer à ce passage
textuel l’étiquette « Ironie » :
« Il lui faut, dit-on moqueusement, cinq épithètes ! » [Corpus OBVIL]
Les règles dans Excom2 peuvent être organisées selon un ordre de priorité et
utiliser en entrée les résultats d’autres règles. Avant l’étape de l’annotation,
l’outil procède à la segmentation des textes afin de les découper en sections,
paragraphes et phrases. Pour l’Ironie, nous avons créé 8 règles que nous
avons associées aux différents marqueurs linguistiques.
4. Corpus
Cette présente étude s’appuie sur des textes à argumentation théâtrophile ou
théâtrophobe, et sur des textes adoptant une stratégie « mesurée ». Cette
dernière consiste à dénoncer des abus de la scène pour, ensuite, convaincre le
lecteur à préserver l’utilité intrinsèque du théâtre. Ces trois types de textes
possèdent une logique souvent détournée et déconcertante pour le lecteur :
sous le déroulement des chapitres, on découvre parfois des connexions
implicites, un usage de l’ironie très répandu et des phrases à la forme
négative qui infléchissent notablement la détection des contenus. Avec ses
reprises pour réitérer ou au contraire pour retourner l’argument contre
l’adversaire, ce corpus de controverses théâtrales se prête bien à des analyses
numériques. Le corpus rassemble 59 textes (environ un million de mots) écrits
en langue française depuis le milieu du 16e siècle jusque dans les années

4

JADT’ 18

18401. Ceux-ci ont été préalablement numérisés et édités dans le cadre du
Labex OBVIL de Paris IV-Sorbonne et sont librement accessibles en ligne2.
5. Evaluation
Une première phase de tests sur un échantillon du corpus a été nécessaire
pour stabiliser les règles d’identification et de désambiguïsation. Afin
d’évaluer la qualité des annotations obtenues, nous nous sommes focalisés
dans un premier temps sur le calcul de la précision. Nous avons alors annoté
avec Excom2 une autre partie du corpus (7 articles, 215675 mots) et nous
avons obtenu 416 annotations. Ensuite, nous avons demandé à une personne
qui connait les œuvres de cette période de juger les sorties du système selon
un guide d’annotation. Pour chaque annotation, l’évaluatrice devait choisir
entre : « Correct », « Incorrect » ou « Je ne sais pas ». Le critère d’évaluation
était le suivant : est-ce que l'auteur du texte fait allusion à l'ironie dans la
phrase en question? Nous avons obtenu 93.9 % de précision.
6. Difficultés rencontrées
Nous nous sommes heurtés à plusieurs difficultés. Au niveau du lexique, peu
de changements ont été effectués sur nos marqueurs, comme par exemple le
mot satire qui se trouve avec les deux orthographes satire (88 occurrences) et
satyre (68 occurrences). En français, le dernier désigne le demi-dieu
compagnon de Dionysos ou Bacchus. Cependant, dans certains textes qui
n’ont pas encore été modernisés, et sont en langue française du 16e ou 17e
siècle, ce mot indique plus largement la « satire ». D’un autre côté, certains
marqueurs sont polysémiques et génèrent du bruit, comme ridicule (437
occurrences, le marqueur le plus fréquent), plaisanter (176 occurrences) et comique
(131 occurrences). Exemple [Rousseau, 1758] :
Le ridicule est l'arme favorite du vice. C'est par elle qu'en attaquant dans le
fond des cœurs le respect qu'on doit à la vertu, il éteint enfin l'amour qu'on
lui porte.
Concernant la syntaxe du 17e et 18e siècle, nous avons observé une certaine
complexité au niveau des phrases qui sont parfois très longues (cinq lignes ou

Nous renvoyons à la liste de la bibliographie française qui constitue le corpus
total de la Haine du Théâtre: http://obvil.paris-sorbonne.fr/corpus/hainetheatre/bibliographie_querelle-france/
2 Il s’agit d’une partie du corpus de « La Haine du théâtre », projet dirigé, au sein
du Labex OBVIL, par François Lecercle et Clotilde Thouret (Lecercle et al., 2016),
http://obvil.paris-sorbonne.fr/projets/la-haine-du-theatre.
1

JADT’ 18

5

plus), et au niveau des signes de ponctuation qui ne sont pas stables.
Plusieurs virgules, points virgules, etc. peuvent en effet se succéder dans une
seule phrase. De plus, les auteurs de notre corpus utilisent des tournures
complexes. Très souvent, ces phrases sibyllines sont ironiques, et cela se
passe d’autant plus si elles se trouvent à la forme interrogative.
7. Interprétation des résultats
Dans l’étude des débats sur le théâtre, les expressions de l’ironie sont une
voie d’entrée féconde dans le corpus. On constate d’abord que, tout au long
des siècles couverts par le projet Haine du Théâtre (16e – 19e siècles), l’usage
de l’ironie se situe entre les valeurs de 0,20 à 0,30 % (1265 annotations en
total). Nous avons ensuite analysé les marqueurs de l’ironie en étudiant leur
présence relative selon les siècles et nous avons pris en compte uniquement
ceux ayant un pourcentage supérieur à 5% à l’intérieur d’un même siècle.

Figure 1 : Les marqueurs d’ironie dans le corpus HdT pondérés par siècle

Une baisse considérable a lieu au 17e siècle. S’il est prématuré d’en tirer des
conclusions hâtives, nous pouvons cependant tout de suite constater que cela
est probablement dû à l’affirmation de la religion, de l’ordre du classicisme
ainsi qu’à l’autoritarisme étatique qui s’insinuait dans les esprits des
écrivains de cette époque. En revanche, au fur et à mesure du 17e au 19e
siècle, les valeurs de ces marqueurs augmentent de manière assez stable.
De manière générale, l’ironie est utilisée dans le corpus comme procédé
éthique et stylistique, ce qui rend les auteurs bien efficaces dans l’élaboration
de leur vision de la querelle. Qu’ils soient théâtrophobes ou théâtrophiles, ils
peuvent jouer avec les nuances des marqueurs d’ironie, dissimuler un
double-sens dans leurs phrases, s’exprimer figurément de manière contraire

6

JADT’ 18

à ce qu’ils communiquent littéralement. Par exemple, nous retrouvons une
présence considérable du lemme « mépris » au 17e et 18e siècles. Il s’agit
principalement d’un usage de l’ironie en tant que mécanisme de régulation
de la vie sociale. Notamment, Conti et Voisin utilisent un humour inoffensif
contre les excès de l’art et mettent en avant la bienséance :
Ceux qui vont aux Spectacles, non par hasard, mais de propos délibéré, et
avec tant d'ardeur, qu'ils abandonnent l'Eglise par un mépris insupportable
pour y aller, où ils passent tout le jour à regarder ces femmes infâmes,
auront-ils l'impudence de dire qu'ils ne les voient pas pour les désirer
[Conti 1667, Voisin 1671]
L’ « hypocrisie » commence à être utilisée au 17e et son utilisation se réduit
avec le temps (jusqu’à 1% au 19e). Le lemme en question est essentiellement
appliqué à des phrases où l’ironie n’est qu’un « autre nom du malheur »
(Martin 2009), une manière de renforcer le point de vue de l’auteur.
L’hypocrisie est un vice privilégié, qui ferme la bouche à tout le monde, et
qui jouit en repos d'une impunité souveraine. [Coustel 1694]
Très répandu dans le corpus est l’usage de l’ironie comme écho satirique. Le
lemme « calomnier », présent dans les textes du 17e au 19e siècle, en est
l’exemple :
[…] cessez de calomnier vos contemporains selon l'usage immémorial de
ceux qui profèrent de vaines paroles. [Senancour 1825]

Figure 2 : Valeurs pondérées de l’annotation de l’ironie dans le corpus

JADT’ 18

7

Les premiers résultats nous ont ainsi permis d’effectuer des comparaisons
très intéressantes entre les textes des auteurs défenseurs et les textes des
auteurs pourfendeurs du théâtre. A partir du nombre d’expressions
ironiques correctement identifiées comme telles, nous avons recensé leur
nombre et dressé des statistiques pour chaque auteur du corpus annoté.
On constate qu’en données relatives, les auteurs qui utilisent le plus les
marqueurs d’ironie appartiennent à la « querelle Rousseau » (moitié du 18e
s.). Cela est à analyser en perspective mais, en l’espèce, dans cet article nous
pouvons le mettre en lien avec l’usage de l’ironie au 18e siècle, comme
plusieurs écrits sur Voltaire le témoignent (Loriot, 2015). Les mots de
D’Alembert sont très parlants sur ce sujet et éclairent le rôle de
l’ironie [Alembert, 1759] :
Si la satire et l’injure n’étaient pas aujourd’hui le ton favori de la critique,
elle serait plus honorable à ceux qui l’exercent, et plus utile à ceux qui en
sont l’objet.
Les marqueurs linguistiques qui ont été détectés pour cette période
appartiennent à la sphère sémantique du ridicule, de la satire, de la farce et du
comique3. D’autres marqueurs verbaux, tels que se moquer et plaisanter sont
présents dans cette querelle et sont communs aux écrits de la précédente
controverse datant du milieu du 17e siècle. Les valeurs ironiques de cette
dernière, dont les représentants théâtrophobes sont Conti et Nicole, parmi
d’autres, sont cependant moins importantes (0,06 vs. 0,17). Outre ces
marqueurs verbaux, nous pouvons citer les catégories de substantifs tels que
le ridicule et le faire rire. A la même période, Aubignac, auteur de la stratégie
offensive-défensive, part d’une critique du théâtre pour arriver à sa défense.
Il s’inspire des marqueurs habituels pour la période du 17e siècle et reprend
dans ses phrases les propos de ses collègues, pour ensuite les réfuter. De
plus, il recourt plus spécifiquement à des marqueurs ironiques tel que railler
et idiot. Contemporaine à d’Aubignac, la querelle entre Caffaro et Bossuet
nous donne des résultats surprenants : si Caffaro emploie peu de marqueurs
relevant de l’ironie (0,05), Bossuet est lui chef de file parmi ses contemporains
(valeur de 0,27). Comme les autres auteurs, Bossuet puise dans les marqueurs
du comique et du ridicule, tout comme la forme verbale plaisanter. Néanmoins,
nous retrouvons dans ses résultats des mots appartenant à la catégorie de
marqueurs piquants [Bossuet, 1694]:

3 Signalons que le marqueur « ironie » et toutes ses variantes n’ont que 11
occurrences dans le corpus !

8

JADT’ 18

Il ne faut pas s’étonner que l’église ait improuvé en général tout ce genre de
plaisirs [les spectacles…] à cause que communément, ainsi que nous l’avons
remarqué, par sa bonté et par sa prudence, elle épargne la multitude dans les
censures publiques : néanmoins parmi ces défenses, elle jette toujours des
traits piquants contre ces sortes de spectacles, pour en détourner tous les
fidèles.
Nous comprenons ainsi que pour juger le théâtre incompatible avec la morale
chrétienne, Bossuet privilégie un style vif et mordant, il appuie l’église tout
en dénigrant les défenseurs du théâtre.
La recherche sur les stratégies de la querelle du théâtre, tout en se
questionnant sur les modalités argumentatives et les objectifs circonstanciels
de chaque auteur, nous dévoile également certaines idées récurrentes autour
de la considération du théâtre. Les différents textes partagent un certain
nombre de lieux communs, comme par exemple l’idée de perversion,
l’inflation temporelle, ou les arguments économiques et politiques.
8. Discussion et perspectives
Dans cet article, nous avons présenté une approche à base de règles pour la
détection automatique de l’ironie et des formes apparentées dans un corpus
de débats sur le théâtre (16e – 19e siècle). La méthode que nous avons adoptée
nous a fourni une matière abondante et des données quantitatives pour
mieux cerner l’objet d’étude. Vu la particularité du phénomène langagier
étudié et la simplicité de notre approche par analyse de surface, nous
considérons que ces premiers résultats sont très encourageants (93.9 % de
précision). A ce titre, ils méritent d’être approfondis afin d’en tirer le plus
grand bénéfice en terme d’exploitation et de précision. Nous envisageons de
calculer le taux de rappel dans l’annotation et d’identifier les sources des
segments annotés (les locuteurs). L’un de nos objectifs consiste également à
annoter les phrases négatives et à analyser leur association avec l’ironie
(Mainardi et al., 2015), ce qui nous permettrait de dégager des pistes de
recherche inédites dans le domaine des humanités numériques.
Références
Alrahabi, M. (2010). EXCOM-2: plateforme d'annotation automatique de
catégories sémantiques. Applications à la catégorisation des citations en
français et en arabe. Thèse de doctorat, Université Paris-Sorbonne.
Joshi A., Bhattacharyya P., Carman M. J., (2016). Automatic Sarcasm
Detection: A Survey ACM Comput. Surv. V, N, Article A (January 2016).
Lecercle F., Mainardi C., Thouret C. (2016). Pour une exploration numérique
des polémiques sur le théâtre, RHLF, n°116 / 4 dir. Didier Alexandre,
Littérature et humanités numériques, PUF.

JADT’ 18

9

Loriot C. (2015), Rire et sourire dans l'opéra-comique en France aux 18ème et
19ème siècles, Lyon, Symétrie.
Mainardi C., Sellami Z., Jolivet V., (2015). “A Semantic Exploration Method
Based on an Ontology of 17th Century Texts on Theatre: la Haine du
Théâtre", First International Workshop on Semantic Web for Cultural
Heritage (SW4CH 2015), New Trends in Databases and Information
Systems, 539, pp. 468-476, Communications in Computer and Information
Science.
Martin L. (2009), “Le rire est une arme. L'humour et la satire dans la stratégie
argumentative du Canard enchaîné”, A contrario 2009/2 (n° 12), 26-45.
Riguet M., Alrahabi M. (2017), "Pour une analyse automatique du Jugement
Critique: les citations modalisées dans le discours littéraire du XIXe
siècle", in DHQ: Digital Humanities Quarterly 2017

10

JADT’ 18

Migrants et réfugiés : dynamique
de la nomination de l'étranger
Mohammad Alsadhan, Sascha Diwersy,
Agata Jackiewicz, Giancarlo Luxardo
Praxiling UMR 5267 (Univ Paul Valéry Montpellier 3, CNRS)
muhammad.alsadhan@univ-montp3.fr, sascha.diwersy@univ-montp3.fr,
agata.jackiewicz@univ-montp3.fr, giancarlo.luxardo@univ-montp3.fr

Abstract
Intense debates arose from the migrant crisis experienced by Europe in recent
years, both in the media and in the politics. We address here the issue of
nomination used for the newcomers, that we propose to study based on the
comparison of the two substantivations in French: migrant and réfugié. Using
their combinatory profiles, we seek to highlight the contrast between the two
terms and the changes in their semantics and their axiological charge. In
order to do so, we rely on a large corpus of texts, established over a threeyear period: the French parliamentary debates of the Assemblée Nationale. The
comparative study of the combinatory profiles related to the two terms
shows that both shared and unshared collocatives are encountered, and that
their profiles overall tend to converge.
Résumé
Au cours des dernières années, la crise migratoire en Europe a suscité de vifs
débats politico-médiatiques. Nous nous intéressons ici à la question de la
nomination des nouveaux arrivants, que nous proposons d’étudier par la
comparaison des deux substantivations migrant et réfugié. A partir de leurs
profils combinatoires, nous cherchons à mettre en évidence le contraste entre
ces deux termes, les changements dans leur sémantique et leur charge
axiologique. Pour cela, nous nous appuyons sur un corpus, établi sur une
période d’environ trois ans : les débats à l'Assemblée Nationale. L’étude
comparative des profils combinatoires associés aux deux termes montre que
l’on rencontre à la fois des collocatifs partagés et d’autres non partagés et que
leurs profils tendent globalement à converger.
Keywords: political discourse, cooccurrences, diachronic data and
hierarchical clustering, curve clustering.

JADT’ 18

11

1. Introduction
L'Union Européenne a connu en 2015 une arrivée massive d’étrangers extraeuropéens, qui a donné lieu à des formules telles que « crise migratoire » ou «
crise des réfugiés ». Dans un contexte de net clivage de l’opinion publique,
cette crise a entraîné des positions politiques contrastées dans chaque pays
concerné et des compromis difficiles à trouver.
Les débats politico-médiatiques ont porté d'abord sur la prise en charge des
victimes, le droit « d'asile » à accorder aux nouveaux venus, de même que sur
la lutte contre les filières illégales, avec des positions « pro-immigration » ou
« anti-immigration ». Mais ce phénomène s'expliquant en partie par les
conflits en cours au Sud et à l'Est de l'Europe, la question de la désignation
des intéressés a été posée. Alors que jusque-là les « migrants » étaient
principalement motivés par des perspectives économiques, il a été remarqué
qu'une partie de ces personnes devraient être nommés « réfugiés » ou «
demandeurs d'asile ». D'autres termes, comme « clandestins », ont pu aussi
être évoqués.
Nous cherchons ici à questionner la dynamique de la nomination utilisée
dans les débats politiques. A partir d’un corpus de débats parlementaires
nous mettons en œuvre divers procédés de classification basés sur la nature
diachronique des données.
2. Les corpus de débats parlementaires
Nous faisons l’hypothèse que les discours autour de la crise migratoire font
usage des deux termes migrant et réfugié en partie de façon interchangeable,
en partie dans des contextes où seulement l’un des deux est possible. Cette
distinction entre plusieurs emplois en discours, nous proposons de la mettre
en évidence par le voisinage des deux termes et d’évaluer sa variation
d’abord sur le discours politique et en fonction du temps.
Le corpus traité dans la suite est constitué à partir des transcriptions des
débats en séance publique à l’Assemblée Nationale pour la période qui va de
janvier 2014 à février 2017 (ce qui correspond à la fin de la XIVe législature).
Les données textuelles, publiées en format XML et disponibles en accès libre
sur le site data.assemblee-nationale.fr, représentent environ 28,6 millions de
mots occurrences. Elles ont été transformées et enrichies par des annotations
linguistiques suivant une méthodologie décrite par Diwersy et al. (2018). De
nombreuses métadonnées sont définies sur ce corpus, mais dans la suite nous
nous concentrons sur la date (mois-année) associée à une unité structurelle
de base correspondant au tour de parole (intervention d’un député).

12

JADT’ 18

3. Analyse chronologique
L’évolution du sémantisme des termes migrant et réfugié peut être étudiée par
l’association de méthodes mettant en jeu : (i) les fréquences d’apparition de
ces deux lemmes dans les corpus, (ii) leurs profils collocationnels, qui
peuvent faire émerger des champs sémantiques spécifiques, (iii) la variation
de la similarité de ces profils collocationnels dans le temps et la
caractérisation de la contribution de chaque collocatif à l’évolution des scores
de similarité obtenus.

Figure 1

L’évolution des fréquences relatives des deux lemmes par trimestre dans le
corpus est illustrée par le graphique en figure 1. Il met en évidence une
évolution fréquentielle en parallèle avec un pic d’utilisation des deux termes
autour de septembre 2015. La corrélation de rang entre les deux séries
fréquentielles, mesurée par le taux de Kendall, est ici significative (environ
0,74, pour une p-valeur de 0,0005). Dans la suite, l’unité de temps choisie est
le trimestre ; il en résulte des analyses sur 13 trimestres pour la période
couverte. Afin de produire une périodisation plus précise, nous avons mis en
œuvre une approche combinant annotations en relations de dépendance
syntaxique, création de lexicogrammes représentant les profils
collocationnels par trimestre des deux termes (ordonnés suivant le score
d’application du test exact de Fisher) et application de Classifications
Ascendantes Hiérarchiques par Contiguïtés (CAHC), cf. (Diwersy et
Luxardo, 2016 ; Gries et Hilpert, 2008).
La construction d’une CAHC peut être entreprise suivant deux méthodes :
 pour chaque lemme, en calculant la similarité entre deux trimestres

JADT’ 18

13

successifs d’après le coefficient de Pearson (Pearson product-moment
correlation coefficient),
 en calculant la variation de la similarité entre les vecteurs représentant
les profils collocationnels des deux lemmes, d’après l’écart type cumulé
sur deux trimestres successifs.
La première méthode révèle les variations plus importantes sur les trimestres
initiaux jusqu’au pic de la crise. La deuxième méthode qui permet d’illustrer
la comparaison entre les deux termes par un graphique unique est
représentée par la figure 2.

Figure 2

L’étude de cette classification hiérarchique permet de révéler sept étapes
(représentées par sept zones grises). L’évolution du score de similarité est
illustrée par un graphique qui se superpose au dendrogramme et qui
confirme une croissance globale de 0 à 0,2 (mais avec un pic à 0,6). Le passage
d’une période à l’autre est marqué par une progression jusqu’à P03
(correspondant au 3e trimestre 2015, suivant le pic de la crise) mais avec un
déclin des périodes P03 à P05 et de P06 à P07.

14

JADT’ 18

Figure 3

4. Évolution des profils combinatoires et orientations discursives
Cette section vise à expliciter les facteurs linguistiques à l’origine des
tendances statistiques établies dans la partie précédente. Il s’agit, d’une part,
de mettre en évidence les différences sémantiques entre migrant et réfugié
telles qu’elles se manifestent à travers leurs profils différentiels et, d’autre
part, de relever les points essentiels concernant leur similarité
distributionnelle. Les profils différentiels sont constitués par les collocatifs
exclusifs à chacun des substantifs étudiés et, de ce fait, ne contribuent à
aucun moment à la similarité de leurs profils combinatoires. Le tableau 1 en
donne un aperçu restreint aux collocatifs les plus saillants, situés dans le
premier décile des inventaires collocationnels en termes de score
d’association.
Tableau 1 - Profils différentiels constitués par le premier décile
des collocatifs exclusifs à migrant et à réfugié
migrant

réfugié

Dépendances en aval
(régime)

Dépendances en amont Coordin
(termes recteurs)
ation

Dépendances en
aval

Dépendances en
amont

Epithètes

Compl.
du nom

Compl.
du nom

Epithètes

irrégulier
illégal
clandestin
âgé

Calais
Calaisis
situation

Compl. objet
Compl. circ.
Sujet
dissuader
entasser
refouler
secourir

retour
langue
déferleme
nt
réadmissi
on

politique
guerre
palestinien
afghan
vietnamien
irakien
cambodgien
persécuté
réinstallé

CO
CC
Sujet
affluer

CDN

Coord
inatio
n

CDN

statut
protection
(Haut-)
Commissar
iat
qualité
relocalisati
on
distinction
concubin
défi

apatri
de
bénéfic
iaire
déplac
é
migra
nt

JADT’ 18

15

Parmi les collocatifs saillants du nom réfugié, on notera d’abord la forte
présence d’une série de termes (statut, qualité ; (Haut-)Commissariat ;
protection ; apatride)1 qui renvoient au cadre des dispositions relevant du droit
international qui imposent aux autorités un devoir d’assistance envers des
personnes dont le départ de leur lieu de résidence habituelle est considéré
comme étant contraint par une menace existentielle. Catégoriser une
personne au moyen du terme réfugié revêt donc un enjeu juridique,
administratif et politique, dont l’ampleur peut se voir régulée d’une part, par
des mises en paradigme explicites avec d’autres termes dans le cadre d’une
coordination (cf. les collocatifs apatride, bénéficiaire, déplacé et migrant) et,
d’autre part, par des catégorisations secondaires exprimées par des
expansions nominales (épithètes ou compléments du nom) caractérisant les
causes du départ forcé. A travers les modifieurs du nom réfugié impliquant
une relation causale (politique, persécuté ; (de) guerre) se construit un
paradigme, et finalement une hiérarchie de causes potentiellement légitimes
ou non-légitimes (et de réponses à apporter aux conséquences liées à ces
causes).2 A côté de ces modifieurs, qui dénotent directement la cause du
départ forcé, on trouve toute une série d’adjectifs ethnonymiques (palestinien,
afghan, vietnamien, irakien, cambodgien) qui la dénotent indirectement en
s’appuyant sur le savoir partagé concernant l’histoire troublée de ces pays.
Cet environnement discursif montre que le mot réfugié se présente comme la
nomination d’un statut juridique et qu’il est intégré à une argumentation
orientée positivement.
Les collocatifs de migrant révèlent un profil sémantique bien différent, en ce
sens que ce terme place au centre de l’intérêt la question de la (non)conformité à des dispositions légales imposées à des personnes dont le
séjour sur un territoire différent de celui de leur lieu résidentiel d’origine est
considéré comme étant le résultat d’un déplacement conditionné par des
considérations utilitaires (et en premier lieu économiques). C’est bien à cette
dimension sémantique que se rapporte, dans le profil différentiel de migrant,
de façon saillante la série des collocatifs irrégulier, illégal, clandestin et situation
(qui, quant à lui, s’oppose, de ce point de vue, à statut et qualité, collocatifs
exclusifs à réfugié). Ayant hérité les traits aspectuels du participe en –ant dont

On trouve dans les déciles inférieurs – non documentés ici – d’autres collocatifs
comme statutaire ou conventionnel qui rentrent dans cette même série.
2 On peut observer que cette sous-catégorisation va souvent de pair avec une
modalisation d’appartenance catégorielle, exprimée par l’adjectif épithète véritable qui
constitue avec vrai et authentique une série de collocatifs (appartenant à la catégorie de
l'enclosure) exclusifs à réfugié qui sont néanmoins représentés à des rangs inférieurs
de l’inventaire cooccurrentiel.
1

16

JADT’ 18

il est issu par conversion, le nom migrant présente le séjour momentané de la
personne qualifiée en tant que telle à un endroit donné comme étant
l’épisode d’une série inaccomplie de déplacements3 – séjour et déplacements
qui, à travers des collocatifs tels dissuader, refouler et retour, se voient
caractérisés comme relevant aussi bien de la volonté des personnes en
mouvement, que de la bienveillance ou du refus des autorités qui en ont le
contrôle potentiel. Faut-il voir en cela la motivation inférentielle de
l’évaluation négative que véhicule un terme comme déferlement contrairement
à ses variantes axiologiquement plus neutres afflux, flux ou encore arrivée, qui,
eux, font tous partie des collocatifs partagés des noms migrant et réfugié ?
Pour mieux cerner les collocatifs partagés qui contribuent le plus à
l’évolution de la similarité distributionnelle des deux noms en question, nous
avons mis en œuvre la méthode de classification proposée par Trevisani &
Tuzzi (2016), en l’appliquant aux séries chronologiques des produits de
scores d’association normés propres à chaque collocatif, qui entrent dans la
composition des sommes donnant les produits scalaires lesquels représentent
les indices de similarité retenus.

Figure 4

L’application de la méthode4 fait ressortir, sur l’ensemble des 72 collocatifs
communs à migrant et réfugié, 6 classes de profils évolutifs, dont 5 sont

Contrairement à cela, réfugié, qui résulte de la nominalisation d’un participe
passé, est associé à la représentation d’un seul épisode de déplacement accompli et
envisagé en termes de son origine.
4 Nous remercions Arjuna Tuzzi d’avoir mis à notre disposition le script R
permettant de mettre en œuvre les calculs respectifs.
3

JADT’ 18

17

constituées par un seul terme, à savoir millier, afflux, accueillir, crise et accueil
(cf. figure 3). D’un point de vue sémantique, ces 5 collocatifs, qui, à différents
moments de la série chronologique analysée, occupent les premiers rangs en
termes de contribution aux scores de similarités respectifs, forment tout un
condensé de la trame discursive impliquant les noms migrant et réfugié au
cours de la période étudiée, avec :
 millier et afflux, qui renvoient à une affluence perçue comme massive ;
 crise, qui caractérise ce processus comme ayant atteint un point
culminant à fort potentiel de déstabilisation ;
 ainsi que accueillir et accueil qui se rapportent à la prise en charge des
conséquences immédiates du processus concerné.
Facteurs distributionnels de premier ordre, ces collocatifs placent migrant et
réfugié dans un rapport paradigmatique associé à plusieurs dimensions
sémantiques, qui, en vue des orientations argumentatives fortement
divergentes instaurées par les deux noms (cf. supra), fait de leur choix un
véritable enjeu discursif.
5. Conclusion et perspectives
Les prolongements de cette étude exploratoire sont nombreux. En partant de
Wihtol De Wenden (2016), il nous semble possible de construire un modèle
d’analyse comportant cinq catégories qui sont autant de facettes du
phénomène migratoire actuel : (i) origines et causes des migrations, (ii)
profils des migrants, (iii) situation des migrants, (iv) gouvernance des
migrations, (v) mobilité et restrictions migratoires. L’application de cette
grille de lecture aux collocations impliquant les termes réfugié et migrant (ou
encore leurs équivalents), peut s’avérer une piste de recherche prometteuse
qui permet de donner aux résultats de l’analyse linguistique que nous venons
d’effectuer une dimension transdisciplinaire, comme c’est par exemple le cas
pour la différence entre facteurs « push » (poussant les individus à partir de
leur pays) et « pull » (incitant les individus à venir dans un pays spécifique)
établie par Wihtol de Wenden, différence qui se reflète dans la divergence
fondamentale de l’orientation argumentative des programmes de sens
propres aux noms étudiés, en ce que réfugié implique la notion de départ
forcé alors que migrant évoque l’idée d’un déplacement volontaire.
Si la figure du réfugié ou du migrant est essentiellement une construction
politique (Wihtol De Wenden, 2016, p. 50) – ce que confirme d’ailleurs le
profil collocationnel du terme correspondant tel qu’il se manifeste dans le
corpus de discours parlementaire analysé - les différents (et nombreux)
profils des personnes en déplacement peuvent être étudiés à partir des
témoignages qu’elles livrent à propos de leur expérience migratoire. C’est
l’objet d’une enquête menée auprès de Syriens arrivés en France depuis 2012,

18

JADT’ 18

qui se situe dans le prolongement du présent article et qui comporte à ce
stade un volet uniquement qualitatif, dont les résultats préliminaires
(Alsadhan et Richard, 2018) montrent que, lorsque le choix se présente, c’est
bien le vocable réfugié qui est privilégié en tant qu’auto-désignant.
Références
Alsadhan, M., Richard A. (2018, à paraître). La réception des réfugiés Syriens
du discours médiatico-politique identitaire français, in Sandré M., Richard
A. & Hailon F. : Le discours politique identitaire face aux migrations, No 8 de
la revue Studii de lingvistica.
Diwersy, S., Luxardo, G. (2016). Mettre en évidence le temps lexical dans un
corpus de grandes dimensions : l’exemple des débats du Parlement
européen, in Mayaffre D., Poudat C., Vanni L., Magri V. & Follette P.
(éds.) : JADT 2016 : Actes des 13es Journées internationales d’Analyse
statistique des Données Textuelles, Nice, 2016, URL :
http://lexicometrica.univ-paris3.fr/jadt/jadt2016/01ACTES/83638/83638.pdf.
Diwersy, S., Frontini, F., Luxardo, G. (2018, à paraître). The Parliamentary
Debates as a Resource for the Textometric Study of the French Political
Discourse, in Proceedings of ParlaCLARIN workshop, 11th edition of the
Language Resources and Evaluation Conference (LREC2018).
Gries, S.T., Hilpert, M. (2008). The identification of stages in diachronic data:
variability-based neighbour clustering. Corpora 3 (1), pp. 59-81.
Trevisani, M., Tuzzi, A. (2016). Analisi di dati testuali cronologici in corpora
diacronici: effetti della normalizzazione sul curve clustering, in Mayaffre
D., Poudat C., Vanni L., Magri V. & Follette P. (éds.) : JADT 2016 : Actes
des 13es Journées internationales d’Analyse statistique des Données Textuelles,
Nice, 2016, URL : http://lexicometrica.univ-paris3.fr/jadt/jadt2016/01ACTES/82630/82630.pdf.
Wihtol De Wenden C. (2016). Migrations. Une nouvelle donne, Éditions de la
Maison des sciences de l'homme, Paris.

JADT’ 18

19

Xplortext, a R package.
Multidimensional statistics for textual data science
R. Alvarez-Esteban1, M. Bécue-Bertaut2, B. Kostov3,
F. Husson4, J-A Sánchez-Espigares2
2

1Universidad de León – ramon.alvarez@unileon.es
Universitat Politècnica de Catalunya – monica.becue@upc.edu; josep.a.sanchez@upc.edu
3Institut d'Investigacions Biomèdiques August Pi i Sunyer – belchin3541@gmail.com
4Agrocampus Ouest – husson@agrocampus-ouest.fr

Abstract
We present here the package Xplortext for textual data science which
provides classical and novel features for textual analysis. Starting from the
corpus encoded into a lexical table, aggregate or not, several problems are
dealt with: revealing both document and word structures and their mutual
relationships, by applying correspondence analysis (CA); comparing several
corpora structures by using multiple factor analysis for contingency tables
(MFACT); uncovering complex relationships between words and contextual
variables via CA for a simple or a multiple generalized aggregate lexical table
(CA-GALT and MFA-GALT), clustering documents thanks to a hierarchical
clustering algorithm (HCA); evaluating the evolution of the vocabulary along
time thanks to a chronological constrained hierarchical clustering algorithm
(CCHCA).
Resumé
Nous présentons ici le paquet Xplortext pour la science des données
textuelles qui comprend des méthodes classiques et récentes d'analyse
textuelle. Partant du corpus encodé sous forme tableau lexical, agrégé ou
non, plusieurs problèmes sont traités: révélation des structures sur les
documents et les mots ainsi comme leurs relations mutuelles, en appliquant
l'analyse des correspondances (AC); comparer plusieurs structures de corpus
en utilisant l'analyse factorielle multiple pour les tables de contingence
(MFACT); découvrir des relations complexes entre mots et variables
contextuelles via CA pour une table lexicale agrégée simple ou multiple (CAGALT et MFA-GALT), en regroupant des documents grâce à un algorithme
de clustering hiérarchique (HCA); évaluer l'évolution du vocabulaire au fil
du temps grâce à un algorithme de classification hiérarchique sous contrainte
chronologique (CCHCA).
Keywords: Xplortext, R package, Textual data, Contextual data,
Correspondence analysis, Multiple factor analysis for contingency tables,

20

JADT’ 18

Generalized aggregate lexical table, Hierarchical clustering, Contiguity
constrained hierarchical clustering, Labeled tree.
1. Introduction
R offers numerous tools for textual data science. However, among them,
multidimensional statistics is not so well represented that it should be.
Xplortext, a new R package, intends to fill in the gaps. Its features are based
on the exploratory approach to texts, in the line of the works by Benzécri
(1981) and Lebart et al. (1998). The fundamental choices behind the design of
Xplortext are to offer classical and novel textual analysis methods based on
multidimensional statistics in a same package. The mains issues were to
consider:
 Classical multidimensional statistical methods, in which CA remains
being the core method.
 Novel methods, favoring those able to jointly analyze textual and
contextual data to know not only who says what, taking here the title
of a paper by Lebart, but also why he/she is saying that.
 Numerous graphical outputs providing great flexibility to choose the
elements to be represented.
 Specific methods to deal with chronological corpora.
2. Example
The political speech corpus used as an example consists of 11 documents of
about 10,000 occurrences each one. These are the "investiture speeches" of 6
Spanish presidential candidates who have been pronounced from 1979 to
2011: Suarez (1979), Calvo-Sotelo (1981), González (1982, 1986, 1989 and
1993), Aznar (1996 and 2000), Zapatero (2004 and 2008) and Rajoy (2011).
3. Encoding the textual data and basic statistics
Xplortext takes advantage of functions of the R package tm to import the
corpus. Mainly, plain text files (typically .txt) and spreadsheet-like file (.csv,
.xls) are considered. By default, plain text and CSV files are assumed to use
the local native system (usually latin1) on Windows and utf8 in Mac or
Linux. The encoding of the file can be given in the R command read. If
necessary, the corpus can be saved in a known encoding beforehand. In any
format, one row corresponds to one document. The text to analyze can be
filled in one or several columns; the remaining columns provide information
about the documents and are automatically imported as contextual
(quantitative and/or qualitative) variables. Textual and contextual data must
be located in the same file. Conversion to lower/upper cases, numbers
removal and punctuation are managed by Xplortext depending on the

JADT’ 18

21

arguments of Textdata function. Stopwords can be taken into account using
the lists provided by either Xplortext (issued from tm) or the user. The
importing step ends with the encoding of the corpus into a documents ×
words table (lexical table) and, possibly, a documents × repeated segments
table (segmental table). Another option is to ask for an aggregated lexical
table according to the categories of a variable. Then, elementary indicators,
such as the corpus and vocabulary sizes, are computed and the words and
repeated segments indices are listed and represented by a histogram,
visualizing so their frequency (Fig.1). Classical summaries of the contextual
variables are given.
4. Correspondence analysis as a core method
Correspondence analysis (CA) is a core method in Xplortext revealing both
document and word structures and their mutual relationships.
4.1. CA and content and form of a corpus
The content and form of a corpus are both important as CA results. In fact,
content is better captured when replaced into the form as, "the form is the
bottom that comes back to the surface" in the words of Victor Hugo. Figure 2
shows the factor maps issued from a CA performed on the documents ×
words table.

Figure 1: Most frequent words and repeated segments

The trajectory of the speeches is revealed, enhancing the existence of three
temporal poles. The represented words are the most contributive and have to
be read as seen along the trajectory. In this way, they clearly illustrate the
three poles and allow us to capture the meaning of the evolution. Note that
the confidence ellipses around the documents are very narrow.

22

JADT’ 18

Figure 2: Documents and the most contributive words on the first CA plane

4.2. Multiple factor analysis for contingency tables
When dealing with a multiple contingency table (=juxtaposition of several
contingency tables), the multiple factor analysis for contingency tables
(MFACT; Bécue-Bertaut & Pagès, 2004; Bécue-Bertaut & Pagès, 2008),
extension of CA, turns to be useful. Very different aims can be looked for. For
example, interesting aims would be comparing the documents structures as
issued either from using different thresholds on the word frequency (10, 20,
30 or 50; 4 lexical tables) or from keeping or not the tool words (2 lexical
tables) or the stopwords.

Figure 3: Synthetic representation of the groups as issued from MFATC

JADT’ 18

23

MFACT offers a high number of graphical and numerical results, either
similar to those of any principal component methods (such as PCA or CA) or
specific to the comparison of structures defined on the rows by the groups of
columns. Among the latter, the representation of the groups provides a
synthetic tool by representing each group with one point, revealing the
global dissimilarities between the group structures (Fig. 3).
4.3. Generalized aggregate lexical tables
Correspondence analysis on a generalized aggregated lexical table (CAGALT; Bécue-Bertaut & Pagès, 2015; Bécue-Bertaut, Kostov & Pagès, 2014)
deals with two paired tables (frequency table, contextual variables table)
observed on the same statistical units. In textual analysis, the frequency table
is a lexical table and the statistical units are the documents. This method can
be seen as a canonical correspondence analysis (CCA; ter Braak, 1986)
approach to the texts. It enables to study the relationships between
contextual variables and words but untangling the respective influences of
the variables/categories on the lexical choices to avoid spurious relationships.
MFA-GALT (multiple factor analysis for analyzing a series of generalized
aggregated lexical tables; Kostov, 2015) deals with several paired tables,
possibly defined on several sets of statistical units while the set of variables is
common to all the contextual tables. In textual analysis, MFA-GALT
compares the relationships between words and variables in these several
paired tables. A favored application concerns surveys answered in different
languages by several samples, being common the open-ended and the closed
questions.
5. Clustering algorithms
A classical hierarchical clustering algorithm (HCA) is included in Xplortext.
Clustering starts from the documents coordinates on the CA dimensions. An
exhaustive description of the clusters is provided, extracting their
characteristic words and looking for the differentiated behavior of the
variables in the clusters. The number of clusters is issued from the
hierarchical tree structure. An automatic suggestion is done.
A method for chronological constrained hierarchical clustering algorithm
(CCHCA) is also offered. Only chronological contiguous nodes can be
grouped. Further, the tree is described by the chronological words defined as
follows. The characteristic words of each node are identified but finally a
word is associated to only one node, the one that it better characterizes. These
words are used to label the nodes (Fig. 4). Although the tree could be used to
determine clusters, its main role is to allow for capturing the evolution of the
speeches and their vocabulary through a descending reading of the labels

24

JADT’ 18

and nodes of the tree.

Figure 4: Labeled chronological tree

6. Works in progress
The following features will be included in a next future:
 Chronological clustering (Legendre et al., 1985) has been proposed to
divide a chronological series of species (=species counts operated at
different moments) into homogeneous temporal parts. A same
aggregation criterion as in chronological constrained clustering is
used but a test is performed before aggregating two nodes to ensure
their homogeneity. If homogeneity does not exist, the corresponding
aggregation is not performed. As a result, the series is possibly
divided into non-connected sub-series. This clustering method has
been applied with benefit to the chronological series of words
corresponding to a chronological corpus, allowing for dividing the
corpus into non-connected homogeneous parts (Bécue-Bertaut et al.,
2014).
 Regularized CA (Josse et al.) allows for recovering a low-rank
structure from noisy data, such as textual data, by using
regularization schemes via a simple parametric bootstrap algorithm.
7. Conclusion
Xplortext is published on R CRAN. Bécue-Bertaut, et al. (2018) present a

JADT’ 18

25

series of applications of this package through several examples whose results
are interpreted in details. The corresponding bases and scripts are published
on the website http://xplortext.org.
References
Bécue-Bertaut M. and coll. (2018). Analyse textuelle avec R. Presses
Universitaires de Rennes (PUR), Rennes.
Bécue-Bertaut M., Kostov B., Morin A. and Naro G. (2014). Rhetorical
strategy in forensic closing speeches. Multidimensional statistics-based
methodology. Journal of Classification, 31: 85-106.
Bécue-Bertaut, M. and Pagès, J. (2004). A principal axes method for
comparing multiple contingency tables: MFACT. Computational Statistics
and Data Analysis, 45: 481-503.
Bécue-Bertaut M. and Pagès J. (2008). Multiple factor analysis and clustering
of a mixture of quantitative, categorical and frequency data. Computational
Statistics and Data Analysis, 52: 3255–3268.
Bécue-Bertaut M. and Pagès J. (2015). Correspondence analysis of textual
data involving contextual information: CA-GALT on principal
components. Advances in Data Analysis and Classification, 9: 125–142.
Bécue-Bertaut M., Pagès J. and Kostov B. (2014). Untangling the influence of
several contextual variables on the respondents’ lexical choices. A
statistical approach. SORT – Statistics and Operations Research Transactions,
38: 285–302.
Benzécri, J.-P. (1981). Pratique de l’Analyse des Données. Tome III. Linguistique &
Lexicologie. Dunod, Paris.
Josse J., Sardy S. and Wager S. (2016). denoiseR: A Package for Low Rank Matrix
Estimation. arXiv: 1602.01206
Kostov B. (2015). A principal component method to analyse disconnected
frequency tables by means of contextual information. (Doctoral
dissertation). Retrieved from
http://upcommons.upc.edu/handle/2117/95759.
Lebart, L., Salem, A. and Berry, L. (1998) Exploring textual data. Kluwer.
Legendre, P., Dallot, S. and Legendre, L. (1985). Succession of species within
a community: chronological clustering, with applications to marine and
freshwater zooplankton, American Naturalist, 125: 257–288.
ter Braak CJF. (1986). Canonical correspondence analysis: a new eigenvector
technique for multivariate direct gradient analysis. Ecology, 67: 1167–1179.

26

JADT’ 18

L'evoluzione delle norme: analisi testuale delle
politiche sull'immigrazione in Italia
Elena, Ambrosetti1, Eleonora Mussino2, Valentina Talucci3
1

Associate Professor, Sapienza Università di Roma
2 Associate Professor, Stockholm University
3 Researcher, ISTAT

1. Introduzione
Nei paesi del Sud-Europa, le politiche migratorie tendono a privilegiare le
questioni relative all'ingresso degli immigrati (ad esempio ingressi regolari e
irregolari, sanatorie e ricongiungimento familiare) rispetto agli aspetti legati
all'integrazione (Pastore 2004, Solé 2004). Questo squilibrio nell'azione
politica è imputabile alla volontà dei paesi di immigrazione di poter
controllare i flussi, bloccare gli ingressi non autorizzati e determinare il
numero e la composizione dei migranti. Le politiche migratorie regolano in
modo diretto l’esito dell’ingresso o meno nel Paese di destinazione e
successivamente orientano i percorsi di inserimento nel tessuto economicosociale e culturale degli stranieri ammessi in Italia. Attraverso lo studio delle
politiche dell’immigrazione dall’Unità d’Italia a oggi possiamo analizzare
come il linguaggio istituzionale nel corso degli anni e varie legislature si sia
trasformato tracciando diversi aspetti legati alle migrazioni internazionali nel
nostro paese. Questo argomento assume particolare importanza in quanto la
scelta di un tipo di linguaggio potrebbe influenzare opinioni e atteggiamenti
nei confronti degli stranieri da parte della popolazione italiana.
2. Le politiche migratorie in Italia
L’Italia, sebbene sia diventata un paese di immigrazione negli anni Settanta,
soltanto nel 1986 si è dotata della prima normativa sull’immigrazione a
seguito dell’adesione nel 1975 alla Convenzione alla Convenzione 143
dell'Organizzazione Internazionale del Lavoro (OIL) e dell'aumento dei flussi
di immigrati nel corso degli anni Ottanta. La legge 943/1986 (Legge Foschi)
riguardava in primo luogo lo status dei lavoratori, inoltre includeva il
ricongiungimento familiare e l'accesso allo stato sociale di base (Colombo e
Sciortino, 2004). La legge venne indirizzata ai lavoratori extra-comunitari,
con l’obiettivo di equipararli ai lavoratori italiani e ai lavoratori dell'Unione
europea (Nascimbene, 1988; Colombo e Sciortino, 2004). Inoltre la legge
introdusse una sanatoria per i lavoratori extracomunitari che si trovavano già
nel territorio senza documenti regolari. Nel febbraio 1990, la legge 39/1990
(Legge Martelli) fu approvata dal Parlamento italiano a seguito delle

JADT’ 18

27

pressioni dovute all’incremento degli arrivi dopo la caduta della cortina di
ferro e dalla imminente ratifica del Trattato di Schengen (ratificato nel 1993 e
entrato in vigore nel 1997). Al contrario della precedente legge Foschi, la
legge si rivolgeva a tutte le categorie di migranti e non solo quindi ai
lavoratori, per cui è considerata la prima legge organica sulle migrazioni.
Nonostante ciò essa viene ricordata principalmente per la sanatoria di circa
218.000 irregolari. Ricordiamo anche alcuni altri aspetti significativi coperti
dalla legge Martelli: l’introduzione dell’obbligo di visto, con conseguente
inasprimento del controllo delle frontiere, che rese molto più difficile entrare
in Italia, la programmazione annuale delle quote di lavoratori
extracomunitari attraverso il cosiddetto Decreto Flussi, l’asilo politico, e da
ultimo, l’inasprimento delle condizioni per l’ottenimento ed il rinnovo del
permesso di soggiorno. Nel 1995 fu emanata la legge 489/1995 (Legge Dini):
essa conteneva ulteriori misure restrittive per il controllo delle frontiere, una
nuova sanatoria per i lavoratori stranieri irregolari e la regolamentazione dei
flussi di lavoratori stagionali. A differenza delle misure restrittive, che non
trovarono attuazione in quanto ritenute contrarie alla Costituzione, la
sanatoria rappresentò il vero successo del decreto Dini, con un numero di
stranieri regolarizzati pari a 248.000 persone.
Nel 1997, con l’entrata in vigore dell'accordo di Schengen è stato introdotto
nell’ordinamento italiano l'adeguamento alla politica comune in materia di
visti. Sempre in tema di normative comunitarie, la legge 209/1998 ha
ratificato il trattato di Amsterdam, entrato in vigore in Italia in quell'anno.
Nello stesso anno il governo ha approvato il Testo Unico delle disposizioni
concernenti la disciplina dell'immigrazione e norme sulla condizione dello
straniero, Dlgs 286/1998 (Legge Turco-Napolitano). Obiettivo della legge era
quello di operare una rottura con il passato e di condurre ad una gestione del
fenomeno migratorio strutturale e di lungo periodo. La legge era basata su
quattro pilastri (Zincone e Caponio, 2004): 1. Prevenzione e lotta all’
immigrazione irregolare: da notare in particolare l’introduzione
dell'espulsione immediata di migranti irregolari e di centri di permanenza
temporanea per detenere immigrati clandestini in attesa di espulsione; 2.
Migrazioni da lavoro: i nuovi arrivi di lavoratori stranieri sono regolati con
quote annuali di lavoratori stabilite ogni anno dal Ministero del lavoro; viene
introdotto il meccanismo dello sponsor secondo il quale un cittadino italiano
o uno straniero residente si fa garante dell’ingresso di uno straniero privo di
contratto di lavoro; 3. Promozione dell'integrazione di migranti già residenti
in Italia: creazione del Fondo Nazionale per l’integrazione dedicato al
finanziamento di attività multiculturali e ad azioni antidiscriminazione;
introduzione del permesso di soggiorno di lungo periodo, o carta di
soggiorno per i migranti residenti da almeno 5 anni in Italia; 4. Concessione

28

JADT’ 18

di diritti umani fondamentali, come l'assistenza sanitaria di base, ai migranti
irregolari. La legge Turco-Napolitano si fece carico della regolarizzazione di
217.000 stranieri.
Nel 2002 è stata introdotta la legge Bossi-Fini, con lo scopo di modificare in
maniera restrittiva il Testo unico del 1998. Più specificamente, la legge ha
modificato i primi due pilastri della legge. Con la nuova normativa sono state
adottate una serie di misure volte a scoraggiare l’insediamento permanente
dei migranti tra le quali: l’abolizione del sistema dello sponsor, la riduzione
del periodo di validità del permesso di soggiorno e il collegamento della
validità del permesso di soggiorno a un contratto di lavoro ("contratto di
soggiorno"). Inoltre fu adottata una politica più repressiva nei confronti dei
migranti irregolari che includeva l’applicazione del rimpatrio forzato,
controlli più sistematici della polizia che includevano il pattugliamento delle
coste italiane e la detenzione di coloro che rimanevano sul territorio italiano
più a lungo di quanto previsto dal permesso di soggiorno (over-stayers). In
linea con le precedenti leggi, la legge 189/2002 ha regolarizzato 634.728
immigrati, rappresentando la più grande sanatoria mai adottata in Europa
fino a quel momento (Zincone, 2006). Dopo il 2002 sono state apportate
poche modifiche alla normativa sulla migrazione, si tratta in particolare di:
misure per combattere l'immigrazione clandestina, sanatorie per migranti
irregolari presenti sul territorio italiano e recepimento di direttive UE che
implicano modifiche alla normativa esistente.
L'acquisizione della cittadinanza per nascita (jus sanguinis) e per residenza
(jus soli) era inizialmente regolata dalla legge 555/1912. Le condizioni erano
molto restrittive: la cittadinanza era concessa solo al figlio di un uomo
italiano e sotto condizioni specifiche al figlio di una donna italiana. La legge
123/1983 ha introdotto nella legislazione italiana l'acquisizione della
cittadinanza per matrimonio e ha riformato l'acquisizione della cittadinanza
per nascita, concedendo indifferentemente il diritto di cittadinanza al figlio di
madre o padre italiani. L'acquisizione della cittadinanza italiana è stata
ulteriormente riformata dalla legge 91/1992 riservando particolari diritti ai
cittadini europei rispetto agli extra europei. La cittadinanza per matrimonio è
stata riformata nel 2009 (legge 94 del 15 giugno), prolungando il periodo di
residenza necessario in Italia da sei mesi a due anni dalla data del
matrimonio. Negli ultimi anni si contano diversi tentativi per introdurre una
nuova normativa sulla cittadinanza allo scopo di semplificare e ridurre il
tempo per l’ottenimento della cittadinanza per i migranti di seconda
generazione (nati in Italia). Come primo risultato, l'art. 33 del decreto 69/2013
ha semplificato la procedura di acquisizione della cittadinanza per gli
stranieri nati in Italia. Nonostante ciò, fino ad oggi manca una nuova
normativa in materia.

JADT’ 18

29

La normativa sulle migrazioni in Italia è stata costantemente caratterizzata
dalla mancanza di una politica attiva degli ingressi e dal continuo tentativo
di rallentare ed osteggiare il radicamento giuridico e sociale della
popolazione straniera sul territorio italiano. Il ricorso continuo a strumenti
ex-post come le sanatorie, l’utilizzo delle quote come sistema di emersione di
lavoratori stranieri già presenti sul territorio italiano piuttosto che come
norma di ingresso di nuovi lavoratori, ed il forte accento che la classe politica
e i media pongono sulla lotta all’immigrazione illegale sono esempi
emblematici di come il fenomeno migratorio in Italia venga affrontato in
termini di contenimento e controllo e non di allargamento e integrazione. La
presenza straniera, ancora oggi, è perlopiù considerata transitoria e viene
percepita e gestita in termini di risposta ad eventi contestuali di emergenza.
3. Dati e Metodi
I dati testuali utilizzati per realizzare questo lavoro sono tutti i capi normativi
contenuti nelle leggi approvate in Italia dal 1912 al 2014 in materia di
migrazione. La metodologia di analisi proposta fa capo alla Content Analysis
realizzata attraverso tecniche automatizzate dei dati. Si effettua applicando
un insieme di routine, supportate da specifici software in questo caso
TaLTAC2 – Trattamento automatico Lessico testuale per l’Analisi del
contenuto - che consentono di automatizzarne in parte o del tutto
l’esplorazione, la descrizione e il trattamento di grosse moli di dati; in questo
modo vengono trasformati insiemi di testi non strutturati in insiemi di testi
strutturati. Oltre alla descrizione dei contenuti del testo è possibile analizzare
il corpus in base ad una o più variabili disponibili sui frammenti come l’anno
e la maggioranza di governo1. L’estrazione dell’informazione peculiare
individuata attraverso il test del p-value permetterà di avere, per ogni
variabile esplicativa, una lista di parole chiave sovra o sotto rappresentate
rispetto a un modello di riferimento. Inoltre tramite l’analisi delle

Casa delle libertà: centro destra: Governo Berlusconi II, XIV Legislatura (30
maggio 2001 - 27 aprile 2006); Coalizione di centro destra: Governo Berlusconi IV, XVI
Legislatura (dal 29 aprile 2008 al 23 dicembre 2012); Grande coalizione: XVII
Legislatura Governi Letta e Renzi, centro sinistra e Alternativa popolare; Indipendente:
Governo Dini - (17/01/1995 - 17/05/1996) governo tecnico; Indipendenti: Governo Monti
(dal 16 novembre 2011 al 27 aprile 2013) Governo tecnico, XVI Legislatura; Liberale:
Governo Giolitti (1911-1914), UL - PR - PDC - PD - UECI – CC, centro destra; L’Unione:
centro sinistra, XV Legislatura (28 aprile 2006 - 6 febbraio 2008) Governo Prodi II;
Pentapartito: Coalizione politica: DC - PSI - PSDI - PRI -PLI, IX Legislatura;
Quadripartito: Coalizione politica: DC - PSI – PSDI - PLI, X Legislatura; Ulivo: centro
sinistra, XIII Legislatura.
1

30

JADT’ 18

corrispondenze lessicali cerchiamo un pattern che metta in relazione in modo
sistematico i lemmi e le dimensioni identificate con le caratteristiche associate
ad ogni legge.
4. Risultati
Le leggi sono state analizzate come un unico corpus che soddisfa i criteri
standard di dimensione minima richiesta affinché le analisi siano robuste. Ad
una prima analisi lessicometrica il testo, costituito da 150.714 occorrenze e
8.113 forme grafiche, rassicura sulla sua adeguata estensione: la proporzione
di parole diverse sul totale delle occorrenze (V/N*100= 5,383) si allontana
notevolmente dalla soglia del 20% rispettando, quindi, la soglia minima di
significatività statistica di un corpus (Bolasco, 1999). Sorprendentemente il
livello di ricercatezza del linguaggio non è particolarmente elevato, come si
vede dalla percentuale di hapax (V1/V*100) e dal coefficiente a di Zimpf
rispettivamente 28,350% e 1,325. Guardando il vocabolario, la prima parola
non vuota è comma (1529) seguita da numero (1160) e articolo (1066). Le altre
parole tema, ovvero quei sostantivi che compaiono con maggiore frequenza
nel testo, sono straniero, decreto, Stato, disposizioni, ingresso, territorio e
soggiorno.
Abbiamo poi eseguito un confronto tra il nostro vocabolario e il “lessico del
discorso programmatico di Governo” (Bolasco, 1999) per individuare quanto
fosse peculiare il linguaggio del nostro corpus anche rispetto a un
vocabolario tecnico-legislativo. Da questo confronto abbiamo ottenuto uno
“scarto” che indica quanto la forma in questione sia sovra (postivo) o sottorappresentata (negativo) rispetto al modello di riferimento Bolasco (1999);
più lo scarto è alto più le forme sono definite peculiari rispetto al testo
analizzato, ovvero lo caratterizzano. Senza entrare nel merito delle parole
chiave legate al vocabolario prettamente giuridico (come decreto, lettera),
emerse già dalla gerarchia delle occorrenze, si possono analizzare le altre
principali dimensioni del testo: oltre alla parola straniero la prima
dimensione che emerge è quella di frontiera (ingresso, territorio, frontiera,
accesso, durata) e di esercizio di diritto (regolamento, autorizzazione,
disposizioni). Ma la dimensione più corposa è quella delittuosa (pena, delitti,
reato, reati, tribunale, sentenza, condanna, violazione, esecuzione). Fa riflette
invece come le parole sottorappresentate siano governo, politica, pubblico,
parlamento: ovvero quelle legate alla dimensione legislativa.
Partendo dall’ipotesi che il linguaggio sia cambiato nel tempo abbiamo
effettuato un’analisi delle specificità (vedi tavola 1). Quando una parola è
sovra-rappresentata si parlerà di forma caratteristica (o specificità positiva),
al contrario quando essa è sottorappresentata parleremo di specificità
negativa; le forme prive di specificità in quel gruppo si definiscono banali,

JADT’ 18

31

mentre quelle che non sono specifiche di nessun gruppo sono considerate
appartenenti al vocabolario di base del corpus (Bolasco, 1999).
Tavola 1: Specificitá positive per anno di legislatura
1912
1986
1990
1992
1995
1998
2000
cittadinanza
lavoro
entrata
cittadinanza
lavoro
lavoro
visto
legge
lavoratori
permesso
straniera
soggiorno
soggiorno
ottenimento
Stato
immigrati
materia
Stato
permesso
stranieri
professionale
italiana
extracomunitari frontiera
italiana
entrata
permesso
consente
residenza
sociale
lavoro
figlio
penale
autonomo
seguito
cittadino
lavoratore
previdenza
estero
motivi
attuazione
transito
presidente della
estero
previdenza
extracomunitari cittadino
stagionale
motivi
Repubblica
servizio
autorizzazione cittadini
servizio
tempo
attivita'
autonomo
Governo
collocamento
decreto
militare
sociale
sociale
tempo
figli
materia
previdenza
durata
straniera
extracomunitariostranieri
figlio
italiani
apolidi
italiano
caso
europea
societa'
figli
occupazione
soggiorno
entrata
lire
estero
visti
militare
diritti
interno
residenza
legislativo
modalita'
sussistenza
padre
consulta
quanto
acquista
previdenza
regolarmente
requisiti
matrimonio
entrata
prima
età
pubblica
pubblica
importo
2002
2004
2007
2008
2009
2010
2011
lavoro
convalida
soggiorno
prevenzione
penale
conoscenza
seguente
decreto
successive
permesso
pubblico
codice
test
espulsione
soggiorno
legislativo
periodo
legislativo
procuratore
lingua
questore
testo
euro
sensi
ricerca
entrata
italiana
penale
legislativo
modificazioni
legislativo
penale
interno
permesso
termine
permesso
giudice
familiari
sensi
pubblico
lungo
provvedimento
penale
provvedimento volontariato
procuratore
giudiziario
svolgimento
permesso
asilo
seguenti
unico
procedura
imputato
modalità
periodo
soggiornanti
giudice
interno
commi
ricongiungimentoin presenza di europea
in presenza di decorrere
familiare
funzioni
domestico
europeo
rimpatrio
stagionale
provvedimenti motivi
sicurezza
seguenti
prefettura
respingimento
codice
parole
durata
persona
parole
sistema
lettera
caso
accompagnamen lungo
ricercatore
comma
legislativo
legislativo
autoritÃ
composizione nazionale
sostituite
guida
istruzione
allontanamento
procedura
decisione
rilasciato
antimafia
legislativo
rilascio
misure

Dalle specificità ottenute analizzando l’andamento del linguaggio nel tempo
emerge come si sia iniziato a scrivere di migrazioni parlando di cittadinanza
e residenza, introducendo progressivamente concetti connessi al lavoro e
all’essere extra-comunitario arrivando da un lato a temi di integrazione e
dall’altro a temi di criminalizzazione dello straniero. Il panorama lessicale nel
tempo si è arricchito ma anche “estremizzato”. Questa “estremizzazione”
potrebbe essere il risultato delle diverse coalizioni/maggioranze e quindi non
solo legato ad una dimensione temporale ma ancor di più politica, per questo
motivo é bene analizzare le due dimensioni contemporaneamente.

32

JADT’ 18

5. Dimensioni lessicali
L’analisi delle corrispondenze lessicali2 è stata condotta sui primi 50 lemmi
estratti dal confronto tra i lemmi dei verbi del nostro vocabolario e quelli del
“lessico del discorso programmatico di Governo”. Attraverso l’analisi delle
corrispondenze abbiamo riassunto la diversità del lessico utilizzato nelle
diverse leggi rispetto all’anno e la coalizione di governo. I primi due assi
fattoriali, proiettati in figura 1, rappresentano il 46% della variabilità
spiegata. La prima dimensione, rappresentata dal primo fattore, è
caratterizzata dalla dimensione temporale. Fatta eccezione per il 1992 e il
2007 tutte le leggi approvate dopo il 2002 si contrappongono a quelle
precedenti. Il secondo asse è caratterizzato dalla contrapposizione del partito
Liberale (Governo Giolitti 1911-1914) e Quadripartito (Governo Andreotti VII
1991-1992), in contrapposizione alle altre maggioranze di Governo. Le
coordinate ci permettono di proiettare le classi e le forme grafiche sul piano e
il posizionamento ci permette di individuare e interpretare i profili a seconda
della vicinanza dei punti.

Figura 1: Dimensioni lessicali delle interviste, rappresentazione del primo piano fattoriale

Andando a vedere più in dettaglio i quadranti possiamo notare che nel
primo, in cui si collocano il Quadripartito e gli anni 1992 e 2010, le forme
grafiche che contraddistingono lo spazio fanno riferimento alla dimensione
culturale “lingua” e “conoscenza”. Le forme grafiche proiettate nel secondo
piano, carattarezzato dagli anni 2002 e a seguire e dalla Casa delle libertá, la
grande-coalizione e gli indipendenti del governo tecnico Monti, esprimono

2

Con l’ausilio del programma Spad, nello specifico con il metodo CORBIT

JADT’ 18

33

principalmente gli aspetti legati alla delittuosità (e.g. violazioni, delitti, reato,
pena) e giuridica (e.g. norme, tribunale, giudice, esecuzione). A cavallo del
primo e secondo quadrante troviamo anche la coalizione di Centro-destra.
Nel terzo quadrante troviamo gli anni 1986, 1990, 1995, 1998, 2007 e il
governo tecnico Dini con l'Unione e il Pentapartito. Le forme grafiche su
questo piano identificano le caratteristiche del soggiorno come: carta, durata,
status, temporanea. Mentre la dimensione di frontiera caratterizza il quarto
quadrante: territorio, frontiera, legale, autorizzazione. A cavallo di queste
due dimensioni si trovano il mondo del lavoro e quello associativo che sono
parte integrante del percorso migratorio in Italia; non sorprende quindi che
caratterizzano sia il terzo che il quarto quadrante.
6. Conclusioni
L’obiettivo di questo lavoro era di esplorare il panorama legislativo in
riguardo alle migrazioni in un’ottica statistica, con l’obiettivo di estrarre le
sue caratteristiche e le sue peculiarità. In questa prospettiva le differenze
linguistiche, temporali e soprattutto dei diversi esecutivi rappresentano
un’interessante bacino informativo per investigare l’evoluzione semantica
delle norme. Seppur descrittivo questo lavoro assume una particolare
importanza in quanto la scelta di un tipo di linguaggio potrebbe influenzare
opinioni e atteggiamenti nei confronti degli stranieri da parte della
popolazione italiana. I nostri risultati mostrano che il panorama lessicale
della normativa italiana sull’immigrazione dal 1912 al 2014 è notevolmente
mutato. In primo luogo, dal punto di vista delle specificità ottenute
analizzando l’andamento del linguaggio nel tempo è emerso che
inizialmente, quando l’Italia era un paese di emigrazione, la normativa sulle
migrazioni era caratterizzata da temi quali la cittadinanza e la residenza.
Dagli anni Ottanta del secolo scorso, con l’incremento dei flussi migratori in
entrata nel nostro paese, sono stati introdotti progressivamente concetti
connessi al lavoro e all’essere extra-comunitario. Alla fine degli anni Novanta
del secolo scorso, a seguito del netto incremento degli arrivi di stranieri in
Italia, si inizia a parlare di integrazione e di ricongiungimento familiare.
Infine a partire dagli anni duemila inizia il processo di “criminalizzazione”
dello straniero pertanto entrano nel vocabolario specifico temi quali
sicurezza, respingimento, allontanamento. In secondo luogo, l’analisi delle
corrispondenze fattoriali ha confermato che a partire dal 2002 (Legge BossiFini) vi è stato un netto cambiamento del linguaggio usato nella normativa
dell’immigrazione, il linguaggio è infatti caratterizzato sempre più da temi
legati alla sicurezza e alla legalità. Inoltre il linguaggio usato, è stato
senz’altro influenzato da altri fattori che qui non abbiamo preso in
considerazione come per esempio, il recepimento delle politiche europee

34

JADT’ 18

sull’immigrazione, la situazione geo-politica internazionale, l’incremento
degli atti terroristi di matrice islamista a partire dagli attentati negli Stati
Uniti l’11 settembre 2001. Con questo lavoro abbiamo delineato un panorama
lessicale che ha cambiato direzione orientandosi sempre di più verso temi di
regolamentazione e contenimento (espulsione, allontanamento irregolare).
Esso ha confermato un approccio negativo riguardo alle migrazioni
indipendentemente dalla maggioranza di governo.
Bibliografia
Bolasco S. (1999), Analisi multidimensionale dei dati, Carocci Roma
Colombo, A., & Sciortino, G. (2004). Alcuni problemi di lungo periodo delle
politiche migratorie italiane. Le Istituzioni del Federalismo, 5, 763–788.
Nascimbene, B. (1988). Lo Straniero nel diritto italiano. Milano: Giuffré
Editore.
Pastore, F. (2004). A community out of balance: nationality law and
migration politics in the History of post-unification Italy. Journal of
Modern Italian Studies, 9(1), 27–48.
Solé, C. (2004). Immigration policies in southern Europe. Journal of Ethnic
and Migration Studies, 30(6), 1209–1221.
Zincone, G., & Caponio, T. (2004). Immigrant and immigration policymaking: the case of Italy. IMISCOE Working Paper Country Report.
Amsterdam: IMISCOE.
Zincone, G. (2006). The making of policies: immigration and immigrants in
Italy. Journal of Ethnic and Migration Studies, 32(3), 347–375.

JADT’ 18

35

A bibliometric meta-review of performance
measurement, appraisal, management research
Massimo Aria1, Corrado Cuccurullo2
University of Naples Federico II– aria@unina.it
University of Campania L. Vanvitelli – corrado.cuccurullo@unicampania.it
1

2

Abstract
Performance measurement, appraisal, and management have become one of
the most prominent and relevant research issues in in management studies.
The emphasis on empirical contributions has resulted in voluminous and
fragmented research streams. Thus, synthesizing the research literature is
relevant for effectively using the existing knowledge base, advancing a line
of research, and providing evidence-based insights.
In this paper, we propose a bibliometric meta-review that offers a different
knowledge base for future research agenda with implications also for
teaching and practice. We analyze the performance management literature
through a bibliometric analysis of reviews recently published (2000 - 2017) in
the scientific journals of domains, such as Management, Business and
Operations. The main purpose is to map and understand the intellectual
structure through co-citation analysis.
Keywords: Science Mapping; Content Analysis; Bibliometrix; Performance
Measurement.
1. Introduction
Performance measurement, appraisal, and management have become one of
the most prominent and relevant research issues in in management studies.
They are an ongoing topic of conferences and of books and journal articles as
well as of professional and popular grey literature. Researches on these
topics have been conducted in different sectors and for various
organizations, including public and professional ones. While the number of
academic publications on these topics is increasing at a rapid pace, the
emphasis on empirical contributions has resulted in voluminous and
fragmented research streams that hampers the ability to accumulate
knowledge and actively collect evidence through a set of previous research
papers. So, literature reviews are increasingly assuming a crucial role in
synthesizing past research findings to effectively use the existing knowledge
base, advance a line of research, and provide evidence-based insight into the
practice of exercising and sustaining professional judgment and expertise.
Among the different qualitative and quantitative reviewing, bibliometrics

36

JADT’ 18

has the potential to introduce a systematic, transparent, and reproducible
review process based on the statistical measurement of science, scientists, or
scientific activity.
In this paper, we propose a bibliometric “review of reviews” (meta-review)
that offers a different knowledge base for future research agenda with
implications also for teaching and practice. The goal of this article is to find a
path and to take stock of the existing knowledge in performance
measurement, appraisal, and management research.
2. Research Synthesis on performance measurement, appraisal and
management
2.1 Overcoming semantic ambiguity
‘‘Performance’’ is a complex concept and can be seen from different angles.
It is a multi-dimensional construct, the measurement of which varies
depending on a variety of factors. For example, it is important to determine
whether the measurement objective is to assess performance outcomes or
behavior at organizational or individual levels, in financial terms or
multidimensional ones (e.g. balanced scorecard framework), as
intermediate or final consequence of a managerial action. In very general
terms, performance is the contribution (result and how to achieve the result)
that an entity (individual, group of individuals, organizational unit,
organization, program, or public policy) provides through its action towards
achieving the aims and objectives and also the satisfaction of the needs for
which the organization was formed.
While measurement concerns performance indicators and appraisal is the
process of evaluating the performance of individuals and teams,
performance management is a systematic process for improving
organizational performance by developing the performance of individuals
and teams. It is a means of getting better results by understanding and
managing performance within an agreed framework of planned goals,
standards and competency requirements.
2.2 The need of a meta-review
In this work we analyze the performance management literature through a
bibliometric analysis of literature reviews recently published (2000 - 2017) in
the scientific journals of domains, such as Management, Business and
Operations.
The main purpose is to map and understand the intellectual structure
through co-citation analysis of this recent and evolving macro-topic,
highlighting internal clusters. The main contribution is to understand better
the state of art in terms of gaps, divergences, commonalities and tendencies

JADT’ 18

37

in which the field is going on. So, we provide a map to scholars in
positioning their future research work and to teachers to introduce so vast
topic to students.
This field of research is well suited to a bibliometric meta-review for the
following reasons:
1. there is little consensus among scholars. For example, Franco-Santos et
al. (2007) have counted 17 different definitions for business performance
measurement system, while Taticchi et al. (2010) almost 25 diverse
frameworks.
2. the field is deeply multidisciplinary. The most widely cited authors
come from a variety of different disciplinary backgrounds, such as
accounting, strategy, operations management and research, human
resources. The scholars’ background diversity brings different research
questions, theoretical bases and methodological approaches. The
functional silos, through which research on performance management is
developing, impede to have a coherent and agreed body of knowledge.
Understanding deeply the intellectual structure of the field and its
evolution is a relevant challenge for researchers.
3. there is a community of dedicated scholars around the world that share
the same agenda (cohesion in dominant issues) but use divergent
theoretical approaches and methods.
4. the field is still relatively immature. As in terms of age it is relatively
young, the limited professionalization is not surprising. In addition, there
is not a reference journal as Strategic Management Journal for strategy
scholars. In this case, our study can be contributive, showing the gaps in
literature and providing some guidelines for researchers.
5. common accepted performance management practices do not exist
(Richard et al., 2009). In many contexts performance management is
dysfunctional, although this problem is known since more 50 years
(Ridgway, 1956). We still miss more robust empirical and theoretical
analysis of performance management frameworks and methodologies.
Empirical investigations of the performance impact of frameworks,
including the most diffused balanced scorecard, have failed to offer
uncontroversial findings (Banker et al., 2000; Ittner et al., 2003; Neely et
al., 2004). Some authors call for further and longitudinal studies for
understanding the social influences and implications, but they do not
show which paths follow.
6. some publications assumed seminal roles in the evolution of the
scientific field. These articles, owing to their impact, are accelerating
factors of development of the field (Berry, Parasuraman 1993). It is
therefore important to identify what are the most influential performance

38

JADT’ 18

management articles published between 1991 and 2010, to understand
better the state of art and discover the linkages among authors.
7. there is an extended spectrum of this research field and an increased
intensity of research, but most part of it also confirms the
incompleteness and inconsistence of results. There are still various open
issues and unsolved problems. This depends on the fragmentation of the
field of research, on different disciplinary membership of researchers and
their cultural context. This diversity implies the use of different theories
and methods and therefore also the emergence of different dominant
themes.
8. a profound and rapid evolution is taking place. Not only the research
has shifted from the financial performance to the multidimensional one,
but a shift of scholars’ attention from the organizational to the individual
performance is under way. Moreover, another significant shift is ongoing.
While earlier research was often normative, founded on economic
rationality, more recent research is more analytical and explanatory
(Cuccurullo et al., 2016).
The overwhelming volume and variety of new information, conceptual
developments, and data are the milieu where bibliometrics becomes useful
by providing a structured, more objective and reliable analysis to present the
“big picture” of extant research.
3. Methods
Our bibliometric meta-review is a quantitative research synthesis of the
reviews published on the same topic that we conducted with bibliometrix
(Aria, Cuccurullo, 2017), a unique tool, developed in the R language, which
follows a classic logical bibliometric workflow.
3.1 Data collection
For data retrieval, we used the Social Science Citation Index (Indexes=SCIEXPANDED, SSCI) of Clarivate Analytics Web of Science. It is the most used
database of scientific knowledge by management scholars (Zupic, Čater,
2015). Our search terms were (TS=(("performance manag*") OR
("performance measur*") OR ("performance apprais*"))). We applied our
search keyword to the Timespan=2000-2017 and filtered findings for
language (English) and document types (Review). Therefore, we found 783
reviews. Then, we refined our search by categories (Management or Business
or Operation Research Management Science) and obtained 167 reviews.
Finally, we selected all the reviews published in the most authoritative
journals as ranked as 3, 4, 4* by ABS 2015: We dropped off 31 journals for a
total of 50 reviews. Our final dataset is formed by 117 reviews.

JADT’ 18

39

3.2 Data analysis
Our effort at delineating the intellectual structure of the discipline involves
author co-citation analysis (ACA), a bibliometric technique that uses a matrix
of co-citation frequencies between authors as its input. This matrix is the
basis for various types of analyses.
ACA ability to reveal patterns of association between authors based on their
co-citation frequencies makes it a prospective methodology for
understanding the evolution of an academic discipline. Authors working in a
stream of research often cite one another as well as draw on common sources
of knowledge. Further, their works are likely to be frequently co-cited (i.e.,
cited together) by other authors working on intellectually similar themes. The
citations of seminal authors provide a basis for unraveling the complex
patterns of associations that exist among them as well as to trace the changes
in intellectual currents taking place over time.
4. Findings
4.1 Descriptive analysis
Our dataset includes 117 reviews published in 46 journals since 2000 (table 1
and 3). They received 105 citations on average (table 2). They show
fluctuating growth that reaches its peak every 5 years (table 3).
Table 1: Main Information about data
Articles

117

Sources (Journals, Books, etc.)

46

Keyword Plus – Author's Keywords
Period

770 – 383
2000 - 2017

Average citations per article

Authors

297

Authors of single
authored articles
Co-Authors per Articles

10
2.65

Collaboration Index

2.79

105.1

Table 2: Top manuscripts per citations
Paper

TC

TCperY
ear
71.1

1 BHARADWAJ AS,(2000),MIS Q.
2 DIAMANTOPOULOS A;SIGUAW JA, (2006), BRIT. J. MANAGE.

128
0
588

3 MELO MT et al.(2009), EUR. J. OPER. RES.

587

65.2

4 ZHOU P et al.(2008),EUR. J. OPER. RES.

429

42.9

5 WRIGHT PM;BOSWELL WR, (2002), J. MANAGE.

379

23.7

6 WRIGHT PM et al.(2005), PERS. PSYCHOL.

347

26.7

7 ZACHARATOS A et al..(2005), J. APPL. PSYCHOL.

305

23.5

49.0

40

JADT’ 18

8 ADAMS R et al.(2006), INT. J. MANAG. REV.

302

25.2

9 GIBSON C;VERMEULEN F, (2003), ADM. SCI. Q.

291

19.4

10 CARDOEN B et al. (2010), EUR. J. OPER. RES.

288

36.0

Table 3: Most Relevant Sources
Sources
1 J. OF MANAGEMENT

Article
s
11

2 INT. J. OF OPERATIONS & PRODUCTION MANAGEMENT;INT. J. OF
PRODUCTION ECONOMICS
4 EUROPEAN J. OF OPERATIONAL RESEARCH; INT. J. OF
MANAGEMENT REVIEWS
6 INT. J. OF HUMAN RESOURCE MANAGEMENT; INT. J. OF
PRODUCTION RESEARCH
8 J. OF BUSINESS ETHICS; STRATEGIC MANAGEMENT J.

8

10 BRITISH J. OF MANAGEMENT; J. OF APPLIED PSYCHOLOGY; J. OF
MANAGEMENT
STUDIES;
MANAGEMENT
ACCOUNTING
RESEARCH; OMEGA-INT. J. OF MANAGEMENT SCIENCE; SUPPLY
CHAIN MANAGEMENT

3

7
5
4

4.2 Co-citation network and cluster analysis
The objective of our paper is to identify the intellectual structure of the
performance measurement and management field. More specifically, our
goals are to (1) delineate the subfields that constitute the intellectual structure
of the field; (2) determine the relationships, if any, between the subfields; (3)
identify authors who play a pivotal role in bridging two or more conceptual
domains of research; and (4) graphically map the intellectual structure in a
network space in order to visualize spatial distances between intellectual
themes. In extreme synthesis, figure 1 shows:
1. A first cluster (red bubbles) represented by works concerning the system
of multidimensional performance measurement and evaluation. At its
center, we find prominent authors who contribute with specific
frameworks, such as the balanced scorecard (Kaplan, Norton, 1992,
1996) and performance prism (Neely et al., 1995). Next to them, we find
the contribution of Ittner et al. (2003) about one of the great problems in
the multidimensional measurement: the balance between subjectivity
and objectivity. Always central in the cluster, we find performance
system design (Neely et al., 2000), At the upper and lower extremes of
the cluster, we find other two issues of multidimensional performance
systems: strategic alignment (Chenhall, 2005) and the guidelines to
implement systems (Bititci et al., 1997).

JADT’ 18

2.

3.

4.

5.

41

A second cluster (blue bubbles) concerns the current prevailing
perspective of studying performance measurement and management:
the strategic one. In particular, figure 1 highlights the bridge
contributions of two cornerstones of the resource-based view (Barney,
1991, Wernelfelt, 1984).
In front of this cluster, in the upper-left part of the map, we find another
one (violet bubbles) that deals with theories, such as the agency theory:
(Eisenhardt, 1989; Jensen, 1976) - which are the main method of
investigation - (Carpenter, 2003), and psychology (Kahneman, 1979)
Two other neighboring clusters, located in the lower left part of the map,
concerns human resources. A first cluster (green bubbles) includes
almost entirely works published on Academy of Management. Their
preferred theme is perceptions of organizational performance (Delaney,
Huselid 1996). A second cluster concerns participation in the appraisal
from psychological perspective (Cawley et al., 1998; Keeping, Levy
2000).
One last cluster is isolated and concerns the studies of operation
research on performance measurement

Figure 1: Co-citation network of cited references

42

JADT’ 18

References
Aria, M. & Cuccurullo, C. (2017). bibliometrix: An R-tool for comprehensive
science mapping analysis, Journal of Informetrics, 11(4), pp 959-975.
Barney, J. (1991). Firm resources and sustained competitive
advantage. Journal of management, 17(1), 99-120.
Bititci, U. S., Carrie, A. S., & McDevitt, L. (1997). Integrated performance
measurement systems: a development guide. International journal of
operations & production management, 17(5), 522-534.
Cawley, B. D., Keeping, L. M., & Levy, P. E. (1998). Participation in the
performance appraisal process and employee reactions: A meta-analytic
review of field investigations. Journal of applied psychology, 83(4), 615.
Chenhall, R. H. (2005). Integrative strategic performance measurement
systems, strategic alignment of manufacturing, learning and strategic
outcomes: an exploratory study. Accounting, Organizations and
Society, 30(5), 395-422.
Cuccurullo, C., Aria, M., & Sarto, F. (2016). Foundations and trends in
performance management. A twenty-five years bibliometric analysis in
business and public administration domains, Scientometrics.
Delaney, J. T., & Huselid, M. A. (1996). The impact of human resource
management
practices
on
perceptions
of
organizational
performance. Academy of Management journal, 39(4), 949-969.
Eisenhardt, K. M. (1989). Agency theory: An assessment and
review. Academy of management review, 14(1), 57-74.
Ittner, C. D., Larcker, D. F., & Meyer, M. W. (2003). Subjectivity and the
weighting of performance measures: Evidence from a balanced
scorecard. The accounting review, 78(3), 725-758.
Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial
behavior, agency costs and ownership structure. Journal of financial
economics, 3(4), 305-360.
Kaplan, R. S., & Norton, D. P. (1992). In Search of Excellence. Harvard
manager, 14(4), 37-46.
Kaplan, R. S., & Norton, D. P. (1996). Using the balanced scorecard as a
strategic management system.
Keeping, L. M., & Levy, P. E. (2000). Performance appraisal reactions:
Measurement, modeling, and method bias. Journal of applied psychology,
85(5), 708.
Neely, A. (2005). The evolution of performance measurement research:
developments in the last decade and a research agenda for the
next. International
Journal
of
Operations
&
Production
Management, 25(12), 1264-1277.
Neely, A., Gregory, M., & Platts, K. (1995). Performance measurement system

JADT’ 18

43

design: a literature review and research agenda. International journal of
operations & production management, 15(4), 80-116.
Neely, A., Mills, J., Platts, K., Richards, H., Gregory, M., Bourne, M., &
Kennerley, M. (2000). Performance measurement system design:
developing and testing a process-based approach. International journal of
operations & production management, 20(10), 1119-1145.
Wernerfelt, B. (1984). A resource-based view of the firm. Strategic
management journal, 5(2), 171-180.

44

JADT’ 18

Textual Analysis of Extremist Propaganda and
Counter-Narrative: a quanti-quali investigation
Laura Ascone
Université de Cergy-Pontoise – laura.ascone@etu.u-cergy.fr

Abstract
This paper investigates the rhetorical strategies of jihadist propaganda and
counter-narrative in English and French. Since jihadist propaganda aims at
both persuading the Islamic State’s sympathisers and threatening its enemies,
attention was focused on the way threat and persuasion are verbalised. As
far as jihadist propaganda is concerned, the study was conducted on the
Islamic State’s two official online magazines: Dabiq, published in English, and
Dar al-Islam, published in French. As for the counter-narrative, the corpus
was composed of the articles published on the main English and French
governmental websites. Combining quantitative and qualitative approaches
allowed to examine the general characteristics as well as the specificities of
both jihadist propaganda and counter-narrative. The software Tropes was
used to analyse the corpora from a semantic-pragmatic perspective. The
results’ statistical validity was then verified and synthesised with the
softwares Iramuteq and R. This study revealed that the rhetorical strategies
varied between both jihadist propaganda and counter-narrative, and French
and English.
Keywords: jihadist propaganda, counter-narrative, discourse analysis, threat,
persuasion.
1. Introduction
The recent terrorist attacks by Daesh in Western countries have led
researchers and experts to examine the islamisation of radicalism (Roy, 2016).
Different studies have been conducted on the psychosociological contexts
that may lead someone to adhere to the jihadist ideology (Benslama, 2016;
Khosrokhavar, 2014), as well as on the role played by the Internet in the
radicalisation process (Von Behr, 2013). Yet, even though terrorism would
not exist without communication (McLuhan, 1978), the rhetorical strategies
of the jihadist propaganda have been neglected and remained unexplored.
This research investigates the rhetorical strategies of both jihadist
propaganda and counter-narrative published on the Internet in English and
French. More precisely, this analysis focuses on the way threat and
persuasion are expressed in jihadist discourse, as well as on the way French
government and international institutions face and counter jihadist

JADT’ 18

45

propaganda. From a linguistic perspective, threat and persuasion are
complex speech acts. Therefore, pragmatics and, more specifically, Searle’s
(1969) speech act theory, constituted the basis of this study. As far as jihadist
propaganda is concerned, the analysis was conducted on the Islamic State’s
two official online magazines: Dabiq, published in English, and Dar al-Islam,
published in French. As for the counter-narrative, the corpus was composed
of the articles published on the main French and English institutional
websites such as stopdjihadism.fr or counterjihadreport.com. The fact that
jihadist propaganda and counter-narrative address different readerships, led
us to hypothesise that differences in both content and form might be
identified between the two magazines, as well as among the different
governmental websites. Combining quantitative and qualitative approaches
(Garric and Longhi, 2012; Rastier, 2011) (that is, lexicometry and textometry
for the quantitative approach, and the interpretation of the text according to
the ideology behind it for the qualitative one), allowed to examine the
general characteristics as well as the specificities of both jihadist propaganda
and counter-narrative. Following Marchand’s (2014) work, the software
Tropes was used to analyse the corpora from a semantic-pragmatic
perspective. The results were then investigated in a qualitative way, and their
statistical validity verified with the softwares Iramuteq and R. The
combination of these two approaches allowed to overcome the limitations
imposed by both the software’s automatic analysis and the qualitative
subjective interpretation. By comparing the rhetorical strategies used in both
jihadist propaganda (Huyghe, 2011) and counter-narrative, the aim of this
research was to identify the linguistic differences between these two
discourses and these two languages, in order to determine the rhetorical
strategies that might prove efficient in countering jihadist propaganda. After
having presented the rhetorical pattern of jihadist propaganda, the linguistic
characteristics of English and French counter-narratives will be examined.
The jihadist and governmental rhetorical strategies will then be contrasted.
2. Corpus and methodology
2.1. Jihadist propaganda
The analysis of the rhetorical strategies in jihadist propaganda was
conducted on Daesh’s official online magazines Dabiq, published in English,
and Dar al-Islam, published in French. Since these two magazines address a
readership that has already adhered to the jihadist ideology, their goal is to
both reinforce the reader’s adhesion and incite him/her to act in the name of
the jihadist ideology. The reader is then incited to adopt the behaviour a
good Muslim should have, and to take revenge on who is presented by
Daesh as the responsible for the Muslims’ humiliation, that is the West. As

46

JADT’ 18

far as Dabiq is concerned, the corpus investigated was composed of all the
articles published on the first fourteen numbers (i.e. 377,450 words). As for
Dar al-Islam, the analysis was conducted on the first nine issues (229,762
words). To analyse the rhetorical strategies used in jihadist propaganda, a
quanti-qualitative approach was adopted (Garric and Longhi, 2012; Rastier,
2011). More precisely, this iterative approach was composed of five stages. A
first qualitative analysis of the jihadist ideology, the radicalisation process,
and the linguistic characteristics of hate speech and propagandistic discourse
was essential to the understanding of the jihadist discourse as well as to the
advancement of our first hypotheses. The second stage corresponded to a
quantitative analysis whose goal was to verify the validity of our hypotheses:
the corpus was then examined with the software Tropes, which allows to
investigate a text from a semantic perspective. More precisely, based on a
pre-established lexicon, the software identifies the themes tackled in the text,
and shows how these themes are linked to one another. The most frequent
themes in both magazines are religion and conflict. However, in order to study
the way threat and persuasion are expressed in the two corpora, a deeper
qualitative analysis was conducted on the themes sentiment for the French
corpus, and feeling for the English one (third stage). In other terms, the
quantitative analysis constituted the basis for a qualitative study, which was
then conducted only on the expressions conveying feelings. Because of their
size-difference, the nine issues of the French magazine count 318 sentimentexpressions, whereas Dabiq counts 705 feeling-expressions. Therefore, in order to
contrast the results, a normalisation was applied. Then, a quantitative
analysis was conducted with the software Iramuteq, which is an interface of
the software R and which performs statistical analysis of textual data based
on Reinert’s classification method. This way, it was possible to test the
hypotheses and results issued by the qualitative study (fourth stage).
Furthermore, a qualitative manual analysis of the first number of both Dabiq
and Dar al-Islam allowed to identify the propositions conveying threat,
persuasion, obligation, prohibition, and rewards, that had not been detected
by the software Iramuteq. This way, it was possible to provide a lexicon
specific to the corpus under investigation, that was not detected by the
software because of the special features of the jihadist discourse (fifth stage).
The combination and alternation of both quantitative and qualitative
approaches allowed to examine Daesh’s discourses in relation to the context
in which it is produced (Valette and Rastier, 2006).
2.2. Counter-narrative
The analysis on the rhetorical strategies in French and English counternarratives was conducted on the main governmental and institutional

JADT’ 18

47

websites. The French corpus was composed of the articles published on
www.stop-djihadisme.gouv.fr (the platform created after the first terrorist
attacks in France in 2015), www.interieur.gouv.fr (the website of the Minister of
Interior), and www.cpdsi.fr (the website of the Centre de Prévention contre les
Dérives Sectaires liées à l’Islam). The corpus counts 115,950 words. As far as
the English corpus is concerned, it was composed of the articles published on
www.counterjihadreport.com
(a
news
aggregating
website),
www.consilium.europa.eu (the website of the European Council and of the
European Union Council), www.ec.europa.eu (the website of the European
Commission), and on the Radicalisation Awareness Network (this is a
specific section of the website of the European Commission). The corpus
counts 116,000 words. In order to conduct comparable analyses, the same
quanti-qualitative approach was adopted. The qualitative analysis of the
geopolitical context and of the different campaigns used to face and counter
the jihadist radicalisation was essential to the understanding of both French
and English counter-narratives (first stage). Then, a quantitative analysis was
conducted with the software Tropes, which allowed to identify the most
frequent themes. The themes religion and droit (“law”) were the most present
in the French corpus, whereas the themes education and communication were
the most frequent in the English one (second stage). The third stage
corresponded to the qualitative analysis that was conducted on the category
sentiment, for the French corpus (292 propositions), and feeling, for the
English one (370 propositions). A normalisation was then applied to compare
jihadist and governmental discourse. A second quantitative analysis was
then conducted with the softwares Iramuteq and R to test the results issued by
the qualitative study (fourth stage). The results of the analysis on jihadist
propaganda and counter-narrative were then contrasted to compare the
rhetorical strategies used in jihadist propaganda and counter-narrative.
3. The rhetorical strategies used in French and English jihadist propaganda
The quantitative analysis conducted with the software Tropes, and the
qualitative study conducted on the categories sentiment and feeling, revealed
the components of the jihadist discourse. The propaganda of the Islamic State
is based on five key concepts: threat, persuasion, reward, obligation, and
prohibition. The assessment of interjudge agreement was necessary to
determine these five concepts as well as to categorise the different
propositions selected by Tropes as objectively as possible. Each category was
examined from both quantitative (i.e., its identification and distribution in the
two magazines, Dabiq and Dar al-Islam, using the softwars Tropes and
Iramuteq, and the corpus analysis toolkit AntoConc) and qualitative (i.e.,
analysing each concept in relation to the context in which it was produced)

48

JADT’ 18

perspectives. Yet, these five concepts are not independent from one another.
Rather, they are strongly linked to one another.

Figure 1: the rhetorical pattern of jihadist propaganda

Figure 1 shows the rhetorical pattern of jihadist discourse. Since Dabiq and
Dar al-Islam aim at manipulating the reader’s behaviour, jihadist propaganda
is based on obligations and prohibitions. Rewards as well as guilty feelings
towards the Muslims living in the Middle-East, aim at leading the reader to
respect these prescriptions. Not respecting them would mean facing negative
consequences. Threat may then be expressed against the members of the
Islamic State themselves and, more in general, against any Muslim.
Obligations are also exploited to impose the readership a hostile and violent
attitude against Western countries, which is justified by the feeling of
victimisation. Fighting against the Muslims’ enemy is presented by jihadists
as a heroic and valorising action, and therefore, a persuasive one.
Furthermore, not only are attractive factors rewards for the reader’s
obedience. They are sometimes presented as independent from the reader’s
behaviour. In other terms, persuasion is presented as a positive and
valorising act that, contrary to rewards, does not depend on whether the
reader respects or not the prescriptions imposed The sentence “Jihad is
necessary to obtain Allah’s forgiveness”, for instance, presents an obligation
(“it is necessary”) and a reward that will be granted if the obligation is
respected (“to obtain Allah’s forgiveness”). However, this sentence expresses
more than an obligation and a reward. Jihad, which is interpreted as attractive
by jihadists, tends to be associated with terrorist attacks and, consequently, it
will be perceived as threatening by Western countries. Furthermore, this

JADT’ 18

49

sentence implies that if the obligation is not respected, the individual will not
obtain Allah’s forgiveness. In other terms, this sentence indirectly expresses a
threat against the readership too.
4. The rhetorical strategies used in French and English counter-narratives
The large number of Daesh’s sympathisers and foreign fighters shows that the
communicative and rhetorical strategies adopted in Daesh’s propaganda
have an important and persuasive impact on the readership. On the contrary,
the counter-narrative produced by the different governments to face and
counter jihadist propaganda, has been criticised not to be as efficient as
jihadist propaganda. In the French corpus, 292 propositions conveying
sentiment (“feeling”) were identified, whereas 370 propositions conveying
feelings were identified in the English one.
The frequency of the five categories (i.e. of the propositions conveying threat,
persuasion, reward, obligation, and prohibition) was calculated in the French
and English corpora. The reward-category is the only one that was more
present in the French corpus than in the English one. Contrary to the Islamic
State’s propaganda, the propositions conveying rewards and prohibitions are
almost absent in both French and English counter-narratives. On the
contrary, what these two discourses have in common is the high frequency of
the propositions conveying threat (Example 1).
1. “Terrorist groups will continue to exploit the refugee crisis in their
propaganda, seeking to portray Western mistreatment of Muslims,
and inciting fear by alleging that their supporters are being
smuggled in amongst genuine refugees.” (RAN website)
As Example 1 shows, threat tends to be associated to the other (i.e., the Islamic
State), which implies that Western countries are presented as victims of the
Islamic State. In the English corpus, 355 occurrences of the word victim(s)
were identified. The corpus analysis toolkit AntConc showed that the most
frequent collocation of this term is the word terrorism (57 co-occurrences). On
the contrary, the French corpus, where the word victime/s occurs 70 times
only, presents only 2 co-occurrences of the term terrorisme. Rather, French
counter-narrative tends to talk about rescuing and helping victims
(secours/aide aux victimes). Furthermore, differences were identified between
the different websites in a same language.
Figure 2 shows the under- and overuse of the most representative terms in
two French governmental websites: stopdjihadisme and CPDSI. More
precisely, based on a Chi2 dependence test, the graph shows the words that
are significantly associated or “anti-associated” to the two websites. The
figure revealed that CPDSI website focuses more on the religious dimension.
The words islam, jihad and jihadiste (“jihadist”) are significantly associated to

50

JADT’ 18

this sub-corpus. This implies that jihad and jihadiste are presented and
interpreted as religious terms. On the contrary, the website of the
stopdjihadisme campaign is characterised by an overuse of the words terroriste
(“terrorist”), terrorisme (“terrorism”), Syrie (“Syria”), radicalisation
(“radicalisation”), Irak (“Iraq”), français (“French”), and France (“France”).
The overuse of these specific terms shows that the campaign and,
consequently, its website focus more on the geopolitical dimension, where
the radicalisation process is presented in relation to terrorism and not to
Islam.

Figure 2: under- and overuse of some key-terms in French counter-narrative

5. Conclusion
This comparative analysis revealed that jihadist discourse and counternarrative present both similarities and differences. As far as the differences
are concerned, the frequency of the propositions conveying threats,
persuasion, prohibitions, obligations, and rewards varied between these two
discourses: they were more frequent in counter-narrative than in jihadist
propaganda. The Islamic State’s propaganda aims at reinforcing the reader’s
adhesion to the jihadist ideology, and at inciting him/her to act against its
enemies in the name of the jihadist ideology. On the contrary, counternarrative does not aim at reinforcing an ideology. Rather, it aims at
countering the jihadist radicalisation. This difference was confirmed by the
variation of the different category-frequencies in jihadist propaganda and
counter-narrative. Despite this crucial difference, similarities between these
two discourses were identified. More precisely, both discourses present the
respective speakers’ communities as victims of the other and, consequently,
incite the readership to fight, whether violently or not, against the enemy. As
far as the methodology is concerned, the procedures adopted allowed to

JADT’ 18

51

investigate the general and special features of both jihadist and governmental
discourses. The results obtained in the quantitative analysis constituted the
starting point for a qualitative analysis, which permitted to identify the
features that had not been detected by the softwares as well as to refine
Tropes’s pre-established lexicon.
References
Angenot, M. (2008). Dialogue de sourds. Traité de rhétorique antilogique. Paris :
Mille et une nuits.
Benslama, F. (2016). Un furieux désir de sacrifice : le surmusulman. Paris :
Edition du Seuil.
Garric, N., & Longhi, J. (2012). L’analyse de corpus face à l’hétérogénéité des
données : d’une difficulté méthodologique à une nécessité
épistémologique. Langage, (3) : 3-11.
Huyghe, F.-B. (2011). Terrorismes : violence et propagande. Paris : Gallimard.
Khosrokhavar, F. (2014). Radicalisation. Paris : Editions de la maison des
sciences de l’homme.
Marchand, P. (2014), Analyse avec IRaMuTeQ de dialogues en situation de
négociation de crise : le cas Mohammed Mehra. Communication présentée
aux 12es Journées Internationales d’Analyse statistique des Données Textuelles,
Paris, 25.
McLuhan, M. (1978). The brain and the media: The “Western” hemisphere.
Journal of Communication, 28(4): 54-60.
Rastier, F. (2011). La mesure et le grain : sémantique de corpus. Champion ; diff.
Slatkine.
Roy, O. (2016). Le djihad et la mort. Le Seuil.
Searle, J. (1969). Speech acts: an essay in the philosophy of language. London:
Cambridge University Press.
Valette, M., & Rastier, F. (2006). Prévenir le racism et la xénophobie :
propositions de linguistes. Langues modernes, 100(2),68.
Von Behr, I. (2013). Radicalisation in the digital era: the use of the Internet in 15
cases of terrorism and extremism.

52

JADT’ 18

Analyse de données textuelles appliquée à des
problématiques de sécurité et d'enquête judiciaire
Laura Ascone1, Lucie Gianola1
1

AGORA, Université de Cergy-Pontoise – laura.ascone@etu.u-cergy.fr, lucie.gianola@u-cergy.fr

Abstract
This presentation investigates two cases of textual analysis applied to
security contexts: - the analysis of the rhetorical strategies adopted in the
Islamic State’s official online magazines: Dabiq, published in English, and
Dar al-Islam, published in French; - the use of methods for named entities’
automatic extraction, and the conception of a textual exploration software for
criminal analysis.
Résumé
Nous présentons deux cas d'application de l'analyse de données textuelles
dans des contextes liés à la sécurité :
- l'analyse des stratégies rhétoriques de propagande djihadistes à travers
l'étude des revues Dabiq et Dar-al-Islam,
- l'utilisation de méthodes d'extraction automatique d'entités nommées et la
conception d'un outil d'exploration textuelle pour l'analyse criminelle.
Keywords: analyse de données textuelles, radicalisation, analyse criminelle
1. Introduction
L'essor de préoccupations sécuritaires liées aux actes de terrorisme perpétrés
à travers le monde depuis le début du XXIème siècle pousse les chercheurs,
acteurs publics et sociaux à rechercher de nouveaux moyens d'analyse de ce
phénomène. En France, les sciences humaines et sociales se saisissent de la
question comme le démontre l'organisation de plusieurs journées d'études
sur la question (« Nouvelles figures de la radicalisation », Toulouse, avril
2017, « Les SHS face à la menace », Cergy, septembre 2017, « Des sciences
sociales en état d'urgence : islam et crise politique », Paris, décembre 2017).
Nous souhaitons présenter dans cet article deux sujets d'étude relatifs à ces
préoccupations sécuritaires : une étude de la rhétorique de Daesh du point de
vue du recours aux émotions dans les revues Dabiq (anglais) et Dar al-Islam
(français), ainsi qu'une collaboration entre le Pôle Judiciaire de la
Gendarmerie Nationale (PJGN) et l'Université de Cergy-Pontoise visant à
fournir de nouveaux outils d'analyse textuelle des procédures judiciaires aux
équipes d'analystes criminels. Le phénomène de la radicalisation djihadiste a
amené chercheurs et professionnels à examiner les raisons

JADT’ 18

53

psychosociologiques qui sont à la base de l'adhésion à l'idéologie djihadiste
(Khosrokhavar, 2014) ainsi que les stratégies adoptées par le groupe
extrémiste pour diffuser les messages de propagande (Lombardi, 2015).
Toutefois, bien qu'elles jouent un rôle crucial au sein de la propagande
djihadiste, les stratégies rhétoriques qui visent à menacer ou à persuader les
différents lecteurs restent inexplorées. La première partie de cette étude vise
donc à présenter une analyse quanti-qualitative du schéma rhétorique et des
émotions sur lesquels se base la propagande djihadiste. Dans la continuité
des travaux de Marchand (2014), les logiciels Iramuteq et Tropes ont permis
d’étudier le corpus d’un point de vue quantitatif. Les résultats issus de cette
analyse quantitative ont ensuite constitué le point de départ d’une analyse
qualitative sur les extraits exprimant des émotions, afin d’examiner plus en
détail les stratégies rhétoriques de la propagande djihadiste.
Le cas de l'analyse des procédures judiciaires nous confronte à une
problématique typique d'extraction d'information passant par la
reconnaissance automatique d'entités nommées : notre travail de recherche
consiste notamment à concevoir les bases d'un outil de navigation textuelle
ad hoc. Bien que les besoins des analystes criminels soient similaires à ceux
d'autres domaines d'application (analyse de la voix du client, traitement
automatique de la langue biomédicale, etc.), le contexte de l'enquête
judiciaire pose de nouvelles contraintes de précision dans l'extraction et dans
la mise à disposition des résultats à l'expert, c'est-à-dire à l'analyste criminel.
Le besoin social et institutionnel de nouvelles approches de documents
d'origines variées dans les contextes judiciaires et sécuritaires nous permet de
démontrer la pertinence de méthodes d'analyse de données textuelles déjà
éprouvées dans ces deux cas d'étude.
2. Description de la rhétorique djihadiste : cas des revues Dabiq et Dar alIslam
2.1. Corpus et méthodologie
Cette recherche a été menée sur les deux revues de Daech : Dabiq, publié en
anglais, et Dar al-Islam, publié en français. Dabiq s’adresse aux sympathisants
non arabophones de Daech, tandis que Dar al-Islam, qui n’est pas une
traduction de Dabiq, s’adresse à un lectorat uniquement francophone. Cette
distinction nous conduit à avancer l’hypothèse que les deux revues diffèrent
dans leur contenu ainsi que dans la forme du message qu’elles portent.
Toutefois, l’une et l’autre s’adressent à un lectorat qui a déjà adhéré à
l’idéologie islamiste. Leur objectif n’est donc pas de persuader le lecteur de
s’approcher de l’islamisme, mais de renforcer son adhésion et de l’amener à
agir au nom de cette idéologie. Afin d’analyser les stratégies rhétoriques du
discours jihadiste, une approche quanti-qualitative a été adoptée (Rastier,

54

JADT’ 18

2011). Plus particulièrement, cette approche itérative était constituée de
quatre étapes. Une première analyse qualitative de l’idéologie djihadiste, du
processus de radicalisation et des caractéristiques linguistiques du discours
de haine a été essentielle à la compréhension du discours djihadiste ainsi qu’à
l’avancement des premières hypothèses. La deuxième étape correspond à
une analyse quantitative qui a permis de vérifier les hypothèses avancées : le
corpus a donc été examiné avec le logiciel Tropes (Ghiglione et al, 1998), qui
permet d’analyser un texte d’un point de vue sémantico-pragmatique à partir
d’un lexique préétabli, et d’identifier les thèmes les plus récurrents dans le
corpus ainsi que la manière dont ces thèmes sont liés l’un à l’autre. Afin
d’analyser la manière dont le discours djihadiste arrive à persuader et
menacer les différents lecteurs (Giro, 2014), une analyse qualitative a été
menée sur les thèmes sentiment, pour le corpus français, et feeling, pour le
corpus anglais (troisième étape). En d’autres termes, l’analyse quantitative a
constitué le point de départ pour une étude qualitative, qui a donc été menée
sur les énoncés exprimant des émotions et des sentiments (Caffi et Janney,
1994). Enfin, une dernière analyse quantitative a été menée avec le logiciel
Iramuteq (Ratinaud et Marchand, 2012) qui, basé sur la méthode Reinart,
permet, par exemple, de déterminer le sous- et suremploi de certains termes
au sein des différents corpus (quatrième étape). La combinaison d’approches
qualitatives et quantitatives a permis d’examiner de discours djihadiste en
relation avec le contexte dans lequel il a été produit (Valette et Rastier, 2006).
2.2. Résultats
L’analyse des énoncés exprimant des émotions et des sentiments dans les
deux revues officielles de Daesh a permis de déterminer le schéma rhétorique
sur lequel se construit la propagande djihadiste. Puisque l’objectif de Dabiq et
de Dar al-Islam est de manipuler le comportement du lecteur, la propagande
de Daech se fonde sur l’imposition d’obligations et d’interdictions. L’accord
de récompenses ainsi que le sentiment de culpabilité visent à amener le
lecteur à respecter ces indications. En revanche, tout musulman qui ne
respecte pas ces indications, subira des conséquences négatives : il sera jugé
d’apostat et il sera donc considéré comme un ennemi. On a ici la menace
exprimée par Daech contre les musulmans. Les obligations sont exploitées
également pour imposer au lecteur une action violente contre l’Occident,
justifiée et alimentée par le sentiment de victimisation. Combattre l’ennemi
est présenté comme une action héroïque et valorisante. En participant au
combat contre l’Occident, le lecteur aura l’impression de devenir un héros
qui lutte au nom d’une cause juste et noble (De Bonis 2015), et de voir ses
faiblesses disparaître (Rumman, Suliman et al 2016). En outre, en citant des
versets coraniques concernant la victoire des musulmans, l’auteur assure à

JADT’ 18

55

son lecteur que la communauté musulmane aura la victoire sur l’ennemi ;
l’extrait suivant en est un exemple : « Allah par vos mains les châtiera, les
couvrira d’ignominie, vous donnera la victoire sur eux et guérira les poitrines d’un
peuple croyant » (Dar al-Islam, n° 8). La victoire sur l’ennemi est perçue par les
djihadistes comme persuasive. Toutefois, cet énoncé, perçu comme persuasif
par les djihadistes, le sera comme menaçant par l’Occident. De même, le
djihad, qui est interprété comme persuasif par les membres du groupe
djihadiste puisqu’il permet d'accéder au Paradis, tend à être associé aux
attentats terroristes et donc à être perçu comme menaçant par les
occidentaux. Cette double interprétation rejoint la définition de Perelman et
Olbrechts-Tyteca (1988), qui proposent d’« appeler persuasive une
argumentation qui ne prétend valoir que pour un auditoire particulier » (p.
36). Bien que Dabiq et Dar al-Islam présentent le même schéma rhétorique,
leur contenu varie de manière conséquente. Cette étude a révélé, par
exemple, que la revue française focalise son discours sur la figure de l’autre
(i.e., de l’ennemi). En revanche, la revue anglaise est focalisée sur la figure du
musulman et, plus particulièrement, sur la conduite qu’un bon musulman
devrait avoir.
3. Analyse textuelle des procédures judiciaires
Au sein d'une équipe d'enquête, le travail des analystes criminels consiste à
lire et synthétiser les documents de procédures (auditions de témoins,
données téléphoniques et bancaires, comptes-rendus d'expertise, etc.) afin de
fournir aux enquêteurs et aux magistrats une vision plus globale des
informations collectées, par le biais de schémas de représentation et de
synthèses (Rossy 2011). Leur intervention est requise dans des affaires
complexes comme les cold cases ou les affaires impliquant de larges réseaux,
et permet de fournir de nouvelles pistes d'investigation pour les enquêteurs.
À l'heure actuelle, les analystes s'appuient sur un logiciel de reconnaissance
optique de caractères, des outils de bureautique classique (traitement de
texte, tableur) ainsi que sur le logiciel de représentation graphique d'IBM
Analyst's Notebook. Cet outillage ne les dispense pas d'une phase de lecture
précise et chronophage de la procédure visant entre autres à repérer et
extraire manuellement les informations pertinentes pour l'enquête,
regroupées en différents types d'entités qui une fois extraites sont agencées
en représentation graphique (chronologique ou relationnelle).
3.1. Corpus de travail
Le corpus de travail mis à notre disposition par le PJGN est une procédure
judiciaire complète jugée et résolue concernant un homicide. Le dossier,
comme toute procédure judiciaire, rassemble une variété de documents :

56

JADT’ 18

rapports d'expertise, procès-verbaux d'investigations, procès-verbaux
d'auditions de témoins et de mis en cause, factures téléphoniques détaillées,
données bancaires, planches photographiques, etc. Nous avons choisi de
concentrer notre travail sur le sous-corpus composé des auditions de témoins
et de personnes gardées à vue. Ce choix s'est fait lors de notre prise de
connaissance du corpus et du domaine, les auditions représentant la masse
d'information la plus dense et la plus difficilement accessible d'une
procédure : le nombre des auditions (dans notre cas, 370 auditions pour
environ 600 000 mots) et leur manque de structure gênent leur traitement
avec des outils standards, contrairement par exemple aux données
téléphoniques qui peuvent être intégrées telles quelles dans Analyst's
Notebook ou à d'autres données collectées en gendarmerie sous forme de
formulaires structurés.
3.2. Détection automatique d'entités nommées
La notion d'entité en analyse criminelle correspond à la notion d'entité
nommée (EN) en extraction d'information : une unité linguistique
monoréférentielle qui a la capacité de renvoyer à un référent unique (Nouvel
& al, 2015). D'une manière générale, cinq types d'entités intéressent les
analystes criminels : les personnes, les lieux, les dates et heures, les véhicules
et les numéros de téléphone. Nous avons entrepris d'appliquer des
techniques de détection d'EN éprouvées sur les documents de procédures
judiciaires, tout en variant les approches de manière à répondre au mieux
aux contraintes de chaque type d'entité. Deux fonctionnalités du logiciel
UNITEX (Paumier, 2016) ont été mises en œuvres : l'édition de grammaires
pour la détection des dates, l'utilisation d'un lexique pour la détection des
villes, et la combinaison d'un lexique de prénoms et de règles pour les noms
de personnes. Les numéros de téléphone quant à eux sont détectés à l'aide
d'une expression régulière.
En l'état actuel des choses, nous sommes donc en mesure de détecter :
 Les dates normées : “le 10 janvier 2017”, “l'an deux mille dix-sept, le
dix janvier”, “le 10/01/2017”
 Les noms et prénoms de personnes : “Blanche Rivière”, “Petit
Noémie”, “Michel E. Dupont”
 Plus de 36000 villes figurant dans un lexique1
Le développement d’une approche de détection des véhicules, car leurs
mentions dans le corpus combinent plusieurs types d’informations :
 genre de véhicule : moto, scooter, camionnette, voiture, etc.

1 Disponible
(janvier 2018)

à l'adresse

:

http://sql.sh/736-base-donnees-villes-francaises

JADT’ 18

57

 marque
 mention du modèle ou d’une forme (4X4, citadine, berline, break,
etc.)
 couleurs et signes distinctifs (rouille, sérigraphie, année du modèle,
etc.)
La délimitation de la mention d’un véhicule ne peut se résumer à la
combinaison d’une marque et d’un modèle, comme le montrent les deux
exemples suivants tirés du corpus :
 Il s'agit d'un petit modèle comme une TWINGO pour vous donner le
volume. Il était de couleur orangé. Il est petit car il a un petit coffre.
 M. X. m'a cependant parlé d'un véhicule 4X4 conduit par un individu qui
avait un fusil.
La détection des véhicules nous amènera donc à envisager une approche de
détection plus complexe que celles déjà mises en place.
3.3 Analyse de données textuelles et analyse criminelle, une même
problématique ?
Si la détection automatique des entités nommées dans le contexte de l'analyse
criminelle en gendarmerie constitue une tâche habituelle de TAL, on ne peut
pas pour autant en circonscrire les apports potentiels à des aspects purement
techniques. La méthodologie de travail de l’analyse criminelle repose sur
l'interprétation humaine pour la production d'hypothèses, et en cela nous la
rapprochons de l'analyse des données textuelles (ADT) telle que définie par
(Ho-Dinh, 2017) : « Avec l’ADT, nous nous situons au contraire dans une
perspective de construction des connaissances, par l’interprétation humaine
des résultats obtenus grâce à des outils informatiques de calcul et de
visualisation. La puissance informatique vient donc en assistance de
l’exploration et la fouille des données. Cette différence fondamentale permet
de produire des connaissances qualitatives sur les données et non seulement
quantitatives. » La poursuite de nos travaux s'oriente donc non seulement
vers l'amélioration des résultats de détection d'entités et l'introduction
d'approches statistiques (TF-IDF, clustering de documents, etc) mais
également vers le développement d'une interface d'exploration textuelle
propre, prenant en compte les spécificités du genre textuel de la procédure
judiciaire (tri du texte en fonction de sa nature : texte d'en-tête, informations
d'état-civil), et permettant une navigation efficace entre entités détectées,
mesures statistiques, et texte original. La méthodologie de l’analyse
criminelle et les pratiques du métier pourraient être à revoir en conséquence,
impliquant une phase de formation des analystes criminels aux méthodes
textométriques.

58

JADT’ 18

4. Conclusion
Nous estimons avoir soulevé des perspectives théoriques et techniques pour
l'analyse de données textuelles dans les domaines judiciaires et de la sécurité,
relevant aussi bien de l’analyse de discours que du TAL et de la textométrie.
Dans le cas de la propagande de Daesh, l’analyse et la compréhension du
discours djihadiste pourraient contribuer à la formulation d’un contrediscours qui puisse faire face et contrer la propagande djihadiste. Concernant
les pratiques d'analyse textuelles en analyse criminelle, nous espérons que la
mise en place de techniques d'automatisation et d'un outil d'exploration
textuelle permette de repenser la méthode d'accès à l'information en analyse
criminelle et soit une première étape d'une réflexion plus large sur la collecte
et la circulation de l'information et des documents dans le processus
judiciaire. Ces deux cas d'études illustrent la pertinence d'approches de
sciences humaines et sociales dans le contexte sécuritaire et judiciaire, qui a
jusqu'à présent surtout eu recours à des expertises en sciences dites « dures »
(médecine légale, biologie, chimie, informatique, etc.), regroupées sous
l'appellation de « sciences forensiques ». Nous espérons que de telles
contributions permettront de renforcer les liens et d'ouvrir la voie à d'autres
projets associant institutions judiciaires et de défense et chercheurs en
sciences humaines et sociales.
References
Caffi C., & Janney R. W. (1994). Toward a pragmatics of emotive
communication. Journal of pragmatics, 22(3), 325-373.
De Bonis M. (2015). La strategia della paura. Limes, 11.
Ghiglione, R., Landré, A., Bromberg, M., & Molette, P. (1998). L’analyse
automatique des contenus. Paris, Dunod.
Giro M. (2015). Parigi: il branco di lupi, lo Stato Islamico e quello che
possiamo fare. Limes.
Ho Dinh O. (2017). Caractérisation différentielle de forums de discussion sur le
VIH en vietnamien et en français. Thèse de doctorat, Inalco, Paris.
Marchand P. (2014). Analyse avec Iramuteq de dialogues en situation de
négociation de crise : le cas Mohammed Mehra. Actes des 12èmes Journées
internationales d’Analyse statistique des Données Textuelles (JADT), Paris, pp.
457-471.
Nouvel D., Erhmann M., Rosset S. (2015). Les entités nommées pour le traitement
automatique des langues. ISTE Editions
Paumier S. (2016). Unitex 3.1 user manual, http://www-igm.univ-mlv.fr/
unitex
Perelman C., & Olbrechts-Tyteca L. (1988) (5e éd.). Traité de l’argumentation.
Bruxelles : Edition de l’Université de Bruxelles.

JADT’ 18

59

Rastier F. (2011). La mesure et le grain: sémantique de corpus. Champion; diff.
Slatkine.
Ratinaud P., Marchand P. (2012). Application de la méthode ALCESTE à de
"gros" corpus et stabilité des "mondes lexicaux" : analyse du "CableGate"
avec IraMuTeQ. Actes des 11eme Journées internationales d’Analyse statistique
des Données Textuelles (JADT), Liège, 13-15 juin, p. 835-844.
Rossy Q. (2011). Méthodes de visualisation en analyse criminelle : approche
générale de conception des schémas relationnels et développement d'un catalogue
de patterns. Thèse de doctorat, Université de Lausanne, Faculté de droit et
des sciences criminelles.
Rumman A., Suliman M. et al. (2016). The Secret of Attraction: ISIS Propaganda
and Recruitmenet. Traduit par Ward, W. J. et al. Amman: Friedrich-EbertStiftung.
Valette M., & Rastier F. (2006). Prévenir le racisme et la xénophobie:
propositions de linguistes. Langues modernes, 100(2), 68.

60

JADT’ 18

A two-step strategy for improving
categorisation of short texts
Simona Balbi1, Michelangelo Misuraca2, Maria Spano1
1

Università di Napoli Federico II – simona.balbi@unina.it maria.spano@unina.it
2 Università della Calabria – michelangelo.misuraca@unical.it

Abstract
Text categorisation allows organising a collection of documents with respect
to their content. When we consider short texts – e.g., posts and comments
shared onto social media – this task is harder to achieve because we have few
significant terms. Refer to higher-level structures, representing concepts, or
topics occurring in the collection, can improve the effectiveness of the
procedure. In this paper, we propose a novel two-step strategy for text
categorisation, in the frame of feature extraction. Concepts are identified by
using network analysis tools, namely community detection algorithms.
Therefore, it is possible to organise the document collection with respect to
the different concepts and describe the groups of documents with respect to
terms. A case study about Pope Francis on Twitter is presented for showing
the effectiveness of our proposal.
Keywords: short texts, text categorisation, textual network, community
detection
1. Introduction
The ever-increasing popularity of the Internet, together with the amazing
progress of computer technology, has led to a tremendous growth in the
availability of electronic documents. Therefore, there is a great interest in
developing statistical tools for the effective and efficient extraction of
information on the Web, in a so-called Text Mining perspective.
The most common reference model for representing documents, in Text
Mining, is the so-called vector space model: a document is a vector in the
(extremely sparse) space spanned by the terms. Documents are usually coded
as bag-of-words, i.e. as an unordered set of terms, disregarding grammatical
and syntactical roles. The focus is on the presence/absence of a term in a
document, its characterisation and discrimination power. In the knowledge
discovery process, the core of the majority of procedures is related to
dimensionality reduction, both via feature selection and/or feature extraction.
Statistical tools enable an effective feature extraction. One of the most
interesting tasks in Text Mining is Text categorisation which allows
organising a collection of documents, grouping them with respect to their

JADT’ 18

61

content. Here we propose a novel two-step strategy designed for the text
categorisation of short documents – e.g., posts and comments shared onto
social media – when the task is harder to achieve because we have few
significant terms. The basic idea is that Textual data can be processed at
different levels, e.g. we can consider single terms, or subsets of terms
identifying different concepts, in a feature extraction frame. Concepts are
identified by using network analysis tools, namely community detection
algorithms. Therefore, it is possible to organise the document collection with
respect to the different concepts and describe the groups of documents with
respect to terms. The effectiveness of our proposal is showed by analysing a
set of tweets about the Pope Francis, posted on November 2017.
2. Background and related work
The bag-of-words encoding is characterised by high dimensionality and an
inherent data sparsity. According to Aggrawal and Yu (2000), the
performances of text categorisation algorithms decline dramatically due to
these aspects. Therefore, it is highly desirable a previous dimensionality
reduction.
In pre-processing, feature selection and/or feature extraction are often used
before applying any further analysis. Via feature selection, only a subset the
original vocabulary is considered, according to with some criterions. Several
feature selection techniques are reported in the literature, such as term
strength (Yang, 1995), information gain (Yang and Pedersen, 1997), Chi-squared
statistic (Galavotti et al., 2000), entropy-based ranking (Dash and Liu, 2000).
Feature extraction (also known as feature reduction) is a process for
extracting a set of new features from the original vocabulary by applying
some functional mapping. Common feature reduction techniques include
lexical correspondence analysis (Lebart et al., 1998), latent semantic indexing
(Deerwester et al., 1990). These techniques obtain dimensionality reduction,
by transforming the original terms in fewer linear combinations, spanning
sub-dimensional spaces, that may not have a clear meaning and sometimes
results are difficult to be interpreted.
To cope with this limit, here we consider a different viewpoint. Both feature
selection and feature extraction are basically founded on the analysis of a
documents x terms matrix, in which the generic element is the frequency of a
term in a document, or another related weight representing the importance
of the term. It is possible to get back part of the use context of each term by
constructing a terms x terms co-occurrence matrix. In general, each element of
this latter matrix is the number of times two terms co-occur in the corpus.
This particular data structure can be represented as a network, where each
term is a vertex and each element of the matrix different from 0 is an edge.

62

JADT’ 18

The problem of reducing the original dimensionality and perform a feature
extraction can be seen as a community detection problem: terms used
together define a concept, as in latent semantic indexing, or correspondence
analysis, but without any algebraic transformation. Differently from the
approaches previously described, indeed, this method preserves the original
meaning of the terms and allows a better readability of the results.
A community in a network is a set of nodes where vertices are densely interconnected and sparsely connected to other parts of the network (Wasserman
and Faust, 1994). There is no universally accepted definition for a
community, but it is well known that most real-world networks display
community structures. When we consider networks of terms, communities of
terms densely interconnected can be interpreted as topics. From a theoretical
point of view, community detection is not very different from clustering.
Many algorithms have been proposed. Traditional approaches are based on
hierarchical or partitional clustering (e.g.: Scott, 2000; Hlaoui and Wang,
2004). The most popular algorithm is the one proposed by Girvan and
Newman (2004). The method is historically important because it marked the
beginning of a new era in the field of community detection, by introducing
the notion of "modularity". Originally introduced to define a stopping
criterion, modularity (nowadays refers as Girvan and Newman's modularity)
has rapidly become an essential element of many community detection
methods, as fast-greedy (Clauset et al., 2004), label propagation (Raghavan et al.,
2007), leading eigenvector (Newman, 2006). It measures the difference between
the observed fraction of edges that fall within the given communities and the
expected fraction in the hypothesis of random distribution. For a most
comprehensive review of the community detection literature, it is possible to
refer to Fortunato (2010).
3. Problem definition and proposed method
Text categorisation allows to group documents belonging to a collection with
respect to the textual content of the documents themselves. When we
consider short texts, this task is more difficult to achieve because we have
few significant terms for characterising the different groups. The
identification of high-level structures representing the concepts/topics
occurring in the collection can improve the effectiveness of the grouping
procedure. In this paper, a two-step strategy for improving the automatic
organisation of a collection of documents is proposed.
LetT={d1, …, dn}p be a set of n document vectors in a term space of p
dimension, represented by a documents x terms matrix, where each element tij
is the occurrence of an i term into a j document (i=1, ..., p; j=1, ..., n). For the
purpose of our analysis, we are just interested if the term i occurs in

JADT’ 18

63

document j, or not. Then we consider a binary matrix B, where the generic
element bij is equal to1 if the term i occurred at least once in document j, 0
otherwise. From the matrix B we derive the terms x terms co-occurrence
matrix A by the product ABBT. The generic element aii′ is the number of
documents in which the term i and the term i′ co-occur (ii′). An element aii
on the principal diagonal represents the total number of documents in the
collection containing the term i. A is an undirected weighted adjacency
matrix that can be used to analyse the relations existing among the different
terms.
As each community can be seen as a concept/topic occurring in the collection,
in order to detect a group of terms defining a concept, we perform a
community detection on the matrix A. Each community can be seen as a
concept/topic occurring in the collection.
As we said above, the greedy algorithm is based on the optimisation of a
quality function known as modularity. Suppose the vertices are divided into
communities such that vertex/term i belongs to the community ci. The
modularity Q is defined as

Q=


 i i' 
1

 aii'  s(c ,c )
2 h  i i'
2 h ii' 

where h is the total number of edges in the network, i is the degree of the
term i and the s function s(ci,ci′) is 1 if ci=ci′ and 0 otherwise. In practice, a
value above about 0.3 is a good indicator of an interesting community
structure in a network.
The greedy algorithm falls in the general family of agglomerative hierarchical
clustering methods. Starting with a state in which each term is the sole
member of one of K concepts, the algorithm repeatedly joins concepts
together in pairs choosing in each step the join that results in the greatest
increase in modularity.
At the end of the detection process, we obtain a terms x concepts matrix C, a
complete disjunctive table where the cik element (k=1, …, K) is 0 or 1 when a
term i belongs or not to a community. The text categorisation is performed
with a clustering algorithm on the matrix documents x concepts T*(TTC)DK-1,
where DK-1 is the diagonal matrix of the column marginal distribution of C.
Each cell of T* contains the proportion of terms belonging to a concept.

64

JADT’ 18

4. A case study
Twitter is one of the most popular – and worldwide leading – social
networking service. It can be seen as a blend of instant messaging,
microblogging and texting, with brief content and a very broad audience.
The embryonic idea was developed considering the exchange of texts like
Short Message Service in a small group of users. As of the third quarter of
2017, it has 330 million monthly active users, with an amount of daily sent
tweets close to 500 million (Source: Twitter, Statista). Our aim is to categorise
a set of tweets, generated by the same hashtags, with respect to the different
concepts expressed in the collection itself.
4.1. Data description and pre-processing
By using the Twitter Archiver add-on1 for Google Sheet, we collected 24588
tweets about Pope Francis, published between November 10th and December
7th 2017. We use the hashtag #papafrancesco in the query, with any kind of
restriction on the language of the tweets. Moreover, we do not filter the socalled retweets, so that some texts are replicated in the corpus. The preprocessing was performed in two steps. First, we stripped URLs, usernames,
hashtags, emoticons and RT prefixes, and we normalised the tweets by
removing special characters and any separators than blanks. Second, on the
23915 cleaned tweets, we performed a lemmatisation and a grammatical
tagging. The terms contained in the tweets written in other languages
different from Italian were considered as noise.
In the analysis, we consider only nouns because of their content-bearing role.
Moreover, we delete from the vocabulary the terms occurring less than 10
times. Thus we obtain a documents x terms matrix T with 23915 rows and 1603
columns, and the corresponding terms x terms co-occurrence matrix A.
4.2. Concept identification and categorisation process
We perform the community detection procedure on A in order to identify the
concepts. For better highlighting relations among the terms, we fixed a
threshold of 30 on the value of co-occurrence, deleting isolated terms. The
greedy algorithm detected 38 different concepts. The high value of the
modularity measure (Q = 0.648) supports the effectiveness of our procedure
results. In Table 1, we list as an example the terms belonging to some of the
detected concepts.

1
https://chrome.google.com/webstore/detail/twitter-archiver/pkanpfekacaoj
dncfgbjadedbggbbphi

JADT’ 18

65
Table 1 – Concepts detected in the collection with corresponding terms

Concept
2
7
10
19
23
27
…

Terms
scienza, sperimentazione, accanimento, responsabilità, malato, cura, eutanasia, …
bangladesh, religione, viaggio, cultura, myanmar, discorso, buddista, monaco, …
aborto, perversione, febbraio, don, pieri, colonizzazione, crimine, mafia
pensiero, figlio, papà, cecilia, moser, monte
dramática, miedo, josé, experimentan, condición, maría, marcada, incertidumbre
giornatamondialedeipoveri, aula, giovanni, paolo, preparazione, pranzo
…

It is interesting to note that the algorithm identifies the concepts not written
in Italian (e.g., #23 contains Spanish terms) and the concepts not related to
Pope Francis (e.g., #19 refers to a popular reality show). By selecting only the
terms belonging to the different communities, we obtain a 19799 x 38 matrix
T*. On this matrix, we perform a hierarchical clustering based on the Ward
criterion. In Figure 1 it is shown the histogram of the level indices obtained by
the clustering. The indices represent the loss of inter-class inertia caused by
the aggregation. The maximum gap in the distribution suggests to consider a
partition in 37 clusters.

Figure 1 – Histogram of the level indices calculated on the dendrograms’ nodes

66

JADT’ 18

Because of the unsupervised nature of the approach, the quality of the results
can be investigated only by looking at the clusters’ composition. Due to the
limitation of 140 characters, each tweet can express one to three concepts at
most. In Table 2 we can see the concepts occurring in the different clusters.
The order of the concepts represents their importance in terms of statistical
significance. The preliminary results seem to be very promising, but a deep
investigation has to be considered in order to validate the proposal.
Table 2 – Clusters’ size and composition
Cluster

Tweets

Concepts Cluster

Tweets

Concepts Cluster

Tweets

Concepts

1

120

6

14

8210

4, 7

27

51

30

2

506

15, 6, 9

15

536

1

28

150

36

3

95

9, 15

16

1348

32

29

163

37

4

62

12

17

1379

13

30

41

21

5

179

29

18

677

3

31

51

28

6

93

14

19

2699

2

32

102

22, 4

7

79

16

20

666

8, 7

33

71

26, 22

8

160

10

21

48

24, 20, 13

34

42

17, 11

9

445

5

22

155

20, 4, 24

35

288

11, 34

10

304

19, 18

23

242

38

36

125

34, 11

11

36

18

24

55

25

37

42

23, 11

12

66

31

25

71

33

Total

19799

13

335

27

26

107

35

5. Final remarks
The proposed strategy aims at categorising the documents of a collection by
detecting high-level structures, i.e. concepts, as subsets of terms. The terms
belonging to each concept are retained in the process and can be used for
characterising the identified groups of documents. The tools are given by
network analysis tools, namely community detection algorithms. The
strategy is suitable when we deal with short texts. Future developments of
this work are devoted to set automatically a co-occurrence threshold in the
community detection step and to evaluate alternative similarity indices for
measuring the relation strength among terms.

JADT’ 18

67

References
Aggrawal C.C. and Yu P.S. (2000). Finding generalized projected clusters in
high dimensional spaces. Proceedings of SIGMOD’00, pp. 70-81.
Clauset A., Newman M.E. and Moore C. (2004). Finding community
structure in very large networks. Physical review E, 70(6), 066111.
Dash M. and Liu H. (2000). Feature selection for clustering. Proceedings of
Pacific-Asia Conference on knowledge discovery and data mining, pp. 110-121.
Deerwester S., Dumais S.T., Furnas G.W., Landauer T.K., Harshmanet R.
(1990). Indexing by latent semantic analysis. Journal of the American Society
for Information Science, 41(6): 391-407.
Fortunato S. (2010). Community detection in graphs. Physics Reports, 486(3):
75-174.
Galavotti L., Sebastiani F. and Simi M. (2000). Feature selection and negative
evidence in automated text categorization. Proceedings of KDD-00.
Hlaoui A., Wang S. (2004). A direct approach to graph clustering. Neural
Networks and Computational Intelligence: 158-163.
Lebart L., Salem A., Berry L. (1998). Exploring textual data. Springer
Netherlands.
Newman M.E. (2006). Modularity and community structure in networks.
Proceedings of the national academy of sciences, 103(23): 8577-8582.
Newman M.E. and Girvan M. (2004). Finding and evaluating community
structure in networks. Physical review E, 69(2): 026113.
Raghavan U.N., Albert R. and Kumara S. (2007). Near linear time algorithm
to detect community structures in large-scale networks. Physical review E,
76(3): 036106.
Scott J. (2000). Social Network Analysis: a handbook. Sage, London.
Wasserman S. and Faust K. (1994). Social network analysis. Cambridge
University Press.
Yang Y. (1995). Noise reduction in a statistical approach to text
categorization. Proceedings of the 18th annual international ACM SIGIR
conference on Research and development in information retrieval, pp. 256-263.
Yang Y. and Pedersen J.O. (1997). A comparative study on feature selection in
text categorization. Proceedings of ICML-97, pp. 412-420.

68

JADT’ 18

Appeler à signer une pétition en ligne :
caractéristiques linguistiques des appels
Christine Barats1, Anne Dister2, Philippe Gambette3,
Jean-Marc Leblanc1, Marie Peres1
1

Université Paris-Est, CEDITEC (EA 3119), Créteil, France – christine.barats@parisdescartes.fr,
jean-marc.leblanc@u-pec.fr, marie.leblanc@u-pec.fr
2Université Saint-Louis - Bruxelles, Belgique – anne.dister@usaintlouis.be
3Université Paris-Est, LIGM (UMR8049), Champs-sur-Marne, France – gambette@u-pem.fr

Résumé
L’analyse des 12 522 textes d’appel d’une plateforme de pétitionnement en
ligne permet d’examiner leurs caractéristiques linguistiques. Le recours à des
outils textométriques met ainsi au jour certaines régularités quant aux
modalités d’appel à signer. Nous nous intéressons tout particulièrement aux
régularités lexicales, aux formes d’adresse ainsi qu’aux modalités
d’implication des signataires.
Mots-clés : statistique textuelle, pétition en ligne, textes d’appel
Abstract
The analysis of the 12 522 petition texts of an online petition platform allows
to examine their linguistic characteristics. The use of statistical textual
analysis tools brings to light several regularities as for the modalities of the
call to be signed. We focus on the lexical regularities, the salutations as well
as the modalities of implication of the signatories.
Keywords : statistical textual analysis, online petition, petition texts
1. Introduction
Les plateformes de pétitionnement en ligne prolongent et modifient l’acte de
pétitionnement (Contamin, 2001). Dans la dynamique des recherches sur
l’incidence des dispositifs de participation en ligne sur les formes d’écriture
numérique et d’engagement politique (Boure, Bousquet, 2011 ; Mabi, 2016 ;
Badouard, 2017 ; Contamin, 2017), nous nous proposons d’interroger les
caractéristiques des textes d’appel au regard d’une plateforme numérique de
pétitionnement.
Le corpus que nous avons analysé est issu de l’un des principaux sites
francophones de pétitions en ligne (lapetition.be). Il se compose de plus de 12
500 pétitions ayant récolté au total 3,25 millions de signatures sur la période
comprise entre le 31 octobre 2006 et le 12 février 2015.
Le site propose 9 rubriques parmi lesquelles le porteur de la pétition est tenu

JADT’ 18

69

de classer sa pétition : Art et culture ; Droits de l’Homme ; Environnement,
nature et écologie ; Humour/Insolite ; Loisirs ; Politique ; Protection
animalière ; Social ; Autres. Comme nous l’avons montré ailleurs (Barats et
al., 2016) et rappelé en figure 1, les différentes rubriques connaissent des
variations importantes tant en termes de nombre de pétitions (figure 1) qu’en
ce qui concerne la longueur des textes des appels, le nombre de signatures ou
encore le nombre et le volume des commentaires laissés par les signataires.
Le choix de la rubrique relève du promoteur de la pétition et témoigne d’une
interprétation qui varie selon les porteurs de projet, mais débouche sur des
régularités internes à chaque rubrique qui émergent de classifications
automatisées du corpus.
Dans cet article, nous nous centrerons exclusivement sur les textes des
appels, avec une attention particulière sur leur incipit, afin d’observer quelles
sont les régularités lexicales et syntaxiques qui caractérisent les textes d’appel
sur l’ensemble du corpus, mais également en contrastant les rubriques. Les
12 522 textes constituent un corpus de 2,6 millions de mots.
Humour / Insolite

397

Art et culture

652

Loisirs

795

Environnement, nature et écologie

1034

Protection animalière

1378

Droits de l’Homme

1738

Social

1806

Politique

2276

Autres

2446

Figure 1 - Distribution du nombre de pétitions par rubrique

70

JADT’ 18

2. Les mots les plus fréquents dans les textes d’appel
Afin d’identifier la présence ou non de formes communes aux textes d’appel,
nous avons examiné les débuts des textes d’appel, indépendamment des
rubriques. La répartition du premier mot des appels ne correspond pas à une
loi de puissance (l’habituelle loi de Zipf) car la courbe décroit plus lentement.
Les débuts des textes d’appel font donc apparaitre un vocabulaire fréquent
particulier. Les 20 formes de cette liste sont en première position dans plus
de la moitié des textes de pétitions : nous, pour, bonjour, le, la, je, les, monsieur,
pétition, l, il, a, depuis, non, en, cette, si, madame, contre, suite.
Si l’on se penche maintenant sur le vocabulaire des 200 formes les plus
fréquentes dans l’ensemble des textes d’appel, on constate que les premiers
verbes conjugués sont est, sont, ont, soit, peut, demandons, faut, doit, avons,
sommes, demande, sera et les premiers mots lexicaux pétition, enfants, pays,
personnes, vie, Belgique, France, temps, animaux, monsieur, monde, place, projet,
jour, droit, loi, politique, mois, travail, ville, ministre, gouvernement, citoyens, cas,
Bruxelles, justice, président, lieu, site, chiens, situation, rue.
On le voit dans la figure 2, dix formes apparaissent non seulement parmi les
30 mots les plus fréquents (hors mots vides) des appels mais aussi parmi les
30 les plus fréquents en première position des textes : nous, pour, je, pétition,
non, contre, j, vous, on, notre.
À l’inverse, des mots qui apparaissent avec une fréquence élevée en première
position des textes d’appel ne se retrouvent pas parmi les 200 mots les plus
fréquents, ou très bas dans le classement : bonjour (545), monsieur (313),
madame (141), chers (111), stop (82), signez (80), mesdames (73), appel (60), voila
(53), marre (45), messieurs (41), cher (40), voici (40), lettre (36), voilà (30), trop
(30), oui (29), sauvons (24), test (23), aidez (22), salut (18).
On trouve ici des formes spécifiques de l’interpellation directe : bonjour, salut,
madame et mesdames, monsieur et messieurs ou encore chers. La présence de
bonjour ou salut rend compte de la diversité des modalités d’interpellation qui
renvoient à des niveaux de langue différents et des formulations parfois
inattendues. L’accessibilité en ligne du dispositif facilite le lancement d’une
pétition : notre corpus se décline sur un continuum qui va des pétitions les
plus sérieuses, celles qui trouvent un écho dans la presse, qui auraient sans
doute existé sans le dispositif d’une plateforme en ligne, qui sont signées par
plusieurs dizaines ou centaines de personnes, aux pétitions très
confidentielles, « juste pour rire », dont le texte de l’appel est très réduit et
qui récoltent peu de signatures. Bonjour apparait avec une plus grande
fréquence dans la rubrique « Loisirs ». La forme test, quant à elle, révèle
certaines difficultés liées au dispositif : il s’agit de tester si une pétition peut
être mise en ligne, et le texte de l’appel comprend alors ce seul mot.

JADT’ 18

71

Figure 2 – Visualisation en chaines de fréquences partagées (Lechevrel & Gambette, 2016)
des 30 mots les plus fréquents, hors mots vides, en première position
et parmi les textes des pétitions.

Deux présentatifs (voici : 40 occurrences, voila/voilà : 83 occurrences) sont
fréquemment attestés en première position des appels à pétition, en
particulier dans les rubriques « Loisirs » et « Humour ». La valeur
énonciative de ces deux formes est relativement différente. La forme voilà est
dans un grand nombre d’emplois une marque de l’oralité qui introduit le
propos sans en modifier fondamentalement le contenu, mais qui reste un
présentatif (« Voilà je suis une très grande fan du destin de Lisa », « Voilà les Tokyo
Hôtel refont des tournées »…). D’autres emplois sont le produit d’une réflexion
(« Voilà, j’ai décidé de faire une pétition », « Voilà, je fais cette pétition ») ou ont
valeur de conclusion : (« Voilà pourquoi il faut avoir peur de l’avenir »). Cette
dernière configuration reste plus fréquente lorsque voilà se trouve dans une
position autre dans la phrase (« Voilà le problème », « voilà pourquoi j’ai décidé de
»…). Une deuxième catégorie d’emploi, où voici et voilà revêtent les mêmes
valeurs, avec une fréquence plus importante de voici, concerne les marques
temporelles (« Voilà quelques années que l’on demande l’autorisation de porter des
shorts », « Voici 22 mois que je suis papa »). Enfin voici comme voilà (dans des

72

JADT’ 18

proportions bien moindres pour la seconde forme) prennent une valeur de
présentatif dans un grand nombre d’emplois (« Voilà le but de ma pétition », «
voilà ma propre pétition », « voici une histoire comme tant d’autres », « voici une
pétition à faire suivre », « voici le lien de ma pétition…).
Avec les verbes à l’impératif signez, aidez et sauvons, le porteur de la pétition
entre directement dans le vif du sujet : il s’agit d’inciter les signataires à agir
par l’acte de pétitionnement. Stop, marre, trop, et oui participent du même
mouvement : agir, mettre fin, encourager à, etc. On ajoute à cette liste pour,
deuxième mot le plus fréquent en première position. Avec contre, il est très
clairement une marque caractéristique de la posture pétitionnaire : on
s’oppose, on soutient. Dans la majorité des rubriques, les textes qui
commencent par non ou contre sont moitié moins nombreux que ceux qui
commencent par oui ou pour, excepté dans la rubrique « Environnement » où
ils sont plus nombreux. Nos investigations vont se poursuivre en privilégiant
les fonctionnalités d’annotation du corpus offertes par TextObserver afin de
davantage prendre en compte les différents contextes d’emploi de ces formes
et ainsi renforcer leur désambigüisation.
Les verbes à l’impératif sont un indicateur intéressant d’implication du
signataire que l’on retrouve aussi dans l’emploi des pronoms nous, vous et je
auxquels nous allons maintenant nous intéresser.
3. L’implication des signataires et des porteurs de pétitions
Le pronom nous est particulièrement mobilisé dans notre corpus : mot le plus
fréquent au début des appels, il est aussi le pronom le plus utilisé dans
l’ensemble du corpus. Ce nous se veut mobilisateur : il inclut dès le texte de la
pétition les futures pétitionnaires dans l’acte de pétitionnement. Une
extraction des 10 mots cooccurrents les plus spécifiques du pronom nous
placé en première position, à l’aide de l’outil TextObserver (Barats et al.,
2013), permet de faire émerger par ordre décroissant de spécificité :
demandons, voulons, souhaitons, soussignés, citoyens, soutenons, réclamons,
opposons, déclarons, appris. Ce pronom introduit très souvent une demande ou
une dénonciation, parfois des éléments de contexte (cf. appris).
On ne peut évidemment exclure que certains de ces nous ne réfèrent qu’aux
porteurs de la pétition, sans l’inclusion des signataires. Néanmoins, la
présence des cooccurrents citoyens et soussignés et les retours que nous avons
faits aux textes montrent que la grande majorité des nous incluent les
signataires. Une étude plus approfondie est en cours pour quantifier plus
précisément les différents cas. Une interrogation par rubrique confirme
l’importance quantitative de ce nous inclusif, en particulier dans le cas des
rubriques « Environnement », « Politique » et « Social » comme le montre la
figure 3(a).

JADT’ 18

73

Figure 3 – Nombre de pétitions, par rubrique, dont le texte d’appel contient j’, je ou nous (a)
et nombre médian de mots des textes de pétitions qui contiennent ou non ces pronoms (b).

Le pronom je arrive quant à lui en quatrième position des mots les plus
fréquents en début de texte, et il est le troisième pronom le plus mobilisé sur
l’ensemble des textes après nous et vous. Il n’est pas rare que les deux
pronoms nous et je/j’ soient utilisés dans les textes d’appel, le porteur de la
pétition passant de son expérience personnelle pour ensuite mobiliser les
pétitionnaires, comme dans l’exemple de la pétition suivante intitulée «
Contre la fermeture du Delhaize d’Herstal » (pet 14595) : « Je trouve ça honteux
de fermer un magasin qui est récompensé du meilleur rapport clients-Personnel! Il
est temps de se serrer les coudes et de se battre jusqu’au bout! Ne nous laissons pas
faire!!!!! ».

Figure 4 – Pourcentage de textes de pétitions renvoyant ou non à une URL (a)
et mentionnant facebook (b), par type de rubrique.

Un des moyens de passer d’une implication individuelle à une mobilisation
collective est de faire référence à d’autres espaces de relai d’information sur
le web, ce qui se traduit par la présence d’URL, qui ciblent parfois des

74

JADT’ 18

réseaux sociaux. 11% des appels comprennent des URL. L’incidence des
rubriques se confirme : « Protection animalière » et « Environnement »
comportent le plus grand nombre d’URL (17%), comme le montre la figure
4(a). Afin d’approfondir ce résultat, nous avons prêté attention à la présence
du réseau social Facebook : 1,6% des textes de pétition y renvoient, comme
on le voit en figure 4(b). La rubrique « Protection animalière » est celle qui
fait le plus appel à des relais via des pages Facebook, confirmant un mode de
mobilisation spécifique et transmedia (Barats et al., 2016). La rubrique «
Politique » est celle qui fait le moins appel au réseau social Facebook. Notons
cependant que la pétition la plus signée sur l’unité de la Belgique, d’aout
2007, a proposé, à l’issue de la fermeture de la pétition, de rassembler sur un
site web les photos d’une des manifestations organisées en novembre 2007.
Les textes des pétitions rendent ainsi compte de l’articulation de différents
dispositifs web dans la dynamique de pétitionnement, qu’une approche
strictement quantitative n’indique que partiellement.
On peut s’étonner, en observant la figure 3(a), du nombre relativement
important, dans chacune des rubriques, de pétitions dans lesquelles aucun de
ces deux pronoms n’apparait et qui serait peut-être le signe de pétitions
moins implicantes, plus impersonnelles. En effet, on constate également que
moins de 15% de ces textes sans nous ni je/j’ contiennent le pronom vous. Si
l’on y regarde de plus près, on se rend compte que les textes des pétitions
sans nous ni je/j sont, pour chaque rubrique, beaucoup plus courts que les
textes de celles qui incluent nous et/ou je/j, comme le montre la figure 3(b).
5. Conclusions et perspectives
Notre analyse des premiers mots de textes d’appel de pétitions montre que le
vocabulaire utilisé dans cette position présente davantage de régularités liées
aux particularités de la pétition que la totalité des textes. Elle permet de
repérer quelques caractéristiques linguistiques qui varient parfois selon les
rubriques (pronoms personnels, formes d’adresse, URL, etc.).
L’approche textométrique trouve parfois ses limites, comme avec l'ambigüité
du nous qui peut inclure ou non les promoteurs ou les signataires de la
pétition, ou bien dans le cas de la polarité positive ou négative de
prépositions et de verbes qui ne suffisent pas à repérer si la pétition traduit
plutôt une demande ou une dénonciation.
Ce travail constitue une première étape vers une vérification systématique
d’autres marqueurs qui permettent d’impliquer les signataires, comme par
exemple la présence de verbes à l’impératif ou de déterminants, en vue d’une
mise en relation avec le nombre de signataires et éventuellement de
recommandations pour la rédaction de textes de pétitions en ligne.

JADT’ 18

75

Références
Badouard R. (2017). Le désenchantement de l’internet. Désinformation, rumeur et
propagande. Paris, FYP éditions.
Barats C., Leblanc J.-M. and Fiala P. (2013). Approches textométriques du
web : corpus et outils. In Barats, C., editor, Manuel d’analyse du Web en
sciences humaines et sociales. Paris, Armand Colin.
Barats C., Dister A., Gambette Ph., Leblanc J.-M., Peres M. (2016).
Analyser des pétitions en ligne : potentialités et limites d’un dispositif
d’études pluridisciplinaires, JADT 2016, Nice. http://lexicometrica.univparis3.fr/jadt/jadt2016/01-ACTES/83043/83043.pdf
Boure R. and Bousquet F. (2011). La construction polyphonique des pétitions
en ligne. Le cas des appels contre le débat sur l’identité nationale.
Questions de Communication, vol. 20: 293-316.
Contamin J.-G. (2001). Contribution à une sociologie des usages pluriels des
formes de mobilisation: l’exemple de la pétition en France. Thèse de
doctorat, Université Paris 1.
Contamin J.-G., Léonard T. and Soubiran T. (2017). Les transformations des
comportements politiques au prisme de l’e-pétitionnement. Potentialités
et limites d’un dispositif d’étude pluridisciplinaire, Réseaux, vol. 204(4):
97-131.
Lechevrel N. and Gambette P. (2016). Une approche textométrique pour
étudier la transmission des savoirs biologiques au XIXe siècle. Nouvelles
perspectives en sciences sociales, vol. 12(1): 221-253
Mabi C. (2016). Analyser les dispositifs participatifs par leur design. In
Barats, C., editor, Manuel d’analyse du Web en sciences humaines et sociales.
Paris, Armand Colin.

76

JADT’ 18

Newsgroup e lessicografia: dai NUNC al VoDIM*
Manuel Barbera, Carla Marello
Università degli Studi di Torino – b.manuel@inrete.it; carla.marello@unito.it

Abstract
VoDIM (Vocabolario dinamico dell’italiano moderno - Dynamic dictionary of
modern Italian) represents a new development in recent Italian lexicography.
In this paper we argue that NUNC corpora ( www.corpora.unito.it), which
contain texts from newsgroups that were downloaded at the beginning of
XXI century, display aspects of “written-spoken” Italian. NUNC might offer
instances of new meaning of “old” words and new collocational contexts. We
discuss several examples taken from the corpora, such as the
internationalism Umwelt, the collocation assolutamente sì and the abbreviation
clima for ‘climatizzatore’ ‘air conditioning’.
Abstract
Il VoDIM (Vocabolario dinamico dell’italiano moderno) rappresenta una
grande novità nella lessicografia italiana di questi anni. Qui si argomenta che
i corpora italiani della suite NUNC ( www.corpora.unito.it), ricavati dai testi
presenti nei newsgroup di inizio millennio, sono un buon testimone
dell’italiano “scritto-parlato” e potrebbero essere utili per documentare nel
VoDIM nuove accezioni e l’uso di nuove collocazioni. Si portano come
esempi il caso dell’ internazionalismo Umwelt, della collocazione di
assolutamente con sì e dell’accorciamento clima per ‘climatizzatore’.
Keywords: VoDIM – NUNC – Lessicografia – italiano
1. Introduzione
Il VoDIM (Vocabolario dinamico dell’italiano moderno), progetto capitanato
dall’Accademia della Crusca1 che coinvolge otto gruppi di ricerca di
altrettante università italiane, fra cui anche il gruppo torinese, sarà un
dizionario dell’italiano postunitario online, basato su corpora e su altri
dizionari acquisiti in formato digitale come il Tommaseo - Bellini, la quinta
Crusca ed il Battaglia, e disegnato per poter essere interrogabile anche a

A Manuel Barbera si devono i §§ 2 e 3, a Carla Marello i §§ 4 e 5 ed il § 1 va
ascritto ad entrambi; anche se ovviamente il lavoro è stato concepito insieme ed
entrambi gli autori se ne sentono pienamente responsabili.
1
Cfr. http://www.accademiadellacrusca.it/it/eventi/crusca-torna-vocabolariolesicografia-dinamica-dellitaliano-post-unitario.
*

JADT’ 18

77

“corpus variabile”, definito dall’utente.
I corpora su cui si appoggia diventano quindi essenziali. Un primo corpus di
riferimento base (i cui risultati non sono ancora pubblici:
http://dizionariodinamico.it/prin2012crusca/dictionary) è stato prodotto col
PRIN 2012 dalla medesima Crusca (in collaborazione con le Università di
Catania, Firenze, Genova, Milano, Napoli, Piemonte Orientale, Tuscia e con il
CNR), ma, naturalmente, da solo è insufficiente alla bisogna.
2. I NUNC
Un corpus con cui si suggerisce di completarlo è il NUNC-IT; i NUNC
(homepage: http://www.bmanuel.org/projects/ng-HOME.html), ideati da
Manuel Barbera (in bmanuel.org), ed appannaggio del medesimo gruppo
torinese che partecipa al VoDIM, propriamente sono una suite multilingue di
corpora che vorrebbe documentare il genere testuale “newsgroup” all’inizio
del terzo millennio; molte versioni ne sono state implementate (anche per
tematiche specifiche), tutte reperibili dalla homepage; il risultato non è
ancora del tutto soddisfacente; pure, qualche uso può già esserne fatto2.
Un newsgroup è un forum telematico a libero accesso, gratuito, disponibile
su Internet, che si manifesta nella forma di testi scritti, i post, inviati ad una
“bacheca elettronica” mantenuta presso una rete di server (i newsserver che
costituiscono UseNet). Gli utenti del gruppo possono scaricare, leggere e
rispondere ai post, costruendo catene (thread) di botte e risposte. I
newsgroup sono articolati in una tassonomia precisa, ossia in un sistema di
cornici argomentative che si chiamano “gerarchie”, a base geograficonazionale e/o tematica.
I vantaggi di questa base testuale per la linguistica dei corpora sono
numerosi e sono stati trattati in Barbera, 2007 e Barbera et Marello, 2009; qui
ci interessa in primo luogo il fatto che presentano una Umgangssprache
assolutamente contemporanea, reale e molto variata per registri e temi.
Per quanto riguarda il VoDIM, molte voci, neologismi, tecnicismi, prestiti,
ecc., non sono attestate nel corpus base della Crusca e quindi i NUNC
potrebbero risultare utile serbatoio di contesti.
3. Un case study: Umwelt
Si veda ad esempio un prestito tecnico, il termine Umwelt.
Introdotto (in tedesco) dal biologo (estone, ma di famiglia tedesca del Baltico)
Jakob Johann baron von Uexküll già nel titolo della sua importante opera del
1909 (Umwelt und Innenwelt der Tiere), è entrato presto nella tradizione

2 Come dimostrato da alcuni degli interventi presenti in Barbera et al. 2007; in
Costantino et al. 2009, per non citare che i primi utilizzi di dieci anni fa.

78

JADT’ 18

filosofica (a partire da una recensione di Max Scheler del 1914): usato da
Heidegger in un suo corso del 1929-30, è diventato poi moneta corrente (tra
gli altri) in francese con Gilles Deleuze, Maurice Merleau-Ponty e Jacques
Lacan, nonché in italiano con Giorgio Agamben. Ma è usato soprattutto in
testi di biologia, naturalmente, e poi in semiotica, in cui è stato diffuso negli
anni Sessanta da Thomas Albert Sebeok (born Sebők Tamás) ed è alla base
della moderna biosemiotica (cfr. Kull, 2001).
Nei NUNC il termine è ripetutamente attestato.
Per Gadamer comprendere l ' esistenza3 - e qui c'è ancora Heidegger significa prima di tutto pre-comprenderla , in quanto la comprendiamo
con un linguaggio che non scegliamo , ma che , trascendentalmente ,
definisce già la realtà in cui ci muoviamo : l'Um-Welt , da un lato , e
dall ' altro lato , il Mit-welt . Ma , Gadamer cerca di andare alla radice
del movimento del pensiero del soggetto e tale origine sta nell '
esigenza di comprendere e farsi comprendere , cioè nel muoversi nell '
Umwelt e nel Mitwelt . Il fatto è che per Gadamer l ' Altro è visibile
solo con gli " occhi nostri ", ciò con ciò che " siamo ", con la nostra "
identità ", il nuovo si dà solo nel familiare . E in un certo senso è così . L
' altro è ciò che mi disturba che mi inquieta perchè non riesco a ridurlo
al mio mondo : è un'eccedenza .

Quello precedente è un esempio dell’uso tecnico-filosofico del termine, che
non si discosta molto da quello che si potrebbe trovare nello spogliare i testi
(e le traduzioni) di quella tradizione. Più interessante è l’esempio seguente:
Anche in Italia il consumo di televisione è vertiginosamente aumentato
: […] . Oltre a due effetti di rilevanza individuale : - la caduta verticale
della capacità di fissare l ' attenzione per più di un certo tempo ( se a
un buon insegnante occorre anche un ' ora per sviluppare un dato
argomento , gli spazi televisivi obbligati in novanta secondi troncano
quello stesso argomento in modo irreparabile ) e - la perdita di
interesse per la lettura - aspetti che coinvolgono per mimetismo
inconscio ( vale a dire per l ' inconscio occupazione degli spazi mentali
ad opera non solo delle immagini ma dell ' intera atmosfera televisiva
che foggia l ' Umwelt dell ' uomo moderno ) anche persone che
fruiscono della TV per tempi ben sotto la media - l ' esposizione allo "

3 Le citazioni dal corpus sono nel prosieguo riportate tel quel: in particolare sono
mantenute le tokenizzazioni di interpunzioni ed apostrofi, tutti gli “errori di
digitazione”, e le idiosincrasie ortografiche proprie del genere.

JADT’ 18

79

sbarramento " delle immagni4 televisive ha due rilevanti effetti sociali :
- il conformismo applicato e - l ' ignoranza generalizzata . […]

Si tratta di un traslato, chiaramente fuori dai campi “tecnici” di diffusione del
termine. Lessicograficamente ciò è particolarmente rilevante perché
testimonia il traghettamento del prestito al di fuori del dominio originario di
appartenenza, assicurandone lo sdoganamento all’uso comune, anche se
colto o relativamente tale. Per questo tipo di riscontri i NUNC possono
rivelarsi particolarmente utili.
4. Al di qua e al di là della parola grafica
Il VoDIM oltre che datare la comparsa di particolari lessemi o di determinate
accezioni, si propone anche di attestare la comparsa di accorciamenti e
combinazioni di parole: i NUNC, in effetti, presentano usi incipienti passati
dal parlato a questa forma di scritto di inizio millennio.
Dal punto di vista della frequenza statistica di tali usi, i dati estratti dai
corpora NUNC presentano delle criticità dovute al fenomeno del quoting, ma
costituiscono una ricca miniera di prime attestazioni: si vedano, ad esempio,
lo studio di Onesti et Squartini, 2007 sul modo di dire tutta una serie di o di
Valle, 2006 sulla penetrazione precoce di anglismi (più o meno italianizzati).
Per quanto concerne gli accorciamenti, in particolare, in Allora et Marello,
2008 ne abbiamo dato una nutrita raccolta. Un esempio per tutti è clima come
accorciamento di climatizzatore; Marello l’aveva già fatto oggetto di un breve
articolo5 e ne aveva constatato la presenza in più post del 2002 di NUNCMotori. Si veda il brano di thread in cui compare anche un disinvolto conce
per concessionario6:
Qualcuno e' in grado di dirmi quanti grammi (olio/gas?) servono per la
ricarica del clima per un CRD del 2002? Una spesa approssimativa?
Grazie
Ciao a tutti, scusate se mi intrometto, ma oggi dopo giorni di dubbio ho
chiamato il conce per lo stesso motivo di Massimo,30 km per sentire un
po' di aria fresca con il clima impostato a 5 gradi e macchina lasciata

Come si diceva, le citazioni dal corpus sono riportate tel quel, ivi compresi gli
errori presenti nella fonte. Tantopiù che la maggiore tolleranza alle cattive digitazioni,
e l’aperta accettazione di alcune caratteristiche grafico-ortografiche, sono tipiche di
questo genere di CMR.
5 Apparso sul Corriere del Ticino il 23 settembre 2005
6 Non approdato questo agli onori della registrazione nei dizionari, come invece
accade per clima la cui data di prima attestazione è secondo il dizionario Zingarelli il
2000.
4

80

JADT’ 18
prima all'ombra

Al di là della parola grafica può, ad esempio, essere interessante
documentare gli usi di assolutamente sì7: se ne trovano ben 103 nei
NUNC generali. Ecco due esempi:
Ma ti senti tanto tanto tanto depressa ??? Ci dobbiamo preoccupare ?
[>]… Oggi un pò meno , però devo dire che ho passato veramente dei
brutti momenti. L ' importante è riprendersi , no ? Assolutamente sì !
Riprendersi e ripartire subito !
tu sei un troll ? […] No , perché il flame occasionale non fa di una
persona un troll - werted è un troll ? Assolutamente sì , perché attua
flame , insulti e provocazioni in modo sistematico e con offese che vanno
oltre l ' ambito dello sfottò sportivo . In più utilizza tutte le tecniche
tipiche del trollaggio , dal morphing al faking al flooding .

Stessa indagine si può fare per anche no, constatando che è nella
stragrande maggioranza dei contesti è ma anche no.
5. Conclusioni
Un ulteriore fattore che rende i NUNC apprezzabili per il linguista e il
lessicografo attento all’uso è la dialogicità, che si intravede soprattutto negli
esempi presentati nel § 4. È un fenomeno pervasivo nei NUNC, di solito
declinato nei newsgroup come quoting (cfr. Barbera, 2011 e Marello, 2007).
Computazionalmente ciò crea, è vero, alcuni problemi (ancora non del tutto
risolti), dato che il fenomeno del testo ripetuto, se incontrollato, va
inevitabilmente ad intaccare l’aspetto statistico, vanificando un semplice uso
quantitativo dei corpora; però testualmente è un fenomeno di grande
importanza, specie se valorizzabile, come nei NUNC, con la possibilità di
potere allargare i contesti fino a 2000 parole.
La capacità dei newsgroup di fissare nello scritto usi eminentemente orali, di
trasferire la fluidità dell’oralità ad uno speciale tipo di scrittura, costituendo
una sorta di ponte tra i due media, può rivelarsi particolarmente importante
per il VoDIM, proprio perché i corpora NUNC registrano tendenze
emergenti nella lingua italiana. Sulla peculiarità diamesica di questo
particolare tipo di “scritto-parlato” abbiamo sostato in Barbera et Marello,
2009, ma qui non possiamo non rimarcarne l’opportunità che potrebbe
presentare per il VoDIM.
I NUNC, come dicevamo, non sono ancora perfetti: i prototipi che sono stati

7

Oggetto di un articolo sul Corriere del Ticino del 21 gennaio 2004.

JADT’ 18

81

messi online sono solo delle beta, ma la volontà di perfezionarli c’è: e non è
da escludere che il VoDIM rappresenti l’occasione giusta per farlo.
Bibliografia
Allora A. e Marello C. (2008), “Ricarica clima”. Accorciamenti nella lingua
dei newsgroup, in Cresti E., editor, Atti del IX Congresso della Società
Internazionale di Linguistica e Filologia Italiana (SILFI): “Prospettive nello
studio del lessico italiano” (Firenze, 14-17 giugno 2006). Cesati: vol. II, pp.
533-538.
Barbera M., Per la storia di un gruppo di ricerca. Tra bmanuel.org e
corpora.unito.it, in Barbera M., Corino E. e Onesti C., editors, Corpora e
linguistica in Rete. Guerra Edizioni: pp. 3-20.
Barbera M., Une introduction au NUNC: histoire de la création d’un corpus,
in Ferrari A. et Lala L., editors, Variétés syntaxiques dans la variété des textes
online en italien: aspects micro- et macrostructuraux. Université de Nancy II,
2011: pp. 9-36.
Barbera M. e Marello C. (2009), Tra scritto-parlato, Umgangssprache e
comunicazione in rete: i corpora NUNC, in Antonini A. e Stefanelli S.,
editors, Per Giovanni Nencioni. Convegno internazionale di studi. Pisa Firenze, 4-5 Maggio 2009. Le Lettere: pp. 157-86. Poi in Barbera M., Quanto
più la relazione è bella: saggi di storia della lingua italiana 1999-2014,
Bmanuel.org - Youcanprint, 2015: pp. 157-182.
Costantino M., Marello C. e Onesti C. (2009), La cucina discussa in rete.
Analisi di gruppi di discussione italiani relativi alla cucina, in Robustelli
C. e Frosini G., editors, Atti del convegno ASLI 2007 “Storia della lingua e
storia della cucina. Parola e cibo: due linguaggi per la storia della società
italiana”. Modena, 20-22 settembre 2007. Cesati: pp. .717-727.
Kull K. (2001), Jakob von Uexküll: An introduction. Semiotica, vol. 134 (1/4):
pp. 1-59.
Marello C. (2007), Does Newsgroups “Quoting” Kill or Enhance Other Types
of Anaphors?, in Korzen I. and Lundquist L., editors, Comparing Anaphors
between Sentences, Texts and Languages. Samfundslitteratur Press: pp. 145157.
Onesti C. e Squartini M. (2007), “Tutta una serie di”. Lo studio di un pattern
sintagmatico e del suo statuto grammaticale, in Barbera M., Corino E. e
Onesti C., editors, Corpora e linguistica in Rete. Guerra Edizioni: pp. 271284.
Valle L. (2006), Varietà diafasiche e forestierismi nell'italiano nei gruppi di
discussione in rete, in López Díaz M. et Montes López M., editors,
Perspectives fonctionnelles: emprunts, économie et variations dans les langues.
S.I.L.F. 2004. XXVIII Colloque de la Société internationale de linguistique

82

JADT’ 18

fonctionnelle, tenu à Saint-Jacque-de-Compostelle et à Lugo du 20 au 26
septembre 2004. Editorial Axac: pp. 371-374.
Zingarelli N. (2017), Lo Zingarelli 2017. Vocabolario della lingua italiana. A cura
di Mario Cannella e Beata Lazzarini. Zanichelli.

JADT’ 18

83

Techniques for detecting the normalized violence in
the perception of refugee / asylum seekers between
lexical analysis and factorial analysis
Ignazia Bartholini
Univ. of Palermo - ignazia.bartholini@unipa.it

Abstract 1
The theme of gender violence finds a peculiar declination if linked to the
phenomenon of forced migrations, and intersects historical-cultural variants
of neo-patriarchal nature to cultural-religious orthodoxies the newcomers
often bear with them. Studying gender violence in the context of globalized
migrations allows us to highlight three bias that mark the western discourse
and that concern the way of conceiving its phenomenology as pre-modern
(a); detaching violence interpretation from politics of intervention and
contrast (b); considering gender asymmetries, sexist representations and
practices in the Mediterranean hosting society as residual (c). Subsequently,
the factorial structure of the questionnaire was investigated through the
Principal Components Analysis (ACP) and the subsequent Oblimin rotation
of the factorial axes, as a relation between the dimensions of the
questionnaire was assumed. The reliability of the scales was verified by the
Cronbach alpha coefficient.
Abstract 2
Il tema della violenza di genere trova una declinazione peculiare se collegato
al fenomeno delle migrazioni forzate e interseca le varianti storico-culturali
di natura neo-patriarcale alle ortodossie culturali-religiose che i nuovi
arrivati portano spesso con loro. Studiare la violenza di genere nel contesto
delle migrazioni globalizzate ci consente di evidenziare tre pregiudizi che
segnano il discorso occidentale e che riguardano: il modo di concepire la sua
fenomenologia come premoderna (a); la searazione fra l'interpretazione della
violenza e le politiche di intervento e contrasto (b); il considerare le
asimmetrie di genere, le rappresentazioni sessiste e le pratiche Mediterranee
come residuali (c). Successivamente, la struttura fattoriale del questionario è
stata analizzata attraverso la Principal Components Analysis (ACP) e la
successiva rotazione Oblimin degli assi fattoriali, essendo stata ipotizzata una
relazione tra le dimensioni del questionario. L'affidabilità delle scale è stata
verificata dal coefficiente alfa Cronbach.
Keywords: gender violence, forced migrations, sexist representation

84

JADT’ 18

1. Introduction
Over the last two decades, the field of border and migration management has
been characterized by the increasing interrelatedness of discourses about
control practices and about humanitarian issues (Walters 2011, Fassin 2010).
Today, European policies seek to incorporate strategies to support forced
migrants as key instruments for the protection of refugees (Moro 2012).
Forced migration, which can also be addressed through the lens of gender
(Hans 2008), is grafted onto a broader field of research, which includes
welfare strategies, social representations and intercultural dynamics.
According to the UNHCR, gender-based violence refers to “any act of
gender-based violence that results in, or is likely to result in, physical, sexual
or psychological harm or suffering to women, including threats of such acts,
coercion or arbitrary deprivation of liberty, whether occurring in public or
private life” (UNHCR 2008: 201). It can take, among others, the form of “rape,
forced impregnation, forced abortion, trafficking, sexual slavery, and the
intentional spread of sexually transmitted infections, including HIV/AIDS”
(UNHCR 2008: 7, 10).
Forms of violence happen not only inside the migratory journey by other
refugees, but also by public officers, government employees, aid agencies
crew (Ferris 2007; Freedman 2015).
2. The numbers of the phenomenon
According to data of the Italian Ministry of Internal Affairs, between 2015
and 2016, 154719 migrants disembarked in Italy, of which 82136 asylum
seekers. From January to March 2016 9,307 migrants disembarked in Italy.
Currently, migrants come mostly from Gambia, Senegal, Mali, Guinea, Ivory
Coast, Morocco, Somalia, Sudan and Cameroon (Source: ANSA).
In January 2016 asylum seekers were 7,505, mostly from Pakistan (1510),
Nigeria (1306), Afghanistan (665) and Gambia (625). Among these, 6739 were
men, 766 women, 292 unaccompanied minors and 199 minors. 6507 requests
were reviewed so far with the following outcomes: 190 people (3%) were
granted the refugee status, 698 (11%) obtained a subsidiary permit, 1352
(21%) were granted with a humanitarian protection and 4266 (66%) were
denied (source: Italian Ministry of Internal Affairs).
Only in the 2017, from the Hotspot Trapani-Milo, managed by "Badia Grande
NGO” one of partners of the project " Provide ", have transited 21,478
refugees / asylum seekers (Source - Ministry of Interior), with 21 different
nationalities. These include 16,010 men, 3177 women, 2291 children divided
in 1787 males and 504 females.

JADT’ 18

85

Last year, two researchers from the University of Palermo submitted a
questionnaire of 36 items to 465 women, temporarily hosted at the TrapaniMilo Hotspot in Sicily.
3. Objectives of research
The core question of the research concerns the identification of violence’s
subjective dimensions from the side of the victims and the operators, as well
as the problems in building social multicultural constructions of violence.
The research wants to identify violence’s subjective dimensions from the side
of the victims and the operators, as well as the problems in building social
multicultural constructions of violence.
For this purpose, the research investigates a specific articulation of the
“migratory violence,” which entails cultural specificities and contextual
conditions, such as the journey and the time spent in reception facilities. In
order to highlight topics and problems related to the social construction of
gender violence, attention will be paid to victims’ point of view concerning
the ‘normalized’ procedural violence, even by means of operational
definitions of victims’ first reception treatments in the institutional arenas.
Furthermore, gender relations are biased by the whole migration experience,
and this leads to various forms of direct, indirect and structural violence:
forms of gender-based violence are seen not only among refugees. Finally,
refugees and asylum seekers may suffer structural violence in the form of
social exclusion and discrimination (Jaji 2009, Crisp, Morris & Refstie 2012),
secondary victimization (Pinelli 2011, Tognetti 2016), labour exploitation
(Coin 2004), forced prostitution (Naggujja et al 2014, Krause-Vilmar 2011)
and sexual abuse (Crisp, Morris & Refstie 2012). Therefore, the migratory
violence to which women—as well as minors and LGBT—are subjected,
becomes, a particular mode of reading and interpretation of intra- and
intercultural gender relations.
For the first part of the research's objective, was to assess the perception of
the violence suffered of the women of sample before and during the journey
to the coast of Sicily.
For the second one of the research's objective, was to individuate some
effective interventions for the reduction of the migrant' exposure to different
types of violence and threat, to encourage the access to physical and
psychological services, to assist the violence' victims with integration,
support safe and appropriate cultural instruments , to provide support for
families, stable settlement in host country and to concerted actions for
reducing the inequalities in access to resources.

86

JADT’ 18

4. Methodology
A1. Once the ethnic intersection, socioeconomic gender and status explored,
an internalist perspective will be employed, based on the analysis of the
narrative devices, that is the conversations’ reports that migratory violence
victims conduct with experts (linguistic and intercultural mediators, social
assistants, psychologists and lawyers, but also doctors and police officers) or
with members of the third sector.
A2. Definitions of lived or experienced violence, through interviews to
refugees and operators in the first and second reception centres, that have
particular acquaintance with the phenomenon;
Subsequently, the factorial structure of the questionnaire investigated
through the Principal Components Analysis (ACP) and the subsequent
Oblimin rotation of the factorial axes, as a relation between the three
dimensions of the questionnaire was assumed:
a. the daily life before the trip;
b. the gender dynamics and relationships among the family members;
c. the violence normalized.
The reliability of the scales was verified by the Cronbach alpha coefficient.
In order to verify the hypothesis, that there are statistically significant
differences to the mean scores of the different dimensions, analyzes of the
variance have been carried out. Multivariate analysis techniques on variance,
together with a lexical analysis, allowed us to select:
1. the keywords present in the corpus of the questionnaire using frequency
indexes;
2. the meta-information contained within the text units;
3. the context units through specific data arrays for content analysis
The communication that we propose to present will describe the results of
the research conducted and the methodological opportunity of the text
analysis tools used by the researchers involved.
5. Some Research’s results
To individuate the vulnerabilities of migrants, it was necessary to identify
appropriate instruments of analysis for being able the needs of violence
victims and in order to deal with them in a respectful, sensitive, professional
and non-discriminatory manner. The have explained the need to receive the
proper degree of assistance and a stronger support and protection. The
keywords more frequently used by migrants are been: protection, fear,
opportunity, work, life.

JADT’ 18

87

The content analysis, and the context units involved through specific data,
describe the necessity to acknowledge the women/asylum seekers, who could
be victims by other men after their arrive in reception center too and the
opportunity to put specific procedures to prevent, identify, and respond to
the different forms of proximity gender-based violence.
The content analysis, and the context units involved through specific data,
describe the necessity to acknowledge the women/asylum seekers, who could
be victims by other men after their arrive in reception center too and the
opportunity to put specific procedures to prevent, identify, and respond to
the different forms of proximity gender-based violence.
6. Conclusion
The problems that refugees face require humanitarian responses and
effective interventions (Dal Lago 1999; Colombo 2012; Camarrone 2016), such
as the reduction of exposure to different types of violence and threat in postmigration phase and the access to physical and psychological services
(Shamir 2005; Ambrosini 2010; Bartholini 2017). From this perspective, the
Mediterranean represents a peculiar field of analysis of that normalized
violence – procedural and proximal – that denies refugees/asylum-seekers,
minors and LGBT people to consider themselves as right holders and
subjects of the same dignity and value.
Morevor, the results or content analysis shows the necesity of a stronger
integration, with a support strategies of appropriate cultural s and social
practices and to provide adeguate support for families in a stable settlement
in our host countries (Balibar 2012). Lastly, the research highlights the need
of some concerted action to reduce inequalities in access to resources
(Robinson et al. 2006).
Gender violence related persecution may give rise to claims for international
protection (Gilbert 2009).
Council of Europe Convention on preventing and combating violence against
women (Istanbul Convention of 2011) and the Directive 2012/29/EU in
establishing minimum standards on the rights, support and protection of
victims, contribute to achieve the obligation to "ensure access for victims and
their family members to general victim support and specialist support, in
accordance with their needs".
Although member states are stepping up their work in order to streamline a
gender understanding into public decision making, policy and operations,
this effort is not always reflected in the asylum procedures.

88

JADT’ 18

References
Ambrosini M. (2010). Richiesti e respinti. L’immigrazione in Italia. Come e perché.
Milano: il Saggiatore.
Balibar E. (2012). Strangers as enemies. Walls all over the world, and how to
tear them down. Mondi Migranti, Vol. 6, n. 1: 7-25. DOI: 10.3280/MM2012001001
Bartholini I (2017). Migrations: A Global Welfare Challenge: Policies, Practices and
Contemporary Vulnnerabilities, (with F. Pattaro Amaral; A. Silvera
Samiento; R. Di Rosa), Edition Corunamerica, Barranquilla (Colombia),
p.1-196 (ISBN 978-9588-59812-2-5).
Camarrone D. Hotspot di Lampedusa, la sindaca chiede al Ministero dell’interno
una verifica urgente delle procedure UE, Diritti e frontiere, 8 gennaio 2016, in
http://dirittiefrontiere.blogspot.it/2016/01/la-verita-sul-sistema-hotspot.html
Colombo A. (2012). Fuori controllo? Miti e realtà dell’immigrazione in Italia.
Bologna: Il Mulino.
Coin F. (2004). Gli immigrati, il lavoro, la casa. Franco Angeli: Milano.
Convenzione
di
Dublino
(1990),
in
http://www.camera.it/_bicamerali/schengen/fonti/convdubl.htm
Crisp J., Morris T. & Refstie, H. (2012). Displacement in urban areas: new
challenges, new partnerships. Disasters, 36(1): S23-S42.
Dal Lago A. (1999). Non Persone. L’esclusione dei migranti in una società globale.
Milano: Feltrinelli.
Fassin D. (2010). La raison humanitaire. Une histoire morale du temps present,
Gallimard-Seuil-Hautes Études: Paris.
Gilbert L. (2009). Immigration as Local Politics: Re-Bordering Immigration
and Multiculturalism through Deterrence and Incapacitation. International
Journal of Urban and Regional Research, Vol. 33, n. 1: 26-42. DOI:
10.1111/j.1468-2427.2009.00838.x
Jaji R. (2009). Refugee woman and the experiences of local integration in Nairobi,
Kenya. University of Bayreuth: Bayreuth.
Krause-Vilmar J. (2011). The Living Ain’t Easy, Urban Refugees in Kampala. UN
Report
Ministero dell’Interno, Rapporto sulla protezione internazionale in Italia 2015, in
http://www.interno.gov.it/sites/default/files/t31ede-rapp_prot_int_2015__rapporto.pdf
Naggujja Y. et al (2014). From The Frying Pan to the Fire: Psychosocial Challenges
Faced By Vulnerable Refugee Women and Girls in Kampala, Report of the
Refugee Law Project.

JADT’ 18

89

Osti G. & Ventura F. a cura di (2012). Vivere da Stranieri in Aree Fragili. Napoli:
Liguori.
Palidda S. a cura di (2011). Il discorso ambiguo sulle migrazioni. Messina:
Mesogea.
Pinelli B. (2011). Attraversando il Mediterraneo. Il sistema campo in Italia:
violenza e soggettività nelle esperienze delle donne, Lares, 77: 159-180.
Regolamento (CE) n. 343/2003 (Dublino II), in http://eur-lex.europa.eu/legalcontent/IT/TXT/?uri=URISERV%3Al33153
Regolamento
UE
n.
604/2013
(Dublino
III),
in
http://eurlex.europa.eu/LexUriServ/LexUriServ.do?uri=OJ:L:2013:180:0031:0059:IT:P
DF
Robinson D. & Reeve K. (2006). Neighbourhood Experiences of New Immigration.
Reflections from the Evidence Base. York: Joseph Rowntree Foundation.
Shamir R. (2005). Without borders? Notes on globalization as a mobility
regime. Sociological Theory, Vol. 23, n. 2: 197-217. DOI: 10.1111/j.07352751.2005.00250.x
Tognetti M. (2016). Donne e processi migratori fra continuità e cambiamento.
ParadoXa, X(3): 69-88.
Walters W. (2011). Foucault and Frontiers: Notes on the Birth of the
Humanitarian Border. In: Bröckling U. (Ed.). Governmentality: Current
Issues and Future Challenges. Routledge: London.

90

JADT’ 18

Dal corpus al dizionario: prime riflessioni
lessicografiche sul Vocabolario storico della cucina
italiana postunitaria (VoSCIP)
Patrizia Bertini Malgarini1, Marco Biffi2, Ugo Vignuzzi3
1LUMSA – p.bertini@lumsa.it
Università degli Studi di Firenze – marco.biffi@unifi.it
3Sapienza. Università di Roma – ugo.vignuzzi@uniroma1.it
2

Abstract
The Vocabolario storico della cucina italiana postunitaria (VoSCIP) it is a
historical dictionary of the language of the cooking, which has also had a
considerable importance for identifying a national linguistic model after the
Unity of Italy. The dictionary is based on a representative corpus (today 42
texts), but by its nature it is a work in progress, open, and it is progressively
increasing. The first exemplar entries (such as cappelletti, anolini, tagliatelle,
bagnomaria) had been presented in various conferences and in some articles;
the entries had been based on a restricted corpus (28 texts) and they have
highlighted some critical issues, so it was necessary a further methodological
reflection. The aim of our paper is to propose some aspects of these
investigations and this methodological reflection: a) the structure of the voice
in a differentiated form (“light” and “complex”); b) the treatment of
emerging positions from the statistical analysis tools of the corpus; c) the
lemmatization of compound words in the face of the morphological
polymorph emerging from the diachronic depth of the corpus; d) the correct
balance between the examples mentioned in the voice and the possibility of a
direct interrelation with the database.
Sintesi
Il Vocabolario storico della cucina italiana postunitaria (VoSCIP) è un dizionario
storico di una lingua speciale, quella della cucina, che ha avuto una notevole
importanza anche nel quadro dell’individuazione di un modello linguistico
nazionale soprattutto all’indomani dell’Unità. Il dizionario si basa su un
corpus rappresentativo (attualmente di 42 testi), ma che per sua natura è
elastico, e aperto, e viene quindi progressivamente incrementato. Le prime
voci campione (quali per esempio cappelletti, anolini, tagliatelle, bagnomaria)
presentate in vari convegni e in articoli in volume e riviste, basate su un
corpus ristretto a 28 testi, hanno messo in luce alcune criticità che hanno
spinto a una ulteriore riflessione metodologica. Proprio alcuni aspetti di tali
approfondimenti sono oggetto del contributo che proponiamo: a) la struttura

JADT’ 18

91

della voce in forma differenziata (“leggera” e “complessa”); b) il trattamento
delle collocazioni emergenti dagli strumenti statistici di analisi del corpus; c)
la lemmatizzazione di parole composte a fronte della polimorfia morfologica
emergente dalla profondità diacronica del corpus; d) il corretto equilibrio tra
esempi citati nella voce e possibilità di un’interrelazione diretta con la banca
dati.
Keywords: lingua della cucina, lingue speciali, linguistica dei corpora,
lessicografia, vocabolario, italiano, dizionario storico
1. Il VoSCIP
Il “Vocabolario storico della cucina italiana postunitaria” (VoSCIP) nasce con
lo scopo di documentare il costituirsi e il fissarsi di una cultura e di una
lingua unitaria della gastronomia in Italia dopo l’Unità. Si tratta di
un’esigenza ben presente a tutti gli addetti ai lavori (linguisti, storici
dell’alimentazione, sociologi ecc.) e che nello specifico ha preso le mosse da
una precisa prospettiva di ricerca, quella di esaminare le vie e i modi
dell’affermarsi di un italiano gastronomico “comune”, a partire da Pellegrino
Artusi e dal modello archetipico del suo fortunatissimo La scienza in cucina e
l’arte di mangiar bene. Il progetto “L’Italiano in cucina. Per un Vocabolario
storico della lingua italiana della gastronomia” è stato assunto
dall’Accademia della Crusca che lo ha inserito nell’ambito degli studi che
mirano alla costruzione del suo progetto strategico dedicato alla redazione di
un Vocabolario Italiano postunitario.
Per la realizzazione del VoSCIP si è proceduto preliminarmente a fissare un
corpus rappresentativo di testi, nel quale naturalmente un ruolo nodale
spetta alla Scienza in cucina: corpus che, per motivi di fattibilità pratica, si è
deciso di far arrivare alla Seconda guerra mondiale e dintorni,
nell’auspicabile prospettiva di poter spostare successivamente il terminus ad
quem alla contemporaneità (con l’inclusione, oltre che dei testi a stampa
posteriori al ’50, delle diverse produzioni legate al “trasmesso” nelle sue
varie forme, dai ricettari presenti sul WEB, ai blog ai social media etc.). Il
corpus principale di riferimento comprende al momento oltre un centinaio di
volumi apparsi tra la fine del Settecento (torneremo fra poco sulle ragioni
della scelta di arretrare il terminus post quem) e il 1950: i testi sono stati
selezionati utilizzando le principali bibliografie sulla produzione
gastronomica italiana del periodo considerato (preziosa in primo luogo
quella di Alberto Capatti che correda l’edizione del 2010 della Scienza
artusiana della Rizzoli). Necessariamente si è dovuto tener conto pure di
fattori pratici quali in primo luogo la reperibilità delle opere e soprattutto la
loro disponibilità e/o acquisibilità da parte dell’Academia Barilla, con la
quale è stata a tali scopi stipulata una specifica convenzione da parte

92

JADT’ 18

dall’Accademia della Crusca. Al momento, i testi acquisiti informaticamente
e marcati (XML/TEI) sono quaranta.
Prima di proseguire, una doverosa precisazione (già annunciata) sul terminus
post quem: anche se il nostro obiettivo primario è, come abbiamo detto, quello
di raccogliere e descrivere la lingua della tradizione gastronomica italiana
postunitaria, per meglio documentare le origini di questo italiano in cucina
(soprattutto per l’aspetto della fraseologia, cioè in primo luogo polirematiche
e collocazioni, ma anche detti proverbiali, modi di dire, ecc.) abbiamo deciso
di prendere in considerazione anche alcuni dei testi più significativi tra fine
Settecento e primo Ottocento, a partire dalle due redazioni dell’Apicio
moderno e dal Cuoco galante di Vincenzo Corrado. Sempre al medesimo fine,
stiamo procedendo inoltre allo spoglio sistematico di tutto ciò che è
pertinente all’ambito semantico del cibo nella tradizione lessicografica
italiana, a partire dalle cinque impressioni del Vocabolario degli Accademici
della Crusca, dal Tommaseo Bellini, dal Giorgini Broglio, e soprattutto dal
Dizionario moderno (prima ed., 1905) di Alfredo Panzini. L’interesse di questo
vocabolario, che offre un vero e proprio panorama della vita e della cultura
italiana tra fine Ottocento e Novecento, è costituito dal nostro punto di vista
proprio dallo spazio attribuito a quelle parole nuove, che già nella prima
edizione lo stesso Panzini catalogava in “scientifiche, tecniche, mediche,
filosofiche, [parole straniere, neologismi, parole dello sport,] della moda, del
teatro, della cucina”.
Imprescindibile nell’ambito lessicale del cibo (come è ben noto) è la
dimensione diatopica per la quale il VoSCIP potrà utilizzare gli importanti
risultati delle indagini geolinguistiche del Novecento, in primis degli atlanti
linguistici: l’AIS e l’ALI, ma anche l’ASLEF, l’ALEPO, l’ALT, l’ALLI, e i
preziosi materiali in corso di pubblicazione per l’ALS (tra cui si ricorderà
almeno il paradigmatico volume di Ruffino 1995).
Per verificare la fattibilità del nostro progetto abbiamo realizzato alcune voci
pilota: siamo partiti da tagliatella, cui sono seguite agnelotto, cappelletto e
anolino; in tutt’altro ambito abbiamo recentissimamente elaborato la voce
bagnomaria. Proprio la redazione di queste voci e in particolare dell’ultima,
bagnomaria, ha messo in luce alcune criticità del modello di voce
originariamente elaborato e reso necessario un ripensamento che sfruttasse a
pieno le risorse della lessicografia computer aided (o della lessicografia
computerizzata) e della multimedialità oggi disponibili.
2. La banca dati
I testi del corpus sono stati sottoposti a una marcatura XML/TEI leggera,
mirata soprattutto a finalità lessicografiche. Attualmente sono stati acquisiti,
collazionati e marcati 42 testi che coprono uniformemente l’arco cronologico

JADT’ 18

93

considerato. Per quanto riguarda l’header sono state previste le indicazioni di
autore, titolo, luogo di edizione, editore, anno, tipologia testuale, indicazione
diamesica, in modo che possano costituire la base per filtrare sottocorpora
specifici. All’interno del testo sono state marcate le pagine di ogni volume
(così che le trascrizioni possano essere di volta in volta collegate alla
riproduzione in facsimile dell’originale), le eventuali figure, le parti in lingue
diverse dall’italiano (perché possano essere escluse dall’interrogazione del
lessicografo). Non si è ritenuto di prevedere nessuna marcatura per i
forestierismi, che, al pari degli altri lessemi, devono essere analizzati
opportunamente dal lessicografo in ogni loro contesto. In una seconda fase
della marcatura dei primi 42 testi, in via di attuazione, è prevista anche la
marcatura del testo delle singole ricette e del loro titolo. Lo scopo primario di
questa marcatura è quello di ottenere una lista aperta delle ricette presenti
nel corpus, che possano eventualmente essere messe a confronto tra di loro
con appositi algoritmi legati alle forme presenti nel titolo. In questo modo
sarà possibile individuare una linea diacronica delle singole ricette e seguire
l’evoluzione della lingua in esse contenute. Per quanto concerne il
trattamento informatico va tenuto conto che la banca dati è un esempio di
testualità ibrida: sia in relazione all’acquisizione filologica del testo e alla sua
interrogabilità, sia per quanto riguarda la possibilità di applicazione di
procedure di lemmatizzazione automatica. Trattandosi di testi ottonovecenteschi la possibilità di buoni risultati nell’applicazione degli
strumenti informatico-linguistici realizzati nel panorama nazionale e
internazionale
scema
progressivamente
allontanandosi
dalla
contemporaneità verso il 1861, ma anche per i testi ottocenteschi e primo
novecenteschi si hanno garanzie sufficienti. Vista la particolare natura della
banca dati, la sua cronologia e la sua finalità lessicografica, nell’equilibrio
della gestione delle risorse, si è preferito quindi non investire su una
lemmatizzazione controllata, che avrebbe comportato l’inserimento di
correttivi legati alla lingua ottocentesca e primo-novecentesca sia sui
dizionari macchina che sulle morfologie macchine attualmente in
circolazione (prevalentemente di base anglofona, con tutti i limiti che questo
comporta, e, anche nel migliore dei casi tarati per l’italiano scritto recente; cfr.
Biffi 2016). La banca dati (attualmente in fase di testing nella sua versione
beta) è quindi consultabile con un motore di ricerca per forme, potenziato da
strumenti (caratteri jolly, ricerca fuzzy) che facilitino l’individuazione delle
varianti formali, morfologiche e grafico-fonetiche, e da una lemmatizzazione
automatica basata sulle morfologie macchina attualmente esistenti (e quindi
tarate sull’italiano scritto contemporaneo, ma comunque sufficientemente
funzionali per il reperimento delle forme varianti di testi otto-novecenteschi,
soprattutto se a fini lessicografici). La piattaforma di interrogazione prevede

94

JADT’ 18

specifiche funzioni di ricerca a distanza e collocazioni, e la possibilità di
accedere a dati statistici, sia in versione tabellare, sia in versione heatmap e tag
cloud. Con queste caratteristiche la banca dati può peraltro essere del tutto
omogenea a quelle che gravitano intorno al progetto del Corpus di riferimento
per un nuovo vocabolario dell’italiano moderno e contemporaneo. Fonti
documentarie, retrodatazioni, innovazioni, finanziato su fondi PRIN 2012 e
coordinato da Claudio Marazzini, offrendo così ampi margini di dialogo con
gli strumenti lessicografici a essa collegati.
3. Struttura delle voci e dizionario elettronico
La struttura della voce progettata risente naturalmente delle caratteristiche
dei dizionari storici. Ecco la sua architettura:
LEMMA + categoria grammaticale
0.1. Forme attestate nel corpus dei testi (con tutte le varianti)
La forma lemmatizzata per la voce principale è quella più
diffusa nell’uso odierno: ci si serve del GRADIT, Grande
dizionario italiano dell’uso, di Tullio De Mauro, con i relativi
aggiornamenti.
0.2. Nota etimologica essenziale.
0.3. Prima attestazione nel corpus.
0.3.1. Indicazione numerica della frequenza (per ciascuna
forma; nell’indicazione delle occorrenze, la seconda cifra,
preceduta dal segno +, si riferisce alle forme presenti in
eventuali indici).
0.4. Distribuzione geografica delle varianti.
Per ora si forniscono i dati relativi ai soli AIS e ALI.
Aggiungiamo in nota il riscontro con le forme registrate
da Touring Club Italiano 1931.
0.5. Note linguistiche/merceologiche (forestierismi; italianismi in
altre lingue).
La bibliografia per ora si riferisce solo alle ‘Note
linguistiche’, e, per quanto riguarda gli italianismi in altre
lingue, al DIFIT (consultabile in versione elettronica in
http://www.italianismi.org/difit-elettronico).
0.6. Riepilogo dei significati.
0.7. Locuzioni polirematiche e vere proprie (con la prima
attestazione nel corpus).
0.8. Rinvii (sono previsti soprattutto ‘iperlemmi’, o, se si
preferisce voci ‘generali’, di raccordo).
0.9. Corrispondenze lessicografiche (= riscontri nei dizionari e

JADT’ 18

95

nei corpora lessicografici in rete): si distinguono i vocabolari
etimologici (compreso il LEI) da quelli descrittivi (in ordine
cronologico, a partire dal Tommaseo-Bellini).
1. Prima definizione
Contesti
1.1. Definizione subordinata
Contesti
1.2. Definizione subordinata
Contesti
[...]
2. Seconda definizione
Contesti
[...].
La voce richiama, con gli opportuni adattamenti, quella del TLIO Tesoro della
Lingua Italiana delle origini, dell’Istituto dell’Opera del Vocabolario Italiano
del CNR di Firenze. I primi esperimenti, sui quali è basata ad esempio
l’ultima voce campione relativa a bagnomaria (a partire da una versione
iniziale del corpus, limitata a 28 testi), hanno evidenziato che la struttura
rischia però di essere troppo pesante in vista di una effettiva fattibilità
realizzativa del progetto.
I limiti “dimensionali” emergenti (che bene risultano evidenti in Bertini
Malgarini e Vignuzzi 2017) sono legati soprattutto alla ricchezza degli esempi
e all’ampiezza delle citazioni da altri strumenti lessicografici.
A entrambi questi limiti si pensa però di provvedere aumentando
l’interazione con gli altri strumenti collegati e collegabili.
In primo luogo prevedendo una profonda interazione tra banca dati testuale
e dizionario sia nella fase di redazione della scheda che in quella di
pubblicazione. In questo modo sarà possibile limitare il numero di esempi
citati per poi rimandare a un dossier completo delle occorrenze mediante il
collegamento con il corpus informatizzato. Nell’ottica di creare un accesso
aperto alla banca dati dei testi è opportuno porsi il problema dell’utilizzo
pubblico di testi coperti da diritto d’autore. Il tema è già stato affrontato
all’interno del gruppo PRIN 2008 “Il portale della TV, la TV dei portali” e in
occasione del convegno conclusivo del progetto Marina Pietrangelo –
ricercatrice dell’ITTIG (Istituto di Teoria e Tecniche dell’Informazione
Giuridica) appositamente invitata a parlare sul tema Per un uso legale degli
audiovisivi in corpora di ricerca – ha risposto con un sostanziale via libera
previsto dalla norma nel caso di progetti con esclusiva finalità di ricerca e
senza nessun risvolto economico (Pietrangelo 2017). Anche i riferimenti agli

96

JADT’ 18

altri dizionari vanno poi realizzati attraverso collegamenti con le versioni
elettroniche in rete attualmente disponibili (ad esempio quella del
Tommaseo-Bellini: Tommaseo online; quella delle edizioni del Vocabolario degli
Accademici della Crusca: Lessicografia della Crusca in rete; e infine quella del
vocabolario postunitario che si sta realizzando all’interno del progetto PRIN
2015 “Vocabolario dinamico dell’italiano post-unitario”, coordinato da Claudio
Marazzini). Sono tuttora allo studio procedure per il trattamento delle
collocazioni emergenti dagli strumenti statistici di analisi del corpus, e per la
lemmatizzazione di parole composte a fronte della polimorfia morfologica
emergente dalla profondità diacronica del corpus. All’interno di una vera e
propria stazione lessicografica tutti questi strumenti saranno integrati
all’interno di un sistema di back-office che, tramite fasi di valutazione
progressiva e di controllo, porterà alla diretta pubblicazione della voce in
rete. Infine, proprio la potenziale interazione/integrazione con il citato futuro
“Vocabolario dinamico dell’italiano post-unitario” ha suggerito al gruppo di
ricerca di predisporre una scheda lessicografica variabile: alla scheda
approfondita del dizionario storico si affiancheranno infatti una scheda
strutturata secondo le specifiche di un dizionario sincronico per quelle voci
che facciano ancora oggi parte dell’italiano dell’uso, e strumenti di
calibrazione dei campi che l’utente esperto e non esperto potrà gestire in
modo da avere di volta in volta una voce personalizzata.
In sede di discussione sarà presentata e discussa una voce “esemplare” del
VoSCIP, anche in relazione alla selezione e all’organizzazione del materiale
lessicografico e alla sua pubblicazione (in rete e in forma cartacea).

JADT’ 18

97

Riferimenti bibliografici
Bertini Malgarini, P. e Vignuzzi, U. (2017). Bagnomaria nel Vocabolario storico
della
cucina
italiana
postunitaria
(VoSCIP):
<
http://permariag.wixsite.com/permariagrossmann/vignuzzi>.
Biffi, M. (2016). Progettare il corpus per il vocabolario postunitario, in Marazzini,
C. e Maconi, L. (a cura di), L’italiano elettronico. Vocabolari, corpora, archivi
testuali e sonori. Accademia della Crusca, pp. 259-80.
Pietrangelo, M. (2016). Per un uso legale degli audiovisivi in corpora di ricerca, in
Alfieri, G., Biffi, M. et alii (a cura di), Il portale della TV. La tv dei portali.
Bonanno, pp. 171-185.
Ruffino, G. (1995). I pani di Pasqua in Sicilia. Un saggio di geografia linguistica e
etnografica. Centro di Studi Filologici e Linguistici Siciliani.
Touring Club Italiano (1931). Guida gastronomica d’Italia. Touring Club
Italiano [rist. anast. 2003].
Strumenti
AIS = Jaberg, K. e Jud, J. (1928-1940). Sprach- und Sachatlas Italiens und der
Südschweiz. Ringier, 8 voll. (trad. it. 1987. AIS. Atlante linguistico ed
etnografico dell’Italia e della Svizzera meridionale, Unicopli). Anche in rete:
NavigAIS, <http://www3.pd.istc.cnr.it/navigais/>.
ALEPO = Telmon, T. e Canobbio, S. (1984-). Atlante linguistico ed etnografico del
Piemonte occidentale (vedi <http://www.alepo.eu/>)
ALI = Bartoli, M. G et alii (1995-). Atlante Linguistico Italiano. Istituto
Poligrafico e Zecca dello Stato.
ALLI = Moretti, G. et alii (1982-). Atlante Linguistico dei Laghi Italiani
(vedi <http://www.lettere.unipg.it/ricerca/centri>
ALS = Ruffino, G. (1995-). Atlante Linguistico della Sicilia (vedi
<http://atlantelinguisticosicilia.it/>).
ALT = Giacomelli, G. (2000). Atlante Lessicale Toscano. LEXIS (in CD-ROM);
Ora in rete come ALT-WEB: <http://serverdbt.ilc.cnr.it/altweb/RT_ALTWEB_home.htm>.
ASLEF = Pellegrini, G. B. et alii (1972-). Atlante Storico-Linguistico-Etnografico
Friulano. Istituto di glottologia e fonetica dell’Università Istituto di
filologia romanza della Facoltà di lingue e letterature straniere
dell’Università.
DIFIT = Stammerjohann. H. (2008). Dizionario di italianismi in francese, inglese e
tedesco.
Accademia
della
Crusca.
Anche
in
rete:
<http://www.italianismi.org>.
GRADIT = De Mauro, T. (2007). Grande Dizionario Italiano dell’Uso. UTET.
LEI = Pfister, M. e Schweickard, W. (1979-). Lessico Etimologico Italiano, Edito
per incarico della Commissione per la Filologia romanza. Reichert.
Lessicografia della Crusca in rete = Accademia della Crusca (2004). Lessicografia

98

JADT’ 18

della Crusca in rete. <http://www.lessicografia.it>.
TLIO = Opera del Vocabolario Italiano (1997-). Tesoro della lingua italiana delle
origini. <http://www.vocabolario.org/>.
Tommaseo-Bellini = Tommaseo, N. e Bellini V. (1861-1879). Dizionario della
lingua italiana, Società L’Unione Tipografico-Editrice.
Tommaseo online = Accademia della Crusca (2015). Tommaseo online.
<http://www.tommaseobellini.it>.

JADT’ 18

99

Strumenti informatico-linguistici per la realizzazione
di un dizionario dell’italiano postunitario
Marco Biffi
Università degli Studi di Firenze – marco.biffi@unifi.it

Abstract
The paper focuses on some general problems about representative corpora
for the compilation of dictionaries. It starts from the concrete case of the
Vocabolario dell’italiano post-unitario, which, due to its hybrid nature, offers a
complete view of both the criticalities of synchronic lexicography and of the
historical one. Therefore is introduced the concept of Banca linguistica, that is
a platform in which different types of corpora, a search meta-engine of the
existing databases, and tools of access to existing electronic dictionaries
converge. A final paragraph is dedicated to the concept of “quantum
relativity” of data of computational linguistics.
Sintesi
Il contributo mette a fuoco alcuni problemi generali relativi alla costituzione
di corpora rappresentativi per la redazione di dizionari, partendo dal caso
concreto del Vocabolario dell’italiano post-unitario, che, per la sua natura ibrida,
offre un quadro completo sia delle criticità della lessicografia sincronica sia di
quella storica. Si introduce pertanto il concetto di Banca linguistica in cui
convergono diverse tipologie di corpora, un metamotore di ricerca per la
consultazione delle banche dati esistenti e sistemi di integrazione con i
dizionari elettronici esistenti. Infine ci si sofferma sul concetto di “relatività
quantistica” dei dati estrapolabili dalle ricerche informatico-linguistiche.
Keywords: Linguistica dei corpora, Italiano, Dizionario sincronico,
Dizionario storico, Testo elettronico, Bilanciamento, Metamotore, Banca
linguistica, Relatività quantistica, Informatica linguistica, Linguistica
computazionale
1. Introduzione
In questo contributo cercherò di mettere a fuoco alcuni problemi generali
relativi alla costituzione di strumenti per la redazione di dizionari partendo
da un caso specifico, quello del progetto di un dizionario “ibrido”, insieme
storico e sincronico, su cui sta lavorando un gruppo di ricerca nazionale
coordinato da Claudio Marazzini. Il progetto – che ha come obiettivo finale la
redazione di un vocabolario dell’italiano post-unitario che raccolga il
patrimonio linguistico nazionale della lingua ufficiale dello Stato dal 1861 a

100

JADT’ 18

oggi – ha visto l’avvio con una prima fase finanziata sul PRIN 2012 Corpus di
riferimento per un Nuovo Vocabolario dell’Italiano moderno e contemporaneo. Fonti
documentarie, retrodatazioni, innovazioni; e ha poi potuto continuare con un
secondo finanziamento sul PRIN 2015 Vocabolario dinamico dell’italiano postunitario. Ai due progetti hanno partecipato numerose università italiane:
Piemonte Orientale, Milano, Genova, Firenze, Viterbo, Napoli, Catania (al
progetto sul corpus ha partecipato anche l’Istituto di Teorie e Tecniche
dell’Informatica Giuridica ITTIG del CNR di Firenze; al progetto sul
vocabolario dinamico partecipa anche l’Università degli Studi di Torino);
come partner esterno ha collaborato l’Accademia della Crusca, per la quale il
dizionario post-unitario è uno dei tre progetti strategici attuali, accanto al
Vocabolario dantesco e all’Osservatorio degli italianismi nel Mondo (OIM).
Per quanto le dinamiche di impiego di corpora per la redazione di dizionari
storici siano note, soprattutto dopo l’esperienza del TLIO Tesoro della lingua
italiana delle origini dell’Istituto dell’Opera del Vocabolario Italiano del CNR
di Firenze, meno si è riflettuto sulle implicazioni pratiche della costituzione
di un dizionario sincronico basato su un corpus rappresentativo, e del tutto
nuovo è il caso di uno strumento ibrido come il vocabolario post-unitario, in
cui le criticità della lessicografia informatica storica e sincronica si mescolano,
evidenziando come si debba piuttosto muoversi nella direzione di strumenti
articolati.
2. Criticità di fisionomia di un corpus rappresentativo dell’italiano postunitario
Un primo problema da affrontare per un corpus rappresentativo per un
dizionario è la sua dimensione. Se proviamo a effettuare un rapido controllo
sulla situazione dei corpora di riferimento per altre lingue europee (in
particolare inglese e tedesco, che hanno avuto una maggiore attenzione a
questo tema), sia il British National Corpus (per il 10% costituito da trascrizioni
dell’inglese parlato – cfr. Cresti-Panunzi 2013: 36-37) che il DWDS-Kerncorpus
(testi del XX secolo di cinque tipologie: letteratura, 25%; giornali, 25%; prosa
scientifica, 20%; guide, libri di ricette e testi analoghi, 20%; lingua parlata
trascritta, 10% – cfr. Klein 2013: 18-19) hanno dimensione pari a circa 100
milioni di parole. Questa era la dimensione che nel primo decennio del secolo
individuava corpora di dimensioni standard (cfr. Chiari 2007: 45; secondo la
tabella ivi riportata); anzi, 100 milioni di parole era la soglia che divideva i
corpora standard da quelli di grandi dimensioni. Tenendo conto dei
progressi informatici e metodologici degli ultimi anni, certamente è
opportuno introdurre qualche correttivo; e in effetti sia per l’inglese che per il
tedesco questi correttivi esistono, perché i corpora bilanciati sono affiancati
da thesauri. Al BNC è stata recentemente affiancata la Bank of English (un

JADT’ 18

101

monitor corpus, secondo la terminologia di Sinclair, di testi completi per un
totale di 650 milioni di parole – cfr. Cresti-Panunzi 2013: 36-37); al Kerncorpus
si sono aggiunti alcuni moderni corpora di giornali (successivi al 1995) e altre
raccolte più piccole di testi, per un totale di 2,6 miliardi di parole (e, anche sul
piano diacronico, si sta cercando di completare il quadro con il Deutsche
Textarchiv in allestimento dal 2005 e ormai in via di completamento, che
raccoglie 1500 libri accuratamente scelti, di solito prime edizioni e volumi di
giornali, nell’arco cronologico compreso fra il 1650 e il 1900 – cfr. Klein 2013:
18-19). Per quanto riguarda la raccolta di testi si è già sottolineata
l’importanza di quella che è stata definita “parabola dimensionale dei
corpora” (Biffi 2016: 262).

Figura 1

La rappresentazione geometrica analitica di questa parabola evidenzia il
rapporto tra la lingua nei secoli (nella fattispecie l’italiano) e la possibilità di
rappresentarla con un corpus dell’ordine di grandezza di 100.000 parole
(kiloparole), di milioni di parole (megaparole), di miliardi di parole (gigaparole).
La possibilità di costruire corpora di grandi dimensioni diminuisce tanto
maggiormente quanto più si va indietro nel tempo, mentre aumenta
vertiginosamente per la lingua dei nostri giorni, con dimensioni ormai
veramente molto elevate, che non corrispondono certamente a tutto ciò che si
produce in una certa lingua, perché questo è ovviamente impossibile, ma che
tendenzialmente vi si avvicinano molto. La ridotta dimensione dei corpora
dell’italiano del passato – questo sottolinea la curva – non è soltanto legata al
fatto, oggettivo, che per il passato disponiamo di un minor numero di testi,

102

JADT’ 18

ma, in modo determinante, al fatto che molto più difficilmente riusciamo a
riunire i testi del passato in formato elettronico per poterli interrogare con
efficacia. Le difficoltà sono legate ai limiti di tutti gli strumenti informatici
coinvolti nella realizzazione di corpora informatici, che paradossalmente
convergono nel determinare l’andamento di questa curva: l’efficacia
dell’OCR (il riconoscimento ottico, automatico, dei caratteri), l’efficacia delle
morfologie macchina per la lemmatizzazione, l’efficacia dei motori di ricerca
disponibili con facilità e a costo poco elevato; quindi toccano i processi che
coinvolgono sia l’acquisizione dei testi, sia il loro trattamento, sia la loro
interrogazione e interrogabilità (Biffi 2016: 263-267). Per il passato gli effetti
della parabola rendono gestibile il problema di una reale rappresentatività
del corpus di riferimento. In effetti il TLIO, che si muove in un arco
cronologico che va dalle origini al 1375, può disporre come base di partenza
di un corpus che riunisce una raccolta consistente di testi volgari del periodo
considerato, spaziando a tutto tondo sull’asse diatopico e diafasico (e quindi
garantendo una grande rappresentatività anche in diastratia). Ha
fondamenta molto solide anche a fronte di dimensioni che, sulla scala di
misurazione dei corpora, non sono particolarmente elevate. Le ridotte
dimensioni hanno consentito infatti di abbattere gli effetti “parabolici”
dell’efficacia dell’acquisizione e del trattamento del testo elettronico (i testi,
ricavati dalle principali edizioni critiche hanno potuto essere sottoposti a
un’attenta collazione), così come dell’efficacia delle morfologie macchina (il
corpus è stato lemmatizzato di fatto manualmente con l’ausilio di procedure
semiautomatiche). La possibilità di progettare e realizzare un motore per
lemmi e un motore per forme personalizzato ha poi definitivamente
abbattuto i problemi di interrogazione/interrogabilità. Ma è evidente che
anche salendo di poco nella cronologia, proprio per l’effetto “parabolico”, i
problemi aumentano vertiginosamente. Per quanto riguarda le morfologie
macchina, ad esempio, sarebbe opportuno ricalibrarle in base alle variazioni
diacroniche delle strutture morfologiche e morfosintattiche, seguendo l’asse
del tempo (ed esperimenti si stanno facendo: ad esempio per la morfologia
della lingua di Leonardo in un progetto finanziato dalla Biblioteca
Leonardiana di Vinci e da me curato per la parte linguistica); ma il processo è
lungo e non è mai stato affrontato in modo sistematico, né
metodologicamente né pragmaticamente. Questo perché, ma vale per tutti gli
aspetti della linguistica computazionale e più in generale di quella che
preferisco chiamare linguistica informatica, la tendenza generale è quella di
lavorare per piccole monadi e non creare sistema mettendo in sinergia le
competenze e gli strumenti in modo da ampliare e affinare le tecnologie
disponibili rendendole sempre più potenti. Così oggi disponiamo di vari
strumenti, in parte sovrapponibili, in parte complementari, ma nulla di

JADT’ 18

103

realmente condivisibile da migliorare con un sistema open source, in modo da
concentrare gli sforzi su ciò che realmente manca e o è debole. Il “pezzo”
delle morfologie macchina è particolarmente significativo: costruire un
corpus diacronico per un dizionario storico significa infatti fornire i primi
mattoni per ricalibrare le morfologie macchine esistenti tarandole sul periodo
preso in considerazione; ma in nessun caso si è pensato di usare questi
corpora del passato come punto di partenza per migliorare le procedure di
lemmatizzazione che a loro volta potenzierebbero le possibilità
lessicografiche in un circolo virtuoso destinato a raffinare gli strumenti a
disposizione della comunità scientifica. Per tornare alle specificità del
dizionario dell’italiano post-unitario, il suo carattere ibrido lo colloca in una
posizione particolarmente delicata perché in quanto diacronico, dal 1861 al
2000, risente dei limiti informatici di cui abbiamo parlato (anche se, ad
esempio, in questo segmento cronologico le procedure di riconoscimento
automatico dei caratteri danno ottimi risultati). Ma diventa decisamente
sincronico nel periodo 2000-2014, quando abbiamo la possibilità di creare un
enorme corpus massivo (delle dimensioni delle gigaparole), anche con facilità,
semplicemente attingendo dal web mediante programmi di data crawling (web
crawler, o spider), come dimostra molto bene il caso di RIDIRE (www.ridire.it,
diretto da Emanuela Cresti), un corpus di 1,3 miliardi di parole, realizzato
con un crowler controllato che ha permesso un “bilanciamento” basato su
domini semantici (architettura e design, arti figurative, cinema, cucina,
letteratura e teatro, moda, musica, religione, sport) e domini funzionali
(amministrazione e legislazione, economia e affari, informazione).
3. Dal corpus rappresentativo alla “Banca linguistica”
Da un punto di vista teorico la scelta migliore per il corpus di riferimento del
dizionario dell’italiano post-unitario sarebbe quella di un corpus bilanciato
nell’ordine di megaparole dal 1861 al 2014, da affiancare con un corpus
massivo della dimensione delle gigaparole sul 2000-2014, un risultato, come si
è visto, ormai realizzabile. Però il gruppo di ricerca è partito da una
situazione pregressa di progetti già realizzati e studi già avviati con validi
risultati raggiunti, per cui si è scelto di mettere a frutto al massimo le
esperienze dei componenti del gruppo, recuperando tutti i materiali che
ciascuno poteva portare in dote al progetto per poi ampliarli e consolidarli
con competenze specifiche. La copertura quindi è “a macchia di leopardo”,
ed è pertanto necessario utilizzare al massimo, anche per la zona cronologica
che va dal 1861 al 2000, un approccio massivo, che conduce inevitabilmente
sulla strada della “banca linguistica”, del thesaurus, dal quale poi estrarre un
corpus bilanciato (o più di uno, in modo dinamico anche in relazione alle
esigenze del redattore della voce assegnata).

104

JADT’ 18

Figura 2

La “Banca linguistica” può essere una piattaforma in cui siano disponibili
vari sub-corpora, in cui siano raccolti tutti i materiali con una marcatura
semantica che consenta successivi bilanciamenti, con un “corpus centrale”
che sarà la base primaria del lavoro del lessicografo del vocabolario
postunitario, ma che andrà continuamente tarato grazie ai dati emergenti
dalla consultazione del corpus massivo contemporaneo e dei sub-corpora
diacronici presenti. La piattaforma dovrà anche dialogare con i dizionari
elettronici di cui disponiamo dal 1861 a oggi: il Tommaseo Online e la versione
elettronica della quinta edizione del Vocabolario degli Accademici della
Crusca presente nella Lessicografia della Crusca in rete per la parte diacronica
(nella speranza che l’accordo siglato nel settembre 2017 tra UTET e
Accademia della Crusca per la digitalizzazione del Grande Dizionario della
Lingua Italiana maturi frutti rapidi); le versioni dei dizionari sincronici
presenti in rete (il Sabatini Coletti, il De Mauro, il Treccani, e tutto quanto sarà
disponibile); tutti i corpora dell’italiano presenti sul web, inclusi quelli,
preziosissimi, degli archivi elettronici delle principali testate giornalistiche
nazionali (Biffi 2016: 272-273). Non va dimenticato infatti che il panorama dei
corpora dell’italiano è abbastanza ampio (per un quadro generale si veda
Cresti-Panunzi 2013; ma è necessario perfezionare il censimento). Però è
mancata, come del resto è naturale, una politica organica di costruzione di un
sistema: abbiamo quindi un’estrema eterogeneità di strumenti, piattaforme,
codifiche (per fortuna in anni recenti, almeno per quest’ultimo aspetto, la
forza centrifuga si sta progressivamente contenendo con il ricorso sempre più
frequente, se non totale, alla codifica XML/TEI), che costringe il ricercatore a
collegarsi n volte, su n piattaforme, con n filosofie diverse, con n motori
diversi, per poter effettuare una ricerca a tutto campo. Diventa quindi

JADT’ 18

105

fondamentale un metamotore. Una versione beta di metamotore dei corpora
dell’italiano è stata realizzata dall’unità di ricerca dell’Università degli Studi
di Firenze del gruppo PRIN 2012 da me diretta (www.metaricerche.it). Come
si legge nella sezione del portale intitolata “Il metamotore”: «Gli strumenti
individuati sono stati classificati secondo i possibili livelli di integrazione:
corpora liberamente consultabili; corpora liberamente consultabili previa
registrazione; corpora da scaricare. È stato poi predisposto uno studio di
fattibilità per la definizione di una serie di procedure atte ad analizzare gli
strumenti di partenza, determinare il livello di integrabilità (che passa anche
dalla possibilità di poter interagire con lo staff tecnico della singola banca
dati, a seguito di un accordo “strategico” sulla condivisione dei contenuti) e
individuare delle procedure da seguire a seconda del livello. Si è passati poi a
definire l’architettura del sistema, la tecnologia di riferimento e l’interfaccia
di consultazione, almeno per una prima versione prototipale della
piattaforma». La versione beta prevede l’integrazione di 8 banche dati, scelte
come campioni delle principali tipologie di livelli di integrazione:
 Livello massimo (si è trovato accordo con lo staff tecnico che gestisce
la banca dati): LIR (Lessico dell’Italiano Radiofonico), LIS (Lessico
dell’Italiano Scritto) e LIT (Lessico Italiano Televisivo), Accademia della
Crusca.
 Livello base (si è integrata la banca dati in una finestra, in attesa di
una maggiore interoperabilità): MIDIA (Morfologia dell'Italiano in
DIAcronia) Università Roma Tre; CorDIC (Corpora Didattici Italiani di
Confronto) Laboratorio Linguistico Italiano Università degli Studi di
Firenze.
 Livello minimo (si è integrata la banca dati in una finestra senza
possibilità di maggiore interoperabilità): Archivio dei quotidiani
«Corriere della Sera» e «La Repubblica».
Se questo strumento potrà essere potenziato fino a riunire nella lista dei
risultati tutte le banche dati testuali disponibili attualmente per l’italiano,
nella “Banca linguistica” del redattore del Vocabolario post-unitario sarà
disponibile un accesso centralizzato a tutti i corpora esistenti, da integrare,
modulare e bilanciare con il corpus riunito dal gruppo di ricerca PRIN, con il
corpus massivo dell’italiano contemporaneo, con gli strumenti lessicografici
elettronici. Rimangono da considerare alcune criticità che, se rimosse,
consentirebbero un ulteriore potenziamento della “Banca linguistica”, e che
possiamo richiamare in questa sede solo brevemente per punti. a) La gran
parte dei testi (ad esempio quelli letterari recenti) sfuggono alla possibilità di
essere organizzati in corpora interrogabili per le difficoltà legate ai diritti
d’autore. b) Le raccolte di corpora in diacronia, tranne rare eccezioni (come
ad esempio il CEOD, Corpus Epistolare Ottocentesco Digitale) prediligono la

106

JADT’ 18

tradizione letteraria di registro alto. Esistono già campioni rappresentativi di
italiano post-unitario, come il DIACORIS (25 milioni di occorrenze), ma si
devono ancora integrare i vuoti legati alle lingue speciali (come è stato
tentato di fare all’interno del progetto PRIN 2012). c) Resta da indagare
quanto dal web si possano recuperare (in modo più o meno automatico)
materiali per le sezioni in diacronia, grazie soprattutto alla presenza
massiccia di testi ottocenteschi riuniti in biblioteche digitali come Google libri
e Archive.
4. Informatica linguistica e relatività quantistica
Se il punto di partenza per la redazione di un dizionario non è più un corpus
di riferimento omogeneo predisposto allo scopo, ma una “banca linguistica”
in cui si è chiamati a gestire materiali non omogenei ed esogeni, non è inutile
richiamare in questo paragrafo finale l’importanza dei risvolti “quantistici”
della linguistica informatica (Biffi 2017: 545-549).
Consultando banche dati (includendo in questa categoria non solo i corpora
ma anche le edizioni elettroniche dei dizionari) non è difficile imbattersi in
diffrazioni nei risultati quantitativi (e quindi in quelli qualitativi, nella misura
in cui possono determinarsi lacune nella ricerca di determinati contesti), che
sicuramente in parte si spiegano con errori umani inseriti nelle varie fasi
realizzative delle banche dati (dovuti ai moderni copisti digitali, ai
programmatori, al progetto), ma anche per il concorso di fattori precisi e
individuabili. Nel contributo citato (Biffi 2017) le diffrazioni riguardano i
risultati relativi al numero dei lemmi nelle tre versioni elettroniche del
Vocabolario degli Accademici della Crusca del 1612, e sono da ricondurre a
diversità di tokenizzazione, diversità di approccio nella restituzione alle voci
dell’intrinseca struttura di base di dati, diverse priorità nella restituzione del
testo elettronico. In altre banche dati i fattori di diffrazione saranno
probabilmente da ricondurre ad altro, ma si dovrà sempre tener conto delle
caratteristiche e dell’architettura della banca dati così come degli strumenti di
ricerca a essa applicati. Come nelle scienze esatte da Heisenberg in poi si
deve tener conto dell’indeterminazione introdotta dallo strumento di
misurazione, consultando le banche dati sarà opportuno ricordare che le
caratteristiche dello strumento di conoscenza (in questo caso la banca dati)
perturbano il risultato della ricerca costringendoci a un’inevitabile
approssimazione “quantistica”; una perturbazione però dominabile, giacché
si possono ricostruire le cause di diffrazione e quindi correggere il risultato
finale, come avviene con la meccanica quantistica laddove è necessario
sostituirla alla meccanica classica.
E allora, per poter ottenere risultati scientifici consultando una banca dati, è
necessario conoscere a fondo le caratteristiche dello strumento, e tenere conto

JADT’ 18

107

della sua variabilità “quantistica” nel momento in cui leggiamo i dati. E,
quando si leggono e gestiscono i risultati, è necessario non solo essere
consapevoli di quale strumento si è usato, ma anche delle specifiche modalità
di ricerca applicate; in altre parole si deve tener conto continuamente del
contesto filologico della ricerca informatica, esattamente come, quando si
consulta l’edizione critica di un testo, si tiene conto anche delle varianti
dell’apparato.
Riferimenti bibliografici
Biffi, M. (2016). Progettare il corpus per il vocabolario postunitario, in Marazzini,
C. e Maconi, L. (a cura di), L’italiano elettronico. Vocabolari, corpora, archivi
testuali e sonori. Accademia della Crusca, pp. 259-80.
Biffi, M. (2018). Tra fiorentino aureo e fiorentino cinquecentesco. Per uno studio
della lingua dei lessicografi, in Belloni, G. e Trovato, P. (a cura di), La Crusca
e i testi. Lessicografia, tecniche editoriali e collezionismo librario intorno al
Vocabolario del 1612. libreriauniversitaria.it, pp. 543-560.
Chiari, I. (2007). Introduzione alla linguistica computazionale, Laterza.
Cresti, E. e Panunzi, A. (2013). Introduzione ai corpora dell’italiano, Il Mulino.

108

JADT’ 18

Comparaison de corpus de langue « naturelle » et de
langue « de traduction » : les bases de données
textuelles LBC, un outil essentiel pour la création de
fiches lexicographiques bilingues
Annick Farina, Riccardo Billero
Università degli Studi di Firenze – annickfarina@unifi.it; riccardo.billlero@gmail.com

Abstract
The aim of this paper is to describe the work done to exploit the LBC
database for the purpose of translation analysis as a resource to edit the
bilingual lexical sections of our dictionaries of Cultural Heritage (in nine
languages). This database, made up of nine corresponding corpora, contains
texts whose subject is cultural heritage, ranging from technical texts on art
history to books on art appreciation, such as tour guides, and travel books
highlighting Italian art and culture. We will illustrate the different questions
with the SketchEngine LBC French corpus, made up at the moment of
3,000,000 words. Our particular interest here is in research that not only
orients lexical choices for translators but that also precedes the selection of
bilingual quotations (from our Italian/French parallel corpus) and that we
rely on for editing an optional element of the file called "translation notes."
We will rely on this as much for works on "universals of translation" already
described by Baker (1993) as for studies aimed at improving Translation
Quality Assessment (TQA). We will show how a targeted consultation of
different corpora and sub-corpora that the database allows us to distinguish
("natural language" vs "translation”, "technical texts" vs "popularization
texts" or "literary texts") can help us identify approximations or translation
errors, so as to build quality comparative lexicographical information.
Keywords: electronic lexicography, multilingual lexical resources, corpus
linguistics
Résumé
Cet article a pour but de décrire notre travail sur la base de données LBC
pour ce qui concerne l’analyse de traductions comme ressources pour la
rédaction de la partie bilingue de nos dictionnaires du Patrimoine (dans les
neuf langues du projet). La base de données contient des corpus distincts de
neuf langues composés de textes qui sont tous reliés au patrimoine italien :
des textes techniques des différents domaines artistiques, des ouvrages de
critique d’art ou d’histoire de l’art, des guides touristiques, des récits de

JADT’ 18

109

voyages, etc. Nous illustrerons différentes interrogations du corpus français
(actuellement composé d’environ 3 millions de mots) dans SketchEngine. En
particulier, nous nous intéresserons à des recherches qui nous guident non
seulement vers la sélection de traduisants pour certains termes mais qui
précèdent aussi la sélection de citations bilingues (extraites de notre futur
corpus parallèle italien/français) et sur lesquelles nous nous appuyons pour
la rédaction d’un élément facultatif de la fiche appelé « notes de traduction ».
Nous nous appuyons pour ce faire tant sur les travaux sur les « universaux
de traduction » (Baker 1993) que sur études qui visent à l’amélioration de la
qualité des traduction (TQA : Translation Quality Assessment). Nous
montrerons comment une consultation ciblée des différents corpus et souscorpus que la base nous permet de distinguer (textes en « langue naturelle »
vs « en traduction », « textes techniques » vs « de vulgarisation » vs «
littéraires ») peut nous aider à repérer des approximations ou des erreurs de
traduction, nous aidant à construire une information lexicographique
comparative de qualité.
Keywords: lexicographie, ressources lexicales plurilingues, corpus
linguistiques.
1. Introduction
Un des principaux buts du projet Lessico dei Beni Culturali est de constituer
des dictionnaires monolingues de neuf langues différentes en fonction d’un
usage précis relié à un objet particulier : la description (et traduction de
descriptions) du patrimoine toscan principalement dans des textes de
vulgarisation (guides touristiques, sites de musées, etc.). Pour ce faire, nous
avons constitué des bases de données textuelles, que nous complétons sous la
forme d’un Work in progress, qui nous serviront pour différentes tâches, de la
création
de
nomenclatures
à
la
rédaction
de
fiches
lexicographiques/terminologiques monolingues et de fiches de traduction
reliant les nomenclatures des différentes langues entre elles (pour la
description de ces bases cfr. Billero et al. 2017). C’est l’utilisation de ces bases
de données textuelles pour la rédaction de fiches bilingues de traduction que
nous illustrerons ici1, en nous basant sur l’analyse de différentes
interrogations sur SketchEngine (principalement statistiques et de contexte)
de notre corpus LBC français, composé actuellement d’environ trois millions

Pour l’utilisation de nos bases pour la réalisation des dictionnaires
monolingues, voir l’article de Nicolás et Lanini dans ce volume. Nous constituons en
effet actuellement les nomenclatures des différentes langues en suivant le modèle
qu’elles ont défini pour l’italien. Le lien bilingue entre ces différentes nomenclatures
ne sera possible que lorsque nous aurons constitué nos bases de données parallèles.
1

110

JADT’ 18

de mots. Nous comparerons en particulier des données provenant de
plusieurs sous-corpus comparables de textes « en langue naturelle » et de
textes « en traduction ». Nous proposerons aussi une première comparaison
de résultats provenant d’un sous-ensemble du corpus italien avec un sousensemble contenant les traductions françaises des mêmes textes, qui
constituent un matériau fragmentaire pour le moment parce que nous
travaillons encore à l’insertion des textes dans le but de créer des bases
parallèles de traduction de l’italien vers toutes les langues du projet. Nous
montrerons comment une consultation ciblée des différents corpus et souscorpus que la base nous permet de distinguer (italien « langue naturelle » vs
français « langue naturelle », français « en traduction » vs français « langue
naturelle », français « textes spécialisés » vs français « vulgarisation » vs
français « littéraire ») peut nous aider à repérer des approximations ou des
erreurs de traduction, nous aidant à construire une information
lexicographique comparative de qualité.
2. Comparaison entre corpus « en langue naturelle » et « en traduction » :
une perspective à mi-chemin entre traductologie descriptive et prescriptive
Nous appuyant sur des analyses qui ne considèrent pas la langue de
traduction comme un « troisième code » (Frawley 1984), nous estimons pour
ce que des textes traduits trouvent parfaitement leur place à l’intérieur d’une
base textuelle unique d’une même langue, aux côtés de textes « en langue
naturelle ». Cependant, sur le modèle de propositions d’utilisation de corpus
de traduction dans un but didactique, tant pour l’enseignement des langues
que pour celui de la traduction, il nous semble nécessaire d’offrir la
possibilité d’une consultation de la base dans des sous-corpus distincts
regroupant des textes des deux types et de définir des critères d’évaluation
des textes traduits à intégrer dans la base, en constituant des corpus séparés
de textes traduits dans toutes les langues du projet. Ces corpus nous sont
utiles comme outils de mémoire de traduction pour travailler sur la partie
bilingue de nos fiches lexicographiques dans une perspective plus
prescriptive que descriptive. Comme le montrera notre comparaison de
résultats provenant de notre base française LBC « en langue naturelle » et «
en traduction » avec un corpus de près de 100.000 mots actuellement non
intégré dans la base composé de traductions d’ouvrages de « vulgarisation »
traduits en français (guides touristiques de la Toscane et sites de musées
surtout), certains des textes qui nous intéressent présentent des
caractéristiques que l’on peut assimiler à du « translationese » et ne
pourraient que fausser des interrogations de la base visant à attester des
formes ou structures typiques du français tel qu’il est écrit et parlé par la
majorité des locuteurs de cette langue sans interférence avec une autre
langue.

JADT’ 18

111

2.1 Information descriptive et prescriptive dans les dictionnaires LBC :
universaux et écarts
A la suite de Baker (1993), nous partons du principe qu’il existe des
universaux de traduction qui nous serviront de canevas pour l’illustration
des différents types d’interrogation effectués à l’intérieur de nos sous-corpus
et de comparaison des résultats obtenus. C’est sur ces universaux que nous
nous basons pour fournir la partie descriptive de l’information
lexicographique comparative détaillée présente dans la partie bilingue de nos
dictionnaires. Cette information correspond d’abord à l’observation des
corpus parallèles, qui fournissent des attestations de traduction des lemmes
(mots ou collocations) décrits par le dictionnaire, apparaissant dans des
citations bilingues à l’intérieur de la partie bilingue de l’article. Nous
analyserons en particulier :
- la simplification (principalement, pour ce qui concerne notre corpus, le
choix d’hyperonymes pour traduire certains termes plus spécifiques) qui
donne lieu dans nos dictionnaires à l’introduction d’une information
sémantique ajoutée qui accompagne le traduisant proposé : les traits
distinctifs particuliers au lemme qui ne sont pas rendus par le traduisant
seront indiqués avec ou sans parenthèses après le traduisant (par ex. tavola
traduit par peinture (sur bois) et tavoletta traduit par (petite) peinture (sur bois))
- le nivellement (non-respect du registre, par exemple le choix de
technicismes plutôt que de mots de la langue générale et vice versa). Toutes
les entrées ont une indication de marque d’usage. Dans le cas d’une
traduction qui implique un changement de registre, ce changement sera
relevé dans la partie « note de traduction » ou apparaitra dans la partie
réservée aux indicateurs sémantiques distinctifs dans le cas où plusieurs
traductions du même lemme seraient possibles avec ou sans perte de registre.
C’est le cas par exemple de tondo italien (non marqué) par rapport à médaillon
(non marqué) et à l’italianisme tondo (technicisme utilisé principalement par
les historiens de l’art). Baker analyse aussi l’explicitation qui est
particulièrement fréquente dans les textes qui nous intéressent parce qu’elle
est quasi systématiquement utilisée lors de l’usage d’un italianisme, en
particulier pour les realia qui ont un traitement particulier dans nos
dictionnaires (cfr. Farina 2014, 2016). Il serait possible de rechercher d’une
manière systématique ce type de données dans notre corpus en extrayant
toutes les occurrences de « type de » ou « sorte de » ou les éléments indiqués
entre parenthèses, mais nous avons volontairement laissé de côté cette
catégorie qui est trop fortement reliée à l’objet décrit par nos textes et à des
choix stylistiques partagés entre les auteurs de textes « en langue naturelle »
et les traducteurs dans le contexte de notre base, et ne nous permettrait donc
pas d’illustrer par une comparaison des deux types de ressources des

112

JADT’ 18

contraintes linguistiques reliées aux opérations de traduction2. Nous avons
laissé de côté aussi la « normalisation » ou « conservatisme » qui s’adapte
peu à notre matière, peu propice à la variation ou à l’exploration sur le plan
lexical et stylistique. Contrairement à Baker (1993 : 243), qui définit les
universaux de traduction comme des « features which typically occur in
translated text rather than original utterances and which are not the result of
interference from specific linguistic systems », nous avons adopté une
perspective plutôt prescriptive, ou mieux didactique, en prenant en
considération les phénomènes d’interférence (influence de la langue source
sur la langue cible) fréquents dans des opérations de traduction qui
concernent deux langues proches comme l’italien et le français et dans des
textes dont la qualité est loin d’être homogène. L’interférence est en effet
selon nous à la source non seulement de nombreux cas de simplification et
d’écarts de nivellement trouvés dans nos comparaisons mais d’autres
manifestations assimilables à des pertes découlant de l’opération de
traduction, voire à des erreurs ou inexactitudes de traduction. Le modèle du
TQA (Translation Quality Assessment) et, en particulier, les différents types
de mesures de qualité qui peuvent orienter le traducteur vers une
amélioration de la fluidité et de la précision peuvent nous servir de référence
pour ce faire (cfr. « Multidimensional Quality Metrics », Uszkoreit et al.
2013). Ces analyses nous orientent principalement vers le choix d’une
position qui peut sembler aller à l’encontre d’une exploitation de corpus
descriptive comme celle de Baker. De fait, elle se présente comme un
accompagnement permettant à l’utilisateur de nos dictionnaires d’effectuer
des choix, sur la base d’une exploitation descriptive des ressources consultées
telle que nous l’avons déjà décrite et de l’indication de données statistiques
résultant d’analyses de fréquence comme celles que nous les présenterons cidessous. Le rédacteur des fiches lexicographiques pourra de plus décider, le
cas échéant et lorsque nos analyses de ces données le pousseront à repérer
des erreurs ou écarts qui pourraient être réduits, de ne pas proposer une
forme qui apparait dans la base comme traduisant (tout en l’indiquant dans
la partie de l’article fournissant des indications statistiques sur les traduisants
trouvés) ou de rédiger la partie « note de traduction », facultative dans nos
articles bilingues, pour conseiller les utilisateurs dans leurs choix en
expliquant pourquoi certaines formes peuvent être préférées à d’autres.

2 L’utilisation abondante d’italianismes est une caractéristique dominante dans
les guides touristiques analysés, assimilable à une volonté de leurs auteurs de donner
à ces textes une « touche d’italianité» (Farina 2014 : 61)

JADT’ 18

113

3. Langue naturelle vs langue traduite : observation du corpus
La différence de fréquence de mots ou de collocations présents dans des
corpus comparables contenant des textes français en « langue naturelle » et
des textes qui proviennent d’une traduction en français peuvent nous
permettre de repérer des formes choisies sous l’influence de la langue source.
3.1 Fréquence zéro dans les textes en langue naturelle
Nous avons comparé la liste des mots présents dans le sous-corpus LBC de
textes de vulgarisation écrits en français contenant 270.000 mots avec un
corpus non intégré à la base pour le moment de textes de même type mais en
traduction 93.000 mots en réalisant une liste des mots présents exclusivement
dans le sous-corpus « en traduction ».
- fautes
La majorité des formes rencontrées sont assimilables à des fautes : absence
d’accent (cloitre), influence de l’orthographe italienne le français (baroche),
« francisation » excessive au niveau orthographique (Caliari) ou par
l’utilisation d’une traduction française là où l’usage préconise la forme
italienne (Sainte-Réparate désigne en français la personne ou la cathédrale de
Nice mais pas l’église Santa Reparata de Florence, la forme française n’est
attestée nulle part dans la base LBC) ou l’inverse (Giove n’est jamais utilisé en
italien dans notre corpus, où il est traduit par Jupiter), utilisation de mots qui
n’ont rien à voir avec la description du patrimoine florentin, probablement
parce qu’ils correspondent à un sens du mot-source qui s’applique à d’autres
contextes (coursive dans une description du Dôme de Florence, ou panonceau
pour se référer aux compartiments des portes du Paradis). Ce genre d’erreurs
ne donne pas lieu à la réalisation d’une information ciblée à l’intérieur des
dictionnaires sauf dans le cas d’une grande fréquence de l’erreur (par ex.
pour panonceau présent dans plusieurs sources avec un total de 8 occurrences
mais pas coursive qui n’a qu’une attestation).
- nivellement
On peut distinguer des formes qui correspondent à une différence
« pragmatique » ou stylistique entre français et italien qui ne nous intéressent
pas d’un point de vue lexicographique, comme l’utilisation de mentionnons
dans plusieurs textes en traduction qui ne se retrouve dans aucun des textes
de la base complète, ou de certaines formes du passé-simple (décora, succéda)
qui ne sont pas utilisées dans les textes de vulgarisation en français
« naturel ». Il s’agit de formes qui correspondent à des normes différentes
relatives aux types de texte du corpus : une analyse plus approfondie
pourrait probablement nous permettre d’observer un usage peu ou pas
attesté du « nous » dans les guides touristiques, et l’usage peu fréquent de
formes au passé-simple par rapport au passé-composé ou au présent, etc.

114

JADT’ 18

Ce qui nous intéresse beaucoup plus dans cette comparaison c’est de repérer
des formes qui, tout en étant parfaitement « correctes » en français, peuvent
être considérées comme hors contexte par rapport aux usages attestés dans le
même type de contexte en langue naturelle. La différence dans l’usage d’un
mot non attesté peut faire l’effet d’un « anachronisme » (différence dans la
fréquence d’usage en synchronie). C’est le cas par exemple de l’adjectif grandducal et du participe passé paraphé dont les équivalents italiens sont plus
fréquents dans la langue d’aujourd’hui que ne le sont leurs traductions
littérales françaises. L’écart dans le registre peut aussi s’appliquer dans le cas
d’une différence de « technicité ». L’adjectif autographe présent dans plusieurs
sources de vulgarisation en traduction est absent des textes de même type de
notre corpus en langue naturelle, mais on en trouve quelques occurrences
dans des textes plus spécialisés du corpus général. La différence de registre
donnera lieu à un marquage différencié entre lemme en langue source et sa
traduction attestée.
3.2 Différence de fréquence dans les textes-source par rapport aux textes-cible
Pour illustrer les phénomènes de simplification, nous avons interrogé deux
sous-corpus de notre base LBC constitués de 51 vies de l’ouvrage Le vite de'
più eccellenti pittori, scultori e architettori de G. Vasari (1568) et de leurs
traductions en français (traduction Leclanché-Weiss, 1900). Ne pouvant
encore nous baser sur des statistiques provenant des bases parallèles de
traduction (pour la description de ces bases cfr. Zotti 2017), nous nous
sommes concentrés sur des mots français qui avaient une grande fréquence
en comparant cette fréquence à celle du mot le plus proche en italien (même
sens, mêmes traits distinctifs). Ceci nous a permis de relever des écarts de
fréquence qui nous pousseront à une étude plus approfondie dans le but de
définir des réseaux analogiques dans les deux langues qui nous donnent la
possibilité de proposer des liens de traduction permettant d’éviter une perte
de précision. Tableau a par exemple une fréquence de 2232 par million de
mots dans notre sous-corpus français tandis que quadro a une fréquence de
793 par million de mots dans le sous-corpus italien contenant les mêmes
textes en langue originale. Un grand nombre d’hyponymes de quadro sont en
effet traduits par tableau en français. Si cette perte est probablement
compensée par l’ajout de traits distinctifs qui accompagnent le mot, nous
retenons que le traducteur ne pourrait que gagner en précision si nous lui
proposions d’autres formes pour rendre le sens de ces différents hyponymes.
4. Conclusion
La comparaison de résultats qui concernent la fréquence de formes à
l’intérieur du corpus LBC nous a permis d’illustrer l’utilisation de différents

JADT’ 18

115

sous-corpus pour orienter l’information tant descriptive que normative que
nous souhaitons fournir dans la partie bilingue de nos dictionnaires LBC.
« Nous considèrerons, même si cela reste à démontrer […] qu’une sur- ou
une sous-représentation d’un phénomène linguistique donné peut
correspondre à une violation de la contrainte d’usage […] et qu’une bonne
traduction se doit de tendre vers une homogénéisation entre la langue
originale et la langue traduite. » (Loock et al. 2013 : sp)
L’application de méthodes visant à la vérification de la qualité des
traductions et la création d’outils qui se basent sur des analyses critiques de
traductions existantes, en les comparant, en particulier, à des productions qui
ne passent pas par la médiation d’une autre langue devrait permettre une
optimisation du caractère naturel des textes traduits et de la précision,
objectif essentiel pour la diffusion d’une information de qualité.
Bibliographie
Baker M. (1993). Corpus Linguistics and Translation studies. Implications
and Applications. In Baker M. and al. editors, Text and Technology,
Amsterdam/Philadelphie, Benjamins, pp.233–250.
Billero R., Nicolas Martinez, M.C. (2017). Nuove risorse per la ricerca del
lessico del patrimonio culturale: corpora multilingue LBC. In CHIMERA
Romance Corpora and Linguistic Studies, Vol.4, No. 2, pp 203-216, ISSN:
2386-2629, 2017
Farina A. (2016). Le portail lexicographique du Lessico plurilingue dei Beni
Culturali, outil pour le professionnel, instrument de divulgation du savoir
patrimonial et atelier didactique, PUBLIF@RUM, vol. 24
http://publifarum.farum.it/ezine_articles.php?id=335
Farina A. (2014). Descrivere e tradurre il patrimonio gastronomico italiano: le
proposte del Lessico plurilingue dei Beni Culturali. In Chessa F. and De
Giovanni C., La terminologia dell'agroalimentare, Milan, Franco Angeli, pp.
55-66.
Frawley W. (1984). Prolegomenon to a theory of translation. In Frawley W.
editor, Translation: Literary, Linguistic and Philosophical Perspectives,
Newark, Univ. of Delaware Press : 159-175
Loock R., Mariaule M. and Oster C. (2013). Traductologie de corpus et qualité
: étude de cas. Tralogy - Session 5 - Assessing Quality in MT / Mesure de
la qualité en TA http://lodel.irevues.inist.fr/tralogy/index.php?id=188
Johansson S. and Hofland K. (1994). Towards an English-Norwegian parallel
corpus. In Fries U. and al. editors, Creating and using English language
corpora, Amsterdam, Rodopi pp. 25-37.
Loock R. (2016), La Traductologie de corpus. Villeneuve-d'Ascq. Presses
Universitaires Septentrion.

116

JADT’ 18

Uszkoreit H., Burchardt A. and Lommel A. (2013). A New Model of
Translation Quality Assessment Tralogy - Session 5 - Assessing Quality in
MT / Mesure de la qualité en TA
http://lodel.irevues.inist.fr/tralogy/index.php?id=319
Zotti V. (2017). L’integrazione di corpora paralleli di traduzione alla
descrizione lessicografica della lingua dell’arte : l’esempio delle
traduzioni francesi delle Vite di Vasari. In Zotti V., Pano A. editors,
Informatica Umanistica. Risorse e strumenti per lo studio del lessico dei beni
culturali. Firenze University Press.

JADT’ 18

117

Il rapporto tra famiglie di anziani non autosufficienti
e servizi territoriali: un'analisi dei dati esploratoria
con l'Analisi Emozionale del Testo (AET)
Felice Bisogni1, Stefano Pirrotta2
Associazione GAP - SPS Scuola di Psicoterapia Psicoanalitica - felice.bisogni@gmail.com
2Associazione GAP - SPS Scuola di Psicoterapia Psicoanalitica - stefanopirrotta@gmail.com
1

Abstract
In this paper the authors present a research committed by a local authority to
explore the relationship between not self-sufficient elders, their family
members and the community based assistance services they uses. The
exploratory data analysis, conducted with the Emotional Text Analysis (ETA)
(Carli, Paniccia, 2002), was used to identify emotional and cultural factors
related to the experience of assisting and being assisted at home and within
the community based services. The ETA has been realized on an assembled
text corpus produced transcribing 45 audio recorded interviews to not selfsufficient elders and their family members, patients of general practitioners
and/or users of the community based services (home-based and halfresidential). The interviews has been processed with T-Lab statistic software
(Lancia, 2004) and ETA has been applied to produce a clusters analysis. Four
clusters of dense words related to each others on 3 factorial axes emerged.
From the factorial axes emerges a emotional representation of elderlness as a
continuos allert related to the risk of dyng and as a depressive prescription to
survive related to the pretension to be assisted within their own family in
virtue of “blood ties”. The reciprocal control and contentiousness, and the
desirers to transgress the obligation of care giving and being cared are some
relevant emotions emerging by the ETA. The research's results shows also a
demand of a new assistance model emerges, founded on the possibility to
talk, to play and to have fun with others. Finally it emerges a demand of
services not only dealing with medical problems but also providing
psychological support and training to the families to develop relational
competences and to build reliable relationship out of the family. In the
conclusions of the paper some considerations regarding the relationships
between the clusters on the factorial axes and between clusters and
illustrative variables are highlithed.

118

JADT’ 18

Abstract
In questo articolo gli autori presentano una ricerca, condotta con la
metodologia dell'Analisi Emozionale del Testo (AET) (Carli, Paniccia, 2002),
commissionata da un ente locale al fine di esplorare i fattori emozionali che
organizzano l'esperienza di relazione tra un gruppo di anziani non
autosufficienti e i loro familiari e alcuni servizi socio-sanitari territoriali.
L'AET è stata realizzata su un corpus di testo assemblato trascrivendo 45
interviste audio registrate ad anziani non autosufficienti e loro familiari, che
utilizzano servizi di medicina generale e/o servizi sociali territoriali (di tipo
domiciliare o semiresidenziale). Le interviste sono state processate con il
software statistico T-lab (Lancia, 2004) e l'AET è stata applicata per produrre
una Cluster analysis. Dall’analisi sono emersi 4 cluster di “parole dense”
(Carli, Paniccia, 2002) in rapporto tra loro su 3 assi fattoriali, che
rappresentano il modo emozionale condiviso con cui gli intervistati parlano
delle loro attese sui servizi Dall’interpretazione dei dati è emerso un rapporto
tra famiglia ed anziano in crisi nel condividere desiderio e piacevolezza nello
stare insieme. Emerge una rappresentazione emozionale dell'anzianità come
allerta continua di fronte al rischio di morire e prescrizione depressiva a
sopravvivere connessa alla pretesa di essere assistiti all'interno della propria
famiglia in virtù di “rapporti di sangue”. A questo si contrappone il desiderio
di trasgredire l'obbligo famigliare ad assistere e farsi assistere. I risultati della
ricerca rilevano una domanda di nuovi modelli di assistenza fondati sulla
possibilità di parlare, giocare e divertirsi. Una domanda di servizi non rivolti
esclusivamente ai problemi medici ma anche a offrire supporto psicologico e
formazione alle famiglie per sviluppare competenze relazionali e relazioni
affidabili all'esterno della famiglia. Nelle conclusioni vengono messe in
evidenza alcune considerazioni riguardanti il rapporto tra cluster sugli assi
fattoriali e tra i cluster e le variabili illustrative.
Keywords: Emotional Text Analysis (ETA), assistance, elders, family,
community based services.
1. Introduzione
Sono circa 2,5 milioni gli anziani non autosufficienti presenti in Italia.
Secondo le più recenti previsioni ISTAT (2017), la percentuale di individui di
65 anni e più crescerà di oltre 10 punti percentuali entro il 2050, arrivando a
costituire il 34% della nostra popolazione. La presenza di un anziano non
autosufficiente in famiglia diventerà sempre più un’esperienza comune per le
famiglie italiane. Diversi studi hanno mostrato come l’organizzazione
dell’assistenza agli anziani non autosufficienti da parte dei propri familiari
comporti significativi problemi emozionali (Haley, 2003). Un recente studio

JADT’ 18

119

ha analizzato il testo di 26 interviste a familiari di anziani non autosufficienti
con esperienza di assistenza da parte di un badante (Paniccia, Giovagnoli,
Caputo, 2015). Dall’analisi del testo, condotta tramite la metodologia AET
(Carli, Paniccia, 2002), è emerso come i sistemi di relazione familiari entrino
in crisi contestualmente all’inattività e alla malattia dell’anziano. L'autrice
afferma che la domanda delle famiglie ai servizi sia quella di non essere
emarginate con il loro problemi entro il solo contesto familiare, per altro in
cambiamento. “Sul piano della ricerca - afferma Paniccia - va sviluppata la
differenza, proposta anche dagli intervistati, tra esplorazione dei vissuti degli
anziani assistiti da un lato, degli altri membri della famiglia dall’altro”. In
quest’ottica, la ricerca-intervento proposta risponde a questo invito,
esplorando il vissuto e le attese di un gruppo di anziani non autosufficienti e
loro familiari nei confronti di alcuni servizi territoriali.
2. Il progetto di ricerca-intervento psicosociale
Il progetto di ricerca-intervento è stato realizzato dagli autori per conto
dell'Associazione GAP, un’organizzazione che si occupa di ricerca e
intervento psicosociale nell'ambito della disabilità. Il committente è stato un
ente locale interessato a coinvolgere anziani non autosufficienti e loro
familiari nella costruzione di nuovi modelli di assistenza coerenti con la
domanda delle famiglie stesse. L'ente locale intendeva sviluppare un'offerta
di servizi d'assistenza innovativi a fronte di cambiamenti sociali e culturali
che stanno profondamente modificando l’organizzazione tradizionale della
famiglia. Famiglia in passato maggiormente attrezzata al proprio interno per
provvedere all'assistenza degli anziani. In tale contesto la ricerca intervento
psicosociale è stato proposta come strumento di esplorazione del rapporto tra
servizi d'assistenza rivolti agli anziani presenti nel territorio di competenza
dell'ente committente e famiglie che a tali servizi si rivolgono. In tale contesto
GAP a un gruppo di familiari e anziani non autosufficienti. Tutte le interviste
sono state audio-registrate e trascritte in modo da ottenere il testo su cui è
stata poi applicata l'Analisi Emozionale del Testo. In questa sede presentiamo
i risultati dell'Analisi Emozionale del Testo applicata al testo prodotto
trascrivendo 45 interviste a familiari e anziani non autosufficienti.
2.1. La raccolta dei dati
Le interviste sono state realizzate a 45 familiari e anziani non autosufficienti
in carico ai servizi di medicina generale o ai servizi di centro diurno per
anziani fragili partner del progetto. Di questi circa il 60 % usufruivano di
servizi di medicina generale insieme al servizio di centro diurno per anziani
fragili. Il restante 40% utilizzava esclusivamente i servizi di medicina
generale. Sono state realizzate 25 interviste ad anziani e 20 interviste a loro

120

JADT’ 18

familiari. Le interviste sono state trattate in un unico corpus e per questo in
analisi è stata inserita la variabile illustrativa “ruolo dell’intervistato”,
differenziando le interviste ad anziani da quelle a familiari. L'età media degli
anziani intervistati è di 79 anni, mentre l'età media dei famigliari è di 60 anni.
Gli intervistati sono stati scelti in ordine al criterio di coinvolgere nella ricerca
chi ponesse ai servizi partner problemi complessi che i servizi stessi
sentivano di avere difficoltà a prendere in carico. Questo nell'ipotesi che gli
intervistati potessero poi partecipare ad un intervento psicosociale fondato
sulla restituzione dei risultati della ricerca e sulla loro discussione critica al
fine di contribuire alla progettazione di modelli di assistenza più in linea con
i problemi sperimentati. Agli intervistati è stato proposto di partecipare a
un'intervista aperta, non strutturata, con una sola domanda stimolo seguita
dall'invito a dire tutto quello che veniva in mente. La domanda stimolo è
stata la seguente: “nell'ambito di un progetto di ricerca-intervento siamo
interessati a esplorare il rapporto tra servizi di assistenza, anziani e famiglie
che a tali servizi si rivolgono. In particolare ci interessa esplorare il punto di
vista dei familiari e degli anziani. Aggiungiamo che stiamo intervistano
anche un gruppo di medici di base e di operatori dei servizi socio-sanitari.
Siamo interessati alla sua esperienza; vorremo ascoltarla e raccogliere ciò che
lei ha da dire”. Gli intervistatori si sono presentanti come psicologi
professionisti membri di un'associazione interessata a costruire servizi per
l’invecchiamento e la non auto-sufficienza. Agli intervistati è stato detto che i
risultati della ricerca sarebbero stati condivisi con tutti gli interessati per
capire quali iniziative sviluppare.
3. Metodologia
L'Analisi Emozionale del Testo (Carli, Paniccia, 2002) è uno strumento
proprio della ricerca-intervento psicosociale, sviluppato per esplorare i modi
in cui i gruppi sociali simbolizzano emozionalmente e in modo condiviso un
contesto o un tema e come queste simbolizzazioni organizzino il
comportamento di quel gruppo. Tale metodologia, fondata sul principio del
conoscere per intervenire, prevede l'attivazione di un processo di esplorazione,
analisi e discussione critica della “cultura locale” condivisa entro un
determinato contesto, in relazione al tema posto ad oggetto della ricerca.
L'utilizzo di AET implica la destrutturazione del processo narrativo e delle
connessioni che costituiscono il senso intenzionale dei discorsi entro un testo
posto in analisi. Questo approccio metodologico è fondato
sull'individuazione di gruppi di parole in rapporto tra loro che più di altre
veicolavano significati emozionali: parole definite “parole dense”.
Operativamente abbiamo realizzato il processo statistico e informatico
attraverso il software T-lab (Lancia, 2004) scegliendo la strategia dell’Analisi

JADT’ 18

121

Tematica dei Contesti elementari non supervisionata. Le interviste realizzate
sono state assemblate entro un unico corpus, composto da 14053 tokens e
4121 types mentre gli hapax rilevati sono stati 230. Per quanto riguarda la sua
ricchezza lessicale, il TTR (Type/Token Ratio) è 0.293. Abbiamo raggruppato
le occorrenze di “parole dense” entro lessemi e in questo corpus ne sono stati
individuati e messi in analisi 856. Il numero di “contesti elementari” di testo
classificati 1423 (= 99.58%; del totale di 1429). Il processo di elaborazione dei
dati seguito dal software comporta i seguenti passi: a) costruzione di una
tabella di dati di unità contesto x unità lessicali (fino a 150,000 righe x 3,000
colonne), con valori di presenza/assenza; b) TF-IDF normalizzazione e
scalaggio dei vettori riga alle unità lunghezza (norma Euclidea); c)
clusterizzazione delle unità contesto (misure: coefficiente coseno; metodo:
bisezione K-means); d) - limatura delle partizioni ottenute e, per ciascuna di
esse: e) costruzione di una tabella di contingenza di unita lessicali x clusters;
f) test del chi quadro applicato a tutte le intersezioni della tabella di
contingenza; g) analisi delle corrispondenze della tabella di contingenza di
unità lessicali x clusters. L’analisi statistica ha permesso di individuare
diversi cluster corrispondenti a raggruppamenti di parole co-occorrenti. I
cluster sono quelli che hanno una ricorsività significativa entro il testo e
rappresentano le dimensioni più trasversali che caratterizzano la cultura
locale esplorata.
4. Risultati
Il corpus delle interviste è stato elaborato con il software T-Lab che ha
proposto come ottimale una partizione a 4 Cluster (CL) in rapporto tra loro
su tre fattori (le cui percentuali di inerzia sono Fattore 1= 41,24%, Fattore 2=
32,68%, Fattore 3= 26,08%). Il cluster 3 e il cluster 2 sono in rapporto su
polarità opposte del primo fattore; il cluster 1 e il cluster 4 sono in rapporto
su polarità opposte del secondo fattore, mentre il cluster 1 e il cluster 3 sono
in rapporto sul terzo fattore. Nella tabella (fig.1) è riporta la lista per cluster
delle “parole dense” e le variabili illustrative relative al gruppo delle
interviste degli anziani (_ruol_anz) e al gruppo delle interviste dei familiari
di anziani (_ruol_fam).

122

JADT’ 18
Tabella 1: Lista parole dense per cluster con i relativi valori di chi2

CLUSTER 1
N. of e.c..: 448
χ2
soit: 31.48%
171,81 problema
167,27 casa
79,08 uscire
71,56 lasciare
57,67 vivere
41,82 bisogno
36,62 h24
27,08 abbandonare
26,38 libero
25,59 badante
23,46 pulire
20,33 costringere
19,05 persona
18,71 autonomo
17,74 perdere
16,72 _ruol_fam

CLUSTER 2
N. of e.c.: 371
χ2
soit: 26.07%
155,86 centro
116,53 persona
83,4 aiutare
68,86 trovare
63,55 malattia
57,09 dottore
55,95 psicologia
52,96 supporto
36,48 municipio
31,94 gruppo
26,73 amicizia
24,09 frequentare
22,05 offrire
21,62 cooperativa
21,28 informazione

CLUSTER 3
N. of e.c.: 383
χ2
soit: 26.91%
408,52 figli
90,02 moglie
87,15 fratello
52,81 sposare
46,33 mangiare
40,44 dormire
37,92 morire
36,57 mamma
35,14 telefono
34,77 marito
31,17 maschio
28,96 nonni
26,96 femmina
26,77 cadere
26,68 soldi
27,45 _ruol_anz

CLUSTER 4
N. of e.c.: 221
χ2
soit: 15.83%
122,43 imparare
109,56 cura
97,87 giocare
61,33 parlare
49,95 fumare
47,67 giardino
44,09 dimenticare
42,25 insieme
36,51 somatizzare
35,9 gita
32,1 simpatia
31,21 riflettere
31,21 sigaretta
25,17 ascoltare
25,17 spazio

Di seguito, una lettura dei raggruppamenti di parole dense e della loro
collocazione sul piano fattoriale.
4.1. Cluster 3: obbligo all'assistenza intra-famigliare e prescrizione alla
sopravvivenza
Il cluster è presente in percentuale statisticamente maggiore entro il testo
delle interviste agli anziani (38,4%). Gli intervistati parlano del rapporto con i
propri famigliari: figli, le mogli, i fratelli. L'assistenza viene inscritta entro il
vincolo obbligante dell’essere una famiglia (etimologicamente da famulo, colui
che serve, che si prende cura): emerge l’attesa che il ruolo famigliare implichi il
dovere di occuparsi di chi non riesce a vivere da solo, preoccupandosi di
garantire la sopravvivenza e occupandosi di bisogni inderogabili come
mangiare e dormire. Emerge una rappresentazione infantilizzante dell’anziano
che sollecita l'instaurarsi di rapporti di dipendenza e accudimento. In tale
contesto la quotidianità, deprivata di desideri ed obbiettivi, sembra scorrere
in modo depressivo in attesa di morire, con il rischio di una chiusura
depressiva all'interno della famiglia. L'anzianità sembra identificata con la
figura del vecchio morente che non ha più nulla da dare o da chiedere alla
vita. L'unico riferimento alla vitalità entro il cluster è quello connesso a
parole come nipoti e telefonare: laddove si allenta l'obbligo dell’assistenza
sembra farsi spazio la possibilità di un rapporto piacevole e gratificante.

JADT’ 18

123

4.2. Cluster 2: ricerca di servizi e domanda alla psicologia
In questo cluster è rappresentato il processo di ricerca di servizi di assistenza.
Si cercano centri, contesti estranei alla famiglia, che aiutino ad occuparsi dei
problemi della persona non autosufficiente. Da un lato si guarda alla sua
soggettività, dall'altro si rappresenta una ricerca affannosa di servizi fondata
sull'angoscia di trovare soluzioni. La non autosufficienza è rappresentata
come malattia. Ciò comporta un vissuto di urgenza e pericolo e la fantasia di
dover contrastare qualcosa che mette a rischio la sopravvivenza. Su questo si
chiama in causa il dottore, in ipotesi il medico di base, cui viene attribuita una
competenza utile. Allo stesso tempo è chiamata in causa la psicologia cui viene
richiesto un intervento di supporto. Si evoca in tal modo una prospettiva di
intervento alternativa alla cura. Si chiede di essere aiutati a prepararsi e di
essere accompagnati, di parlare con qualcuno poiché ci si sente impreparati,
confusi.. A questo proposito i famigliari sembrano portatori di una domanda
di ascolto e consulenza fondata sul parlare. Agli enti locali e del privato
sociale gli intervistati si propongono come clienti, viene domandata
l'articolazione di un'offerta di servizi, valorizzando dispositivi d'intervento di
gruppo.
4.3. Cluster 1: funzione di controllo delegata alla badante e paura del
cambiamento
Il cluster è presente in percentuale statisticamente maggiore entro il testo
delle interviste ai familiari (39%). Gli intervistati parlano del problema che
vivono, situato nella casa, un contesto chiuso che offre riparo e che al
contempo costringe. Da un lato si cercano vie di uscita e d'altro lato c'è
difficoltà a lasciare, ad allontanarsi da rapporti protettivi e vincolanti. Viene
rappresentato un contrasto tra queste emozioni e il vivere: emerge un
sentimento di vita contrastata, per dirla con Canguilhem (1998). In tale
contesto si è presi dalla fantasia di abbandonare: emerge l'emozionalità della
colpa. Ciò avviene entro un contesto in cui la non autosufficienza viene
trattata quale bisogno esclusivamente fattuale e pressante, 24 ore su 24.
L'invecchiamento è rappresentato come evento che non lascia tregua, che
tormenta e angoscia. In tale contesto si chiede l’intervento della badante per
ripristinare il controllo, fare ordine. La badante è rappresentata come una
necessità motivata dal bisogno. L'assistenza all’anziano è qualcosa a cui ci si
sente costretti o da cui liberarsi, tertium non datur. Ma in questo cluster
vediamo come vivendo l'invecchiamento come bisogno continuo e
prescrivendo l'assistenza si generi colpa. Colpa connessa all’impotenza per il
non riuscire a rapportarsi ai cambiamenti con cui la non autosufficienza
confronta.

124

JADT’ 18

4.4. Cluster 4: domanda di costruzione di contesti dove parlare, giocare,
apprendere.
In questo cluster gli intervistati esprimono una domanda di contesti e
rapporti fondati sull'apprendimento, il gioco e sulla parola. Emergono desideri e
si riconoscono risorse che evocano la possibilità di trovare motivi per cui
valga la pena vivere. Emerge una rappresentazione della vecchiaia
caratterizzata da vitalità e desiderio di trasgredire. Si allenta la prescrittività
dell'obbligo della sopravvivenza: la vecchiaia è anche creatività, possibilità di
smarcarsi dagli obblighi rituali della vita sociale. Il riconoscimento del limite
del tempo, l'avvicinarsi della fine, motiva la ricerca di esperienze piacevoli
che diano senso alla vita. Si evoca il divertimento come obbiettivo alternativo
al controllo e alla sorveglianza senza obbiettivi. Sottolineiamo come la
domanda divertimento implichi il riconoscimento di una verità non scontata:
che si è ancora vivi fino a cinque minuti prima di morire.
5. Conclusioni
Per concludere proponiamo alcune considerazioni sul rapporto tra i cluster
sui tre assi fattoriali. Ricordiamo che il cluster 3 e il cluster 2 sono in rapporto
su polarità opposte del primo fattore, il cluster 1 e il cluster 4 sono in
rapporto su polarità opposte del secondo fattore, mentre il cluster 1 e il
cluster 3 sono in rapporto sul terzo fattore. Sul primo fattore emerge come la
dimensione motivazionale che sostiene la domanda di servizi da parte della
famiglia sia il desiderio di uscire dall’obbligo familiare. È il vissuto di obbligo
e l’incapacità di condividere entro i rapporti desiderio ed interessi che spinge
la famiglia in un'affannosa ricerca di interlocutori e professionisti esterni. Sul
secondo fattore emergono diverse modalità di rapportarsi al problema della
non autosufficienza. Su di un polo del fattore (cluster 1) la fattualizzazione
dell'invecchiamento come bisogno continuo di assistenza che mette in
pericolo la sopravvivenza mostra come i problemi associabili alla non
autosufficienza non siano esplorati. Tali problemi sembrano piuttosto
presunti dal familiare in modo autoreferenziale. L'emozionalità della colpa e
la fantasia irrealizzabile di ristabilire il controllo su una situazione in
cambiamento vissuta come persecutoria sono corollari di tale autorefenzialità
sottesa dall'incompetenza a utilizzare i rapporti familiari come contesto di
confronto e scambio sui problemi e sul da farsi. D'altro lato, sull'altro polo
del secondo fattore il riconoscimento di limiti, quali ad esempio il tempo
limitato della vita e l'ineluttabilità della fine, sembra fare spazio al
riconoscimento del desiderio degli anziani di divertirsi anche concedendosi
qualche trasgressione, come alternativa a sopravvivere in modo controllante
e mortifero. Infine il terzo fattore suggerisce una relazione tra la dinamica di
autorefenzialità dei rapporti familiari e la domanda di servizi emergente

JADT’ 18

125

entro la cultura in analisi, a cui si chiede non soltanto di curare ma anche di
aiutare la famiglia a sviluppare competenze e confrontarsi sui propri
problemi. I risultati della ricerca suggeriscono una domanda nei confronti di
servizi di accompagnamento e che sostengano la famiglia – intesa come
contesto di rapporti tra la persona non autosufficiente e i suoi familiari - nel
riconoscimento di desideri e obbiettivi attorno a cui organizzare l'assistenza e
la convivenza nel modo più piacevole, vitale e divertente possibile.
Bibliografia
Carli R., Paniccia R.M. (2002). L’analisi emozionale del testo. Franco Angeli,
Roma.
Haley, W. E. (2003). Family caregivers of elderly patients with cancer:
understanding and minimizing the burden of care. The journal of
supportive oncology, 1(4 Suppl 2), 25-9.
ISTAT (2017), Demografia in cifre, Roma, Istituto Nazionale di Statistica –
www.demo.istat.it.
Lancia, F. (2004). Strumenti per l’analisi dei testi. Franco Angeli, Roma.
Paniccia, R. M., Giovagnoli, F., & Caputo, A. (2015). In-home elder care. The
case of Italy: the badante. Rivista di Psicologia Clinica, (2), 60-83.

126

JADT’ 18

Esperienza di analisi testuale di documentazione
clinica e di flussi informativi sanitari, di utilità nella
ricerca epidemiologica e per indagare la qualità
dell'assistenza.
Antonella Bitetto1, Luigi Bollani2
1

Azienda Socio Sanitaria Territoriale di Monza – a.bitetto@asst-monza.it
2Università di Torino – luigi.bollani@unito.it

Abstract
This study finds reason in the now wide availability of clinical
documentation stored in electronic form to track the patient's health status
during his care path or for sending information to other institutions on the
activities carried out for administrative purposes. The diffusion of these
methods now makes available many biomedical collections of electronic data,
easily accessible at low cost that can be used for research purposes in the
field of observational epidemiological studies, in analogy with what was
historically already practiced in studies based on the reviewing of medical
records. However, since these collections are not organized according to
specific survey schemes, they sometimes do not allow the index events to be
discriminated with the necessary reliability between one source and another.
It has always been believed that the critical re-reading of texts can partially
help these informative shortcomings with the aim of bringing back according to possibility - the words or segments contained in the texts, to
statistically analyzable categories. The recent transfer of these collections
from paper to electronic forms opens the possibility of carrying out this
process automatically, reducing time and costs of the process and perhaps
increasing its reliability. It is proposed to address the problem, showing
study criteria and an example of analysis based on an empirical experience,
consistent with the needs of a biomedical context.
Keywords: textual analysis; electronic health data; medical thesaurus;
analysis of lexical correspondences; emergency in psychiatry
Riassunto
Questo studio trova ragione nella ormai ampia disponibilità di
documentazione clinica archiviata in forma elettronica per tracciare lo stato
di salute del paziente durante il suo percorso di cura o inviare informazioni
ad altri enti sulle attività svolte a scopo amministrativo. La vasta diffusione
di questi metodi mette a disposizione ormai numerose raccolte di tipo

JADT’ 18

127

biomedico, facilmente accessibili a basso costo che possono essere utilizzate a
scopo di ricerca nel settore degli studi epidemiologici osservazionali, in
analogia con quanto storicamente veniva già praticato negli studi basati sulla
rilettura delle cartelle cliniche. Non essendo però tali raccolte organizzate
secondo schemi di rilevazione specifici a volte non permettono di
discriminare con la necessaria attendibilità tra una fonte e l’altra gli eventi
indice. Da sempre si ritiene che la rilettura critica dei testi possa,
parzialmente soccorrere a tali carenze informative nell’obiettivo di
ricondurre - secondo possibilità - le parole o i segmenti contenuti nei testi
disponibili a categorie statisticamente analizzabili. Il recente passaggio di tali
raccolte dalla forma cartacea a quella elettronica apre la possibilità di operare
per via automatica riducendo tempi e costi del processo e forse
incrementandone l'attendibilità. Ci si propone di affrontare il problema,
mostrando criteri di studio ed un esempio di analisi basato su un’esperienza
empirica, conforme alle esigenze di un contesto biomedico.
Parole chiave: analisi testuale; dati sanitari elettronici; thesaurus medico;
analisi delle corrispondenze lessicali; psichiatria d’urgenza
1. Introduzione
Il progressivo processo di dematerializzazione della documentazione clinica
(valutazioni specialistiche ambulatoriali, verbali di Pronto Soccorso, referti
esami diagnostici) e l’implementazione dei flussi di dati sanitari a scopo
giuridico amministrativo (per il pagamento delle prestazioni erogate o per
l’aggiornamento dell’anagrafe, dell’INPS etc.) hanno reso disponibili
informazioni che possono essere utilizzate anche per obiettivi diversi da
quelli per cui i dati sono raccolti. I dati sanitari informatizzati (EHR “electronic health records”), vengono generalmente distinti in: a) strutturati
(ad es. registrati utilizzando terminologie cliniche controllate come la
Classificazione internazionale delle malattie -10ª revisione (ICD10) o la
nomenclatura sistematica della medicina - Termini clinici (SNOMED-CT), b)
semistrutturati (ad es. esami di laboratorio ed informazioni sulla
prescrizione) che seguono uno schema che varia a seconda delle convenzioni
adottate localmente, c) non strutturati (ad es. testo clinico) e d) binari (ad
esempio file di immagini come Rx e TAC). La sistematicità di queste raccolte
di dati, organizzati in maggioranza per entità individuali, li rende
particolarmente preziosi per diversi scopi di ricerca epidemiologica che
utilizza disegni di tipo osservazionale sia nell’ambito della qualità
dell’assistenza che dell’epidemiologia più classica, che studia rischi ed esiti
delle malattie (Mitchell J. et al., 1994). Per contro essendo tali raccolte
organizzate per scopi altri da quelli del monitoraggio della qualità o della
ricerca scientifica, spesso devono essere “trattate” prima di poter essere

128

JADT’ 18

analizzate con metodi statistici. In passato ciò veniva fatto attraverso la
rilettura delle cartelle cliniche da parte di esperti della materia. Attualmente
si cerca sempre più di ricorrere a metodi di analisi automatica dei testi che
garantisce una miglior standardizzazione e revisione (Denaxas S. et al., 2017).
A titolo di esempio si segnala che l’analisi automatica dei testi di flussi
informativi e della documentazione clinica elettronica ha permesso
d’indagare ambiti terapeutici e di sicurezza fondamentali come la qualità
dell’assistenza infermieristica e l’occorrenza di eventi avversi come – tra i
tanti - gli incidenti domestici, le reazioni allergiche e gli effetti collaterali ai
farmaci (Ehrenberg A. et Ehnfors M., 1999; Coloma P.M. et al., 2011;
Migliardi A. et al., 2004). Sono stati anche prodotti numerosi studi
epidemiologici classici per lo più riferiti a patologie croniche ad alta
prevalenza come le malattie cardiocircolatorie, il diabete o l’asma, all’estero e
in Italia (Gini R. et al., 2016; Vaona A. et al. , 2017), in alcuni casi mettendo in
evidenza bisogni di cura inespressi o complicanze dovute a ritardi o
trattamenti inappropriati (Persell S.D., et al., 2009; Ho M.L., et al., 2012).
Alcune ricerche si sono focalizzate sui disturbi mentali, area medica scelta
per l’esperienza di analisi di testo di seguito presentata. In questo ambito la
documentazione clinica elettronica permette di ottenere informazioni a basso
costo su ampi settori di popolazione che possono ricomprendere casistiche
difficili altrimenti da reclutare: questo è il caso di soggetti in fase prodromica
ad alto rischio di sviluppare psicosi (Fusar-Poli P. et al., 2017) o autolesionisti
(Zanus C. et al., 2017).
2. Metodi
La classificazione dei corpora non ancora studiati in categorie statisticamente
analizzabili rappresenta un argomento controverso ma anche una sfida che
giustifica, a nostro avviso, indagini di approfondimento delle procedure
metodologiche da adottare. Nel seguito si propone un metodo per il
trattamento di testi medici non strutturati di psichiatria, secondo criteri già in
parte utilizzati in precedenti esperienze (Bitetto A. et al., 2017).
2.1. Corpus
Le informazioni provengono dai verbali di consulenze psichiatriche svolte
presso il Pronto Soccorso di un ospedale universitario lombardo di grandi
dimensioni (1250 letti accreditati).
Il corpus è monolingua - in italiano - composto da brevi testi scritti dallo
psichiatra di turno alla fine della consulenza in urgenza. I referti sono
verificati e quindi conservati dal servizio informativo ospedaliero, certificato
ISO 9001/2015, che ha fornito il corpus dei dati, in forma anonima. Si sono
analizzati 1721 referti, relativi al periodo 01/01/2012 – 31/12/2012.

JADT’ 18

129

2.2. Pretrattamento di filtraggio linguistico
Il corpus è stato sottoposto ad un pretrattamento di filtraggio linguistico.
Dalle 177349 parole presenti nei referti originali sono state eliminate la
punteggiatura, i numeri, i pronomi, gli articoli, le proposizioni, i nomi propri
- anche dei farmaci- e le parole con una ricorrenza inferiore a 10. Ne è
risultato un elenco di 1679 parole distinte, che è stato rivisto manualmente da
un esperto per selezionare i termini in grado di descrivere i problemi/bisogni
di salute mentale secondo il modello strutturale utilizzato dalla scala HoNOS
(Wing J.K. et al., 1998; Lora A. et al., 2001). Si tratta di un modello di
valutazione dello stato di salute mentale impostato per problemi e non sulle
diagnosi, che difficilmente sono riportate nei referti di pronto soccorso. Il
modello distingue 12 “problemi” riconducibili ai seguenti concetti:
item H1- COMPORTAMENTI IPERATTIVI, AGGRESSIVI; item H2 COMPORTAMENTI DELIBERATAMENTE AUTOLESIVI; item H3 PROBLEMI LEGATI ALL’ASSUNZIONE DI ALCOOL O DROGHE; item H4 PROBLEMI COGNITIVI; item H5 - PROBLEMI DI MALATTIA SOMATICA;
item H6 - PROBLEMI LEGATI AD ALLUCINAZIONI E DELIRI; item H7 PROBLEMI LEGATI ALL’UMORE DEPRESSO; item H 8 - ALTRI PROBLEMI
DA ALTRI SINTOMI PSICHICI; item H9 - PROBLEMI NELLE RELAZIONI
SIGNIFICATIVE; item H10 - PROBLEMI NELLO SVOLGIMENTO DI
ATTIVITÀ DELLA VITA QUOTIDIANA; item 11- PROBLEMI NELLE
CONDIZIONI DI VITA; item H12 - PROBLEMI NELLE ATTIVITÀ
LAVORATIVE E RICREATIVE.
In questo modo è stato creato un thesaurus composto da 214 locuzioni brevi e
81 parole singole riconducibili a 11 categorie cliniche (esclusa la H10, data la
mancanza di locuzioni in grado di ricondurre ad essa). Nel thesaurus si sono
inoltre considerate parole e acronimi che individuano accesi legati al “rifiuto
delle cure”. La procedura di filtraggio dei testi, basata sul thesaurus (ponendo
anche attenzione a non includere contesti dove la parola chiave è negata), ha
permesso di riclassificare 1629 referti che rappresentano la base dell’analisi.
2.3. Analisi statistica
I diversi referti sono stati esaminati per la presenza/assenza di ciascuna
parola o locuzione chiave esaminata, in modo da introdurre per ogni parola
una codifica binaria rispetto al complesso dei testi considerati.
Successivamente tale codifica è stata estesa agli item della classificazione
HoNOS valutando – in ogni referto – la presenza di ciascun item,
determinata dalla presenza di almeno una parola chiave ad esso associata
(l’assenza dell’item si determina per contro in mancanza di parole chiave ad
esso associate). Per rappresentare l’associazione tra i diversi item, rispetto ai
referti studiati, si è quindi condotta un’analisi delle corrispondenze

130

JADT’ 18

(Benzécri, Jean-Paul, 1973) sulla tabella testi x item HoNOS (in aggiunta ad
essi si è anche incluso concetto di rifiuto/interruzione delle cure); per poter
apprezzare inoltre le relazioni tra parole e comportamenti/problemi, espressi
dalla classificazione introdotta, si sono aggiunte le parole e locuzioni chiave
in forma supplementare.
3. Risultati
La tabella 1 mostra la distribuzione di frequenza delle aree problematiche
descritte e riclassificate secondo i criteri della scala HoNOS.
Tabella 1 – Item HoNOS e percentuale di presenza del comportamento/problema
riscontrato nei referti
Item HoNOS

H1

H2

H3

H4

H5

H6

H7

H8

H9

H11 H12 RifiutoCure

% di presenza
30.82 15.22 12.22 7.18 20.32 18.35 32.72 59.55 5.10 1.23 7.31
nei referti

18.97

Come atteso i referti riferiscono soprattutto le manifestazioni cliniche del
disagio attraverso descrizioni dettagliate dei sintomi osservati rispetto ad
altri fattori di tipo ambientale (H9, H11, H12). Tra i sintomi, quelli di più
frequente riscontro sono l’umore depresso (H7) e la classe che raccoglie tutte
le manifestazioni cliniche non specificate “altri sintomi psichici” (H8). Molto
frequente è anche la descrizione di problemi di natura organica (sintomi fisici
H5) come atteso, visto che la gestione delle urgenze psichiatriche avviene
presso il pronto soccorso generale in cui la richiesta di parere su accesi legati
a problematiche fisiche è più alta che presso un ambulatorio di secondo
livello. Molto elevata è anche l’occorrenza di comportamenti violenti ed
iperattivi (H1), una delle urgenze più tipiche dell’ambito psichiatrico.

Figura 1 – A sinistra : rappresentazione congiunta dei primi 8 item HoNOS (sintomi psichici
e fisici); A destra : sintomi comportamentali (H1, H2, H3), sintomi psichici (H6, H7, H8) e
fattori ambientali precipitanti (H9, H11, H12)

Nella figura 1 – grafico di sinistra - sono rappresentati i risultati dell’analisi
delle corrispondenze sulle categorie dei sintomi, l’area problematica di

JADT’ 18

131

maggior riscontro nei testi. Il primo piano fattoriale – mostrato nel grafico –
spiega il 34.17 % della varianza totale. Rispetto alla dimensione 1, lungo
l’asse delle ascisse, le categorie di sintomi si suddividono in due gruppi: sulla
destra troviamo i problemi legati all’umore depresso (H7) vicino ad altri
sintomi (H8), di cui come già detto l’ansietà rappresenta l’area più vasta, e i
sintomi fisici (H5), confermando la probabile origine psicosomatica di parte
di essi. Nel medesimo raggruppamento si collocano i comportamenti
deliberatamente autolesivi e suicidari, che sono secondo la letteratura spesso
associati a problemi di depressione. Su valori elevati di ascissa, sono invece
raggruppati i sintomi psicotici (H6), i comportamenti agitati (H1), in
relazione con il rifiuto delle cure, cui spesso infatti si associano. Risultano
invece indipendenti dalle altre categorie di sintomi i problemi legati all’abuso
di alcool e droghe (H3) e quelli dovuti alla presenza di problemi cognitivi di
origine neurologica (H4), che occupano gli estremi della dimensione 2,
individuate dall’asse delle ordinate. La stessa analisi è rappresentata nella
figura 2, proiettando anche le parole pertinenti del thesaurus utilizzato.

Figura 2 – Rappresentazione congiunta degli item HoNOS relativi ai primi 8 item e
rappresentazione supplementare delle parole/locuzioni chiave utilizzati per individuare i
diversi item

132

JADT’ 18

Riprendendo la figura 1 – grafico a destra – si trova una seconda analisi delle
corrispondenze condotta sulle categorie di sintomi psichici e
comportamentali insieme ai fattori precipitanti di tipo ambientale. In questo
caso il primo piano fattoriale spiega il 30.33% della varianza totale.
La distribuzione dei sintomi psichici lungo l’asse delle ascisse conferma,
come atteso, i risultati dell’analisi del primo subset di categorie. In questo
caso è possibile notare la tendenza dei problemi legati all’abuso di alcool e
droghe (H3) a disporsi verso il centro del grafico in prossimità della categoria
altri sintomi (H8), con cui è possibile che certe manifestazioni siano in
relazione. Per quanto riguarda i fattori ambientali emerge dai dati una
relazione tra problemi di lavoro (H12), sintomi dello spettro depressivo (H7)
e condotte deliberatamente autolesive (H2). È possibile che il Pronto Soccorso
rappresenti un primo punto di accesso per un’utenza con forme reattive
anche gravi, secondarie a fattori di stress occupazionale (burnout,
depressioni reattive). Le altre categorie relative a problematiche ambientali
(H9 e H11) si collocano agli estremi della dimensione 2, mostrando un certo
grado di indipendenza rispetto all’occorrenza di sintomi comportamentali e
psichici.
5. Conclusioni
L’esperienza empirica di analisi testuale automatica di referti del Pronto
Soccorso conferma la sua utilità nell’indagare fenomeni complessi come le
manifestazioni cliniche e i fattori di rischio dell’urgenza psichiatrica. L’analisi
delle corrispondenze si dimostra un metodo semplice e utile per esplorare le
relazioni tra le diverse dimensioni in esame.
Emergono per altro alcuni problemi legati alla qualità delle informazioni che,
in quanto raccolte per altri scopi, presentano un eccesso di informazione
rispetto ad alcune aree (manifestazioni sintomatologiche) mentre sono
carenti in altre, come il grado di disabilità del soggetto non analizzabile come
fattore precipitante dell’urgenza. È possibile che tali carenze possano essere
superate acquisendo informazioni da altre fonti come alcuni ricercatori
hanno fatto (Fusar-Poli P. et al., 2017). Resterebbe comunque aperto il
problema di condividere e standardizzare i metodi di trattamento dei dati
nelle diverse fasi dell’indagine, dalle modalità con cui sono raccolte le
informazioni e compilati i referti, alla creazione di un thesaurus di parole e
locuzioni chiave standard per la psichiatria sulla base di concetti teorici e
criteri condivisi.
Bibliografia
Benzécri, J.P. (1973). L'analyse des données. Vol. 2. Paris: Dunod.
Bitetto A., et al. (2017). La consultazione psichiatrica in Pronto Soccorso come

JADT’ 18

133

fonte informativa sui bisogni inespressi di salute mentale. Nuova rassegna
studi psichiatrici vol. 15 novembre 2017
Coloma P.M. et al. (2011). Combining electronic healthcare databases in
Europe to allow for large-scale drug safety monitoring: the EU-ADR
Project. Pharmacoepidemiol Drug Saf.; 20(1):1–11. 40.
Denaxas S., et al. (2017).Methods for enhancing the reproducibility of
biomedical research findings using electronic health records. Bio Data
Mining;10:31
Ehrenberg A. et Ehnfors M. (1999). Patient problems, needs, and nursing
diagnoses in Swedish nursing home records. Nursing Diagnosis; 10(2), 6576.
Fusar-Poli P, et al. (2017). Diagnostic and Prognostic Significance of Brief
Limited Intermittent Psychotic Symptoms (BLIPS) in Individuals at Ultra
High Risk. Schizophr Bull; 43(1):48-56
Gini R. et al. (2016). Automatic identification of type 2 diabetes, hypertension,
ischaemic heart disease, heart failure and their levels of severity from
Italian General Practitioners' electronic medical records: a validation
study. BMJ Open; 6(12): e012413.
Ho ML, et al. (2012). The accuracy of using integrated electronic health care
data to identify patients with undiagnosed diabetes mellitus. J Eval Clin
Pract. ;18(3):606–11.
Lora A. et al. (2001). The Italian version of HoNOS (Health of the Nation
Outcome Scales), a scale for evaluating the outcomes and the severity in
mental health services. Epidemiology and Psychiatric Sciences; 10.3: 198-204.
Migliardi A. et al. (2004). Descrizione degli incidenti domestici in Piemonte a
partire dalle fonti informative correnti. Epidemiologia & Prevenzione ; 28.1:
20-26.
Mitchell J. et al., (1994). Using medicare claims for outcome research. Medical
care; 35:589-602
Persell S.D. et al. (2009). Electronic health record-based cardiac risk
assessment and identification of unmet preventive needs. Med Care;
47(4):418–24.
Vaona A. et al. (2017). Data collection of patients with diabetes in family
medicine: a study in north-eastern Italy. BMC Health Serv Res.;17(1):565
Wing J.K. et al., (1998). Health of the Nation Outcome Scales (HoNOS).
Research and development. The British Journal of Psychiatry; 172 (1) 11-18
Zanus C. et al. (2017). Adolescent Admissions to Emergency Departments for
Self-Injurious Thoughts and Behaviors. PLoS One.;12(1): e0170979.

134

JADT’ 18

Exploring the history of American philosophy in a
computer-assisted framework
Guido Bonino1, Davide Pulizzotto2, Paolo Tripodi3
2

1Università di Torino – guido.bonino@unito.it
LANCI, Université du Québec à Montréal – davide.pulizzotto@gmail.com
3Università di Torino – paolo.tripodi@unito.it

Abstract
The aim of this paper is to check to what extent some tools for computerassisted concept analysis can be applied to philosophical texts endowed with
complex and sophisticated contents, so as to yield results that are significant
not only because of the technical success of the procedures leading to the
results themselves, but also because the results, though highly conjectural,
are a direct contribution to the history of philosophy
Sommario
Lo scopo di questo articolo è di verificare in che misura la computer-assisted
concept analysis possa essere applicata a testi filosofici di contenuto
complesso e sofisticato, in modo da produrre risultati significativi non solo
dal punto di vista del successo tecnico delle procedure, ma anche in quanto i
risultati stessi, sebbene altamente congetturali, costituiscono un contributo
diretto alla storia della filosofia.
Keywords: philosophy, history of philosophy, paradigm, necessity, idealism,
Digital Humanities, Text Analysis, Computer-assisted framework
1. Computer-assisted concept analysis
The development of artificial intelligence poses a methodological challenge
to the humanities. Many traditional practices in disciplines such as
philosophy are increasingly integrating computer support. In particular,
Concept Analysis (CA) has always been a common practice for philosophers
and other scholars in the humanities. Thanks to the development of Text
Mining (TM) and Natural Language Processing (NLP), computer-assisted text
reading and analysis can provide the humanities with new tools for CA
(Meunier and Forest, 2005), making it possible to analyze large textual
corpora, which were previously virtually unassailable. Examples of
computer-assisted analyses of large corpora in philosophy are Allard et al.,
1963; McKinnon, 1973; Estève et al., 2008; Danis, 2012; Sainte-Marie et al.,
2010; Le et al., 2016; Meunier and Forest, 2009; Ding, 2013; Chartrand et al.,
2016; Pulizzotto et al., 2016; Slingerland et al., 2017. The use of computer-

JADT’ 18

135

assisted text analysis is also relevant for the distant reading approach,
developed by Franco Moretti in the context of literature studies (Moretti,
2005; Moretti, 2013), but which we are convinced can be usefully extended to
different fields (for the application to philosophy see the Conference “Distant
Reading and Data-Driven Research in the History of Philosophy” held in
Turin in 2017, http://www.filosofia.unito.it/dr2/).
The main aim of this paper is to check to what extent some tools for
computer-assisted CA can be applied to texts endowed with complex and
sophisticated contents, so as to yield results that are significant not only
because of the technical success of the procedures leading to the results
themselves, but also because the results, though highly conjectural, are a
direct contribution to the humanities. Philosophy, in particular the history of
philosophy, seems to be a good case to be considered, because of the
sophistication of its contents. Our main purpose is that of illustrating some of
the different kinds of work that can be done in history of philosophy with the
aid of computer-assisted CA.
2. Method
2.1. The corpus
To understand how TM and NLP can assist the work in history of
philosophy, some standard methods have been applied to a specific corpus,
which is provided by Proquest (www.proquest.com). The corpus is a
collection of 20,751 PhD dissertations in philosophy discussed in the US from
1981 to 2015. It therefore contains 20,751 documents: each document is a text,
comprising the title and the abstract of a dissertation, which are dealt with as
a single unit of analysis. The corpus also contains some metadata, such as the
author of the dissertation, the year of publication, the name of the supervisor,
the university, the department, and so forth. In the present paper we are not
going to exploit fully the wealth of information provided by these metadata,
which are certainly worth being the subject of further research. However, we
will use the crucial datum of the year of publication, which allows us to
assume a diachronic (that is, historical) perspective on the investigated
documents.
2.2. Data preprocessing
A preliminary step consists in a set of four preprocessing operations that
allow us to extract the linguistic information needed for the analysis: 1) Part
of Speech (POS) tagging; 2) lemmatization; 3) vectorization; 4) selection of the
sub-corpora responding to Keyword In Context (KWIC) criteria.
The POS tagging and the lemmatization process are performed on the basis
of the TreeTagger algorithm described by Schmid, 1994 and 1995. This

136

JADT’ 18

operation consists in the annotation of each word for each document
according to its morphological category. Some irrelevant categories (such as
determinants, prepositions and pronouns) are eliminated. Nouns, verbs,
modals, adjectives, adverbs, proper nouns and foreign words are taken into
account. The lemmatization process reduces a word to his lemma, according
to the correspondent POS tag. At the end of this process, we can identify
17,750 different lemmas, which are called types.
The mathematical modeling of each document into a vector space is called
vectorization. In such a model, each document is encoded by a vector, whose
coordinates correspond to the TF-IDF weighting of the words occurring in
that document. This weighting function calculates the normalized
frequencies of the words in each document (Salton, 1971). At the end of the
process, a matrix M is built, which contains 20,571 rows corresponding to
each document, and 17,750 dimensions, corresponding to the types.
Finally, three sub-corpora are created on the basis of the KWIC criterion.
These sub-corpora correspond to the set of all the text segments in which one
of these three lexical form, each of which convey the meaning of a concept,
appears: ‘necessity’, ‘idealism’, and ‘paradigm’. The three concepts have been
chosen because of the considerable diversity of their statuses: ‘necessity’ has
always been a keyword of several sub-fields of philosophy; ‘idealism’ refers
both to a philosophical current, historically determined, and to an abstract
position in philosophy; ‘paradigm’ entered the philosophical vocabulary in
relatively recent times, mainly after the publication of Kuhn, 1962, as a
technical term in the philosophy of science. We obtain a set of 719 documents
for ‘necessity’, 450 documents for ‘idealism’, 975 documents for ‘paradigm’.
2.3. Word-sense disambiguation process
For each sub-corpus, we identify the semantic patterns (usually, word cooccurrence patterns) associated to each lexical form, so as to discover the
most relevant semantic structures of that concept. This is done by using
clustering, a common method in Machine Learning for pattern recognition
tasks (Aggarwal and Zhai, 2012). Clustering techniques applied to texts are
based on two hypotheses: a contiguity hypothesis and a cluster hypothesis. The
former states that texts belonging to the same cluster form a contiguous
region that is quite clearly distinct from other regions, while the latter says
that texts belonging to the same cluster have similar semantic content
(Manning et al., 2009, p. 289 and 350). For our purposes, clustering is an
instrument for semantic disambiguation. In our experiment, we use the Kmeans algorithm (Jain, 2010, p. 50), a widely employed algorithm for WordSense Disambiguation tasks (Pal and Saha, 2015).
The main parameter that needs to be tuned in the K-means algorithm is the k

JADT’ 18

137

parameter, which determines the number of centroids to be initialized. Each
execution of the K-means algorithm generates a partition Pk having a number
of clusters equal to k. Since each centroid is the “center vector” of each
cluster, it can also be used to identify the most “prototypical” documents in a
given cluster. To complete this operation, a tool generally used to select
relevant documents in Information Retrieval is employed, that is, the cosine
computation among a query vector and a group of “document vectors”
(Manning et al., 2009). In this context, each centroid of a Pk partition can be
used as a query in order to identify documents with a higher cosine value.
Clustering has first been applied synchronously on the Si matrices with k = {2,
3, 4, …, 50}, thus obtaining the most recurring semantic patterns; then it has
been applied diachronically, dividing each matrix into three different periods
(1981-1993, 1994-2003, 2004-2015) in order to obtain sets of documents with
similar cardinality. On each sub-matrix of Si several clusterings with k = {2, 3,
4, ..., 50} were performed, in order to identify the temporal evolution of the
most important semantic patterns associated to the three concepts under
study. For each generated Pk partition, we also perform the cosine
computation in order to obtain a set of the most relevant PhD dissertations
belonging to each cluster.
3. Analyses
In this section, we are going to present three analyses, focusing on three
different concepts: paradigm, necessity and idealism. Each case illustrates a
different kind of historical-philosophical result.
3.1. Necessity
After exploring both synchronically and diachronically several clusters (with
different k) associated to the concept of necessity, we have focused on a
clustering with k=18 in the period 1981-2015 (the clusters are not significantly
different from one another in the three decades). It turns out that there are at
least 16 clearly distinct and philosophically interesting meanings of
‘necessity’: two (maybe distinct) theological notions; physical necessity;
political necessity; necessity as investigated in modal logic and possible
world semantics; moral necessity; necessity as opposed to freedom in debates
over determinism; the necessity of historical processes; metaphysical
necessity; two notions of causal necessity (attacked by Hume); the necessity
of life events; logical necessity; phenomenological necessity; necessity of the
Absolute (Hegel); necessity of moral duty (Kant); ancient concept of
necessity; the necessity of law. In addition to these, there is also a rather big
cluster in which ‘necessity’ seems to occur mainly with its ordinary, not
strictly philosophical meaning.

138

JADT’ 18

If the clustering we applied to ‘necessity’ were extended to a large number of
philosophical words (chosen in our corpus by domain experts), that would
be the first step for the construction of a bottom-up vocabulary of
philosophy, and ultimately of a data-driven philosophical dictionary, in
which the different (though related) meanings of philosophical terms would
be determined on the basis of actual use, rather than merely on the
lexicographer’s discernment. This lexicographic work is also an
indispensable step if one wants to overcome the “concordance approach”: it
seems to us that this bottom-up lexicography could be a promising starting
point for the construction of semantic networks.
3.2. Idealism
Unlike ‘necessity’, the term ‘idealism’ has different distributions in the
decades 1981-1993, 1994-2003 and 2004-2015. We have only considered the
largest clusters (> 10 documents), since for our purpose (that of
reconstructing the main historical developments of American academic
philosophy), isolated cases and minor tendencies are not relevant.
The evolution of some clusters over decades suggests interesting historical
reflections. First, the cluster “Kant” is persistently important. In fact, it
becomes more and more important, even in wider contexts, that is, in
documents that are not directly devoted to Kant. This is shown by the rising
trend of the cluster “Transcendental” (a term typically, but not always
directly connected with Kant). Second, the cluster “Hegel” disappears in the
second decade, then it reappears: is this a real phenomenon, rather than a
statistical artefact? How can it be explained? Third, the cluster “Realism”
disappears in the third decade: is there a relationship between the return of
“Hegel” and the disappearance of “Realism”? This is not the kind of
question, which comes naturally to the mind of the historian of philosophy,
on the basis on his/her knowledge of well-known developments of the
history of recent American philosophy. This hypothesis can be formulated
only thanks to some sort of defamiliarization (ostranenie) with respect to the
received views in history of philosophy. Yet, it seems unlikely that
philosophers in the last decade gave up speaking of realism. The received
view may after all be correct, that realism is more and more central in late
analytic philosophy (think, for example, of the centrality of David Lewis)
(Bonino and Tripodi forthcoming). Such a view is confirmed by other data,
such as the number of occurrences of ‘realis-’ in the abstracts of the corpus.
1981-93: 373 (5,76% of 6,471); 1994-2003: 465 (6,31% of 7,361); 2004-2015: 482
(5,6% of 8,585). Thus the focus on realism is still there, in the third decade.
One is therefore led to formulate an alternative hypothesis: philosophers
ceased to speak of idealism in relation to realism: perhaps the contrast realism-

JADT’ 18

139

idealism has become less important than many used to think; perhaps after
Dummett, realism is contrasted with anti-realism, rather than with idealism;
perhaps some sort of “interference” is here produced by the presence of a
further opposition, that between realism and nominalism.
The moral of this example is that clustering applied to large and conceptually
sophisticated corpora allows the historians of philosophy to concoct
alternative stories to account for the historical facts. This indicates that the
data-driven approach can trigger the production of conjectures one would
not think about. It is usually maintained that statistical techniques are useful
in that they restrict the space of possible interpretations (Mitchell, 1997), but
in other cases, such as the one described in this section, at least in an early
phase of the hermeneutic process, in virtue of their defamiliarizing impact
they can also have the opposite effect: that of broadening that same space
and discovering nouveaux observables (Rastier, 2011).
3.2. Paradigm
This case study deals with the term ‘paradigm’ in the period 1981-2015. After
exploring several k in the three decades, we focus on the synchronic analysis
of the set of clusters with k=16. The first result that immediately stands out is
that ‘paradigm’ occurs rather often: 995 documents, twice as many as
‘idealism’ (450), and considerably more than ‘necessity’ (719), a concept
which is widely regarded as central in the recent history of Anglo-American
philosophy. Using Google Ngram Viewer, and thus taking into account a
generalist, non disciplinary corpus, it turns out that such a high frequency is
peculiar to the philosophical discourse (the lowest value of ‘necessity’ is
0.0025%, which is higher than the highest value for ‘paradigm’, which is
0.0016%).
Why does ‘paradigm’ occur so frequently? On the one hand, one could find
this datum not so surprising, since ‘paradigm’ is a technical term in the
philosophy of science, introduced by Kuhn, 1962 to refer to a set of
methodological and metaphysical assumptions, examples, problems and
solutions, a vocabulary, which are taken for granted, in a given period of
normal science, by a scientific community. On the other hand, moving from a
priori considerations to the examination of the data, a partly different
landscape emerges: ‘paradigm’ seems to be a fashionable concept, which is
used in a variety of contexts as a term that is neither technical nor simply
ordinary. Only in cluster 8 has the term a straightforward technical use,
derived from Kuhn’s philosophy of science. Each of the other clusters (1:
theology, 2: music, 3: philosophy of law, 4: education; 5: nursing; 6:
philosophy of religion; 7: moral philosophy; 9: bioethics, 10: spiritualism; 11:
political theory; 12: self narrative; 13: theology; 14: Kant-Leibniz; 15

140

JADT’ 18

aesthetics; 16: philosophy and language in Wittgenstein, Heidegger etc.) does
not correspond to a different meaning of the term ‘paradigm’, but simply to
the application of the same concept to different fields. In most cases we have
to do with non-technical contexts, in which ‘paradigm’ has neither its
original grammatical meaning nor its ordinary, non-philosophical meaning
(standard, exemplar). It seems to us that its meaning and use are generic and
vague, rather than precise and technical; nonetheless, they evoke Kuhn: a
quasi-Kuhnian vocabulary became fashionable; it entered many
philosophical discourses, often more “humanistic” than “scientific” in spirit,
and much less technical than the philosophy of science.
This case study expresses an especially interesting kind of result obtainable
by using TM and NLP techniques to assist research in history of philosophy:
it shows how the interpretation of clusters fosters the discovery of
terminological fashions as opposed to genuine conceptual developments.
References
Aggarwal C.C., and Zhai C.X. (2012). “A Survey of Text Clustering
Algorithms.” In Mining Text Data, 77–128. Springer.
Allard M. et al. (1963). Analyse conceptuelle du Coran sur carte perforées.
Mouton.
Bonino G. and Tripodi P. (eds.), History of Late Analytic Philosophy, special
issue of “Philosophical Inquiries”, forthcoming.
Chartrand L., Meunier J.-G. and Pulizzotto D. (2016). CoFiH: A heuristic for
concept discovery in computer-assisted conceptual analysis. In Mayaffre
D. et al. (eds.), Proceedings of the 13th International conference on statistical
analysis of textual data, vol. I, pp. 85-95.
Danis J. (2012). L’analyse conceptuelle de textes assistée par ordinateur (LACTAO);
une expérimentation appliquée au concept d’évolution dans l’œuvre d’Henri
Bergson.
Université
du
Québec
à
Montréal
(http://www.archipel.uqam.ca/4641/1/M12423.pdf).
Ding X. (2013). A text mining approach to studying Matsushita’s
management thought. Proceedings of the 5th International conference on
informatin, process and knowledge, pp. 36-39.
Estève R. (2008). Une approche lexicométrique de la durée bergsonienne.
Actes des journées de la linguistique de corpus, vol. 3: 247-258.
Jain A.K. (2010). Data clustering: 50 years beyond K-means. Pattern
Recognition Letters, vol. 31(8): 651-666.
Kuhn T.S. (1962). The structure of scientific revolutions. University of Chicago
Press.
Le N.T, Meunier J.-G., Chartrand L. et al. (2016). Nouvelle méthode d’analyse
syntactico-sémantique profonde dans la lecture et l’analyse de textes

JADT’ 18

141

assistéespar ordinateur (LATAO). In Mayaffre D., et al. (eds.), Proceedings
of the 13th International conference on statistical analysis of textual data.
Manning C.D. at al. (2009). Introduction to Information Retrieval. Online
edition. Cambridge, UK: Cambridge University Press.
McKinnon A. (1973). The conquest of fate in Kierkegaard. CIRPHO, 1(1): 4558.
Meunier J.-G. and Forest D. (2005). Classification and categorization in
computer assisted reading and analysis of texts. In Cohen H. and Lefebvre
C. (eds.), Handbook of categorization in cognitive science, pp. 955-978.
Elsevier.
Meunier J.-G. and Forest D. (2009). Lecture et analyse conceptuelle assistée
par ordinateur: premières expériences. In Annotation automatique et
recherche d’informations. Hermes.
Mitchell T.M. (1997). Machine learning. McGraw-Hill.
Moretti F. (2005). Graphs, maps, trees. Abstract models for a literary history.
Verso.
Moretti F. (2013). Distant reading. Verso.
Pal A.R. and Saha D. (2015). Word sense disambiguation: A survey.
International Journal of Control Theory and Computer Modeling, vol. 5(3).
Pincemin B. (2007). Concordances et concordanciers: de l’art du bon KWAC.
XVIIe Colloque d’Albi. Langages et signification – Corpus en lettres et sciences
sociales: des documents numériques à l’interprétation, pp. 33-42.
Pulizzotto D. et al. (2016). Recherche de “périsegments” dans un contexte
d’analyse conceptuelle assistée par ordinateur: le concept d’“esprit” chez
Peirce. JEP-TALN-RECITAL 2016, vol. 2, pp. 522-531.
Rastier F. (2011), La mesure et le grain. Sémantique de corpus. Champion.
Sainte-Marie M. et al. (2010). Reading Darwin between the lines: a computerassisted analysis of the concept of evolution in the Origin of species. 10th
International conference on statystical analysis of textual data.
Salton G. (1971). The SMART Retrieval System: Experiments in Automatic
Document Processing. NJ: Prentice-Hall, Upper Saddle River.
Schmid H. (1994). “Probabilistic Part-of-Speech Tagging Using Decision
Trees.” In Proceedings of the International Conference on New Methods in
Language Processing. Manchester; UK.
Schmid H. (1995). “Improvements In Part-of-Speech Tagging With an
Application To German.” In Proceedings of the ACL SIGDAT-Workshop, 47–50.
Slingerland E. et al. (2017). The distant reading of religious texts: A “big
data” approach to mind-body concepts in early China. Journal of the
American Academy of Religion: 1-32.

142

JADT’ 18

La classification hiérarchique descendante pour
l’analyse des représentations sociales dans une
pétition antibilinguisme au Nouveau-Brunswick,
Canada
Marc-André Bouchard, Sylvia Kasparian
Université de Moncton – emb1214@umoncton.ca; sylvia.kasparian@umoncton.ca

Abstract
In this article, we apply Jean-Blaise Grize’s theoretical framework and Max
Reinert’s descending hierarchical classification to a corpus composed of
comments published as part of a petition against institutional bilingualism in
New Brunswick. Using Iramuteq, we point to the lexical worlds which
constitute anti-bilingualism arguments.
Résumé
Dans cet article, nous appliquons le cadre théorique développé par JeanBlaise Grize et la classification hiérarchique descendante de Max Reinert à un
corpus constitué de commentaires publiés dans le cadre d’une pétition contre
le bilinguisme institutionnel au Nouveau-Brunswick. Utilisant le logiciel
Iramuteq, nous dégageons les mondes lexicaux qui constituent
l’argumentation anti bilinguisme.
Mots-clés: mondes lexicaux, représentations sociales, schématisation,
classification hiérarchique descendante, pétition en ligne
1. Introduction
Toute analyse de discours, comme l’admet Jean-Blaise Grize dans Logique
naturelle et communications (1998; 144-145), est confrontée au problème de la
correspondance entre discours et représentations. Celui-ci serait attribuable
notamment à l’importance que donne l’analyse du discours à la situation de
communication, un facteur qui complique la relation de correspondance
entre ce qu’on dit et ce qu’on pense « vraiment ».
Dans le cadre de cet article, nous proposons d’explorer l’intersection entre
analyse de discours et étude des représentations et nous tenterons de
montrer que, bien que le problème de la correspondance entre discours et
représentations individuelles reste difficile à résoudre, les corpus de pétition
en ligne homogénéisent le discours et jouent sur la schématisation que
construit le locuteur, de façon à ce que les analyses logométriques puissent

JADT’ 18

143

accéder à certaines représentations sociales en jeu. À cet effet, nous aurons
recours à la méthode Reinert (une classification hiérarchique descendante
originalement popularisée par le logiciel ALCESTE) (1990) implantée dans le
logiciel Iramuteq (Ratinaud, 2009), qui consiste à relever les mondes lexicaux
d’un corpus. Plusieurs auteurs, dont Max Reinert lui-même, ont déjà établi
des liens entre cette méthode et le champ d’étude des représentations sociales
(1993; 13). Notre contribution à la conversation sera celle d’appliquer la
méthodologie issue de la logométrie et le cadre théorique développé par
Grize à un nouveau type de corpus qui gagne en popularité depuis le début
du 21e siècle, celui des pétitions en ligne. L’exemple par lequel nous
illustrerons notre exposé théorique sera celui de l’analyse, à l’aide
d’Iramuteq, des mondes lexicaux d’une pétition en ligne lancée au NouveauBrunswick (Canada) en 2013, sur la plate-forme www.change.org, contre
l’exigence du bilinguisme comme critère d’emploi dans la fonction publique
provinciale.
2. Cadre théorique
Selon Denise Jodelet, on peut définir la représentation sociale comme « une
forme de connaissance socialement élaborée et partagée, ayant une visée
pratique et concourant à la construction d’une réalité commune à un
ensemble social » (1997; 53). Ainsi, comme le remarque Serge Moscovici, leur
étude demande des méthodes d’observation plutôt que d’expérimentation
étant donné qu’elle se manifeste « comme une "modélisation" de l’objet
directement lisible dans, ou inférée de, divers supports linguistiques,
comportementaux ou matériels » (idem; 61). Bien qu’elle soit forme de
connaissance, la représentation se distingue de la connaissance scientifique
en ce qu’elle découle de ce que Jean-Blaise Grize nomme la logique naturelle
(Grize, 1997; 171-172), donnant ainsi sur un « savoir de sens commun »
(Jodelet, 1997; 53). Il faut entendre par « logique naturelle » qu’il est question
d’une logique d’ordre logico-discursif, manifestée dans le discours par la
schématisation, qui « prend en compte les contenus et non les seules formes
de la pensée » (Grize, 1997; 171-172). Selon Grize, la schématisation compte
cinq notions articulant son ensemble ainsi :
[1] Une schématisation est la mise en discours [2] du point de
vue qu’un locuteur A [3] se fait – ou a – d’une certaine réalité R.
[4] Cette mise en discours est faite pour un interlocuteur, ou un
groupe d’interlocuteurs, B [5] dans une situation d’interlocution
donnée (idem).

144

JADT’ 18

Ainsi, Grize propose que toute communication est situation d’interlocution,
dans laquelle l’orateur construit une schématisation en fonction de son
préconstruit culturel, de ses représentations de l’objet en question, et de sa
finalité; cette schématisation est constituée d’images de l’orateur, de
l’auditeur et de l’objet dont il s’agit, et elle est ensuite reconstruite par
l’auditeur en fonction de ses propres représentations, préconstruit culturel et
finalité (Grize, 1993; 7). La schématisation est donc partielle et partiale : « elle
est partielle dans la mesure où son auteur n’y fait figurer que ce qu’il juge
utile à sa finalité, à l’effet qu’il veut produire; elle est partiale puisqu’il
l’aménage de telle façon que B la reçoive » (Grize, 1997; 175). En termes de
finalité, selon Patrick Charaudeau, les discours, plus particulièrement ceux
de type argumentatif ont une double quête, soit le vraisemblable et
l’influence, le succès de celle-ci étant fonction des « représentations
socioculturelles partagées par les membres d’un groupe donné au nom de
l’expérience ou de la connaissance » (1992; 784). C’est donc dire que, compte
tenu de la « double quête » du mode de discours, les représentations d’objets
sur lesquelles le locuteur construit sa schématisation sont choisies en raison
du partage, supposé par le locuteur, de ces représentations chez le(s)
destinataire(s). Dès lors, l’analyse des mondes lexicaux communs à un
groupe de locuteurs dans une même situation de communication peut nous
donner des indices des représentations sociales que se fait le groupe d’un
objet du monde social. En effet, selon Max Reinert, dans un corpus collectif,
un monde lexical serait indicateur d’un espace de référence commun à un
groupe et « l’indice d’une forme de cohérence liée à l’activité spécifique du
sujet-énonciateur » (Reinert, 1993; 13). La méthode de classification
hiérarchique descendante (Reinert, 1990) propose une représentation de ces
mondes lexicaux (ou thématiques) sous la forme de tableaux de classification
obtenus par voie du croisement des unités de contexte (ou segments) et des
lexèmes d’un corpus. L’hypothèse à la base de cette méthode est que « dans
la mesure où une représentation collective exprime une certaine régularité de
structure dans une classe de représentations singulières […] cette régularité
est due aux contraintes de ce que nous appelons "un monde" » (Reinert, 1993;
29-30). La prise en compte de la fréquence et de l’environnement des formes
d’un corpus permet non seulement de relever les formes lexicales les plus
propices à constituer des indices de représentations sociales, mais aussi de
définir ces formes lexicales en fonction de leur cotexte.
3. Corpus
Le corpus que nous analysons dans la présente recherche est issu d’une
pétition en ligne. Contrairement à la pétition classique, la pétition en ligne
permet à ceux qui y apposent leur nom d’y publier, s’ils le désirent, un

JADT’ 18

145

commentaire justifiant leur appui au titre et à la description de celle-ci. Celle
dont il est question ici, Stop the hiring discrimination against citizens who speak
English only1, a été lancée en 2013 au www.change.org. Ses commentaires, en
plus d’être signés par leurs auteurs, sont accessibles publiquement sur la
page même. Cette particularité du canal de communication, que Contamin
(2001) appelle « un paradoxe classique des pétitions », a une incidence sur le
destinataire de la mise en discours en ce que ce dernier n’est pas seulement le
gouvernement de la province, mais aussi le grand public. Ainsi, les corpus de
pétitions en ligne homogénéisent les discours selon le modèle de la
communication de Grize. D’abord, le groupe de locuteurs se trouve dans la
même situation d’interlocution (monologues, à l’écrit, mode argumentatif) et
est invité à partager son point de vue sur une même réalité (en l’occurrence,
le bilinguisme institutionnel de la province du Nouveau-Brunswick). Ces
mises en discours sont faites pour un public général, et la nature engagée de
la pétition fait en sorte que, en théorie du moins, seuls les locuteurs
partageant le point de vue énoncé dans le titre sont représentés.
Le point de vue partagé par les intervenants, dans notre corpus, est que
l’exigence du bilinguisme anglais-français pour des emplois dans la fonction
publique provinciale constitue une discrimination envers les NéoBrunswickois anglophones, qui sont largement unilingues (moins de 15% de
ceux-ci se considèrent bilingues, comparativement à un taux de plus de 70%
dans la communauté minoritaire francophone). Ces discours s’inscrivent
dans un long débat au sein de la population néo-brunswickoise sur le
bilinguisme institutionnel, et historiquement le clivage se fonde sur la base
linguistique : les francophones sont en faveur du bilinguisme de l’État et de
l’avancement des droits linguistiques, alors que les anglophones y sont plus
réticents. En tout, à son terme à la fin de l’année 2013, la pétition Stop the
hiring discrimination against citizens who speak only English récolte 7758
signatures, pour un total de 2372 commentaires, la longueur de chacun
variant d’un mot (« jobs ») à 304 mots, pour une moyenne de 37,66 unités
linguistiques par commentaire. Ce corpus compte 4 425 formes différentes
représentant un total de 89 338 occurrences. Le corpus nettoyé et uniformisé
a été soumis à l’analyse du logiciel Iramuteq qui nous donne le
dendrogramme des classes constituant les mondes lexicaux des
commentaires présentés dans la section suivante.

1https://www.change.org/p/the-government-of-new-brunswick-stop-the-hiringdiscrimination-against-citizens-who-speak-only-english

146

JADT’ 18

4. Analyse
Les 89 338 occurrences (4425 formes différentes) qui constituent notre corpus
sont regroupées en 3492 lemmes, soit 2954 formes actives et 538 formes
supplémentaires. Et l'ensemble du corpus est segmenté en un total de 2423
parties constituées d'un nombre plus ou moins égal de formes (en moyenne
36.87 formes par segment). L’analyse de la classification hiérarchique
descendante avec Iramuteq produit le graphe présenté dans la Figure 1.

Figure 1 : Classification sur segments de textes simples

La lecture de la Figure 1 révèle que la première segmentation du corpus
donne lieu à la Classe 1 (en rouge), formant une classe représentant 30.3 %
des segments classés et constituée d'un lexique que nous nommons l’axe
sociopolitique : on y aborde d'abord la dynamique « majority » / « minority »,
qui, à se fier à cette liste de formes, jouerait un rôle d'avant-plan dans les
représentations du Canada et des provinces de ce pays. On remarque aussi,
en plus de quelques formes relevant de la culture et de la langue, un champ
lexical qui semble indiquer la présence de positionnements politiques dans le
corpus (« right », « common », « sense », « rule », « vote », « political », «
equal »), alors que les verbes (« fight », « cater », « stand », « stop », « start », «
push »), de nature politique aussi, renforcent l'hypothèse que cette classe est

JADT’ 18

147

constituée de segments exprimant des représentations au sujet de la société
canadienne. Une fois la Classe 1 constituée, le calcul divise le deuxième
segment en deux classes : la Classe 2 (en vert), contenant 31.7 % de ceux-ci;
contre 38 % dans la Classe 3 (en bleu). On observe que, collectivement, cellesci se démarquent de la Classe 1 par leur lexique relevant de l'expérience
personnelle plutôt que de l’opinion politique.
Cette caractéristique personnelle se manifeste dans la Classe 2 par des formes
comme « home », « family », « child », « young », et « daughter ». Les verbes,
quant à eux, précisent le contexte de cette expérience : « move », « find », «
leave », « work », « live », « stay », « raise », « love », et « born »; tout comme
quelques adjectifs évaluatifs et/ou axiologiques : « hard », « good », « decent
», et « impossible ». On observe aussi quelques formes, en plus de « [new]
brunswick », qui réfèrent à une province canadienne, soit à l'Alberta. Le
contenu de la Classe 2 constitue donc l’axe biographique, rejoignant souvent
le thème de l'exode vers l'Ouest canadien.
La troisième et dernière classe du corpus (en bleu) gravite autour du thème
du travail, voire plus précisément de la recherche d'un emploi. C'est aussi
dans cette classe qu'on trouve les seules références directes à la langue, mise
à part la forme « language » dans la Classe 1 : « bilingual », « speak », et «
french ». Certaines formes spécifiques à la Classe 3 laissent entendre que
celle-ci est, en partie, plus impersonnelle que la Classe 2 : « employee »,
« person », « applicant », et « individual ».
À partir de la classification sur segments de texte, on peut parcourir, de façon
automatisée, l'ensemble des segments de chaque section et leur attribuer un
score selon le nombre de mots représentatifs de la classe où ils se trouvent;
on tient aussi compte du degré de représentativité de ces formes.
Ainsi, les deux segments qui suivent sont caractéristiques de la Classe 1: «
discrimination of the english[-]speaking white majority populace should stop
with the democratic system becoming more in play with majority rules as a
true reflection of the people »; « we as a province cannot afford duplicate
books in 2 languages to support a minority and the need to speak french in a
majority speaking english province to have a job is ridiculous »
Il apparait, dans les segments caractéristiques de la Classe 1, un
renversement du rapport de pouvoir classique entre un groupe majoritaire et
un groupe minoritaire : les anglophones sont ici opprimés, alors que ce sont
les francophones qui sont avantagés, qui ont l'oreille attentive du
gouvernement, et, ultimement, qui détiennent le marché du travail bilingue.
Cette oppression serait apparente dans la difficulté pour les anglophones
unilingues de se trouver un emploi, dans la fonction publique notamment,
mais peut-être aussi dans le secteur privé. On remarque d'emblée une
représentation de la démocratie se résumant à la règle de majorité (telle que

148

JADT’ 18

définie par H. B. Mayo (1957; 50) comme : « the principle that when there is a
majority on a matter, then the wishes of the majority should prevail »), ce qui
est explicitement communiqué au premier segment caractéristique de la
Classe 1. En ce qui concerne la Classe 2, voici deux des segments les plus
caractéristiques : « it is very important to me because my daughter like 1000s
of other working children here in new brunswick have had to leave their
home province in order to find work because they only speak their own
language of english. »; et « i have been out of work for over a year. Unable to
find a full time job due to bilingualism restrictions. Going to have to move
west. ». Il apparait donc qu’il y a un motif récurrent dans la Classe 2 : pour
trouver un bon emploi, voire un emploi tout court, il faut être bilingue, faute
de quoi on s’exile, notamment dans l’Ouest canadien. On remarque que ces
segments témoignent d’un sentiment d’impuissance mais aussi de réticence
face à l’idée de quitter sa province natale. Certains segments caractéristiques
de la Classe 2 traitent de l’expérience personnelle du commentateur, qui a dû
ou qui croit avoir à déménager dans une province non bilingue, alors que
d’autres racontent l’exode, accompli ou prévu, de leur(s) enfant(s). On
remarque que, dans les segments de la Classe 2 qui précèdent, on attribue
volontiers la pauvreté du marché de l’emploi pour les anglophones au
facteur linguistique. Ensuite, les segments caractéristiques de la Classe 3 sont
les suivants: « because this is a problem, i have 17 years’ experience and 2
degrees and i can’t even apply for the jobs i qualify for because it’s
mandatory bilingual positions when over 90% of the day is dealing in
english, they won’t even interview you unless you speak french »; et « the
most qualified person for the job is not always hired because they are not
bilingual ». Les différentes formes du concept de « qualification », et d’autres
qui y sont liées sémantiquement, sont omniprésentes dans ces segments
caractéristiques. Il apparait d’emblée qu’on exclut les compétences
linguistiques de ce concept. En effet, une personne qui parle seulement
l’anglais est présentée comme potentiellement aussi qualifiée, et à l’occasion
plus qualifiée, qu’un candidat bilingue à un emploi qui demande le
bilinguisme. Le scénario, souvent hypothétique, qui est donné à voir tend à
mettre en jeu une personne unilingue qui serait plus qualifiée qu’une autre
chez qui le bilinguisme est présenté comme le seul atout.
5. Conclusion
En somme, dans le cadre de cette pétition, les locuteurs ont mis en discours
des représentations du bilinguisme institutionnel au Nouveau-Brunswick par
l’entremise de trois mondes lexicaux, présentant ainsi trois facettes de la
discrimination perçue envers les anglophones dans la fonction publique. Le
premier monde lexical est sociopolitique et énonce des principes généraux

JADT’ 18

149

sur ce qui est juste; le deuxième est biographique et relate les effets
personnels de cette discrimination; et le troisième porte sur des exemples de
la façon dont se manifeste cette discrimination dans le monde du travail.
Ainsi, l’échantillon des représentations sociales du bilinguisme institutionnel
constituant notre corpus donne à voir un lien de causalité entre l’exigence du
bilinguisme pour certains emplois et les difficultés du marché du travail de la
province. Dans le but de convaincre un public général, ce point de vue est
présenté sous un angle à la fois idéologique, personnel ou pratique,
renvoyant ainsi à certaines images de la démocratie, de l’exode et de la
compétence; images qui, bien que relativement homogènes dans notre
corpus, ne seraient pas nécessairement partagées dans les représentations
sociales des anglophones bilingues et des francophones.
Bibliographie
Charaudeau, Patrick (1992). Grammaire du sens et de l’expression. Hachette.
Contamin, J.-G. (2001). Contribution à une sociologie des usages pluriels des forms
de mobilization : l’exemple de la petition en France. Thèse de doctorat de
l’Université Paris 1.
Grize, Jean-Blaise (1998). Logique naturelle et communications. Presses
Universitaires de France.
Jodelet, Denise (1997). Les représentations sociales. Dans Jodelet ed. Les
representations sociales (5e ed.). Presses Universitaires de France.
Mayo, H. B. (1957). Majority Rule and the Constitution in Canada and the
United States. Political Research Quarterly, vol. 10(1) : 49-62
Ratinaud, Pierre (2009). Iramuteq : interface de R pour les analyses
multidimensionnelles de textes et de questionnaires. http://www.iramuteq.org.
Reinert, Max (1990). Alceste une méthodologie d’analyse des données
textuelles et une application. Bulletin de Méthodologie Sociologique, vol.
26(1): 24-54
Reinert, Max (1993). Les “mondes lexicaux” et leur “logique” à travers
l’analyse statistique d’un corpus de récits de cauchemars. Langage et
société, vol. 66(1) : 5-39
Reinert, Max (1997). Postures énonciatives et mondes lexicaux stabilisés en
analyse statistique de discours. Langage et société, no. 121/122 : 189-202

150

JADT’ 18

Analysing occupational safety culture through
mass media monitoring
Livia Celardo1, Rita Vallerotonda2, Daniele De Santis2,
Claudio Scarici2, Antonio Leva2
2

1 Sapienza University of Rome
INAIL Research – Headquarters for Research of the Italian National Institute for Insurance
against Accidents at Work

Abstract 1
In the last years, a group of researchers within the National Institute for
Insurance against Accidents at Work (INAIL) has launched a pilot project
about mass media monitoring in order to find out how the press deal with
the culture of safety and health at work. To monitor mass media, the Institute
has created a relational database of news concerning occupational injuries
and diseases, that was filled with information obtained from the newspaper
articles about work-related accidents and incidents, including the text itself of
the articles. In keeping with that, the ultimate objective is to identify the
major lines for awareness-raising actions on safety and health at work. In a
first phase of this project, 1,858 news articles regarding 580 different
accidents were collected; for each injury, not only the news texts but also
several variables were identified. Our hypothesis is that, for different kind of
accidents, a different language is used by journalists to narrate the events. To
verify it, a text clustering procedure is implemented on the articles, together
with a Lexical Correspondence Analysis; our purpose is to find language
distinctions connected to groups of similar injuries. The identification of
various ways in reporting the events, in fact, could provide new elements to
describe safety knowledge, also establishing collaborations with journalists in
order to enhance the communication and raise people attention toward
workers' safety.
Abstract 2
Negli ultimi anni un gruppo di ricercatori all’interno dell’Istituto Nazionale
per l’Assicurazione contro gli Infortuni sul Lavoro e le malattie professionali
(INAIL) ha lanciato un progetto pilota riguardante il monitoraggio dei mass
media con lo scopo di analizzare come la stampa tratta la salute e la sicurezza
sul lavoro. A tal fine, l’Istituto ha istituito un database relazionale delle
notizie riguardanti gli infortuni e le malattie, incluso il testo stesso delle
notizie. L’obiettivo finale del progetto è dunque quello di identificare le
direttrici principali su cui muoversi per azioni di sensibilizzazione su salute e

JADT’ 18

151

sicurezza sul lavoro. Nella prima fase del progetto, 1,858 articoli di giornale
riguardanti 580 infortuni sono stati raccolti; per ogni evento, non solo il testo
della notizia ma anche diverse variabili sono state individuate. La nostra
ipotesi è che per diversi tipi di infortunio un diverso linguaggio viene usato
dai giornalisti per narrare l’accaduto. Per verificare ciò, una procedura di
Text Clustering è stata implementata sugli articoli, insieme ad una Analisi
delle Corrispondenze Lessicali; il nostro obiettivo è quello di individuare
delle differenze nel linguaggio in relazione a diversi gruppi di infortuni.
L’identificazione di diversità nel modo in cui viene riportata la notizia al
lettore può fornire nuovi elementi per descrivere la cultura della sicurezza, al
fine di instaurare delle collaborazioni con i giornalisti stessi per rendere
migliore la comunicazione e accrescere l’attenzione del cittadino verso la
sicurezza del lavoratore.
Keywords: Occupational safety; Work-related accident; Text mining; Mass
media.
1. Introduction
The study described here grew out of the collaboration between the
Department of Social Sciences and Economics of Sapienza University of
Rome and the Headquarters for Research of INAIL (Italian National Institute
for Insurance against Accidents at Work) where, since 2012 a team of
researchers has developed the idea of monitoring the mass media in view of
prevention against accidents at work (INAIL, 2015).
With this in mind, those researchers achieved the so-called “Repertorio
Notizie SSL” (News Repository on Occupational Safety and Health), that is a
relational database of media news related to occupational injuries and
diseases. The objective of this project is to observe the culture of occupational
safety and health communicated by mass media agencies in order to identify
new elements for increasing prevention against accidents at work. In this
study we focus on the hypothesis that there are some asymmetries in the
language used to describe the injuries depending on the characteristics of the
event. To test it, we performed on the repository data some Automatic Text
Analysis procedures.
The article is structured as follow: in section no.2, the News Repository is
presented; in section no.3, data are presented and the methodology is
exposed; in section no.4, the results of the analyses are shown; in section no.5,
conclusions are drawn.
2. The tool
News Repository on Occupational safety and health (NeRO) is a tool created to
allow analyses of news contents and texts related to occupational diseases

152

JADT’ 18

and injuries. In fact, our strategic objective is to increase public awareness
and safety culture through a different approach, which will be also based on
the study of news articles, their composition and communication dynamics.
So, the first operational purpose is to understand:
- which kind of terms are used in news articles about accidents at work or
occupational diseases;
- what inspires a title;
- how the same news is treated by different sources/media;
- how the news text could be interpreted in different ways due to who
communicates the news itself;
- whether or not some specific aspects of the events are considered by
media.
Our study plans to analyze the cultural characteristics of mass media
communication regarding occupational safety and health (OSH), observing
the attitude of mass media (and journalists) towards the subject and the way
users perceive the news depending on which words are used. As mentioned
before, NeRO is an ad hoc relational database, centred on the gathering of
newspaper articles regarding accidents at work, but it is also arranged to
gather news on near misses, occupational diseases and incidents from all
kind of sources (press, television or radio). It involves several digital
interconnected tables, which contain structured – i.e. based on appropriate
classifications – and unstructured – i.e. textual – information. Information
retrieval regards events happened in Italy and it could contain both online
and directly consulting newspapers, since we exploited Google Alert Service
(using some suitable keywords) and a daily-newspaper subscription (“la
Repubblica”). The reference unit is the event (right now, we are restricting
events to accidents) and different aspects and information are linked to it:
one or more articles about it, one or more workers injured, and so on. The
data-entry interface consists of a series of thematic screens, starting from the
opening one, which covers the list of already recorded events. These screens
allow to enter the following data, step by step:
 [Screen “Event”] Text containing event description, date of the event,
venue, company where accident occurred (if appropriate), economic
activity;
 [Screens “News”] Texts of each article related to the event, newspaper
name (or press affiliation), news title, web url, date of the article;
 [Screens “Worker” and Sub-screens “Accident” and “Harms, disorders
or diseases”] Injured worker’s biographical data, information about
accident, type of injury, physical implication or resulting disease.

JADT’ 18

153

3. Methodology and data
The repository, at the end of data collection, was composed of 1,858 news,
related to almost six hundreds different accidents. In order to analyse the
content of the news texts in connection with the characteristics of the
different events, we performed a content analysis using the Reinert’s method
(Reinert, 1983) for a descendant hierarchical partition. This algorithm,
starting from the co-occurences matrix, generates groups of lexical units – i.e.
words – that more co-occur in the texts. Then, the lexical groups were
projected on the factorial axes, together with the variables modalities, using
the Lexical Correspondence Analysis (Lebart, Salem and Berry, 1997); in this
way, we could observe how the language is connected to the accidents
features. Finally, to better understand the differences between news texts we
analysed the specificities related to the modalities of the variables.
4. Main results and discussion
The cluster analysis made on news texts using the Reinert’s method–
choosing as segments the articles – produced three lexical groups (in order,
the red, the blue and the green ones, in Figure 1):
- Cluster 1 (56.5%): in this group are included the words related to the
description of the events, in terms of what happened;
- Cluster 2 (26.5%): here we have the terms connected to the road
accidents;
- Cluster 3 (17%): this group is about the emotional aspects connected
to the events.
We projected the lexical groups (Figure 1) and the modalities of the variables
related to the events (Figure 2) on the first two factors obtained using the
lexical correspondence analysis.
As shown in the figure no. 2, there are some interesting characterizations of
the language used in newspapers. Some variables, like the economic activity
and the accident site, present a strong lexical differentiation among the
modalities; this means that who is narrating the event - i.e. the journalist uses a specific language to describe the accident, on the basis of these
characteristics. The other variables presented no particular specificities,
except for the one related to the mortality of the accident. In fact, as shown in
the figure no. 2, on the second factor the variable “accident mortality” is best
represented because of the position and the distance of the modalities “yes”
and “no” from the origin. To better understand the lexical differences, we
analysed also the specificities (Bolasco and De Mauro, 2013; Lafon, 1980;
Lebart, Salem and Berry, 1997) for this particular variable.

154

JADT’ 18

Figure 1 Lexical groups

Figure 2 Lexical correspondence analysis

Starting from the results showed in table no.1, we can observe that there is a
significant difference in the language utilized when the accident is fatal or
not. The terms used in the case of a non-fatal event are related to the
description of the injury, while in the case of a mortal accident the situation is
completely different: the words utilized refer to the emotional sphere of the
event, so concepts like the family or the unpredictability are very often used
to describe what was happened.

JADT’ 18

155

Table 1 Analysis of the specificities – Variable: “accident mortality”
Fatal accident - No
Fatal accident - Yes
z = test-value
z = test-value
Hospital
59.17 Tragedy
35.68
Serious
58.84 Family
27.17
To transfer
54.90 Useless
23.62
Dangerous
28.38 To leave
19.84
Rescue
24.13 Victim
18.68
Ambulance
24.09 Tragic
17.71
Leg
23.12 Friend
14.95
Injury
22.06 Band
14.89
Trauma
20.55 Condolence
12.65
Hand
18.84 Province
12.15
Fracture
16.70 Son
11.49
Helicopter
13.70 Wife
11.48
Bus
12.23 Escape
10.63
Crossroad
10.20 Mayor
9.11

5. Conclusions
The project here presented showed how News Repository on OSH (NeRO)
can contribute to analyse occupational safety and health, although in some
institutions there are already databases dedicated to newspaper articles
dealing with OSH. Actually, in addition to news texts, NeRO provides
several systematized information, enabling to filter news according to
various search criteria and, above all, to carry out a number of studies and
organized analysis on textual data, too. In this paper, we showed one of the
study we implemented on Repository data using Automatic Text Analysis.
The results revealed that a large amount of information is contained within
these data; anyway, some information asymmetries are present. For that
reason, it will be essential to set up a discussion with a network of journalists
and other experts, in order to improve and enhance the media
communication. The challenge is to get out from the inner circle of
prevention practitioners and build a bridge that could connect the Institution
to a more general public, also contemplating liaison organizations (such as
trade unions and employers' associations).
References
Bolasco S. and De Mauro T. (2013). L'analisi automatica dei testi: fare ricerca con
il text mining. Carocci Editore.
Iezzi D. F. (2012). Centrality measures for text clustering. Communications in
Statistics-Theory and Methods, 41(16-17), 3179-3197.
INAIL. (2015). Il monitoraggio dei mass media in materia di salute e sicurezza:
Strumenti per la raccolta e l’analisi delle informazioni.
Lafon P. (1980). Sur la variabilité de la fréquence des formes dans un

156

JADT’ 18

corpus. Mots, 1(1), 127-165.
Lebart L., Salem A. and Berry L. (1997). Exploring textual data(Vol. 4). Springer
Science & Business Media.
Reinert M. (1983). Une méthode de classification descendante hiérarchique:
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, 8(2), 187-198.

JADT’ 18

157

Is the educational culture in Italian Universities
effective? A case study
Barbara Cordella, Francesca Greco, Paolo Meoli,
Vittorio Palermo, Massimo Grasso
Sapienza University of Rome – barbara.cordella@uniroma1.it; francesca.greco@uniroma1.it;
paolomeoli3@libero.it; vittorio.palermo2511@gmail.com; massimo.grasso@uniroma1.it

Abstract 1
The paper explores the professors and students’ representation of
professional training in Clinical Psychology in the faculty of Medicine and
Psychology of the Sapienza University of Rome in order to understand
whether the educational context supports students in developing their ability
to enter the job market. To this aim, an Emotional Text Mining of the
interviews of 30 students and 17 teachers of the Clinical Psychology Master
of Science was performed. Both corpora underwent the analysis procedure
performed with T-Lab, i.e. a cluster analysis with a bisecting k-means
algorithm followed by a correspondence analysis on the keyword per cluster
matrix, and the results were compared. The results show 4 clusters and 3
factors for each corpus, highlighting a relationship between student and
professor representations. Both of them split the training process,
distinguishing the educational process from the professional one. The
emotional text mining of the interviews turned out to be an enlightening tool
letting their latent dimensions emerge, setting the process and outcome of the
academic training, and it proved to be very useful for educational purposes.
Abstract 2
La ricerca ha esplorato la rappresentazione della formazione in Psicologia
Clinica dei professori e degli studenti della facoltà di Medicina e Psicologia
della Sapienza Università di Roma al fine di comprendere se il contesto
formativo supporti gli studenti nello sviluppo di competenze utili
all’inserimento nel mercato del lavoro. A questo scopo è stata effettuata
un’Emotinal Text Mining delle interviste di 30 studenti e di 17 professori del
Corso di Laurea Magistrale in Psicologia Clinica con T-Lab (analisi dei cluster
con algoritmo bisecting k-means seguita da un’analisi delle corrispondenze
sulla matrice cluster per parole-chiave). I risultati mostrano 4 cluster e 3
fattori in entrambi i corpora, evidenziando una relazione tra le
rappresentazioni degli studenti con quelle dei professori per quanto concerne
il processo di apprendimento, distinguendo e mantenendo separati gli aspetti
formativi da quelli professionali. L’Emotional Text Mining risulta essere uno

158

JADT’ 18

strumento utile ad evidenziare le dimensioni latenti che organizzano il
processo e i risultati dell’apprendimento accademico.
Keywords: Education, Clinical Psychology, Job Market, Youth
Unemployment, Emotional Text Mining.
1. Introduction
The problem of youth unemployment is relevant nowadays. In Italy, 25% of
young people under 30 years of age are unemployed and this percentage
grows to 40% for under 25s (Mckinsey & Company, 2014). But why is this
percentage so high? According to Mckinsey’s study (ibidem), it shows that the
figure of 40% for youth unemployment does not rely on the economic cycle
but on “structural causes”. Among other causes, education is one of the
relevant factors of youth unemployment, and is a protection factor for
poverty and quality of life, as stated by ISTAT (2017). Graduates are less
likely to become poor although the employability and the wages depend on
the type of degree. 80% of young graduates in psychology are employed after
four years (Anpal Servizi, 2017). Psychologists are more likely to become
entrepreneurs than employees. Most probably, the length of time needed to
get into the job market is connected to the mismatch between the educational
system and enterprise (McKinsey & Company, 2014). Young people’s skills
are considered appropriate by 70% of Schools and Universities, but only by
42% of employers. The effectiveness of education depends in part on the
representation of the professional training characterizing the University.
Several studies were performed in order to investigate students’
representation in the Psychology Faculty in order to improve the training
process (e.g., Carli et al., 2004; Paniccia et al., 2009). Due to the change in the
educational plan that took place over the past decade, this study aims to
understand whether the present educational context supports students in
developing their ability to enter the job market, performing an emotional text
mining (Cordella et al., 2014; Greco, 2016) of the interviews of students and
teachers of the Master Degree in Clinical Psychology at the Sapienza
University of Rome.
2. Methodology
We know that a person's behaviour depends not only on their rationale
thinking but also, and sometimes most of all, on their emotional and social
way of mental functioning (Carli, 1990; Moscovici, 2005). Namely, people
consciously categorize reality and, at the same time, unconsciously symbolize
it emotionally (Fornari, 1976). These two thinking processes are the product
of the double-logic way of the functioning of the mind (Matte Blanco, 1981)
which allows people to adapt to their social environment. According to this

JADT’ 18

159

socio-constructivist approach, based on a psychodynamic model, the
unconscious processes are social, as people generate interactively and share
the same emotional meanings. The socially shared emotional symbolization
sets the interactions, behaviours, attitudes, expectations and communication
processes, and for this reason, the analysis of the narrations allows for the
acquisition of the latent emotional meaning of the text (Salvatore & Freda,
2011). If the conscious process sets the manifest content of the narration,
namely what is narrated, the unconscious process can be inferred through
how it is narrated, that is to say, the words chosen to narrate and their
association within the text. We consider that people emotionally symbolize
an event, or an object, and socially share this symbolisation. The words they
choose to talk about this event, or object, is the product of the socially-shared
unconscious symbolization (Greco, 2016). According to this, it is possible to
detect the associative links between the words to infer the symbolic matrix
determining the coexistence of these terms in the text. To this aim, we
performed a multivariate analysis based on a bisecting k-means algorithm
(Savaresi et Boley, 2004) to classify the text, and a correspondence analysis
(Lebart et Salem, 1994) to detect the latent dimensions setting the cluster per
keywords matrix. The interpretation of the cluster analysis results allows for
the identification of the elements characterizing the emotional representation
of education, while the results of correspondence analysis reflect its
emotional symbolization (Cordella et al., 2014; Greco, 2016). The advantage
connected with this approach is to interpret the factorial space according to
words polarization, thus identifying the emotional categories that generate
professional training representations, and to facilitate the interpretation of
clusters, exploring their relationship within the symbolic space.
3. Data collection and analysis
In order to explore the emotional representation of the education in the
Master of Science in Clinical Psychology, we interviewed 30 students (13% of
students) and 17 teachers (71% of teachers) of the Sapienza University of
Rome accordingly to their voluntary participation. We used an openquestions interview for students and teachers. Students’ interviews resulted
in a medium size corpus of 57.387 tokens, and teachers’ interviews resulted
in a small size corpus of 28.746 tokens. In order to check whether it was
possible to statistically process data, two lexical indicators were calculated:
the type-token ratio and the hapax percentage (TTRstudents = 0,09; Hapaxstudents =
50,3%; TTRteachers = 0,147; Hapaxteachers = 53,8%). According to the size of the
corpus, both lexical indicators highlight its richness and indicate the
possibility to proceed with the analysis. First, data were cleaned and preprocessed by the software T-Lab (Lancia, 2017) and keywords were selected.

160

JADT’ 18

Due to the size of the corpus and the hapax percentage, in order to choose the
keywords, we used the selection criteria proposed by Greco (Cordella et al.,
2014; Greco, 2016). In particular, we used stem as keywords instead of type,
filtering out the lemma of the open-questions of the interviews. Then, on the
context units per keywords matrix, we performed a cluster analysis with a
bisecting k-means algorithm (Savaresi et Boley, 2004) limited to ten partitions,
excluding all the context units that did not have at least two keywords cooccurrence. The eta squared value was used to evaluate and choose the
optimal solution. To finalize the analysis, a correspondence analysis on the
keywords per clusters matrix was made (Lebart et Salem, 1994) in order to
explore the relationship between clusters, and to identify the emotional
categories setting professional training representations both for students and
teachers.
4. Main results and discussion
The results of the cluster analysis show that the keywords selected allow the
classification on an average of 96% for both corpuses. The eta squared values
was calculated on partitions from 3 to 9, and they show that the optimal
solution is four clusters for both corpora. The correspondence analysis
detected three latent dimensions. In table 1 and 2, we can appreciate the
emotional map of the professional training emerging from the interviews of
the teachers and the students and cluster location in the factorial space.
Table 1  Cluster coordinates on factors of the teachers’ corpus (the percentage of explained
inertia is reported between brackets above each factor)
Cluster
(CU in Cl %)
1
2
3
4

Training Group
(22,3%)
Clinical Training
(33,7%)
Institutional Obligations
(20,2%)
Student Orientation
(23,8%)

Factor 1 1
(26,53%)
Motivation
Group
-0,21
Institution
0,33
Institution
0,65
Group
-0,79

Factor 2
(19,03%)
Outcome
Competence
0,51
Competence
0,23
Degree
-0,66
Degree
-0,39

Factor 3
(14,56%)
Role
Teacher
-0,50
Professional
0,39
Teacher
-0,38
Professional
0,16

CU in Cl = context units classified in the cluster.
The teachers’ corpus first factor (table 1) represents the motivation in
teaching, focusing on the group of students and their specific needs or on the
Institutional generic scopes; the second factor focuses on the training
outcome, the degree or the professional skills; and the third factor reflects the
role of the academic professor that could represent oneself as a teacher or a

JADT’ 18

161

professional. As regards the students corpus (table 2), the first factor
represents the approach to university experience, which can be perceived as
an individual experience or a social one (relational); the second factor
explains how students experience vocational training, perceiving it as the
fulfilment of obligations or the construction of professional skills that
requires personal involvement; and the third factor reflects the outcome of
the educational training that can focus on professional skills development or
on the achievement of qualifications.
Table 2  Cluster coordinates on factors of the students’ corpus (the percentage of explained
inertia is reported between brackets above each factor)
Cluster
(CU in Cl %)
1
2
3
4

Idealized Product
(27,6%)
Professional Education
(20,8%)
Group Identity
(26,3)
Empty Degree
(25,3%)

Factor 1
(23,2%)
Approach
Individual
-0,56
-0,04
Relational
0,69
Individual
-0,32

Factor 2
(15,3%)
Training
Fulfilment
0,45
Construction
-0,63
Fulfilment
0,22
0,01

Factor 3
(14,0%)
Outcome
Skills
-0,43
Skills
-0,24
-0,01
Qualifications
0,59

CU in Cl = context units classified in the cluster.
Table 3  Teachers’ Cluster (the percentage of context units classified in the cluster is reported
between brackets)
Cluster 1 (22,3%)

Cluster 2 (33,7%)

Training Group
CU
keyword
studente
59
cercare
43
corso
43
teoria
32
lezione
21
modalità
21
20
organizzazione
intervento
19
relazione
17

Clinical Training
keyword
CU
psicologia
94
lavoro
81
clinico
54
insegnare
36
contesto
29
problema
27
intervento
27
diverso
25
conoscenza
22

modello

16 interno

Cluster 3 (20,2%)
Institutional
Obligations
keyword
CU
scuola
29
persona
28
laurea
19
università
18
trovare
17
specializzazione
16
importante
16
entrare
15
14
scegliere
percorso
14

22

CU = context units classified in the cluster.

Cluster 4 (23,8%)
Student Orientation
keyword
CU
domanda
42
idea
40
33
organizzazione
aggiungere
32
processo
30
rispetto
29
orientare
21
parlare
21
Corso di laurea
20
Attività
18
didattiche

162

JADT’ 18

The four clusters of both corpuses are of different sizes (tables 1 and 2) and
reflect the representations of the professional training (table 3 and 4).
Regarding the teachers’ corpus (table 3), the first cluster represents the group
of students as a tool to teach professional skills, focusing on the group
process where relational dynamics are experienced; the second cluster
focuses on clinical training, teaching skills marketable in the job market; the
third cluster focuses on the teachers’ institutional obligations regardless of
the students’ training needs; and the fourth cluster represents students’
orientation as a way to support students in managing their academic training
regardless of professional skills. As regards the students’ corpus (table 4), in
the first cluster the good training involves students’ adherence to lesson tasks
regardless of critical thinking on the theoretical model proposed; in the
second cluster, learning professional skills is strictly connected to the ability
to get and respond to market demand; the third cluster reflects the relevance
of belonging to a group of colleagues supporting the construction of a
professional identity that, unfortunately, seems unconnected to professional
skills development; and the fourth cluster represents professional training as
a process in which the degree achievement is the main goal, regardless of the
job market demand.
Table 4  Students’ Cluster (the percentage of context units classified in the cluster is reported
between brackets)
Cluster 1 (27,6%)
Idealized Product
CU
keyword
esperienza
116
triennale
44
percorso
43
professione
41
università
37
possibilità
35
capire
33
diverso
31
senso
30
vivere
25

Cluster 2 (20,8%)
Professional Education
keyword
CU
pensare
89
esame
71
psicologia
65
seguire
55
realtà
55
vedere
55
iniziare
53
triennale
53
lavoro
44
interessante
44

Cluster 3 (26,3)
Group Identity
keyword
CU
scelta
154
studiare
153
frequentare
104
rapporto
102
piacere
98
colleghi
97
parlare
74
organizzare
68
domanda
55
aggiungere
36

Cluster 4 (25,3%)
Empty Degree
keyword
CU
vivere
26
trovare
85
tesi
20
sentire
91
riuscire
30
prendere
33
persone
105
maniera
23
livello
35
laboratorio
18

CU = context units classified in the cluster.

Students and teachers seem to have similar representations of the training
process: the academic need of building a network, highlighted by the
students’ cluster on group identity, and the teachers’ cluster on training group
and student orientation; the relevance of achieving a qualification, highlighted
by the students’ cluster on empty degree and the teachers’ cluster on
institutional obligation; and the development of professional skills marketable
in the job market reflected by the teachers’ cluster on clinical training and the

JADT’ 18

163

students’ cluster on professional education in line with what it was found by
Carli and colleagues (2004) and Paniccia and colleagues (2009) by means of a
similar methodology, the emotional textual analysis (Carli et al., 2016). The
awareness of the psychological demand of the labour market is an indicator
of the professional training process’s effectiveness. Nevertheless, students
and teachers split the academic achievement from the development of
professional skills. This could be a critical aspect, possibly explaining young
graduates’ difficulty in entering the job market, focusing more on academic
context rather than on market demand. As a consequence, during the
training process, students do not develop the connection between
professional training (what they are learning) and professional skills (what
they are going to do in the future).
5. Conclusion
Although the study results could not be generalized, due to the participants’
selection criteria and the methodology we used, they highlight professional
training representation characteristics, which are the elements influencing the
rate of unemployment among young psychologists. Even though it is not
possible to quantify the relevance of the characteristics of the representation,
the emotional text mining, allowing for the identification of the words
association explanatory of the education representation, allows for
hypotheses definition and the identification of the resources and the issues
pertaining the professional training in a specific context.
The interpretation of the text mining results lets the social unconscious
process emerge, setting the education useful to defining the type of
psychological intervention able to support the representation transformation
toward a more effective training process. In this particular case study, the
intervention would aim to develop the connection between professional
qualification achievement and the professional skills development, which are
currently split.
References
Anpal Servizi (2017), L’inserimento occupazionale dei laureati in psicologia,
dell’università La Sapienza di Roma, Direzione e studi analisi statistica - SAS.
Carli R. (1990). Il processo di collusione nelle rappresentazioni sociali. Rivista
di Psicologia Clinica, 4: 282-296.
Carli R., Dolcetti F. and Dolcetti (2004). L’Analisi Emozionale del Testo
(AET): un caso di verifica nella formazione professionale. In Purnelle G.,
Fairon C. and Dister A., editors, Actes JADT 2004: 7es Journées
internationales d’Analyse statistique des Données Textuelles, pp. 250-261.
Carli R., Paniccia R.M., Giovagnoli F., Carbone A. and Bucci F. (2016).

164

JADT’ 18

Emotional Textual Analysis. In L. A. Jason and D. S. Glenwick, editors,
Handbook of methodological approaches to community-based research:
Qualitative, quantitative, and mixed methods. Oxford University Press.
Cordella B., Greco F. and Raso A. (2014). Lavorare con Corpus di Piccole
Dimensioni in Psicologia Clinica: Una Proposta per la Preparazione e
l’Analisi dei Dati. In Nee E., Daube M., Valette M. and Fleury S., editors,
Actes JADT 2014 (12es Journées internationales d’Analyse Statistque des
Données Textuelles, Paris, France), pp. 173-184.
Fornari F. (1976). Simbolo e codice: Dal processo psicoanalitico all’analisi
istituzionale. Feltrinelli.
Greco F. (2016). Integrare la disabilità. Una metodologia interdisciplinare per
leggere il cambiamento culturale. Franco Angeli.
ISTAT (2017). Rapporto annuale 2017. ISTAT
Lancia F. (2017). User’s Manual : Tools for text analysis. T-Lab version Plus 2017.
Lebart L. and Salem A. (1994). Statistique Textuelle. Dunod
Matte Blanco I. (1981). L’inconscio come insiemi infiniti: Saggio sulla bi-logica.
Einaudi
McKinsey & Company (2014). Studio ergo Lavoro, come facilitare la
transizione scuola lavoro per ridurre in modo strutturale la
disoccupazione giovanile in italia. Report di Ricerca "Studio ergo Lavoro",
McKinsey & Company,
https://www.mckinsey.it/file/2785/download?token=a3VfesjU.
Moscovici S. (2005). Le rappresentazioni sociali. Il Mulino.
Paniccia R.M., Giovagnoli F., Giuliano S., Terenzi V., Bonavita V., Bucci F.,
Dolcetti F., Scalabrella F. and Carli R. (2009). Cultura Locale e
soddisfazione degli studenti di psicologia. Una indagine sul corso di
laurea “intervento clinico” alla Facoltà di Psicologia 1 dell’Università di
Roma “Sapienza”. Rivista di Psicologia Clinica, Supplemento n. 1: 1-49.
Salvatore S. and Freda M. F. (2011). Affect, unconscious and sensemaking: A
psychodynamic, semiotic and dialogic model. New Ideas, Psychology, Vol.
29, pp. 119–135.
Savaresi S. M. and Boley D. L. (2004). A comparative analysis on the bisecting
K-means and the PDDP clustering algorithms. Intelligent Data Analysis
8(4): 345-362.

JADT’ 18

165

Profiling Elena Ferrante: a Look Beyond Novels
Michele A. Cortelazzo1, George K. Mikros2, Arjuna Tuzzi3
2

1University of Padova – cortmic@unipd.it
National and Kapodistrian University of Athens – gmikros@isll.uoa.gr
3University of Padova – arjuna.tuzzi@unipd.it

Abstract
Elena Ferrante represents rather a peculiar editorial and journalistic
phenomenon: Today, she enjoys a wide international audience, though, on
the other hand, there is surprisingly little scientific literature that discusses
her works. Since Elena Ferrante is a pseudonym for an anonymous writer,
some investigators have already dealt with the pursuit of her real identity
and, at the moment, the main suspects that emerged are Domenico Starnone,
Marcella Marmo and Anita Raja. Corpora collected in order to analyze Elena
Ferrante's works and compare them with the works of other authors are
usually composed of novels, however Marcella Marmo and Anita Raja are
not novelists and their works are not ascribed to genres comparable with
novels. One of Elena Ferrante's books, La Frantumaglia, is useful to collect
corpora of texts of different genres (letters, essays, interviews, etc.) and they
might include texts by authors that have never been taken into consideration
in research studies based on novelists. Nevertheless, these texts raise specific
questions that concern their exploitability in traditional authorship
attribution procedures due to their limited size. This study aims at working
on a corpus of texts other than novels by means of a machine learning
approach, in the frame of methods for authorship attribution and profiling.
Riassunto
Elena Ferrante costituisce un fenomeno editoriale e giornalistico italiano
molto particolare: attualmente gode di grande visibilità internazionale ma,
allo stesso tempo, c'è sorprendentemente poca letteratura scientifica che si
occupa delle sue opere. Siccome Elena Ferrante è lo pseudonimo di un/una
autore/autrice ancora anonimo/anonima, alcuni si sono già confrontati con la
ricerca della sua vera identità e i maggiori sospettati emersi, finora, sono
Domenico Starnone, Marcella Marmo e Anita Raja. I corpora che vengono
utilizzati per studiare la produzione di Elena Ferrante e confrontarla con
quella di altri autori sono costituiti normalmente da romanzi ma Anita Raja e
Marcella Marmo non sono scrittrici e i loro lavori non si possono ascrivere a
generi confrontabili con i romanzi. Una delle opere di Elena Ferrante, La
frantumaglia, può essere utilizzata per costituire corpora con testi di generi

166

JADT’ 18

diversi (lettere, saggi, interviste, ecc.) che possono includere materiali di
autori non ancora considerati nelle ricerche basate su romanzieri. Tuttavia,
questi testi presentano specifiche problematiche legate alla ridotta
dimensione e parziale utilizzabilità con strumenti di attribuzione d'autore
tradizionali. Questo lavoro ha come obiettivo studiare un corpus di testi
diversi dai romanzi con un approccio machine learning nell'ambito dei
metodi per l'attribuzione d'autore e il profiling.
Keywords: authorship attribution, machine learning, profiling, stylometry,
support vector machine
1. Introduction
In previous works the novels signed by Elena Ferrante have already been
studied in the panorama of Italian contemporary literature and they have
displayed that this author has a peculiar writing style and shows relevant
individual traits. Moreover, in previous investigations the Italian writer that
showed the highest level of similarity with Elena Ferrante is Domenico
Starnone (Galella, 2005; 2006; Gatto, 2016; Cortelazzo et Tuzzi, 2017; Tuzzi et
Cortelazzo, 2018). In this study we aim at testing further hypothesis and look
at texts that are not ascribed to the genre "novels". In this way we have the
opportunity to consider for authorship attribution and profiling experiments
new candidates, i.e. writers that are not exclusively novelists. A first
reference can be made to Marcella Marmo and Anita Raja, two Italian
women, that have been suspected to be the hand that hides behind the penname of Elena Ferrante, respectively, by Marco Santagata (2016) and Claudio
Gatti (2016). The corpus collected for this new study has a specific focus on
three main suspects (Marcella Marmo, Anita Raja, Domenico Starnone) and
includes further suspected authors (Goffredo Fofi, Mario Martone, Valeria
Parrella, Francesco Piccolo), authors that in previous analysis showed some
common traits with Elena Ferrante's works (Gianrico Carofiglio, Clara
Sereni), authors that provocatively claimed to be Elena Ferrante (Laura
Buffoni) and members of the E/O publishing house (Sandro Ferri, Sandra
Ozzola and the editorial board that is supposed to be the collective editor of
the publishers' web pages).
2. Corpus
The corpus includes letters, interviews and further material written by
different authors (tab. 1) that can be compared with texts included in the
book La Frantumaglia by Elena Ferrante (2016). An innovative perspective has
been adopted for analyzing texts: a Machine Learning (ML) approach based
on a Support Vector Machine (SVM) method that takes into consideration 13
authors for a classical Authorship Attribution (AA) and different variables

JADT’ 18

167

(gender, age, geographical area) for profiling tasks.
The whole corpus adopted for this study is composed of 113 texts and
includes 143,695 word tokens and 19,020 word types. In the classical ML
perspective, the corpus is arranged into two groups: a "training set" and a
"testing set". The training corpus (tab. 1) includes 86 texts (87,458 word
tokens), 78 written by 12 authors and 8 by a collective subject (EO) that
represents the editorial staff of E/O publishing house. The corpus is balanced
in terms of gender and partly balanced for age and geographical area (tab. 2).
Information about gender and age is not available (n.a.) for E/O, as it is
presumed to be a group. The testing corpus includes 27 texts (6 essays, 7
interviews, 14 letters for a total of 56,237 word tokens in size) signed by Elena
Ferrante and collected in her book La Frantumaglia. Five texts are chapters of
the same large essay that has been written as an answer to Giuliana Olivero
and Camilla Valletti's questions (Ferrante 2016).
Table 1. Authors and categories of texts included in the training corpus
Authors
Category
texts
tokens
texts
Laura Buffoni
3
4,477
article
53
Gianrico
6
4,940
essay
9
Carofiglio
E/O
8
3,955
interview
12
Sandro Ferri
2
3,838
letter
4
Goffredo Fofi
9
7,378
web
8
Marcella
5
12,991
Marmo
Mario Martone
10
9,320
Sandra Ozzola
4
1,879
Valeria
7
4,676
Parrella
Francesco
6
5,529
Piccolo
Anita Raja
4
13,617
Clara Sereni
2
2,271
Domenico
20
12,587
Starnone
Tot
86
87,458
Tot
86

tokens
42,124
22,926
15,480
1,611
5,317

87,458

Since most stylometric measures and linguistic features are heavily
influenced from text size, we decided to split our texts into equal sized text
chunks. Both the training and the testing corpus were segmented into 200
words text chunks. After the chunking procedure, the training corpus
inflated from 86 texts to 386 chunks of 200 words in length and the testing

168

JADT’ 18

corpus from 27 texts to 259 chunks of 200 word tokens in length. This
enlargement had also the positive effect of making our sample space larger,
giving us the opportunity to use a wider spectrum of linguistic features.
Table 2. Descriptive variables of texts included in the training corpus
Gender
n.a.

Age

authors

texts

tokens

1

8

3,955

Naples Area

authors

texts

tokens

n.a.

1

8

3,955

authors

texts

tokens

f

6

25

39,911

>60old

7

46

54,561

Naples

6

52

58,720

m

6

53

43,592

60young

5

32

28,942

NoNaples

7

34

28,738

Tot

13

86

87,458

Tot

13

86

87,458

Tot

13

86

87,458

3. Method
In order to investigate our research aims, we developed a feature-rich
document representation model comprised by the following features groups:
1) Author Multilevel N-gram Profiles (AMNP): 1,500 features, 500 features
of each n-gram category (2-grams and 3-grams at the character level, and
2-grams at the word level);
2) Most Frequent Words in the corpus (MFW, 500 features).
The first feature group (AMNP) provides a robust document representation
which is language independent and able to capture various aspects of
stylistic textual information. It has been used effectively in authorship
attribution problems (Mikros et Perifanos, 2011; 2013) and gender
identification focused on bigger texts (e.g. blog posts, cfr. Mikros, 2013).
AMNP consists of increasing order n-grams in both character and word level.
Since character and word n-grams capture different linguistic entities and
function complementary, we constructed a combined profile of 2, 3
characters n-grams and 2 words n-grams. For each n-gram we calculated its
normalized frequency in the corpus and included the 500 most frequent
entries resulting in a combined vector of 1,500 features. The second feature
group (MFW) can be considered classic in the stylometric tradition and it is
based on the idea that the MFWs belong to the functional words class and are
beyond the conscious control of the author, thus revealing its stylometric
finger print. In this study we used the 500 most frequent words of the corpus.
The above described features have been exploited for training a classification
machine learning algorithm, Support Vector Machines (SVM, Vapnik, 1995),
in both a standard authorship classification task and in three different author
profiling tasks (author’s gender, age, and geographical area). SVM is
considered a state-of-the-art algorithm for text classification tasks. The SVM
constructs hyper-planes of the feature space in order to provide a linear
solution to the classification problem. For our trials we experimented with

JADT’ 18

169

various kernels and we ended up choosing the polynomial one as this was
the most accurate in our dataset. All statistical models developed have been
evaluated using 10-fold cross validation (90% training set – 10% testing set)
and the accuracies reported represent the mean of the accuracies obtained in
each fold. Since the feature space was sparse, we eliminated all features that
showed a variance close to zero, using the two following rules: the
percentage of unique values was less than 20%, and the ratio of the most
frequent to the second most frequent value was greater than 20. The nearzero variance feature removal shrank the number of the employed features
and led to a reduction of 47.4% (from the initial 2,000 available features we
kept 1,052 features).
4. Results
4.1. Authorship Attribution Results
For the standard authorship classification task (tab. 3), first we worked with
the whole corpus as training dataset and obtained an accuracy of 0.7098 on
average (71%). Among the set of 13 candidates included in the corpus, a large
share of testing text chunks resulted attributed to Domenico Starnone (32%),
Anita Raja (21%) and Mario Martone (21%).
Table 3. Attribution of text chunks included in the testing corpora (whole and reduced corpus)

Authors
Starnone
Raja
Martone
E/O
Buffoni
Parrella
Fofi
Carofiglio
Ferri
Marmo
Piccolo
Ozzola
Tot

whole corpus
No. chunks
84
55
55
18
16
15
7
2
2
2
3
0
259

%
32%
21%
21%
7%
6%
6%
3%
1%
1%
1%
1%
0%
100%

reduced corpus
Authors
No. chunks
Starnone
115
Raja
73
Martone
39
E/O enlarged
32

Tot

259

%
44%
28%
15%
12%

100%

170

JADT’ 18

Table 4. Cross-classification matrix in authorship attribution task (whole and reduced corpus)
reduced corpus
whole corpus
Starnone
Raja
Martone
E/O enlarged
Tot
77
Starnone
2
0
5
84
48
Raja
3
0
4
55
30
Martone
14
2
9
55
15
E/O
1
2
0
18
Buffoni
6
5
2
3
16
Parrella
8
7
0
0
15
Fofi
4
3
0
0
7
Piccolo
2
0
0
1
3
Carofiglio
0
2
0
0
2
2
Ferri
0
0
0
2
Marmo
0
2
0
0
2
Ozzola
0
0
0
0
0
Tot
115
73
32
39
259

We deemed useful to reduce the candidates to Starnone, Raja, Martone and
rearrange the E/O collective author into a new enlarged version of the E/O
group, i.e. we pool together all the members of the E/O publishing house
(Sandro Ferri, Sandra Ozzola and the E/O staff). As an effect of this selection
we obtained an improvement in the performance of the ML algorithm (+13%)
since the accuracy rose up to 0.8408 on average (84%). With reference to this
reduced version of the training corpus, that includes only four candidates,
again most text chunks seem to belong to Domenico Starnone (44%) and
Anita Raja (28%). From a cross comparison of the results achieved (tab. 4)
with the whole and reduced versions of the training corpus we observed that
the text chunks of the testing corpus that have been attributed to Domenico
Starnone and Anita Raja proved more stable and consistent if compared to a
more unstable and weak role of Mario Martone. The existence of an action of
the publishing house was confirmed in both versions, although in some cases
a confusion of the E/O editors with Starnone and Raja's hands is somewhat
visible.
4.2. Profiling Results
Results achieved with profiling tasks are more schematic since the algorithm
is called to work with simpler dichotomous variables (tab. 5).
With respect to gender, the ML algorithm obtained an accuracy of 0.8000 on
average (80%) and the results achieved with the automatic classification of
the text chunks of the testing corpus suggested that among the fragments of
La Frantumaglia we might have different hands: at least a man (54%) and a
woman (46%). If compared with the case of gender profiling, the ML

JADT’ 18

171

algorithm achieved a similar performance in terms of accuracy for both the
classification by age (0.8027, 80%) and geographical area (0.7850, 78%) but for
the most part the text chunks appeared to be written by an old author (76%)
from Naples (90%).

f
m
Tot

Table 5. Profiling of text chunks included in the testing corpus
Gender
Age
Naples area
No.
%
No.
%
No.
chunks
chunks
chunks
141
54% >60 old
197
76% Naples
233
118
46% 60
62
24% NoNaples
26
young
259
100% Tot
259
100% Tot
259

%
90%
10%
100%

5. Discussion and conclusions
Among limitations and constrains of this method, first and foremost we have
to take into account that we have different genres among the texts of this
corpus (essays, interviews, newspapers articles, letters) and this feature
surely affects our results. Texts show similarities when they are written by
the same author or belong to the same text genre and these two effects are
not easy to disentangle in our text corpus. Secondly, when the SVM
prediction is called to assign testing chunks to authors and/or categories it
always leads to an attribution that is the result of a formula generated by the
ML algorithm (in other words it never answers "do not know"). Results
depend both on quality of texts and basket of opportunities offered during
the training phase. As a consequence, we have to refer to the accuracy of the
model and consider the classification as the best attribution among options
given by the set of reasonable candidates and available categories. Thirdly, La
Frantumaglia represents an interesting set of texts signed by Elena Ferrante
that are not ascribed to the genre "novels" and it enables new analyses to
compare and contrast the author's writing style with the one of authors that
are not strictly novelists. Nevertheless, we cannot be sure that all texts
included in La Frantumaglia are written by the same hand and, moreover, we
do not know whether these texts are written by the author that actually wrote
also the novels signed by Elena Ferrante. From the authorship attribution
viewpoint more than one hand emerged as likely and we can formulate some
hypothesis. If we take into account only main suspected authors mentioned
in our Introduction, Domenico Starnone and Anita Raja are confirmed; on the
contrary, Marcella Marmo seems not believable. Mario Martone's role is an
interesting suggestion since similarities of chunks taken from La Frantumaglia
with his texts might be the indirect outcome of direct interactions between
Martone and Ferrante (e.g. letters and interviews where they are both

172

JADT’ 18

speaking about the movie L'amore molesto). Also the E/O staff's role is
engaging as it is easy to imagine the effect on the writing style of one or more
editors that work as proofreaders, copyreaders and ghostwriters when Elena
Ferrante has to answer many interviews and letters collected by the
publishing house. From profiling experiments a composite picture of La
Frantumaglia emerges. The procedure reveals the existence of different hands
once more, suggested the involvement of at least a man and a woman, and
draws the portray of an author (single or collective) from Naples that is over
60 years old.
Does the mystery about Elena Ferrante's work remain a mystery?
Acknowledgements
We thank Arianna Menin for providing us with the corpus of texts of La
Frantumaglia collected for her first level (B.A.) 3-years degree thesis in
Communication (University of Padova, a.y. 2016/2017, supervisor prof.ssa
Arjuna Tuzzi).
References
Cortelazzo M.A. and Tuzzi A. (2017). Sulle tracce di Elena Ferrante: questioni
di metodo e primi risultati. In Palumbo, G. (ed), Testi, corpora, confronti
interlinguistici: approcci qualitativi e quantitativi, EUT – Edizioni Università
di Trieste, pp. 11-25.
Ferrante, E. (2016). La Frantumaglia. Roma: E/O.
Galella, L. (2005). Ferrante-Starnone. Un amore molesto in via Gemito, La
Stampa, 16 January 2005, pp. 27.
Galella, L. (2006). Ferrante è Starnone. Parola di computer. L'Unità, 23
November 2006.
Gatti, C. (2016). Elena Ferrante, le «tracce» dell'autrice identificata, Il Sole 24
Ore – Domenica, 2 October 2016, pp. 1-2.
Gatto, S. (2016). Una biografia, due autofiction. Ferrante-Starnone: cancellare
le tracce, Lo Specchio di carta. Osservatorio sul romanzo italiano
contemporaneo, 22 October 2016. www.lospecchiodicarta.it
Mikros, G.K. (2013). Authorship Attribution and Gender Identification in
Greek Blogs. In Obradović, I., Kelih, E. and Köhler R. (eds.), Selected papers
of the VIIIth International Conference on Quantitative Linguistics (QUALICO)
in Belgrade, Serbia, April 16-19, 2012, Belgrade: Academic Mind, pp. 21-32.
Mikros, G.K. and Perifanos, K. (2011). Authorship identification in large
email collections: Experiments using features that belong to different
linguistic levels Proceedings of PAN 2011 Lab, Uncovering Plagiarism,
Authorship, and Social Software Misuse held in conjunction with the CLEF 2011
Conference on Multilingual and Multimodal Information Access Evaluation, 19-

JADT’ 18

173

22 September 2011, Amsterdam.
Mikros, G.K. and Perifanos, K. (2013). Authorship attribution in Greek tweets
using multilevel author’s n-gram profiles. In Hovy, E., Markman, V.,
Martell, C. H. and Uthus D. (eds.), Papers from the 2013 AAAI Spring
Symposium "Analyzing Microtext", 25-27 March 2013, Stanford, California.
Palo Alto, California: AAAI Press, pp. 17-23.
Santagata M. (2016). Elena Ferrante è …, La lettura – Corriere della Sera, 13
March 2016, pp. 2-5.
Tuzzi, A. and Cortelazzo, M.A. (2018), What is Elena Ferrante? A
Comparative Analysis of a Secretive Bestselling Italian Writer, Digital
Scholarship in the Humanities (on line first version).
Vapnik, V. (1995). The nature of statistical learning theory. New York: SpringerVerlag.

174

JADT’ 18

Word Embeddings: a Powerful
Tool for Innovative Statistics at Istat
Fabrizio De Fausti1, Massimo De Cubellis1, Diego Zardetto1
1

ISTAT – Italian National Institute of Statistics
(defausti, decubell, zardetto)@istat.it

Abstract 1
In recent years, word embedding models have proven useful in many
Natural Language Processing problems. These models are generated by
unsupervised learning algorithms (like Word2Vec and GloVe) trained on
very large text corpora. Their main purpose is to map words to vectors of a
metric space in a very smart way, so that the resulting numeric
representation of input texts effectively captures and preserves a wide range
of semantic and syntactic relationships between words. In this paper we
discuss word embedding models generated from huge corpora of raw text in
Italian language, and we propose an original graph-based methodology to
explore, analyze and visualize the structure of the learned embedding spaces.
Abstract 2
Il lavoro illustra le potenzialità dei modelli Word Embedding nell’analisi di
grandi collezioni di dati testuali e propone un originale metodo basato sui
grafi per l’esplorazione della struttura semantica catturata dai modelli.
Keywords: Word Embeddings, Word2Vec, Graphs, Text Summarization,
Italian Tweets, NLP.
1. Introduction
Word embedding models represent a powerful tool that can be used as input
for subsequent machine learning tasks, like text classification, topic modeling
and document similarity. This work shows how we built, tested and used
word embedding models (based on the Word2Vec algorithm, see Section 2.1)
to achieve the following objectives:
 Istat is currently collecting streaming Twitter data on a large scale. Word
embedding models helped us devise domain-specific ‘filters’, namely sets
of keywords that we used to filter out off-topic tweets with respect to the
intended statistical production goal. Here we will show the case of the so-

JADT’ 18

175

called “Europe filter”, meant to measure people’s mood about the
European Union.
 Istat is currently exploiting textual data automatically scraped from the
websites of Italian enterprises in order to predict whether or not they
perform e-commerce. Given the huge corpus of noisy and unstructured
texts derived from this web-scraping procedure, word embedding models
allowed us: (i) to automatically create an “e-commerce pseudo-ontology”
and to smartly summarize the input texts, (ii) to encode the summarized
texts into a rich numeric representation in order to feed a Deep Learning
classifier.
2. Methodology
In recent years, new successful algorithms for natural language modeling
have been proposed, based on Neural Networks (e.g. Word2Vec and Glove).
These algorithms, starting from very large corpora of raw text, are able to
create models that map words to low-dimensional vector spaces, called word
embeddings (Mikolov et al., 2013a). Although these algorithms do not rely on
any linguistic domain-knowledge, nor on handcrafted syntactic and semantic
relationships between words, they are surprisingly able to learn both of them
from raw data. Indeed, words that are strongly related from a syntactic
and/or semantic point of view are mapped to vectors that are almost parallel
to each other; conversely, words that are syntactically and/or semantically
loosely related are mapped to nearly perpendicular vectors. Moreover, these
models perform amazingly well when it comes to solving analogies between
words, just like a human would do. For example, if one asks a trained word
embedding model «which word X completes the analogy: [ ‘Paris’ : ‘France’ =
‘Madrid’ : X ]», the answer will very likely be X = ‘Spain’. We mention here
only one type of relationship (capital-nation), but word embedding models
are able to capture a wide variety of relationships, such as: male-female,
singular-plural, superlative-comparative, synonym-antonym, politicianparty, etc.
2.1 Word2Vec
Word2Vec (Mikolov et al., 2013b) is one of the most influential word
embedding algorithms. It consists of a neural network trained to solve a
predictive problem according to one of the following two approaches:
predicting the central word given the other words of a context (Cbow), or
predicting the words of the context given the central word (Skipgram). At the
end of the training the predictive ability of the network is not used; instead,

176

JADT’ 18

its internal structure (weights of the network) is exploited to represent the
coordinates of each word of the dictionary in the embedding space.
While a large text corpus is the main input to Word2Vec, the algorithm
allows also for several hyperparameters which can be tuned to improve the
quality of the learned model. Some scholars (e.g. Levy et al., 2015) consider
these hyperparameters as key points to understand Word2Vec’s superiority
as compared to previous language modeling techniques.
The main hyperparameters of Word2Vec are:
 Embedding space dimension: the dimension of the vector space to which the
words of the corpus are mapped;
 Window size: the width of the sliding window used to process the corpus.
It defines how large the context is;
 Iteration: how many times the weights of the neural network are updated
during training;
 Learning model: the approach used to train the neural network, either
Cbow or Skipgram.
Of course, further factors affect the performance of a Word2Vec model:
 Size of the corpus: bigger corpora perform better than small ones;
 Quality of the corpus: very noisy, fragmented and poorly curated texts
generally produce lower quality embedding spaces.
At the end of the training phase, the quality of the learned word embedding
model can be assessed through standard test functions. Classical examples
are the word-similarity and the word-analogy functions (see e.g. Pennington et
al., 2014).
2.2 Exploring and visualizing big embedding models through graphs
As sketched in Section 2, word embedding algorithms transform words into
vectors of a low-dimensional metric space. The dimension of this numeric
space is usually set to values in the range 100-300 (see e.g. Mikolov et al.,
2013a). When input corpora are huge, taking into account inflected forms of
words, the output embedding model can contain hundreds of thousands of
vectors. As a consequence, the full structure of the embedding model is very
hard to analyze. Exploration and visualization of such models requires to (i)
reduce the dimensionality of the embedding space, and to (ii) focus on just a
subset of vectors, namely those derived by the most relevant words for the
analysis at hand. While traditional solutions exist for the first task, like PCA
and t-SNE (van der Maaten, Hinton, 2008), no standard methods are
available for the second one. We propose here a new technique, based on
graphs (Gibbons, 1985), that simultaneously addresses both needs. It selects

JADT’ 18

177

just a subset of relevant words, adopting a clever filtering criterion based on
their semantic proximity, and allows visualizing the resulting sub-model in a
two-dimensional graph.
2.3 Building the graphs
a

Given a “node” vector/word v in the embedding space, let’s define

. To build
, we connect v to its W nearest
base graph of width
vectors/words in the embedding space (the cosine distance is used). The base
graph

will thus have W + 1 nodes. Node v can be either the image of

an actual word , i.e.

, or the vector resulting from the sum of multiple

and , i.e.
. The idea is that, within the
words, say
embedding space, the sum of word vectors can be exploited to disambiguate
the meaning of polysemous words. An example is provided in Table 1, where
the 5 closest words to the vector V(‘rome’) are reported on the left panel, and
the 5 closest words to the vector V(‘rome’) + V(‘colosseum’) + V(‘ancient’) are
reported in the right panel. Evidently, the addition of words ‘colosseum’ and
‘ancient’ to the polysemous word ‘rome’ moves the semantic area explored
by the base graph

from a geographical to an historical sense.

Table 1. Word disambiguation by sum of vectors: the polysemous word is ‘rome’.
Closets 5 Words from
V(rome)

Cosine
Similarity

Closest 5 Words from
Cosine
V(rome) + V(colosseum) + V(ancient) Similarity

turin
palermo
naples
milan
bologna

0.6818
0.6377
0.6212
0.6129
0.5857

roman
archeological
pompei
trastevere
trajan

0.5822
0.5318
0.5250
0.5217
0.5189

Our approach builds a full output graph by iteratively combining N base
. We devised three different methods to combine base graphs
graphs
according to different exploration strategies. We called these methods
Geometric, Linear and Geometric-Oriented: the corresponding pseudo-codes
are provided in Table 2. Besides the width parameter W and the number of
iterations N, all the three methods require as input a set of seed words [seeds]
to define the starting point for the exploration of the embedding model.

178

JADT’ 18
Table 2 Pseudo codes of the proposed graph generation methods. Function find_leaves()
returns all the nodes with zero outdegree; function shortestPath() calculates
the shortest path between two nodes.

Geometric ([seeds], N, W) Linear ([seeds], N, W)

Geometric-Oriented ([seeds], N, W)

v = V(seed1) + V(seed2) +
…
G_w(v)
for iteration in [1, …, N]:
for leaf in
find_leaves():
G_w(V(leaf))

v = V(seed1) + V(seed2) + …
G_w(v)
for iteration in [1, …, N]:
for leaf in find_leaves()
virtualNode_leaf = 0
addEdge(leaf,
virtualNode_leaf)
for node in shortestPath(v,
leaf):
virtualNode_leaf =
virtualNode_leaf + node
G_W(virtualNode_leaf)

v = V(seed1) + V(seed2) + …
G_w(v)
for i in [1, …, N]:
virtualNode_i = 0
for leaf in find_leaves():
addEdge(leaf,
virtualNode_i)
virtualNode_i =
virtualNode_i + V(leaf)
G_w(virtualNode_i)

As will be shown in Section 3, the Geometric method tends to expand the
exploration range very quickly, rapidly losing the initial semantic focus
provided by the seed words; the Linear method stays much more focused,
but explores just a narrow sub-model; the Geometric-Oriented method
provides a satisfactory compromise between the previous two methods.
3. Application
3.1 Building word embedding models on large corpora of Italian tweets
Istat is currently collecting streaming Twitter data on a large scale. Italian
tweets are captured provided that they pass at least one active ‘filter’. Filters
are simply sets of keywords deemed to be relevant for specific statistical
production goals. For instance, the ‘Social Mood on Economy’ filter involves
60 keywords borrowed from the questionnaire of the Italian Consumer
Confidence Survey, and collects about 40,000 tweets per day.
We used a large collection of about 100 million Italian tweets to train
Word2Vec with different settings of hyperparameters, therefore generating
different embedding models. We subsequently analyzed the obtained models
and tested their quality as discussed in Section 3.1.2. This way we managed
to identify the best performing set of hyperparameters to be used for the
applications described in Sections 3.2 and 3.3.
3.1.1 Process
The data processing pipeline we implemented consists of the following steps:
 Collection of Italian tweets through Twitter’s streaming API as JSON files;

JADT’ 18

179

 Parsing of JSON files and storage of the tweets in a relational database;
 Extraction from the database of the textual content of about 100 million
tweets and export to a raw text file (corpus);
 Preprocessing of the raw text (text cleaning and normalization);
 Setting of Word2Vec hyperparameters;
 Training of Word2Vec on the tweets’ corpus;
 Test of the learned word embedding model.
3.1.2 Benchmark and selection of the best hyperparameters
With the aim of identifying the best hyperparameters, we customized
benchmark word-analogy tests contributed by the Stanford University
(Pennington et al., 2014), translating them in Italian and adding new word
analogies involving specific terms of the Economics field. Note that our tests
involved many groups of analogies, encoding a wide range of different
relationships between words, of both the syntactic and the semantic kind. As
a measure of model goodness, we adopted the so called “Top-1 accuracy”
criterion. According to this criterion an analogy [a : b = c : x] is successfully
solved by the learned model if and only if the closest (i.e. Top-1) embedding
vector to V(c) - V(a) + V(b) is exactly V(x). We evaluated against our
customized word-analogy tests many output models generated by diverse
settings of hyperparameters, and eventually found the following optimal
values: embedding space dimension = 200, window size = 8, iteration = 15, learning
model = Cbow.
3.2 Design of the “Europe” filter
As already mentioned in Section 3.1, Istat collects only Italian tweets that
match at least one active filter. So far, the keywords defining the filters have
been designed by subject-matter experts. In this section, instead, we illustrate
how word embedding models can be exploited to automatically develop new
filters in a data-driven way. The idea is to leverage our graph-based
exploration methodology to select the best keywords, starting from few
relevant seed words. In particular, on the occasion of the 16th anniversary of
the Treaties of Rome, our objective was to capture the sentiment of Italian
Twitter users about European Union. In Figures 1 and Figure 2 we show the
graphs resulting from the Geometric-Oriented and Geometric methods
respectively. Note that both graphs were generated using the same seed
words, namely: ‘europa’, ‘ue’, ‘bruxelles’, ‘europea’, ‘unione’, ‘euro’. The
Geometric-Oriented graph appears more compact and the words are indeed
closely related to the semantic area of the seed words. The Geometric graph,

180

JADT’ 18

instead, finds many more words, which are clearly grouped in coherent
clusters and represent a valuable semantic enrichment with respect to the
original seeds. Given its richness, this second graph has been considered by
subject-matter experts as a very good candidate to play the role of “Europe”
filter.

Figure 1: Geometric-Oriented
([‘europa’, ‘ue’, ‘bruxelles’, ‘europea’, ‘unione’, ‘euro’], 8, 8)
3.3 Text Summarization and Encoding
One ongoing Istat’s Big Data project aims at exploiting textual data
automatically scraped from the websites of Italian enterprises in order to
predict whether or not they perform e-commerce. To address this task, Deep
Learning techniques are being used. Since input scraped texts are huge and
Deep Learning algorithms are computationally intensive, a preliminary text
summarization step is in order. Besides increasing efficiency, the
summarization algorithm should hopefully improve accuracy by reducing
the signal-to-noise ratio of input data. Word embedding models allowed us
to achieve this goal with a purely data-driven approach.
To guide the summarization, we leveraged word embeddings trained on the
whole web-scraped corpus. We used the Linear-graph illustrated in Figure 3
to select a set of marker words with high discriminative power for the
detection of e-commerce, adopting as initial seeds the words: ‘carrello’,
‘shopping’, ‘online’. (These marker words constitute what we called an “ecommerce pseudo-ontology” in the Introduction.) To summarize the texts,
only input sentences containing marker words have been retained. This way,
we obtained a 92.2% reduction of the original noisy text, along with a
substantial improvement in the performance of the Deep Learning classifier
(+20%, as compared to marker words defined by subject-matter experts).
Lastly, we relied again on word embeddings to encode the summarized texts
and feed the Deep Learning classifier. Once more, our experiments show that

JADT’ 18

181

word embedding models outperform more traditional text encoding
approaches, like bag-of-words.

Figure 2: Geometric([‘europa’, ‘ue’, ‘bruxelles’, ‘europea’, ‘unione’, ‘euro’], 3, 8)

Figure 3: Linear ([‘shopping’, ‘online’, ‘carrello’], 11, 8)
4. Conclusions
The techniques for dealing with large corpora of texts can greatly benefit
from recent technology advancements. Word Embeddings are an example of
this opportunity. Extensive evidence shows that Word Embedding models
are indeed superior to more traditional text encoding methods like, e.g., bagof-words. Ongoing works on textual Big Data at Istat make extensive use of
these new approaches with very promising results.
References
Mikolov T., Yih W., Zweig G. (2013a). Linguistic Regularities in Continuous
Space Word Representations. Proceedings of NAACL-HLT 2013, pp. 746751.
Mikolov T., Chen K., Corrado G., Dean J. (2013b). Efficient Estimation of
Word Representations in Vector Space. CoRR abs/1301.3781.

182

JADT’ 18

Levy O., Goldberg Y., Dagan I. (2015). Improving Distributional Similarity
with Lessons Learned from Word Embeddings. Trans. of the Association
for Computational Linguistics, vol.(3): 211-225.
Pennington J., Socher R., Manning C.D. (2014). GloVe: Global Vectors for
Word Representation. Proceedings of EMNLP 2014, pp. 1532-1543.
van der Maaten L.J.P. and Hinton G.E. (2008). Visualizing High-Dimensional
Data Using t-SNE. Journal of Machine Learning Research, vol(9): 2579-2605.
Gibbons A. (1985). Algorithmic Graph Theory. Cambridge University Press.

JADT’ 18

183

Analisi di dati d’impresa disponibili online: un
esempio di data science tratto dalla realtà economica
dei siti di e-commerce
Viviana De Giorgi, Chiara Gnesi
Istat – degiorgi@istat.it; gnesi@istat.it

Abstract
This work describes the process of extracting, organising and analysing
detailed information on firms that trade electronic equipment on the
Alibaba.com site. The first part concerns how translating unstructured
information into variables organised in a statistical database by using
dimensional classes, indices, indicators and classifications. A companyproduct matching is realised by encoding a textual variable with an
international classification, and an automated analysis is applied in order to
explore, describe and analyse the corpus retrieved from the Internet. In the
second part a descriptive and econometric analysis shows how demographic
and economic information on enterprises from Alibaba.com are very
significant for competitiveness on the foreign market.
Keywords: encoding, classification, textual analysis, regression model.
Sommario
Il presente lavoro consiste nello sviluppo di un modello che consenta di
trattare, organizzare ed analizzare informazioni dettagliate sulle imprese che
commerciano apparecchiature elettroniche sul portale Alibaba.com.
La prima parte riguarda il processo di trasformazione dell’informazione
destrutturata in variabili organizzate in un database statistico attraverso l’uso
di classi dimensionali, indici, indicatori e classificazioni. Si è realizzato un
abbinamento impresa-prodotto utilizzando una classificazione internazionale
attraverso la codifica di una variabile testuale, su cui è applicata un’analisi
automatizzata al fine di esplorare, descrivere e analizzare il corpus testuale
tratto da Internet. Nella seconda parte è svolta un’analisi descrittiva ed
econometrica, i cui risultati mostrano la presenza sul portale cinese di
informazioni demografiche ed economiche sulle imprese altamente
significative per la competitività sul mercato estero.
Parole chiave: codifica, classificazione, analisi testuale, regressione.

184

JADT’ 18

1. Introduzione
Questo lavoro nasce dagli spunti di riflessione e studio offerti nel corso delle
lezioni di un Master universitario in Data Science1 e si rivolge in particolare
alle tecniche di trattamento, gestione ed analisi dei dati provenienti da fonti
recuperabili on line2 e fruibili in maniera gratuita. L’approccio adottato è
quello della singola impresa che vuole migliorare la propria competitività nel
mercato di riferimento, analizzando i dati generati dai processi aziendali nel
settore in cui è presente o mira a posizionarsi. A tal fine sono preziose le
informazioni dettagliate e aggiornate sui volumi prodotti, transazioni,
struttura e demografia delle imprese concorrenti, presenti nei siti di
commercio elettronico. Il presente lavoro è stato sviluppato utilizzando i dati
estratti attraverso un’intensa attività di web scraping dal portale Alibaba.com,
con riferimento alle imprese operanti nel settore delle apparecchiature
elettroniche.
2. Dai dati destrutturati alle variabili statistiche: costruzione del database
Nel processo di trasformazione dell’informazione destrutturata acquisita
online in variabili statistiche, un ruolo centrale riveste laclassificazione delle
imprese a partire dal principale prodotto commercializzato. La variabile
testuale – che corrisponde alla descrizione non codificata del prodotto
commercializzato dalla società – è stata codificata secondo una classificazione
di attività economica standardizzata a livello internazionale. Si è scelto
l’elenco Prodcom con riferimento alle divisioni 26, 27 e 28, per un totale di
989 sottocategorie di prodotti3.
L’attribuzione del codice Prodcom alla singola impresa è stata effettuata
implementando un sistema di codifica ad hoc4 strutturato in step successivi.
La fase iniziale consiste nella normalizzazione dei testi attraverso lo sviluppo

Master universitario in Data Science, Università Tor Vergata, Dipartimento di
Ingegneria dell’impresa "Mario Lucertini", anno accademico 2015/2016. Si ringraziano
Francesco Borrelli, Valentina Talucci e Domenica Fioredistella Iezzi per gli utili
suggerimenti.
2 L’acquisizione dei dati è stata effettuata nell’arco temporale che va dal 26
novembre 2016 al 7 gennaio 2017 dalla dott. Antonella Miele attraverso una attività di
web scraping. I dati utilizzati sono relativi a 2.349 imprese presenti sul sito
Alibaba.com e operanti nel settore delle apparecchiature elettroniche.
3http://ec.europa.eu/eurostat/ramon/nomenclatures/index.cfm?TargetUrl=LST_C
LS_DLD&StrNom=PRD_2011&StrLanguageCode=EN&StrLayoutCode=HIERARCHI
C#
4Non avendo a disposizione software già sviluppati utilizzabili, è stato
implementato un sistema di codifica ad hoc utilizzando il software SAS.
1

JADT’ 18

185

di un parser applicato alla variabile testuale e alle descrizioni della
classificazione utilizzata. Successivamente si è realizzato un matching tra i
due campi, attraverso un algoritmo che identifica l’abbinamento tra stringhe,
sfruttando il dizionario al massimo livello di dettaglio possibile5. Infine si è
realizzato l’abbinamento impresa-prodotto, assegnando a ciascuna impresa
un codice Prodcom che identifica univocamente il principale prodotto
commercializzato6. Il sistema di codifica ha permesso la classificazione del
95% delle imprese: un 30% circa vende “computer e prodotti di elettronica e
ottica, apparecchi elettromedicali, apparecchi di misurazione e orologi”, un
quarto vende “apparecchiature elettroniche e apparecchiature per
usodomestico non elettriche” e il 40% circa vende “apparecchiature elettriche
diverse dalle precedenti” (Tavola 1).
Tavola 1: Imprese per divisioni Prodcom, valori assoluti e percentuali
Divisione prodcom
n
26 – computer e prodotti di elettronica e ottica
718
27 – apparecchiature elettroniche e apparecchiature
per uso domestico non elettriche
618
28 – fabbricazione di macchinari ed apparecchiature n.c.a
893
non classificati
120
Totale complessivo
2349

%
30,6
26,3
38,0
5,1
100,0

L’analisi dei residui ha rivelato che la causa principale del mancato
abbinamento deriva dalla presenza sul mercato di Alibaba di prodotti,
elettrici e non, altamente specializzati ovvero sulla frontiera della tecnologia,
non presenti nella Prodcom. Tuttavia, l’abbondanza di acronimi,
abbreviazioni, slang hanno reso l’attività di standardizzazione
particolarmente complessa.
In seguito alla codifica della variabile testuale, si è proceduto a una sua
analisi automatizzata al fine di esplorare, descrivere e analizzare il corpus

In questa fase si è utilizzato il dizionario al massimo livello di dettaglio
possibile – 8 digit – in modo da abbracciare la descrizione del maggior numero di
prodotti possibile. L’abbinamento prodotto/dizionario si è realizzato per molte
sottocategorie di Prodcom; dopo aver analizzato i risultati ottenuti, si è scelto di
utilizzare i 4 digit come il massimo livello di disaggregazione compatibile con una
soglia di accuratezza ritenuta accettabile.
6L’assegnazione del codice è stata realizzata attribuendo all’impresa il codice
Prodcom corrispondente alla classe in cui si è realizzato in maggior numero di match
prodotto – dizionario, pesata per la frequenza più alta riscontrata in una determinata
categoria di prodotto.
5

186

JADT’ 18

tratto da Internet. L’analisi testuale7 consente di esplorare la struttura del
testo sia come corpus – raccolta di frammenti testuali fra loro confrontabili –
sia in relazione alla codifica ad esso attribuita. A tal fine, si è utilizzato
TaLTaC2, particolarmente adattoallo studio di informazioni testuali non
strutturate di grandi dimensioni e di informazioni strutturate a queste ultime
collegate. Un primo approfondimento è offerto dalle misure lessicometriche,
che consistono in una serie di misure e di indici statistici calcolati sul
vocabolario e sulle sue classi di frequenza (Bolasco, 1999).Il corpusè costituito
da 25.295 occorrenze, che corrispondono al numero totale di forme grafiche
intese come unità di conto(Giuliano, 2004). L’ampiezza del vocabolario, pari
a 4.363 forme grafiche distinte, riflettela specificità settoriale a cui attiene
l’analisi. Coerentemente, l’indice di estensione lessicale percentuale, pari a
17,2, e l’indice di Guiraud normalizzato, pari a 27,4,confermano come la
dimensione del vocabolario sia affetta da un bias determinato dalla specificità
delle imprese analizzate. Tuttavia, nel settore è presente una gamma di
prodotti piuttosto diversificata, come suggerito dal numero di hapax, pari a
50,2 (tavola 2).
Tavola 2: misure lessicometriche sul corpus
Misure lessicometriche
Occorrenze - N
Forme grafiche distinte - V
Type/Token (V/N)*100
% di Hapax (V1/V)*100
Frequenza media generale - N/V
G di Guiraud - V/sqrN
Coefficiente a

Valori
25.395
4.363
17,2
50,7
5,8
27,4
1,2

L’analisi lessicale, svolta a partire dall’analisi delle specificità, ha consentito
di verificare, all’interno di singole classi, la rilevanza dei prodotti attraverso
la sovra o sotto rappresentazione rispetto alla classificazione internazionale.
L’utilizzo del dizionario della Prodcomcome risorsa statistica-linguistica
esterna, ha permessoanalisi in parallelo. In effetti, l’indice Term Frequency
Inverse Document Frequency (TFIDF) calcolato anche sul dizionario, ha
consentito di evidenziare le caratteristiche peculiari dei prodotti venduti
dalle imprese rispetto al panorama delle stesse che commercializzano
prodotti elettronici.Inoltre, attraverso il confronto tra le forme grafiche del

7A tal fine, si è utilizzato TaLTaC2, un software per l’analisi automatica del testo
nella duplice logica di Text Analysis e di Text Mining (TM), quindi sia come analisi
del testo che come recupero e estrazione di informazione all’interno dello stesso

JADT’ 18

187

corpus e quelle del dizionario della Prodcom, si è potuto operare un controllo
indiretto sulla qualità della codifica di cui al precedente paragrafo
utilizzando lo scarto standardizzato come proxy di significatività8. Tale
misura consente, infine, di caratterizzare le imprese rispetto alla peculiarità
dei prodotti che le contraddistinguono all’interno del settore di riferimento
(figura 1).

Figura 1: Parole chiave del corpus in base allo scarto standardizzato

Ulteriori elaborazioni sui dati reperiti dal sito hanno consentito la creazione
di ulteriori variabili statistiche. Tra queste: tenure – una proxy dell’anzianità
dell’impresa, costruita a partire dall’anno di iscrizione al portale; addetti e
fatturato medi – a partire dal valore medio delle classi di riferimento; qualità
– una variabile dummy che segnala la presenza di una certificazione di
prodotto; propensione all’export – come quota percentuale di esportazioni
sul fatturato ; ricerca e sviluppo – in termini di addetti medi impiegati nelle
attività innovative; efficienza – capacità di risposta dell’impresa alle esigenze
dei clienti. Il database finale è costituito da 18 variabili, che afferiscono
all’Anagrafica dell’impresa, all’Attività economica, al Commercio estero, alla
Dimensione economica, alla Competitività e alla Ricerca & Sviluppo.
3. Analisi descrittiva ed ecometrica dei dati
Ai dati descritti precedentemente sono state applicate le tecniche largamente
adottate della ricerca statistica: un’analisi descrittiva del collettivo di
riferimento, un’analisi multivariata di tipo esplorativo per la ricerca delle
variabili da utilizzare in un modello econometrico e un modello di
regressione che tenga conto della specificità dei dati9. Si riportano di seguito i
principali risultati.

Si è utilizzata la formula classicadellamisura di specificità in cui fi* è la
frequenzarelativadella forma graficanell’elencoProdcom.
9Le informazionisulleimpresepresentisulsitovengonoaggiornate, anche se non si
sa bene quando e come, e l’informazione dell’anno di riferimento è presente talvolta e
solo per alcune variabili (per esempio il fatturato)
8

188

JADT’ 18

Per tutti i settori di attività, più della metà delle imprese si dichiara
produttrice e venditrice, forse perché tale caratteristica tende a essere un
parametro di scelta da parte di chi deve acquistare. Sono per lo più imprese
medio-grandi, giovani, che in genere interagiscono con i clienti, con alte
percentuali di export sul valore del fatturato, con presenza di dipendenti
dedicati alla ricerca e sviluppo, disponibilità del certificato dei prodotti
venduti. Cumulano un volume di esportazioni maggiore dell’80% le imprese
che hanno più di 50 dipendenti, oppure sono nelle classi più elevate di
fatturato, oppure rispondono almeno all’80% di richieste dal sito, o infine si
dichiarano produttrici dei prodotti venduti. L’analisi condotta, e quindi il
modello di regressione studiato, riguarda la dipendenza che il volume di
esportazioni ha con le variabili presenti nel data set. Al fine della scelta delle
variabili da utilizzare nel modello è stata effettuata un’analisi cluster
gerarchica (SAS Institute Inc., 1999), scegliendo la variabile con minimo
valore di 1-R2ratio10, e individuando le seguenti variabili: la produttività
d’impresa, la variabile dimensionale data dal numero dei dipendenti
occupati in ricerca e sviluppo e le tre variabili categoriche percentuale di
risposta a richieste, attività economica e tipologia d’impresa. Le prime due
risultano avere nel proprio cluster, nella suddivisione in 5 gruppi, il valore
minimo di 1-R2ratio; tra le variabili categoriche invece si evidenziano quelle
aventi minore correlazione own cluster con le altre variabili. Il modello
implementato consente di stimare i valori della variabile dipendente
“volume delle esportazioni” sulla base dei valori assunti/osservati da/per
alcune variabili indipendenti. Anche come conseguenza dei risultati
dell’analisi cluster descritta precedentemente, si è scelto di includere tra
queste: il fatturato per dipendente, il numero di dipendenti di ciascuna
impresa, la tipologia di prodotto a 2 cifre, la percentuale di risposta alle
richieste di possibili acquirenti, la quota di dipendenti d’impresa occupati in
ricerca e sviluppo, la tipologia di impresa e il numero di anni di attività.
È stato stimato il seguente modello di regressione lineare (Rencher e Schaalje,
2008):
+
dip+
dove: (a)

è il logaritmo naturale del volume di esportazioni; (b)

101-R^2
ratio=(1-R^2
own
cluster)/(1-R^2
nextclosest),
dove
own
cluster=correlazione con il proprio gruppo di variabilie nextclosest=correlazione con il
gruppo più vicino

+

JADT’ 18

189

è il logaritmo della produttività; (c)
della percentuale di risposta; (d)

è il logaritmo

è il logaritmo della quota di

è la tipologia di impresa, (f)
dipendenti occupati in ricerca e sviluppo; (e)
ate è la tipologia di prodotto, (g) dip è il numero dei dipendenti.
In presenza di una variabile dipendente con distribuzione log-normale11,
l’applicazione di una trasformazione logaritmica alla variabile dipendente e
alle variabili indipendenti continue ha come primo obiettivo di ottenere una
distribuzione assomigliante a quella di una normale. Ciò implica, per i
modelli lineari, la possibilità di estensione di tale ipotesi distributiva anche ai
residui (ε) del modello e quindi consente di condurre in modo corretto i
necessari test di significatività sui coefficienti stimati. Inoltre, la
contemporanea trasformazione logaritmica delle variabili indipendenti
(continue) consente di interpretare i valori dei coefficienti stimati
direttamente in termini di elasticità. L’introduzione della variabile dip2 è utile
per verificare l’esistenza di eventuali relazioni non lineari tra dip e la
dipendente, ovvero per capire se all’aumento del numero di dipendenti
corrisponda
una
crescita
delle
esportazioni
progressivamente
superiore/inferiore. È stato inoltre studiato un secondo modello (modello2)
introducendo l’interazione tra la quota di dipendenti occupati nella ricerca e
sviluppo e la variabile categoriale relativa alla tipologia d’impresa. Tale scelta
è coerente con l’idea che il livello di attività in ricerca e sviluppo possa
rappresentare una fonte di valore aggiunto maggiore per le imprese che
producono rispetto a quelle che vendono soltanto. I risultati ottenuti e
riportati nella tavola 3 vengono di seguito descritti: (1) la relazione tra la
variabile dipendente e la misura di produttività utilizzata è
significativamente positiva; a una variazione dell’1% del fatturato per
addetto corrisponde, mediamente, un variazione di oltre l’1% del volume
delle esportazioni; (2) queste sono correlate positivamente anche con la
percentuale di risposta a richieste dal sito e con il numero di anni di attività
dell’impresa (coefficienti sempre significativi); (3) la stima dei due coefficienti
relativi alla dimensione d’impresa evidenziano che questa accresce (come era
logico aspettarsi) il volume delle esportazioni, ma con tassi progressivamente
decrescenti all’aumentare del numero dei dipendenti (rendimenti decrescenti

11

La variabile aleatoria

segue la distribuzione logaritmica

solo se
segue la distribuzione normale
densità di probabilità è f(x)=e^(-〖(lnx-μ)〗^2/〖2σ〗^2)/(x√2πσ)

. La sua funzione di

190

JADT’ 18

di scala); (4) sembrano esistere effetti differenziali tra il volume di
esportazioni e le tipologie di prodotti venduti per settore di attività
economica, ma non sempre i coefficienti sono significativi; (5) le dummy
relative alla tipologia d’impresa mostrano coefficienti sempre non
significativamente diversi da zero in assenza di interazione con la proxy di
ricerca e sviluppo (modello1); (6) se fatte interagire (modello2) emerge invece
come le due tipologie impresa produttrice e produttrice/venditrice abbiano
un effetto positivo sulle esportazioni (rispetto alla modalità di riferimento
impresa solo venditrice) e l’intensità di ricerca e sviluppo sembra accrescere
significativamente le esportazioni solo per il settore delle imprese produttrici;
(7) la variabile in oggetto risulta infatti correlata negativamente con la
dipendente nei casi di imprese operanti esclusivamente nel settore del
commercio e positivamente per quelle manifatturiere o contemporaneamente
anche venditrici.
Tavola 3: Stima dei parametri del modello lineare (modello 1 e modello 2)
nel data set iniziale e nel data set integrato
Variabile
ln(fattxdip)
Resp
num_anni
ate26 (rif,)
at 27
ate28
Others
Dip
dip^2
type venditrice (rif,)
produttrice
produttrice/venditrice
ln(dip_in_rd/dip) x venditrice
ln(dip_in_rd/dip)) x produttrice
ln(dip_in_rd/dip) x produttrice/venditrice
Costante
N
r2_ajusted

modello1
1,024***
0,002***
0,022***

modello2
1,025***
0,002***
0,023***

-0,067*
-0,094***
0,067
0,012***
-0,001***

-0,061
-0,085**
0,063
0,012***
-0,001***

-0,008
-0,055
-0,097***

0,358***
0,161*
-0,212***
0,188***
0,125***
2,084***
1.913
0,866

2,291***
1.913
0,865

*p<0,1; **p<0,05; ***p<0,01

Le funzioni di densità della variabile dipendente osservata e stimata
mostrano entrambe una forma distributiva approssimativamente normale:
non emergono significative differenze tra i due modelli, ce forniscono
entrambi una buona approssimazione.

JADT’ 18

191

Riferimenti bibliografici
Bolasco S. (1999). L’analisi multidimensionale dei dati, Roma, Carocci.
Giuliano L. (2004), L’analisi automatica dei dati testuali. Software e istruzioni per
l’uso, Milano, LED.
Rencher A.C, Schaalje G.B. (2008). LinearModels in Statistics. Second Edition.
Wiley.
SAS Institute Inc. (1999), LogisticRegressionModeling Course Notes, Cary, NC:
SAS Institute Inc., pages 56-57.

192

JADT’ 18

The use of textual sources in Istat: an overview
Alessandro Capezzuoli, Francesca della Ratta, Stefania Macchia,
Manuela Murgia, Monica Scannapieco, Diego Zardetto1
ISTAT – Istituto Nazionale di Statistica – nome.cognome@istat.it

Abstract 1
Text Mining techniques allow a more widespread use of textual materials
also in Official Statistics. We show implementations and current pilots
realized in Istat, with a focus on both techniques and applications. Initially,
text mining techniques were used to manage complex taxonomies or conduct
open question analysis, while at the moment Big data frameworks allow to
expand the different sources of data also to merge several data sources and to
reduce response burden.
Abstract 2
Le tecniche di Text Mining consentono un ampio utilizzo di dati testuali
anche nella Statistica Ufficiale. Sono descritte le implementazioni e le
sperimentazioni realizzate in Istat in questo ambito, focalizzando sulle
tecniche utilizzate e le applicazioni realizzate. Inizialmente il Text Mining
veniva effettuato per gestire le tassonomie o effettuare analisi testuale delle
riposte aperte, mentre più di recente il contesto dei Big data ha consentito di
ampliare le fonti utilizzate e di integrarle tra loro anche in funzione del
contenimento del response burden.
Keywords: text mining, official statistics, sentiment analysis
1. Automatic coding and semantic search of taxonomies
The first use of text mining techniques in Italian official statistics was
finalized to manage complex classifications. Indeed, classifications are
defined, which consist of structured lists of concepts, mutually exclusive,
corresponding to codes that allow to produce a partition of the population.
When the identification of the code corresponding to the concept does not
present any ambiguity, it is possible to use closed questions with lists of
items among which the one matching with the response is selected.

1 This work comes from a common effort; paragraph 1.1 is written by Manuela
Murgia and Stefania Macchia, par. 1.2 by Alessandro Capezzuoli; par. 2 by Francesca
della Ratta, par. 3 by Monica Scannapieco and Diego Zardetto.

JADT’ 18

193

On the other hand, when codes belong to classifications that are complex in
terms of structure, criteria and hierarchies, then the management of
taxonomies is a very difficult task that implies the knowledge of the
classification. Let us think, for example, of the classification of Occupation: in
order to identify the code corresponding to each occupation it is necessary to
consider different aspects, like the level of competences, their scope or the
activities managed. In this paragraph, it is described how, with the evolution
of technologies, this activity has been performed in different ways, using
different software tools.
1.1. Automatic coding
Up to some years ago, statistics survey questionnaires rarely used open
questions allowing textual answers because of the difficulties in processing
them in order to provide a measure of the phenomenon. On the other hand,
this could not often be avoided for some variables, like occupation, economic
activity, education level that have necessarily to be coded according to
official classifications for either national or cross-national data comparison.
In the past, verbal responses were manually coded, but this was very timeconsuming, costly and error prone, especially for large amount of data
(Macchia et Murgia, 2002). For this reason Istat decided to adopt automated
coding systems that consist of two main parts: i) a database (dictionary) and ii)
a matching algorithm. The dictionary is made of texts associated with
numeric codes. Codes are those of official classifications and represent the
possible values to be assigned to the verbal responses entering the coding
process, while texts are the textual labels expressing the concepts that the
classifications associate to codes. In order to improve the coding results,
dictionaries are enriched with common language descriptions, resulting from
answers to previous surveys. The matching algorithm is a ‘weighting
algorithm’ that assigns a weight to each word of the verbal response to be
coded. The weight indicates how much a word is informative and it depends
on the word’s frequency inside the dictionary: the higher its frequency the
minor its weight. Then the algorithm compares the input response with all
the texts inside the dictionary looking for a perfect match. If no exact match is
found then it looks for a partial match with the most “similar” description,
choosing the one with the highest weight.
The efficiency of the automated coding systems allowed Istat to use them not
only to code responses of statistical surveys, but also to offer the coding
service to a larger public such as governmental or private institutions, private
citizens, who need to associate free text descriptions to official classifications
codes, let’s think, for instance, to businesses which have to identify their
economic activity code for declarations to Chambers of Commerce. The

194

JADT’ 18

coding service was then made available on the Istat web site for the ATECO
(the Italian version for Nace, the Economic Activity classification) variable.
The software used for many years was ACTR (1998-2015) developed and
distributed by Statistics Canada. In 2015 ACTR was not working anymore on
the new Istat IT platform and it was substituted by CIRCE that behaves like
ACTR but it is developed in house and based on R (Murgia et al., 2016). The
choice of R made it possible to create a coding package freely downloadable
from the website and also to offer a web service for the coding of the ATECO.
The web service can be easily incorporated in any other software
applications: electronic questionnaires of Istat surveys or in software systems
of external organizations.
1.2 Semantic search within taxonomies
The evolution of technology allowed to explore also other software solutions
suitable to represent the Statistical classifications logical structure, described
within the Generic Statistical Information Model (GSIM). To this end, it was
possible to exploit a very simple JSON object, to which then associate the
metadata related to the classification (family, series, level, etc.). PUT and GET
methods, related to the HTTP protocol, permit an easy acquisition of
classification items that can then be organized through ad hoc procedures, on
the basis of GSIM model, and stored into a relational database.
Being a JavaScript Object Notation, JSON is the natural environment for the
construction of web applications using programming languages like e.g.
Ajax/JavaScript combined with ad hoc frameworks as appropriate.
Elasticsearch and Solr are the main frameworks used to search and share
data. In particular, Elasticsearch provides a set of powerful and complete
tools/plugins for data dissemination and the use of REST resources.
Elasticsearch is well suited for the solution of some critical issues related to
the use of statistical classifications in different fields (surveys, administrative
registers, information systems, etc.), such as:
• acquisition, storage, management and updates of classifications;
• multilingual semantic search for coding;
• sharing and dissemination of coding tools.
Textual search is a very popular technique for users who seek information on
the web. It does not require any special skill and users have already acquired
through surfing the web and it is also suitable to search within statistical
classifications and facilitate coding. The most common problem related to
semantic searches within taxonomies concerns false-positive and falsenegative results. The search is usually done through SQL queries allowing
users to perform two types of operations: "exact match" and "full text". String
parsing algorithms can be associated to the SQL queries.

JADT’ 18

195

A statistical classification can be indexed within Elasticsearch to perform
complex and differentiated textual searches through DSL (Domain Specific
Language) in JSON format. This solution permits to simplify the formulation
of complicated SQL queries and makes the search system from any
programming language usable. Elasticsearch allows users to manipulate
large volumes of data thanks to an internal document management,
completely independent from relational databases, and the opportunity to
create distributed cluster.
Istat experience in using this methodology has been very satisfactory. The
coding systems related to the main statistical classifications (ISCO, NACE,
ISCED, COFOG, COICOP) were included in several Istat surveys ("Labour
Force Survey", multi-purpose survey "Aspects of daily life", "Consumer
prices", etc.) and Information system on occupation. Easy to use, widgets
have been developed to include coding systems within web questionnaires
and web applications.
2. Open questions analysis
Social research uses open questions also when category answers are not
known or when researchers prefer to explore interviewees’ different points of
view using their own categories. This approach offers a great opportunity to
realize analysis in depth, but it is difficult to be applied with the largest
sample used in official statistics. So it is generally preferred using open
questions only in pilot survey or small samples, to explore the possible list of
answers and to obtain the closed-end list for the final survey. As an example,
Istat used this approach in a survey on the female participation in
parliamentary life: in 2000 an open question was introduced in a quarterly
Multipurpose survey and the list of answering categories obtained with
textual analysis was used in the 2005 annual Multipurpose survey.
However, in the early 2000s Text mining tools made it possible to analyse
open questions also when codes does not belong to pre-defined
classifications. The first example was introduced in Istat by Sergio Bolasco,
who analysed the daily diaries collected in 2002-2003 Time use Survey to
obtain a classification of some daily life actions (Bolasco et al., 2007). This
classification was obtained using the Entity Research by Regular expression
(RE) inserted in the tool Taltac2, a function that represents a very important
turning point for the use of textual data in statistical surveys, because it made
possible to pass from the simple description of words contained in a corpus
(Lexical analysis) to the classification of single records on the basis of

196

JADT’ 18

words that are contained in each of them (Textual analysis2). The single
word is no more the unit of analysis as the RE function searches or counts
within the entire record a particular word or a combination of words, putting
the result in a new customized variable.
This function was afterwards used in other Istat surveys. First it was used in
the Survey on Occupations, developed in 2005-2006 and aimed at describing
Italian labour market occupations, providing detailed information on each
Occupational Unit. Researchers were interested also in tasks in which
workers are daily involved, which was asked through an open question:
“What does your job consist of? Which are the activities you are involved in during
your working day?”. Our aim was to provide each Occupational Units with a
list of semi-standardized activities, labelling in the same way similar
activities expressed in different ways by respondents. So, we used a strategy
of text categorization adding in final dataset an extra column variable with a
synthesis of the activities stated by interviewees: the final result was a list of
over 7,000 specific activities (della Ratta, 2009).
A similar approach is currently used to check and correct the coding of
economic activity carried out by interviewers in the Labour Force Survey:
every quarter, 1500 records out of 24000 responses collected in the survey
referred to specific Nace section are analyzed. The correctness of the codes
assigned is verified from a double perspective: not only by comparing
respondents’ vocabulary reported in the response field of the question on
economic activity with the specific dictionary of the official classification
(Nace rev-2), but also considering other extra information connected with
this variable collected in the same survey questionnaire. The process is
completed with a thorough examination of data consistency in each session,
to validate the corrections made and to assign the definitive proper code. At
the end errors are transmitted to interviewers during specific training
sessions in order to improve the all process of data collection, from the
interview to the coding assignment (della Ratta et Tibaldi, 2014).
Other uses of Text Mining tools regarded the classification of open questions
of the online survey on the dimensions of well-being (della Ratta et Tinto,

The search for the textual information is run by complex queries using regular
expressions with Boolean operators (AND, OR, NOT), lexeme reductions (wildcards
as “*” and “?”, e.g. contact* and customer? ) and distances (LAGgxx) between
consecutive words, that allow to identify different expressions used to convey the
same concept (contact*LAG3 customer? is able to identify series such as “to contact
the customer”, “contacts with customers”, “I contact my main customers”; the value
of the new variable could be “to contact customers”).
2

JADT’ 18

197

2012), or the analysis of residual answers inserted in single questions
(“Others”, please specify) that can improve the exhaustiveness of
questionnaires and can be used in training activity for interviewers.
In conclusion, the availability of Text Mining tools made it possible to
process open questions independently by the size of the text, being free in
this way to use un-structured data in official statistics, especially in recursive
analysis in which text categorization strategies can be repeated several times.
3. Dealing with Textual Big Data
Since recent years, in line with European-level strategic directives, Istat has
been exploring the potential of Big Data sources for Official Statistics. Many
of such sources – and notably those that seem the most promising so far – are
made up of huge collections of unstructured and noisy texts.
In current Istat’s projects, two types of unstructured sources were taken into
account, namely: (i) textual data collected from the websites of Italian
companies, obtained through automatic procedures of access and extraction
performed on a large scale (hundreds of thousands of sites); (ii) messages in
Italian language publicly available on Social Networks, typically collected in
streaming after a preliminary selection step performed using ‘filters’ (i.e. sets
of keywords that a message must match to be deemed relevant).
The contexts of use of textual data from company websites include the
enrichment of information in statistical business registers and the potential
replacement of questions from surveys questionnaires. The possible uses of
data from Social Network mainly concern the production of high-frequency
(e.g. daily) sentiment indices.
At the moment the experiments with Social Networks data focused on the
Twitter platform and on the development of “specific” sentiment indices: the
goal is to measure the Italian mood about topics or aspects of life that might
be relevant for Official Statistics (like the economic situation, the European
Union, the migrants’ phenomenon, the terrorist threat, and so on). The hope
is that such sentiment indices can improve the quality of Istat’s economic
forecasting models, enrich existing statistical products (for example the BES)
or create new statistical outputs in their own right.
Among the processing techniques used for these sources, a particularly
promising type consists of the Word Embedding models. These models are
generated by unsupervised learning algorithms (such as Word2Vec (Mikolov
et al., 2013) and GloVe (Pennington et al., 2014), both based on neural
networks) trained on large collections of text documents. Their main
objective is to map natural language words into vectors of a metric space, in
such a way that the numerical representation of texts captures and preserves
a wide range of syntactic and semantic relationship existing between words.

198

JADT’ 18

Istat successfully tested Word Embedding models in both the application
scenarios sketched above. In the first scenario, Word Embeddings have been
exploited to automatically summarize the huge text corpora scraped from
company websites, and to subsequently encode the summarized texts in
order to feed a Deep Learning algorithm for downstream analysis (e.g. to
predict whether a given enterprise performs e-commerce). In the second
scenario, Word Embedding models have been leveraged both to design the
‘filters’ used to select relevant messages from Twitter and to evaluate the
actual performance of the same ‘filters’ after data collection.
In the following of this section a specific focus will be provided on data
scraped from enterprises websites3. The Istat sampling survey on Information
and Communication Technologies (ICT) in enterprises aims at producing
information on the use of Internet and other networks by Italian enterprises
for various purposes (e-commerce, e-skills, e-business, social media, egovernment, etc.). In 2013, an Istat project started with the purpose of
studying the possibility to estimate some indicators produced by the survey
directly from the websites of the enterprises; these indicators included online
sale rate, social media presence rate and job advertisement rate. The idea was
to use web scraping techniques, associated, in the estimation phase, to text
and data mining algorithms, with the aim of replacing traditional
instruments of data collection and estimation, or to combine them in an
integrated approach (Barcaroli et al., 2015). The recently achieved results are
very encouraging with respect to the use of such techniques (Barcaroli et al.,
2017).
The whole pipeline that has been set up for this project includes:

A scraping activity performed by an ad-hoc developed software
(RootJuice4).

A storage step in which scraped data are stored in a NoSQL database,
i.e. Apache Solr.

A data preparation and text encoding step, performed in two different
ways:
1. tokenization, word filtering, lemmatization, generation of a termdocument matrix
2. word filtering and word embeddings.

An analysis step, performed via machine learning methods on each of
the text encodings resulting from the previous step.

3 A more detailed focus on the processing of Twitter data is presented in the
paper “Word Embeddings: a powerful tool for innovative statistics at Istat”,
submitted to this conference.
4 Available on GitHub : https://github.com/SummaIstat/RootJuice/.

JADT’ 18

199

4. Conclusions and remarks
The techniques for dealing with large corpora of texts can greatly benefit
from recent technology advancements. Word Embeddings are an example of
this opportunity, giving additional possibilities to use un-structured data in
official statistics for the purpose of integrating analyses or reducing response
burden. Extensive evidence shows that Word Embedding models are indeed
superior to more traditional text encoding methods like, e.g., bag-of-words.
Ongoing works on textual Big Data at Istat make extensive use of these new
approaches with very promising results.
References
Barcaroli G., Nurra A., Salamone S., Scannapieco M., Scarnò M.and Summa
D. (2015). Internet as Data Source in the Istat Survey on ICT in
Enterprises. Journal of Austrian Statistics, vol. 44, n. 2.
Barcaroli G., Scannapieco and M. Summa D. (2017). Massive Web Scraping of
Enterprises Web Sites: Experiences and Solutions. 61st World Statistical
Congress, ISI.
Bolasco S., Pavone P., D’Avino E. (2007). Analisi dei diari giornalieri con
strumenti di statistica testuale e text mining. In: Romano. I tempi della vita
quotidiana, Istat, Roma, Argomenti, n. 32.
della Ratta Rinaldi F. (2009). Il trattamento dei dati, in F. Gallo, P. Scalisi, C.
Scarnera. L’indagine sulle professioni. Anno 2007, Contenuti, metodologia e
organizzazione. Collana Metodi e Norme, n. 42, Roma, Istat.
della Ratta-Rinaldi F.and Tinto A. (2012). Le opinioni dei cittadini sulle
misure del benessere. Risultati della consultazione online. Roma, IstatCnel.
della Ratta-Rinaldi F. and Tibaldi M. (2014). Sperimentazione di un sistema
di controllo e correzione per la codifica dell’attività economica. Istat
Working Paper, n. 4, 2014.
Macchia S. and Murgia M. (2002). Coding of textual responses: various issues
on automated coding and computer assisted coding. Proc. of JADT 2002:
6es Journées Internationales d’Analyse Statistique des Données Textuelles.
Mikolov T., Chen K., Corrado G. and Dean J. (2013). Efficient Estimation of
Word Representations in Vector Space. Proceedings of Workshop at ICLR.
Murgia M. and Prigiobbe V. (2016). La nuova applicazione di codifica web
dell’ATECO 2007: WITCH, un web service basato sul sistema di codifica
CIRCE. Istat Working Papers n. 19.
Pennington J., Socher R. and Manning C. D. (2014). Glove: Global Vectors for
Word Representation. Proceedings of the 2014 Conference on Empirical
Methods in Natural Language Processing, 1532–1543.

200

JADT’ 18

Twitter e la statistica ufficiale:
il dibattito sul mercato del lavoro
Francesca della Ratta, Gabriella Fazzi, Maria Elena Pontecorvo,
Carlo Vaccari, Antonino Virgillito1
Istat – Istituto Nazionale di Statistica, Rome – Italy

Abstract
The goal of the paper is to show the potential and the benefits of the
integration between the big data analysis techniques and techniques used
for the textual analysis, through the analysis of a corpus extracted from
Twitter. The analysis is the development of a method already
experimented in other works (della Ratta, Pontecorvo, Virgillito, Vaccari,
2016 and 2017), in which we started from the collection of selected tweets
through a list of hashtags defined according to the theme of interest. This
procedure allows to obtain in a reasonable time a selection of tweets of
interest, on which to apply textual analysis techniques to describe the
contents of the text and to identify its main semantic contents. The paper
analyzes the role of the National Institute of Statistics in the discussion on the
labor market in the periods when ISTAT spreads the monthly and quarterly
press releases on employment. The analysis, already conducted at the end of
2016, has been replicated and refined in the same period of 2017, in order to
show the distinctive elements of the labor market debate and to understand
the changes in the perception of public opinion, also taking into account the
changes in terms of the economic situation and the political scenario.
Key words: big data; text mining; twitter; Istat, labour market
1. Big data e Twitter
I dati provenienti dai Social Network sono una delle sorgenti di Big Data più
utilizzate dai ricercatori: l’enorme diffusione di questi siti web, nei quali gli
utenti generano grandi quantità di informazioni, li rende potenzialmente una
delle fonti più interessanti anche per i dati testuali. Twitter è un Social
Network nel quale gli utenti scrivono e leggono corti messaggi chiamati

1 Questo lavoro è frutto della riflessione condivisa degli autori; il paragrafo 1 è
stato redatto da Carlo Vaccari e Antonino Virgillito, il paragrafo 2.1 da Francesca
della Ratta, il 2.2 da Gabriella Fazzi e Maria Elena Pontecorvo, le conclusioni da tutti
gli autori.

JADT’ 18

201

“tweet”, normalmente visibili da tutti gli utenti, che possono anche
“iscriversi” ai tweet di altri utenti (diventando “follower”), inoltrare
(“retweet”) singoli tweet ai propri followers o aggiungere “mi piace” ad altri
tweet. Twitter è oggi uno dei Social Network più diffusi, e ha superato nel
2017 i 300 milioni di utenti attivi. Secondo Alexa (2018) Twitter è oggi il
tredicesimo sito più visitato al mondo, l’ottavo negli USA. Scopo di questo
lavoro è applicare le tecniche dell’analisi testuale a un corpus estratto da
Twitter, unendo i due mondi dei Big Data e dell’Analisi Testuale. La raccolta
dei dati da Twitter è stata effettuata utilizzando una piattaforma, la
“Sandbox”2, che è il risultato finale del progetto “Big Data in Official
Statistics”, portato avanti nell’ambito dell’High Level Group on
Modernisation of Official Statistics (HLG-MOS). La Sandbox è un ambiente
web-based utilizzato per numerosi esperimenti basati su diverse sorgenti dati
come le visite alle pagine di Wikipedia, i dati sul Commercio Estero del sito
Comtrade dell’ONU, i siti delle imprese per ricercare annunci di lavoro e,
appunto, i tweet raccolti in varie nazioni del mondo. La Sandbox è oggi
ancora utilizzata per portare avanti le sperimentazioni della ESSnet on Big
Data3, un progetto europeo coordinato da Eurostat per l’utilizzo dei Big Data
nella produzione di statistiche ufficiali. I tweet analizzati sono stati raccolti
attraverso uno strumento online messo a disposizione gratuitamente da
Twitter (Streaming API), interrogato attraverso programmi scritti in R ed
eseguiti all’interno della Sandbox. Questa soluzione, per quanto semplice da
utilizzare e di immediata implementazione, presenta limitazioni sia per
l’ammontare dei dati che possono essere estratti, sia per la non completa
aderenza dei dati ottenuti rispetto ai filtri impostati in fase di estrazione,
come spiegato nella sezione successiva. I tweet acquisiti sono stati
memorizzati su Elasticsearch, un database installato nella Sandbox
specializzato in dati semi-strutturati, che permette di memorizzare grandi
quantità di documenti ed estrarre velocemente dei sottoinsiemi attraverso
query basate su parole chiave.
2. L’analisi dei post sul mercato del lavoro: l’impatto dell’Istat
2.1 Creazione del corpus
Per analizzare i dati estratti da Twitter si è replicato il metodo testato in
occasione di precedenti lavori (della Ratta, Pontecorvo, Virgillito, Vaccari;
2016 e 2017). Si è deciso, in questo contesto, di focalizzare l’analisi sul ruolo

2 I risultati del progetto Sandbox, coordinato da Virgillito nel 2014 e da Virgillito
e Vaccari nel 2015, sono illustrati in Unece (2014 e 2016).
3

https://webgate.ec.europa.eu/fpfis/mwikis/essnetbigdata/index.php/ESSnet_Big_Data

202

JADT’ 18

ricoperto dall’Istat nella diffusione delle informazioni sulla tematica del
lavoro, estraendo automaticamente un primo set di tweet nelle settimane in
cui l’Istat diffonde i dati mensili e trimestrali sul mercato del lavoro. Tale
estrazione, già effettuata a fine 2016, è stata replicata nello stesso periodo del
2017, partendo da una query piuttosto ampia4 che ha consentito di ottenere
un corpus di 58.277 tweet relativo al periodo 28 novembre-12 dicembre 2017.
Da questo corpus sono stati estratti tutti gli hashtag con occorrenza maggiore
di 14 (facilmente identificabili nel testo grazie alla presenza del simbolo #) tra
i quali sono stati individuati quelli strettamente connessi alla discussione sul
mercato del lavoro (Tabella 1). È quindi stato estratto, utilizzando il software
Taltac2, un corpus di 19.398 tweet contenente almeno uno degli hashtag di
interesse. Questo corpus è stato ulteriormente ripulito eliminando i tweet
relativi alle offerte di lavoro (presenza degli hashtag #offertalavoro
#annunciolavoro), considerati non pertinenti. Si è così arrivati a un corpus
composto da 17.419 tweet, composto da 283.000 occorrenze, 18.000 forme
grafiche e una ricchezza lessicale (rapporto type/token) del 6,7%.
Poco più di un terzo dei tweet sono originali, mentre il volume dei retweet
costituisce il 63% del corpus complessivo, in misura maggiore rispetto al
corpus del 2016. Per “misurare” l’impatto dell’Istat nel dibattito sul lavoro
sono stati etichettati tutti i tweet in cui compare la forma “Istat”: il 13,9% del
totale, una misura quasi triplicata rispetto a quanto osservato nel 2016 (5%).
Se da un lato nel 2016 l’impatto del concomitante dibattito referendario aveva
ridimensionato il peso del commento del dato Istat nella discussione sul
mercato del lavoro, nel 2017 i temi della ripresa occupazionale e delle sue
caratteristiche sembrano aver attirato maggiormente l’attenzione degli utenti.
Inoltre, la prima uscita del rapporto annuale integrato sul mercato del lavoro
ha probabilmente accresciuto il peso dei commenti sui dati.
Se nel 2016 la presenza dei riferimenti a Istat si addensava in corrispondenza
delle uscite ufficiali, nel 2017 è distribuita in maniera più uniforme, con un
picco in corrispondenza del comunicato trimestrale del 7 dicembre (nel quale

La query iniziale utilizzata è la seguente: "(istat OR inps OR #istat OR #inps OR
#lavoro OR #occupati OR #disoccupati OR #disoccupato OR #jobsact OR
#occupazione OR #disoccupazione OR #mercatodellavoro OR #poletti OR
#cassaintegrazione)". Sul primo corpus di 58.277 tweet estratto dall’API di Twitter è
stata rieseguita la stessa query in Elasticsearch, che ha consentito di effettuare una
selezione ulteriore, eliminando moltissimi tweet che pur estratti attraverso la stessa
query non contenevano le parole chiave, evidenziando una non completa accuratezza
dell’API gratuita di Twitter nell’applicazione dei filtri di estrazione. Alla fine si è
ottenuto un corpus di circa 26 mila tweet, su cui è stata effettuata la selezione
successiva.
4

JADT’ 18

203

ha avuto molta eco la notizia del record assoluto di lavoratori a termine Figura 1).
Tabella 1 – Selezione di hashtag
HASHTAG

OCC

HASHTAG

#lavoro

14.172

#licenziamento

#jobsact

1.225

OCC

HASHTAG

190

#occupati

OCC

HASHTAG

48

OCC

#MercatoDelLavoro

19

#disoccupato

173

#Occupazione

42

#precarizzazione

19

#occupazione

948

#Thyssenkrupp

164

#Cococo

42

#precarietà

19

#JobsAct

861

#contaillavoro

158

#Discoll

41

#Smartworking

18

#Jobsact

587

#lavoratori

156

#orientamento

41

#voucher

17

#disoccupazione

463

#Disoccupazione

149

#cassaintegrazione

40

#freelance

17

#povertà

414

#Melegatti

139

#mercatodellavoro

37

#Art18

15

#Poletti

278

#GaranziaGiovani

124

#JobsActSempre

32

#dipendente

15

#precari

265

#precariatodistato

110

#smartworking

31

#ScuolaLavoro

15

#LAVORO

205

#precariato

109

#thyssen

31

#ContailLavoro

201

#pandoro

98

#RelazioniIndustrialiA

20

#disoccupati

196

#articolo18

53

#poletti

20

37.1

2017

2016

26.8
25.1
23.5
21.7
20.3

13.6
11.5
9.3

8.2
6.7
5.1

4.8

0.8
28

0.5
29

4.9

3.7
1.6

1.4

0.4
30

Novembre

4.8

3.8

2.3
0.2
1

2

3

4

5

6

7

8

9

2.2
1.0
10

0.5
11

12

Dicembre

CALENDARIO DIFFUSIONI ISTAT: 28/11 Natalità e fecondità; 1/12 Occupati e disoccupati mese di ottobre *, Conti economici trimestrali; 5/12 Nota trimestrale sull'andamento dell'economia; 6/12 Condizioni di vita,
reddito e carico fiscale delle famiglie;
7/12 Il mercato del lavoro (III trimestre); 11/12 Il mercato del lavoro (rapporto annuale integrato)**.
(*) nel 2016 uscito il 30/11; (**) solo nel 2017

Figura 1 – Incidenza riferimenti a Istat per giorno. Anno 2016 e 2017

Più modesto l’impatto del comunicato mensile sull’occupazione, i cui dati
erano risultati sostanzialmente stabili (al contrario nel 2016, data la
concomitanza con il referendum costituzionale, il mensile aveva registrato la
quota massima di citazioni). Un volume consistente è stato registrato in
occasione del comunicato sulle condizioni di vita e di reddito (6/12) e della

204

JADT’ 18

presentazione del primo rapporto integrato sul mercato del lavoro (11/12).
Anomalo il picco del 3 dicembre, domenica, alimentato da un notevole tasso
di retweet molto critici sulle politiche del lavoro dell’attuale governo dovuti
probabilmente all’intervento del segretario del PD Matteo Renzi in una
popolare trasmissione serale (Che Tempo Che Fa) centrato anche sulle
politiche del mercato del lavoro degli ultimi anni. Il più citato è stato un
tweet critico sul meccanismo di conteggio degli occupati, insieme ad altri più
politici sull’aumento del lavoro a termine dell’ultimo periodo.
2.2 Il contenuto del corpus
Il contenuto del corpus può essere descritto utilizzando le parole chiave,
calcolate rispetto all’italiano standard (Bolasco, 2013) che consentono di
delimitare gli ambiti di contenuto: si incontra innanzitutto contratti, con
riferimento all’aumento dei contratti a termine o a tempo determinato.
Contribuiscono alla sovrarappresentazione del termine un numero limitato
di tweet (13) che ricevono tuttavia numerosi retweet e che, riprendendo il
dato Istat sulla durata dei contratti evidenziano l’aumento del precariato,
anch’esso termine sovrautilizzato (Figura 2). Altri termini molto presenti nel
testo sono disuguaglianze ed esclusione, utilizzati soprattutto in un post della
Caritas che riprende il dato sulla povertà pubblicato il 6 dicembre.

Figura 2 - Tagcloud delle parole chiave

Colpisce la presenza di termini molto forti, che connotano un dibattito dai
toni pesanti: trucco, fraudolenta, infamia, tossico, truffa, schiavitù. Analizzando i
contesti d’uso si riscontra che ciascuno di questi termini è riferito a episodi
diversi (trucco dei dati sulla definizione di occupazione); manodopera
fraudolente; truffa del Governo sulle pensioni; accordo tossico con riferimento

JADT’ 18

205

al CETA; infamia contro il lavoro in riferimento al Jobs Act) e che proprio i
tweet più forti siano quelli in grado di generare un numero elevato di
retweet. Significativi anche termini utilizzati in tweet in cui si evocano storie,
e in cui il dato statistico è sostituito dal caso esemplare, capace di generare
empatia e, di conseguenza, retweet. Non è un caso che i termini
maggiormente sovrarappresentati facciano riferimento ad un unico tweet, su
un lavoratore colpito da leucemia che guarisce ma viene comunque licenziato.
Fra gli esempi anche quello di una madre separata, licenziata dall’Ikea a
Milano. Anche i riferimenti al record degli occupati e a quanti esultano per i
dati sull’occupazione sono riportati talvolta in maniera critica; fa eccezione il
riferimento al tasso di disoccupazione giovanile, che viene ripreso in maniera
neutra dall’agenzia Ansa e retweettato numerose volte.
Prendendo in considerazione i segmenti ripetuti (ossia le sequenze di parole
ripetute nel testo), si possono delimitare quattro aree semantiche principali a
cui fanno riferimento i tweet (Tabella 2). In primo luogo ci sono le espressioni
che rimandano alla pura diffusione delle notizie che ruotano intorno alla
tematica “lavoro” e che hanno un peso rilevante anche in termini di
occorrenze. In particolare emerge da un lato il riferimento ai dati diffusi
dall’Istat su povertà, natalità e occupazione, dall’altro spiccano due segmenti
che si riferiscono agli episodi di attualità già citati: il licenziamento da parte
di Ikea di una madre separata con due figli piccoli e quello di un dipendente
di una fabbrica di vernici, avvenuto dopo un lungo periodo di assenza per
malattia. Accanto ai segmenti relativi alle notizie, vi sono poi i segmenti
riconducibili ai commenti degli esponenti politici, ai provvedimenti
legislativi e alle prime avvisaglie di campagna elettorale. A questi fanno da
contraltare i tweet caratteristici del dibattito pubblico tra cui non mancano
note polemiche o sarcastiche. Infine, nonostante il file sia stato in parte
ripulito dagli hashtag riconducibili agli annunci di lavoro, emergono
comunque alcuni segmenti inerenti la ricerca di particolari profili
professionali.
Come è facilmente intuibile, peraltro, alcuni contenuti caratterizzano
maggiormente i frammenti in cui si fa esplicito riferimento all’Istat. Rispetto
all’analisi effettuata nello stesso periodo dello scorso anno, l’analisi delle
specificità mostra la prevalenza di un linguaggio più tematico che tecnico
quando si cita l’istituto (dati, contratti, #povertà, disoccupazione), mentre i tweet
che parlano di lavoro senza citare l’Istat fanno riferimento ai fatti di cronaca e
alla politica (#jobsact, #pensioni, legge, licenziato ecc.), con minori riferimenti
personali ai soggetti che nel 2016 erano in prima linea nella campagna
referendaria. Inoltre l’analisi delle concordanze mostra che lo stesso
riferimento all’Istat viene utilizzato in differenti contesti.

206

JADT’ 18
Tabella 2 – Segmenti ripetuti principali

Le notizie
Segmento
Occ

Riferimenti politici
Segmento
occ

Dibattito pubblico e polemica
Segmento
occ

dati #Istat

419

Missione
compiuta\\#JobsAct

67

Come ti trucco i dati

348
119

Annunci di lavoro
Segmento
occ
#lavoro #roma
#romalavoro
152
#lavorare
#lavoro
144
#adnkronos

a
rischio
#povertà
guarisce
e
viene
licenziato
esclusione
sociale

392

Ministro #Poletti

60

continuano a produrre
sfruttamento

253

#jobsact funziona

54

tutto da rifare

55

kijiji lavoro

53

236

Fedriga Presidente

47

essere licenziati

40

cerca socio

32

madre
separata

195

campagna elettorale

43

#Bonus
dipendenza

86

manovra finanziaria

28

si sono rivelate tutele
inesistenti

27

80

Liberi e Uguali

18

conti non tornano

18

71

presidenta #boldrini

11

politici hanno distrutto
tre generazioni

3

56

Politiche Attive

9

dovremmo ribellarci

2

55

Lavori usuranti

2

giovani
andati
grazie a te

tempo
determinato
tempo
indeterminato
terzo
trimestre
crollo
della
natalità
#Algoritmi
#BigData

creano

via

31

2

Commessa IV
livello
part
time
#lavoro
professionale
diventare
#psicoterapeuta
ufficio acquisti
dirigente
medico
Concorsi
Pubblici
Gazzetta
Ufficiale

Oltre alla stretta diffusione delle notizie e al commento del dato sull’aumento
dei contratti a termine, non manca l’uso strumentale dei dati come metro di
giudizio delle politiche sul mercato del lavoro [#Istat "record di occupati
a_termine: sono 2,8 milioni". ecco l' unico risultato oggettivo del #jobsact..";
continua a calare la #disoccupazione-i nuovi dati #Istat confermano le previsioni,
un' altra ventata di ottimismo...]. Rispetto al 2016 il tono sarcastico di alcuni
tweet è meno rivolto esplicitamente all’Istat ma in generale alla situazione
del Paese [«record di #precari in Italia, 2,8 milioni. va tutto ben, madama la
marchesa.. #lavoro #Istat #occupazione”»]. Resta però un residuo polemico su
alcune definizioni di occupazione e disoccupazione [«Ricordiamo che per #istat
se si lavora un'ora retribuita a settimana si è considerati occupati. #supercazzola»;
«Come ti trucco i dati #Istat sulla disoccupazione: il 14; 6% dei contratti dura meno
di 3 giorni, il 31% un_mese»]. Infine, di interesse la valutazione del tono del
testo, possibile con l’analisi degli aggettivi positivi e negativi, riconosciuti
all’interno di Taltac2. Il rapporto tra aggettivi negativi e positivi è del 50,2%,
un valore che denota una criticità media, pari a quella che si riscontra nel
linguaggio della stampa (Bolasco, della Ratta, 2004). Il livello di criticità è
variabile nelle diverse giornate: è più basso nei giorni di diffusione dei

27
21
13
8
7

3

JADT’ 18

207

comunicati, specie quello mensile, mentre è particolarmente elevato il 3
dicembre, a causa del “rumore” prodotto dai retweet (i retweet presentano
una criticità del 63,6%), probabilmente a causa del maggiore successo dei
tweet polemici. Tra gli aggettivi negativi più frequenti precari, fraudolenta,
dannoso, fallito5.
3. Conclusioni
L’analisi effettuata ha consentito di affinare una metodologia di trattamento
dei tweet: dal punto di vista della loro estrazione, la procedura utilizzata ha
consentito di ottenere in partenza un file più pulito su cui operare una
selezione a partire dalla lista degli hashtag. L’analisi del testo ha poi
consentito di evidenziare i diversi contesti in cui si fa riferimento al dato
della statistica ufficiale. Particolarmente interessante il confronto tra i risultati
dello stesso corpus a un anno di distanza. Infatti, nello stesso periodo
dell’anno precedente la discussione era fortemente condizionata dal dibattito
referendario che ha probabilmente “stravolto” la discussione sulle tematiche
del lavoro. Nei tweet di un anno prima i livelli di criticità erano più elevati e
il ruolo dell’Istat più ridimensionato (13% la presenza odierna contro il 5% di
un anno prima). Il tono del testo appare in generale più neutro, con maggiori
richiami all’Istat nella sua veste ufficiale di diffusore di dati e meno come
oggetto di scherno e polemica. Riguardo ai contenuti, nella discussione di
fine 2017 sembra avere avuto più peso la discussione sugli effetti del Jobs Act
e della diffusione del lavoro precario. Il corpus odierno è inoltre
caratterizzato da un più ampio ricorso al retweet.
Riferimenti
Alexa (2016). Twitter site overview, at
http://www.alexa.com/siteinfo/twitter.com.
Bolasco S. (2013). L’analisi automatica dei testi. Fare ricerca con il text mining.
Roma, Carocci.
Bolasco S., della Ratta-Rinaldi F. (2004). Experiments on semantic
categorisation of texts: analysis of positive and negative dimension. In
JADT 2004 - Le poids des mots, Actes des 7es Journées internationales d’Analyse
Statistique des Données Textuelles. UCL. Louvain.
della Ratta-Rinaldi F., Pontecorvo M.E., Virgillito A., Vaccari C. (2016). Big
data and textual analysis: a corpus selection from twitter. Rome between
the fear of terrorism and the Jubilee. In JADT 2016 - Statistical Analysis of

5Sono stati comunque eliminati i termini tecnici (riferiti a specifici aggregati
statistici) che hanno una connotazione negativa, come disoccupato, scoraggiato o
povero.

208

JADT’ 18

Textual data – Vol.2. Nice.
della Ratta-Rinaldi F., Pontecorvo M.E., Virgillito A., Vaccari C. (2017). The
Role of NSIs in the Job Related Debate through Textual Analysis of
Twitter Data. NTTS 2017. Brussels.
UNECE (2016). Big Data in Official Statistics.
http://www1.unece.org/stat/platform/display/bigdata/Big+Data+in+Official+S
tatistics
UNECE (2014). Big Data in Official Statistics.
http://www1.unece.org/stat/platform/display/bigdata/Big+Data+in+Official+S
tatistics
Vaccari C. (2014). Big Data and Official Statistics. PhD Thesis, School of Science
and Technologies. University of Camerino.

JADT’ 18

209

Gauging An Author’s
Mood Using Hidden Markov Chains
Sami Diaf
Hildesheim Universität – sami.diaf@uni-hildesheim.de

Abstract
This paper aims to gauge the mood of an author using a text-based approach
built upon a lexicon score and a hidden Markov model. The text is tokenized
into sentences, each given a polarity score, yielding three evaluative factors
(positive, neutral and negative) which represent the observable states. The
mood of the author is considered a latent state (good, bad) and is estimated
via a hidden Markov model. Tested on a psychological fiction, Franz Kafka’s
novel Metamorphosis, this methodology shows an interesting linkage
between the author’s feelings and the intent of his writing.
Keywords: Sentiment analysis, hidden Markov model, polarity.
1. Introduction (Times Bold 14 pt, left)
Sentiment analysis is defined as the general method to extract subjectivity
and polarity from a text, while semantic orientation refers to the polarity and
strength of words, phrases, or texts, meaning a measure of subjectivity and
opinion in the text, capturing an evaluative factor and potency or strength of
a given corpus toward a given subject (Taboada et al., 2011).
Extracting sentiment automatically usually involves two main approaches
(Taboada et al., 2011): a lexicon-based approach built on computing
orientation for a document from the semantic orientation of words or
sentences, and a text-classification approach stemming from supervised
machine learning techniques and involves building classifiers from labeled
instances of texts or sentences. Lexicon-based models stress out the
importance of adjectives as an indicator of a text’s semantic orientation and
have been preferred in the linguistic context as classifiers yielded changing
results regarding their areas of application (Taboada et al., 2011).
Among many lexicon-based approaches adopted in the academic field, the
one implemented by Hu and Liu (Hu and Liu, 2004) remains popular. It was
built upon two hypotheses concerning the semantic orientation:
independence of context (prior polarity) and being expressed as a numerical
value suing an opinion lexicon.
This article uses the polarity approach of Hu and Liu to build a sequence of

210

JADT’ 18

evaluative factors (positive, neutral and negative), considered as the
realization of an observable state x, and supposes the mood of the author
could be approached via a two-state latent variable z taking two hidden
states (good and bad). For this aim, hidden Markov models (Murphy, 2012)
will be used to estimate the transition probabilities between hidden and
observed states, to better estimate long-range correlations among the
sequence of data than standard Markov models.
2. Polarity function
Polarity is defined as the measure of positive or negative intent in a writer’s
tone (Kwartler, 2017) and can be calculated by sophisticated or fairly
straightforward methods, usually using two lists of words: one positive and
one negative. Hu and Liu set up the architecture for the polarity function
used to tag polarized words in the English language (Hu and Liu, 2004) and
Rinkler (2017) provided a detailed description of the polarity function and its
computation. A context cluster of words is pulled around a polarized word
to be considered as valence shifters. Words in this context cluster are tagged
as neutral, negator, amplifier or de-amplifier. Each polarized word is then
weighted according to a dictionary of positive/negative words and weights,
and then further weighted by the number of position of the valence shifters
directly surrounding the positive or negative word. Final computation step is
the sum of the context clusters divided by the square root of the word count,
which yields an unbounded polarity score.
2. Application
To illustrate this framework, we took the English version of the novella
Metamorphosis written by Franz Kafka published in 1915 under the name «
Die Verwandlung » and freely available at the Project Gutenberg database. This
work was translated to English by David Wyllie in 2002 and belongs to the
psychological fiction category.
The novella is broken down into sentences, a process called tokenization, and
then we compute the polarity function for each sentence, to construct a
sequence of evaluative factors (positive, neutral or negative) according to the
polarity score, as shown in Figure 1.

JADT’ 18

211

Figure 1. Sequence of data corresponding to the polarity score of each sentence.

This step generates 812 sentences where the positive and negative polarity
scores represent respectively 29.1% and 28.6% of the total. The remaining
sentences (42.3%) correspond to the neutral evaluative factor. Statistical tests
show that the generated time series has the first two autocorrelations
significantly different from zero and exhibits a slightly persistent memory as
the estimated Hurst exponent is 0.587, significantly different from the value
of 0.5 which corresponds to the case of a Brownian motion (Mandelbrot and
Hudson, 2006). The estimated probability transition matrix of evaluative
factors via the maximum likelihood shows the associated Markov chain is
irreducible with no persistent states, as shown in Figure 2.

Figure 2. Probability transition matrix of the evaluative factors.

We assume the mood of the author could be modeled via a latent variable Z
taking two states (good and bad). Hence, we can build a hidden Markov
model explaining the interactions between observable states (positive,
neutral and negative) and latent, unobservable states (good and bad). To
estimate the hidden Markov model, the transition matrix of the latent state is
set uniformly, that is all its elements equals 0.5, the same applies also for the
initial latent vector. However, the emission matrix which describes the links
between the latent and the observable states is set arbitrarily as in Figure 3.

212

JADT’ 18

Figure 3. Prior probability transition of the emission matrix.

Given these priors, the estimated hidden Markov model using the BaumWelch algorithm (Murphy, 2012) yields a starting probability vector slightly
skewed to good mood (51%) than bad mood (0.49). The estimated transition
and emission matrices are reported in Figure 4 and 5 respectively.

Figure 4. Estimated transition matrix via Baum-Welch algorithm.

Figure 5. Estimated emission matrix via Baum-Welch algorithm.

Results demonstrate significant links between writing without intent (neutral
state) and being in a good mood, and between negative intent and the bad
mood. The most probable states estimated via the Viterbi algorithm
(Murphy, 2012) clearly show the dominance of the good state (71.4%) over
the bad (28.6%) as shown in Figure 6.
These findings help clarify the nature of the story (thriller, roman, novella,
…) and the author’s narrative style which could be confirmed by analyzing
the remaining works. Finally, it is worth noticing that this methodology
could also be used to assess the accuracy of translations with respect to the
original work, by comparing the similarities of the transition and the
emission probabilities of the hidden Markov models.

JADT’ 18

213

Figure 6. Most probable states estimated via Viterbi algorithm
(Bad in red and Good in blue).

4. Conclusion
This works expands the application field of semantic orientation to explore a
new probabilistic approach based on hidden Markov models and evaluation
factors. The resulting outcomes help understanding the author’s mood by
examining the linkage between the evaluative factors which express the
author’s mindscape through his writing. The emission probabilities between
the latent states and the evaluative factors helped identifying hidden
structures linked to the psychological state of the author and the
development of the facts. This approach could be used as a controller of
translation accuracy under the condition of having a precise list of positive
and negative words in the original language, to be able to compute the
polarity score.

214

JADT’ 18

References
Hu M. and Liu B. (2004). Mining and summarizing customer reviews.
Proceedings of the ACM SIGKDD, pp. 168-177.
Kwartler T. (2017). Text Mining in Practice with R. Wiley.
Mandelbrot B. and Hudson R.L (2006). The Misbehavior of Markets: A
Fractal Review of Finance Turbulance. Basic Books.
Murphy K.P. (2012). Machine Learning: A Probabilistic Perspective. MIT
Press.
Project Gutenberg [www.gutenberg.org]
Rinkler T. (2017). Polarity score (Sentiment Analysis)
[https://www.rdocumentation.org/packages/qdap/versions/2.2.9/topics/po
larity]
Silge J. and Robinson D. (2017). Text mining with R: A Tidy Approach.
O’reilly.
Taboada M., Brooke J., Tofiloski M., Voll K. and Stede M. (2011). Lexiconbased Methods for Sentiment Analysis. Computational Linguistics Vol.
37, Issue. 2, pp. 267-307.

JADT’ 18

215

Les hémistiches répétés
Marc Douguet
Université Grenoble Alpes – marc.douguet@univ-grenoble-alpes.fr

Abstract
In this paper, we propose to use the syllabic structure of classical alexandrine
in order to automatically identify textual recurrences in French 17th-century
theater. The two hemistichs of 6 syllables each present a syntactical unity:
consequently, extracting recurrent hemistichs is a way, on the one hand, to
hightlight idiomatic expressions characteristic of this period, and, on the
other hand, to evaluate the influence of metric constraints on writing.
Résumé
Dans cet article, nous proposons d’utiliser les caractéristiques métriques de
l’alexandrin classique afin de repérer automatiquement des récurrences
textuelles dans le corpus du théâtre français du XVIIe siècle. Les deux
hémistiches de 6 syllabes chacun qui le constituent possèdent en effet une
unité syntaxique : dès lors, les réemplois fréquents des mêmes hémistiches
permettent d’une part de faire émerger les éléments langages propres à ce
style d’écriture, et d’autre part d’évaluer l’influence des contraintes
métriques sur l’écriture.
Keywords: repeated segments, metre, verse, textual recurrences
1. Introduction
La détection des segments répétés dans un corpus est un outil
particulièrement précieux pour l’analyse stylométrique : elle permet à la fois
de caractériser le style propre à un auteur, un genre ou une période, et
d’évaluer l’originalité d’un auteur par rapport à ses contemporains, sa
capacité à s’affranchir ou non des éléments de langage de son époque (cf.
notamment Salem, 1987 ; Legallois, 2009 ; Delente et Legallois, 2016). De ce
point de vue, l’alexandrin classique présente une caractéristique qui nous
semble n’avoir pas encore été totalement exploitée. La césure divise en effet
le vers en deux hémistiches d’égale longueur (6 syllabes) qui constituent des
unités à la fois rythmiques et syntaxiques. Or ces unités font l’objet de
nombreuses répétitions. Par rapport à l’approche qui consiste à extraire tous
les segments de n mots pour détecter les récurrences, cette approche (qui la
complète) a, pour la stylistique computationnelle de la poésie, un triple
avantage :
– elle permet de n’extraire que des segments qui constituent déjà des unités

216

JADT’ 18

syntaxiques et évite d’avoir à trier manuellement les résultats pertinents ;
– elle permet d’extraire des segments qui, quel que soit leur nombre de mots,
ont le même nombre de syllabes, et sont donc, en régime poétique,
d’importance strictement comparable ; – elle permet de mettre en rapport
réflexion sur la répétition et analyse de la versification et d’apprécier,
notamment, la contrainte que le mètre fait peser sur l’écriture.
2. Méthodologie
Nous avons travaillé sur un corpus de 200 pièces de théâtre en alexandrins
publiées entre 1630 et 1680, représentatif de la diversité des genres
dramatiques de cette période (tragédie, comédie, tragi-comédie1). Le corpus
est édité en XML-TEI, avec un balisage qui décrit le découpage en actes, en
scènes, en répliques et en vers2.
Nous avons développé un syllabeur capable de césurer les vers et d’en
extraire séparément chacun des hémistiches. Celui-ci est plus modeste que
d’autres outils développés en analyse automatique du vers (notamment
Beaudouin, 2002 ; Delente et Renault, 2015 ; Salvador, 2016), puisqu’il n’a pas
pour ambition de placer avec exactitude la limite entre deux syllabes à
l’intérieur d’un mot. Afin de produire un dictionnaire de diérèses et de
synérèses, nous l’avons préalablement entraîné en vérifiant manuellement les
résultats. Le syllabeur reconnaît automatiquement comme des vers de 12
syllabes 99,98% des 55 031 vers de Corneille dont on a préalablement vérifié
qu’ils étaient des alexandrins. La marge d’erreur est uniquement due à
l’ambiguïté de certains mots, dont la prononciation change en fonction de la
catégorie grammaticale (par exemple « content » et « fier », selon qu’il s’agit
de verbes ou d’adjectifs).
Le corpus est composé de 332 938 vers, soit en théorie 665 876 hémistiches.
Nous n’en avons retenu que 624 597, après avoir exclu ceux qui était
distribués sur plusieurs répliques. Le nombre d’occurrences de chaque
hémistiche est calculé après avoir supprimé les ponctuations et les
majuscules.

La liste des pièces, les scripts utilisés ainsi que les résultats complets sont
disponibles sur https://github.com/marcdouguet/dheform.
2 Les textes sont disponibles sur https://github.com/dramacode/tcp5. Ils nous ont
été fournis par le projet « Bibliothèque dramatique » (http://bibdramatique.parissorbonne.fr/), dirigé par Georges Forestier, et le projet « Théâtre classique »
(http://theatre-classique.fr/), dirigé par Paul Fièvre. Nous les remercions tous deux
d’avoir rendu accessibles leurs sources XML, sans lesquelles ce travail n’aurait pas été
possible.
1

JADT’ 18

217

3. Fréquence des hémistiches répétés
Le phénomène de la reprise textuelle des hémistiches est sans commune
mesure avec celui, similaire, qui concerne les vers entiers. Dans notre corpus,
499 vers sont répétés au moins une fois, soit seulement 0,1%. Pour quelqu’un
qui a une connaissance approfondie du corpus, ces répétitions sont souvent
repérables manuellement, et les éditions critiques en soulignent certaines (on
connaît notamment le célèbre « Je suis maître, je parle, allez, obéissez » dans
La Mort de Pompée de Corneille, repris dans L’École des femmes de Molière).
Les enjeux de ces reprises mériteraient d’être étudiées (plagiat, parodie,
citation d’un personnage par un autre, phénomène de refrain, etc.).
La répétition d’hémistiches possède des enjeux différents, à la fois en raison
de la brièveté des segments répétés et du très grand nombre de répétitions :
16% des hémistiches du corpus sont répétés au moins deux fois, et un
hémistiche y apparaît en moyenne 1,11 fois. L’écriture en vers utilise donc un
certain nombre d’éléments de langage et d’idiomatismes préexistants, que le
dramaturge combine de manière originale.
En complément des relevés quantitatifs, nous avons également développé
une interface de lecture (accessible sur http://obvil.lip6.fr/dheform) :
l’utilisateur peut entrer un texte, dont les hémistiches répétés seront mis en
évidence à l’aide d’un code couleur.
4. Analyse des hémistiches les plus fréquents
À titre d’exemple, le tableau suivant liste les 10 hémistiches les plus fréquents
du corpus, avec leur nombre d’occurrences et deux exemples en contexte :
en cette occasion

119

en l’état où je suis

98

pour la dernière fois

87

à votre majesté

87

que votre majesté

70

en cette extrémité

68

je vous l’ai déjà dit

55

une seconde fois

51

les armes à la main

42

de votre majesté

41

Que me donne l’amour en cette occasion
N’offrez donc point, Seigneur, en cette occasion
Que ferai-je, Philante, en l’état où je suis ?
Je ne réponds de rien en l’état où je suis.
Dites-lui de ma part pour la dernière fois
Pour la dernière fois je me jette à vos pieds.
Le respect que je dois à votre Majesté
Je me livre, grand Prince, à votre Majesté,
Que votre Majesté le rappelait près d’elle.
Ah ! Grand Roi, se peut-il que votre Majesté
Mettre tout en usage en cette extrémité ;
Quoi ? vous m’abandonnez en cette extrémité,
Je vous l’ai déjà dit, sans vous parler de moi,
Je vous l’ai déjà dit, j’estime votre flamme,
Je renonce à choisir une seconde fois ;
J’en ferais un ingrat une seconde fois.
Les armes à la main, venez si bon vous semble,
Laissez-nous lui parler les armes à la main,
Qui vient offrir aux pieds de votre Majesté
Il tira des bienfaits de votre Majesté :

218

JADT’ 18

Si l’on élargit l’analyse aux 470 hémistiches qui possèdent plus de 10
occurrences, on peut distinguer plusieurs catégories de récurrences. De
nombreux hémistiches sont composés d’un substantif de trois syllabes ou
plus, précédé de prépositions, de conjonctions et de déterminants, et placé en
position de sujet, de complément de nom ou d’objet. Dans cette
configuration, on repère plusieurs variations autour d’un même substantif :
« à votre majesté » (87 occurrences – nous indiquerons désormais
systématiquement le nombre d’occurrences d’un hémistiche entre
parenthèses), « que votre majesté » (70), « de votre majesté » (41), « de
générosité » (40), « la générosité » (26), « à ma confusion » (30), « cette
confusion » (15), etc. Les substantifs concernés relèvent principalement d’une
thématique morale ou politique, caractéristique du style d’écriture
dramatique du XVIIe siècle.
Plus intéressants sont les compléments circonstanciels qui insistent sur le
caractère exceptionnel de la situation et sur l’état émotif du locuteur et
renforcent ainsi le pathos du discours : « en cette occasion » (119), « en l’état
où je suis » (98), « en cette extrémité » (68), « en ce malheur extrême » (23),
« en cette conjoncture » (22). De nombreuses expressions modalisent
l’énoncé : insistance agacée (« je vous l’ai déjà dit » (55)), certitude (« il n’en
faut point douter » (37), « il n’en faut plus douter » (25)), prétérition (« je ne
vous dirai point » (40)). On notera également la série « pour la dernière fois »
(87), « une seconde fois » (51), « pour la première fois » (29), qui relie une
situation dramatique à d’autres, passées ou à venir.
Certains syntagmes figés possèdent au contraire une fonction référentielle :
violence des relations (« les armes à la main » (42), « un poignard dans le
sein » (27)), instinct (« la voix de la nature » (19), pouvoir (« la suprême
puissance » (25), « une entière puissance » (24), « un absolu pouvoir » (22)),
etc. Les expressions temporelles sont quant à elle nombreuses, et peuvent
être associées à une sentence générale décrivant les mœurs du temps (« dans
le siècle où nous sommes » (17)) ou à l’urgence d’une situation (« sans tarder
davantage », (19)). La fréquence élevée d’« avant la fin du jour » (31) montre
à quel point le dramaturges explicitent le respect de l’unité de temps dans
leurs œuvres afin d’accroître la tension dramatique. Les expressions spatiales
renvoient elles aussi à l’universalité (« sur la terre et sur l’onde » (16)) ou au
contraire aux lieux fréquemment convoqués dans le théâtre classique (« dans
son appartement » (20), « dans la chambre prochaine » (16)).
Ces expressions figées peuvent souvent être considérées comme des
« chevilles », où l’on sent clairement que l’invention verbale se soumet aux
contraintes de la métrique. On peut ici identifier deux cas de figure. D’une
part, le sémantisme de certains hémistiches circonstanciels est parfois très
faible : « en cette occasion », « en l’état où je suis » pourraient aussi bien être

JADT’ 18

219

supprimés sans nuire au sens du texte, ou greffés sur n’importe quel énoncé.
D’autre part, même si elles sont mieux ancrées dans l’énoncé, les expressions
figées que nous avons relevées (« la suprême puissance », « la voix de la
nature ») doivent certainement leur succès au fait qu’elle rentrent facilement
dans le moule de l’alexandrin. C’est ici l’apposition récurrente d’un adjectif
(la puissance sera « entière » ou « suprême »), ou l’utilisation d’une formule
imagée (« la voix de la nature », au lieu de « la nature ») qui se justifie par les
contraintes de la versification. Il serait intéressant de poursuivre cette analyse
en la croisant avec la théorie de la fonction poétique du langage de Jakobson,
que résume en partie l’exemple suivant : « Without its two dactylic words the
combination “innocent bystander” would hardly have become a hackneyed phrase. »
(1960 : 358)
5. Vers et prose
Afin d’évaluer la spécificité de l’écriture poétique, nous avons constitué un
corpus de pièces en prose de la même époque (11 tragédies de d’Aubignac3,
Baro et Puget de La Serre, et 9 comédies de Molière). Nous avons compté le
nombre d’occurrences de chacune des expressions correspondant à un
hémistiche récurrent, en le rapportant à la taille respective des deux corpus,
calculée en nombre de mots. Certains « hémistiches » (les guillemets
s’imposent ici) sont aussi fréquents en vers qu’en prose, mais il n’existe pas
de corrélation nette entre les deux corpus, alors même que l’on reste dans le
genre dramatique. Or les « hémistiches » que l’on trouve aussi fréquemment
en prose qu’en vers, voire plus fréquemment, sont ceux qui reposent à la fois
sur un substantif unique (suffisamment long pour occuper les six syllabes
avec les déterminants, les prépositions et les conjonctions qui le précèdent) et
qui n’ont pas une fonction de complément circonstanciel. Le fait qu’ils
figurent parmi les hémistiches les plus fréquents dans le corpus en vers
s’explique simplement par le fait que le substantif en question est lui-même
extrêmement fréquent. En revanche, les formules figées qui reposent sur une
association de plusieurs termes et qui ne font qu’apporter une modalisation
sont bien surreprésentées en vers (par exemple « je vous l’ai déjà dit » : 17
occurrences pour un million de mots en vers, 0 en prose ; « il n’en faut point
douter » : 12 en vers, 0 en prose ; « pour la dernière fois » : 28 en vers, 9 en
prose). Ces expressions, spécifiques au théâtre en vers, semblent donc bien
devoir leur suremploi à la nécessité de couler la phrase dans le moule de
l’alexandrin.

3 Nous tenons à remercier ici Bernard J. Bourque, qui nous a fourni la version
numérique de son édition Abbé d’Aubignac, Pièces en prose, Tübingen, Gunter Narr
Verlag, coll. « Biblio 17 », 2012.

220

JADT’ 18

6. Premiers et seconds hémistiches
Un des défauts de cette approche est de surévaluer la césure au détriment de
l’unité du vers, et de la considérer comme une coupure, une pause entre
deux segments indépendants. Deux écueils se profilent. D’un côté, on risque
d’oublier que l’hémistiche ne constituent pas toujours, au sein d’un vers, une
unité syntaxique pertinente. Les dramaturges du XVIIe siècle pratiquent
souvent le rejet, le contre-rejet ou l’enjambement internes (par exemple : « Le
temps de cet orgueil me fera la raison », dans La Galerie du Palais de
Corneille). Cependant, notre projet est avant tout lexical, et non prosodique.
Isoler les hémistiches n’est qu’une manière de faire émerger des
idiomatismes, en se fondant sur le fait que, malgré des exceptions, la césure à
l’hémistiche reste le plus souvent la plus forte rupture syntaxique du vers.
Il ne faudrait pas non plus oublier que l’élocution fond les deux hémistiches
dans un même mouvement, et que ceux-ci ne se situent donc pas sur le même
plan : un poème en alexandrins n’est pas une suite d’hémistiches. Ici,
l’analyse automatique à laquelle nous nous sommes livré donne justement
des arguments en faveur de l’unité du vers, car elle nous permet de faire
émerger plusieurs différences entre les premiers et les seconds hémistiches,
qui complètent et confirment les analyses de Beaudouin (2002 : 275-319)
concernant la répartition des phonèmes et des catégories morphosyntaxiques en fonction de la position métrique.
Ils diffèrent tout d’abord dans le taux de répétition. 13% des hémistiches
placés en première position sont employés ailleurs dans notre corpus (soit en
première, soit en seconde position), ce qui est moins que le pourcentage
global de récurrences. Au contraire, ce pourcentage monte à 18% quand on
considère les hémistiches placés en seconde position. Cette divergence
s’explique facilement par le fait que le second hémistiche n’est pas seulement
soumis à la contrainte du mètre, mais aussi à celle de la rime.
Si l’on considère la proportion d’hémistiches qui commencent par un son
vocalique, on constate également un déséquilibre : 27% des premiers
hémistiches, mais 30% des seconds. La différence est faible, mais elle nous
semble permettre de quantifier la contrainte que pose la présence d’un e à la
fin du premier hémistiche, qui serait fautive si le second commençait par un
son consonantique. Ainsi, tandis que le premier hémistiche peut commencer
par n’importe quel son, un hémistiche commençant par un son vocalique est
plus facile à placer en seconde position qu’un hémistiche commençant par un
son consonantique.
Enfin, les hémistiches les plus fréquents ne sont pas les mêmes selon que l’on
considère ceux placés en première et en seconde position. Certains sont
utilisés aussi bien à l’une ou l’autre place (par exemple, « en l’état où je suis »
apparaît 40 fois en premier, 58 fois en second), mais on observe souvent une

JADT’ 18

221

répartition nette : les hémistiches de modalisation de l’énoncé sont plus
souvent en premier (« je ne vous dirai point » : 39 pour 1, « je vous l’ai déjà
dit » : 52 pour 3 ; « je vous le dis encor » : 20 pour 2), les hémistiches ayant
fonction de compléments, en second (« à votre majesté » : 85 pour 2 ; « de
votre majesté » : 40 pour 1 ; « à mon ressentiment » : 37 pour 0).
7. Conclusion et perspectives
La détection automatique des récurrences d’hémistiches permet donc de
mettre en valeur les contraintes spécifiques qui pèsent sur l’écriture en vers.
Même si les conclusions que l’on peut tirer ne font que confirmer un savoir
déjà existant, cette méthode nous offre aussi un point d’entrée original dans
le corpus du théâtre classique. Elle nous amène à lire autrement ces textes et
rend particulièrement sensible, derrière la voix d’un auteur, la voix diffuse
d’un style d’époque. À travers ces expressions et ces associations d’idées
transparaît tout un imaginaire qui constitue en quelque sorte le « dictionnaire
des idées reçus » du XVIIe siècle.
Nous n’avons fait là que jeter quelques pistes de réflexion. Un examen
quantitatif et qualitatif plus précis est nécessaire pour mieux cerner les enjeux
de ce phénomène, tout comme la prise en compte de textes versifiés non
dramatiques. Il restera également à étendre le corpus de référence des textes
en prose et à définir d’autres principes de comparaison pour évaluer
l’influence de la métrique sur la diversité syntagmatique des textes.
Envisager les récurrences au niveau, plus abstrait, du motif syntaxique (dans
la lignée des travaux de Ganascia, 2001 ; Longrée et al., 2008 ; Mellet et
Longrée, 2013 ; Legallois et Prunet, 2015), nous permettra par ailleurs de
regrouper des occurrences présentant une structure syntaxique semblable
(« la voix de la nature », « le flambeau de la guerre », « les fruits de la
victoire ») ou centrées sur les mêmes termes (« qu’on le/la/les fasse venir »).
Enfin, la fréquence relative de ces hémistiches récurrents nous paraît être un
outil statistique particulièrement prometteur pour évaluer la spécificité du
style d’écriture propre à un genre ou un auteur, ainsi que pour observer
l’évolution de ces éléments de langage dans le temps.
Références
Beaudouin V. (2002). Mètre et rythmes du vers classique. Corneille et Racine.
Honoré Champion.
Delente É. et Legallois D. (2016). La répétition littérale dans Les RougonMacquart : présentation d’un phénomène peu connu. Excavatio, vol.28.
Delente É. et Renault R. (2015). Outils et métrique : un tour d’horizon.
Langages, vol.199 : 5-22.
Ganascia J.-G. (2001). Extraction automatique de motifs syntaxiques. Dans

222

JADT’ 18

Maurel D. (éd), TALN - RECITAL 2001 : 8e conférence annuelle sur le
Traitement Automatique des Langues Naturelles.
Jakobson R. (1960). Closing statements: Linguistics and Poetics. Dans Sebeok
T. A. (éd), Style in Language. The Technology Press of MIT/John Wiley and
Sons, inc.
Legallois D. (2009). À propos de quelques n-grammes significatifs d’un
corpus poétique du XIXe siècle. L’Information grammaticale, vol.121 : 46-52.
Legallois D. et Prunet A. (2015). Sequential patterns: a new corpus-based
method to inform the teaching of language for specific purposes. Journal of
Social Science, vol.44 : 127-140.
Longrée D., Luong X. et Mellet S. (2008), Les motifs : un outil pour la
caractérisation topologique des textes. Dans Heiden S. et Pincemin B.
(éds), JADT 2008. 9es Journées internationales d’Analyse statistique des Données
Textuelles, pp. 733-744.
Mellet S. et Longrée D. (2013). Le motif : une unité phraséologique
englobante ? Étendre le champ de la phraséologie de la langue au
discours. Langages, vol.189 : 65-79.
Salem A. (1987). Pratique des segments répétés. Essai de statistique textuelle.
Klincksieck.
Salvador X.-L. (2016). Versification : outil d’analyse du mètre français
(http://www.projetprada.fr/versification et
https://gist.github.com/xavierLaurentSalvador).

JADT’ 18

223

«Mangiata dall’orco e tradita dalle donne». Vecchi e
nuovi media raccontano la vicenda di Asia Argento,
tra storytelling e Speech Hate
Francesca Dragotto1 Sonia Melchiorre2
1

Università di Roma Tor Vergata – dragotto@lettere.uniroma2.it
2Università della Tuscia – melchiorresmr@unitus.it

Abstract 1
Re-enacted and dissected in the National and International news, the
narration of the rape denounced by Italian actress Asia Argento has triggered
several coming outs revealing the violence perpetuated against other actors
and actresses by prominent personalities of the Hollywood star system.
Textually molded between diffused narration and the blink of a tweet, the
story has hooked the public displaying, in the Italian media in particular, a
morbid legitimation of Victim Blaming. Asia Argento has become the object
of Hate Speech revealing, in turn, a cultural palimpsest of lies and guilty
silences deriving from stereotypes represented in comments of the most crass
and basest order. The present discussion starts therefore from a quantitative
and qualitative analysis of texts, in English and Italian, reporting the story
and aims to reveal the similarities and differences between language
practices substantiating the discourse of violence. Another corpus derived
from the social networks will also reveal the righteous indignant reactions of
cybernauts concerning this story which will help identify the language
patterns at the core of gender-based violence.
Abstract 2
Spolpata dalle cronache nazionali e internazionali, la narrazione della
violenza sessuale denunciata dall’attrice italiana Asia Argento ha funto da
detonatore di una esplosiva sequela di coming out rivelatori di episodi
analoghi subiti, da altre attrici e, seppur in misura inferiore, attori, da parte di
personaggi di spicco dello Star System hollywoodiano. Colata in tutti gli
stampi testuali compresi tra la narrazione diffusa e il succinto tweet, la trama
di questa vicenda ha tenuto e ad oggi ancora tiene significativo banco
mediatico, alimentando un dibattito che, nel caso italiano, si è dimostrato
spesso più interessato all’individuazione di ragioni utili a legittimare il
Victim Blaming che a ricostruire le coordinate del contesto in primis
psicologico nel quale si sarebbe consumata la violenza. Oggetto di
innumerabili discorsi di odio, il racconto rappresentato dalla cronaca italiana

224

JADT’ 18

costituisce un oggetto utile a investigare il sentimento sociale nei confronti di
storie di violenza con protagoniste persone (in special modo donne) famose,
nei confronti delle quali si attivano reazioni di sdegno frammisto alla
colatura dei più beceri stereotipi di genere. Muovendo dall’analisi
quantitativa e qualitativa di un corpus di testi incentrati su questa vicenda,
prodotti in lingua inglese e in lingua italiana, chi scrive si ripropone di far
emergere luoghi di contatto e di separazione tra le diverse forme della
cronaca, unitamente alle costellazioni lessicali, semantiche e pragmatiche che
le hanno sostanziate. Correderà questa analisi quella di un secondo corpus,
stavolta estrapolato dalla ricca produzione social riconducibile ad account
ora individuali, ora di gruppi noti per l’indefessa attività di comunicazione
indignata intorno a vicende dell’attualità. Scopo ultimo del lavoro, sarà
l’intercettazione dell’eventuale pattern linguistico e concettuale della
violenza di genere, del quale si testeranno i limiti di validità all’interno di
sistemi diversi e di varietà diverse dello stesso sistema.
1. La narrazione
Umiliata e offesa. Questo il destino toccato all’attrice italiana Asia Argento,
tra le prime a denunciare la violenza subita dal produttore cinematografico
hollywoodiano Harvey Weinstein. La donna ha avuto il coraggio di esporre
pubblicamente il suo stupratore assieme a una ottantina di altre, che come lei,
hanno subito prima un oltraggio fisico e successivamente un’esposizione
mediatica senza precedenti. Appare significativo da un punto di vista
narrativo, che la vicenda sia stata innescata da un tweet e che sia
successivamente rimbalzata nei media di tutto il mondo. Nel breve lasso di
un cinguettìo Asia Argento rivela i nomi di tutte le donne che con coraggio
hanno denunciato la violenza perpetrata nei loro confronti da un uomo che si
credeva potente e intoccabile. Ed ecco che dal racconto delle vittime
scaturisce una nuova narrazione in cui le donne diventano survivors, dando
voce alla loro rabbia contro un sistema patriarcale, sessista e misogino,
condensato in uno slogan già storico: Me too, “anche io”, nel quale tutte le
donne del mondo vittime di violenza si sono riconosciute. È accaduto poi che
due parole si transustanziassero nella Person of the Year 2017 guadagnando la
copertina del Time, che si incarnassero nei corpi abbigliati di nero di tutte le
attrici che hanno partecipato al Golden Globe 2018 e che, infine, si
trasformassero nel Time’s Up, “Il tempo è scaduto”, refrain che si propone
come impulso trasformatore della rabbia in forza (ri)costruttrice e che,
probabilmente, accompagnerà l’afro-americana Ophra Winfrey nella corsa
per la Casa Bianca. In Italia, nel frattempo, si fatica, e molto, ad ammettere
perfino che le parole usate dai media nel caso Asia Argento dimostrino
l’esistenza di un grave problema culturale. Nel nostro paese parole tossiche,
nell’insieme dette hate speech, hanno condotto a un vergognoso victim blaming

JADT’ 18

225

nei confronti di Asia Argento: una etichetta eufemistica per le orecchie
italiane che finisce però per assumere la forma testuale di un testo
argomentativo dalle cui trame scaturisce violenza e accanimento mediatico –
ironia sprezzante e spregiativa nei casi migliori – non già nei confronti degli
aggressori, bensì delle persone vittime di violenza sessuale. Questa tendenza
ben si evince dalla disamina, anche solo cursoria, di testi recuperabili dal
web. In questa sede ne è stata raccolta una selezione, in lingua italiana e
inglese, successivamente sottoposta ad analisi contrastiva. Dall’analisi è
emersa la tendenza all’uso di una terminologia, sistematicamente sostenuta
da toni aggressivi, rivelatrice di un sistema più complesso di collusione
culturale con un sistema che sarebbe frettoloso liquidare come fallocentrico e
misogino e percorso da una omosocialità maschile da spogliatoio. Portatrice
di significato per quanto e come dice, ma anche per quanto non dice, la
lingua di questi testi (e in generale di ogni testo), costituisce infatti una porta
di accesso all’architettura ideologica che la sorregge e che sorregge le
coordinate di chi se ne serve: una architettura che cela un mondo
sclerotizzato, che nel caso in questione prevede un pendant tra atteggiamento
aggressivo di chi offende e lesione della dignità di chi offeso/a, su cui è
necessario gettare luce se si vogliono comprendere le dinamiche che guidano
l’agire in questa porzione di tempo che vede la vita sociale e comunicativa
governata dalle strutture dei social media. In attesa dei risultati dell’analisi di
un corpus meglio strutturato e più tendente alla sistematicità – con tutti i
limiti che la sistematicità applicata al testo inteso in senso cognitivo può
avere – in questa prima fase si procederà con l’esposizione dei nuclei più
significativi ottenuti per carotaggio. I frammenti proposti sono stati scelti
perché rappresentativi ciascuno di un corpus dalle caratteristiche analoghe.
1.1 Victim blaming
Queste alcune delle domande proposte ad Argento da G.M. Tammaro, de La
Stampa (15 ottobre 2017), a immediato ridosso della denuncia pubblica
dell’attrice. Difficile non rintracciarvi lo schema narrativo plurisecolare
dell’interrogatorio della vittima di violenza (si pensi, uno su tutte, al primo
processo per stupro della storia, quello nei confronti della pittrice Artemisia
Gentileschi, e non dell’aggressore Agostino Tassi, del 1612): il testo-genere si
compone di domande alle quali chi ha subito violenza deve rispondere in
maniera dettagliata per non essere tacciata di collusione con il predatore.1 In
grassetto gli elementi che si ritengono rilevanti per il discorso.

1http://www.lastampa.it/2017/10/15/italia/cronache/un-orco-mi-ha-mangiata-lacosa-pi-sconvolgente-i-tanti-attacchi-dalle-donnehUwq9t9TFgRHkmcjU8yhAL/pagina.html (ultimo accesso 11/01/2018).

226

JADT’ 18

1. Perché ha deciso di rivelare questa storia a distanza di tanti anni?
2. Non pensa che parlare prima avrebbe evitato che altre donne subissero
come lei?
3. Che cosa l’ha ferita maggiormente?
4. E lei come reagisce?
5. Come ha vissuto questi anni di silenzio?
6. Si sente ancora in colpa per questo?
7. Che cosa temeva che le potesse accadere, in caso di denuncia all’epoca dei
fatti?
8. Fabrizio Lombardo, ex capo di Miramax Italia, nega di averla portata da
Harvey Weinstein, come lei invece sostiene.
9. Dopo il primo incontro in un hotel in Costa Azzurra, lei iniziò una
relazione con Weinstein?
10. Weinstein cercò di contattarla ancora?
11. Lei accettò?
12. Qual era l’atteggiamento di Weinstein nei suoi confronti?
13. Come cambiò il suo comportamento, nei confronti di Weinstein?
14. Quindi vi incontraste altre volte?
15. Poi però ha deciso di farsi avanti in prima persona: come mai?
16. In Italia non tutti la pensano così. Non tutti le credono. Non tutti stanno
dalla sua parte.
17. La accusano anche di aver firmato la petizione a favore di Roman
Polanski, indagato per pedofilia.
18. Si è pentita?
19. Dopo essersi fatta avanti insieme alle altre donne e aver raccontato quello
che le è successo, cosa spera che accada?
Poste una di seguito all’altra, le domande assumono la forma di una
narrazione a se stante, caratterizzata da una costellazione di termini e da una
semantica incentrata sulla vittima non in quanto tale ma in quanto teste che
deve fornire spiegazioni per quanto accaduto, per giustificare il suo silenzio.
Quelle a seguire sono invece alcune delle frasi pronunciate, a vario titolo, da
Mario Adinolfi, Vittorio Feltri e Vittorio Sgarbi, rimbalzate tra numerosi siti e
quotidiani del mondo, tra i quali il New Yorker, per primo, il Guardian e
l’Independent. L’articolo di The Guardian riporta, per esempio, le seguenti
parole: “Far from being hailed as brave, Argento’s allegations were initially
treated in some Italian media outlets with a mix of scepticism and scorn”
dove colpisce il pendant tra il brave, ‘coraggiosa’, utilizzato dalla giornalista
per definire Asia Argento, e l’atteggiamento generalizzato di ‘scetticismo’ e
‘disprezzo, disdegno’ (scorn rimanda anche all’idea di ‘rifiuto’, di non
accettazione di qualcosa che viene proposto). La giornalista riporta poi le

JADT’ 18

227

parole di Asia Argento: “Here people don’t understand. They’ll say, ‘oh it’s
just touching tits’. Well yeah, and this is a very grave thing for me. It is not
normal. You can’t touch me, I am not an object”. Il pezzo non omette la
descrizione dettagliata della violenza subita dall’attrice e il commento
offensivo di Vittorio Feltri, che sminuisce l’atto sessuale poiché solo sesso
orale (licking e non oral sex nella sua interpretazione). L’elemento più
rilevante dell’articolo resta una delle frasi conclusive della giornalista: “For
now, not a single fellow female actor who is well known has spoken out in
support of her, even though the Italian film industry is rife with abuse”, dove
rife with abuse rimanda da un lato alla reiterazione di atti, dall’altro,
significando rife ‘pieno zeppo’, allude anche a un atteggiamento collusivo di
quanti con comportamento omertoso non denunciano. In un altro articolo,
sull’Independent, sempre in Gran Bretagna, Lydia Smith scrive: “But she was
subsequently criticised by some sections of the Italian media for not coming
forward sooner about the alleged assaults, despite hesitation being common
among survivors for fear of reprisals, among other reasons. […]”. Riporta poi
gli interventi di Renato Farina apparsi su Libero e i suoi commenti Victim
Blaming volti cioè alla colpevolizzazione della vittima e tipico di chi è rimasto
molto, troppo indietro rispendo a un mondo che va veloce:2 “Conservative
newspaper Libero published an op-ed by Renato Farina, with the headline:
‘First they give it away, then they whine and pretend to repent’”.3
1.2 Hate speech
“Se denunci uno stupro in Italia sei tu la troia”. E, ancora, “Solo in Italia
vengo considerata colpevole del mio stupro perché non ne parlai quando
avevo 21 anni”, denuncia Asia Argento dopo le critiche e le aggressioni
verbali ricevute sui media italiani, anche da parte di star, che insinuano o
apertamente dichiarano che “Si può sempre dire di no...”. Il 13 ottobre Asia
Argento torna sul caso Weinstein con un tweet amaro: “Ho denunciato uno
stupro e per questo vengo considerata una tr...”. Ma il mondo dello
spettacolo affronta la questione in modo che è eufemistico definire prudente.
“Conosco bene Asia Argento e la stimo”, rivela Vladimir Luxuria. “Quando
ho letto che raccontava di essere stata costretta a un rapporto orale, la prima
reazione è stata di solidarietà. Ma quando ho letto che, dopo aver subito
questa violenza, ha fatto un film con lui, è andata con lui sul red carpet a

http://www.liberoquotidiano.it/news/opinioni/13264032/harvey-weinsteinrenato-farina-scandalo-sessuale-hollywood.html (ultimo accesso 08/01/2018).
3
http://www.independent.co.uk/arts-entertainment/films/news/harveyweinstein-sexual-assault-asia-argento-flees-italy-public-condemn-speaking-outa8012511.html (ultimo accesso 08/01/2018).
2

228

JADT’ 18

Cannes, l’ha frequentato per cinque anni, allora mi sono detta che c’era
qualcosa che non andava. Purtroppo in queste vicende bisogna avere una
credibilità totale, altrimenti basta una sola fake news a mettere in discussione
tutto: […]”. Ottavia Piccolo, stimatissima attrice di teatro e cinema, preferisce
sorvolare: “Sono cose che sono sempre accadute, non voglio parlarne perché
rischierei di dire solo banalità”. Mentre Rita Dalla Chiesa affronta senza
timore l’argomento: “Sicuramente la paura di perdere il lavoro può esserci.
Se però una persona si è sentita realmente offesa e traumatizzata ma poi,
invece di scappare, resta all’interno di questo cerchio negativo, prende treni e
aerei e va agli appuntamenti in albergo, non parlerei più di stupro, ma di un
rapporto cosciente”. Cita poi le parole di Barbara Palombelli, con le quali
afferma di concordare: “[…] Sei stata violentata? E perché lo dici dopo anni?
Troppo comodo. Non facciamo battaglie femministe su cose che col
femminismo non c’entrano niente”. “Sarò una mosca bianca”, rivela invece
Alba Parietti, “ma a me non è mai capitato niente del genere. A volte basta
l’atteggiamento per scoraggiare un uomo. Il punto centrale del problema è la
paura: l’eterna paura delle donne nei confronti degli uomini, del loro potere,
di non essere credute. Conosco potenti donne manager che quando tornano a
casa si lasciano menare dal marito. Perché questo tipo di atteggiamento non
riguarda solo il mondo dello spettacolo, ma tutti gli ambiti lavorativi. Con
un’aggravante: nello spettacolo non insegui un posto da 1200 euro al mese,
ma fama e successo”. Il 26 ottobre 2017 Guia Soncini, editorialista della
rivista Gioia, commenta sul New York Times il fallimento del femminismo
italiano, riferendosi alla vicenda di Asia Argento:4 “This episode is another
example of my country just being male-run, sexist Italy […] This, in a country
that has a total of zero national newspapers edited by women and zero
female columnists in its main national papers. […] Where the reaction to Ms.
Argento’s account has been truly vicious has been on social media. And
there, it has primarily come from women. […] What this tells us about Italian
feminism isn’t clear, but it’s certainly ugly. […] There’s something underripened about the state of feminism in my country”. Peccato che Soncini
avesse postato un tweet decisamente poco femminista (“Sogno un pezzo su
Weinstein d’una sola riga. Quello sarà un vecchio porco, ma voli gliela
tiravate con la fionda, finché pensavate servisse”) qualche giorno prima (10
ottobre 2017), cosa che non sfugge propria ad Asia Argento. L’attacco più
diretto è quello sferrato, via Facebook, da Selvaggia Lucarelli in un post

4
https://mobile.nytimes.com/2017/10/26/opinion/italian-feminism-asia-argentoweinstein.html?partner=IFTTT &_r=0&referer=https://t.co/pj6FLcp4Fx (ultimo accesso
10/01/2018).

JADT’ 18

229

molto lungo:5 “Ora. Francamente. Vai a letto con un bavoso potente per anni
e non dici di no per paura che possa rovinare la tua carriera. Legittimo. Frigni
20 anni dopo su un giornale americano raccontando di tuoi rapporti da
donna consenziente tra l’altro avvenuti in età più che adulta, dovendo
attraversare oceani, con viaggi e spostamenti da organizzare, dipingendoli
come “abusi”. Meno legittimo. Ad occhio, sono abusi un po’ troppo
prolungati e pianificati per potersi chiamare tali. E se tu sei la prima a dire
che lo facevi perché la tua carriera non venisse danneggiata, stai ammettendo
di esserci andata per ragioni di opportunità. Nessuno ti giudica, Asia
Argento. Però ti prego. Paladina delle vittime di molestie, abusi e stupri,
anche no. Facciamo che sei finita in un gorgo putrido di squallidi do ut des e
te ne sei pentita. Con 20 anni di ritardo però”.6
1.3 La sindrome di Stoccolma
All’inizio di quest’anno i media hanno riportato la notizia dell’ennesimo
femminicidio in Italia. Si scopre e ci si meraviglia che la donna, bruciata viva
dal suo convivente, abbia più volte difeso il suo aggressore. Questo
atteggiamento ha un nome: Sindrome di Stoccolma, una sindrome che
sembra colpire tante donne e il cui effetto andrebbe per lo meno valutato
anche per spiegare le reazioni delle tante donne che hanno reagito
attribuendo la responsabilità di quanto accaduto alla Argento, chiamando del
tutto fuori il suo aggressore. Natalia Aspesi, femminista e donna di cultura,
ha sostenuto che “Se mi chiedi un massaggio in ufficio e io te lo concedo, poi
non mi posso stupire su come va a finire”. E, ancora, “Che i produttori,
almeno da quando ho memoria di vicende simili, hanno sempre agito così. E
le ragazze, sul famoso sofà, si accomodavano consapevoli. Avevano fretta di
arrivare. E ancor più fretta di loro avevano le madri legittime che su quel
divano, senza scrupoli di sorta, gettavano felici le eredi in cerca di un ruolo,
di un qualsiasi ruolo”.7 “L’eccezione alla regola proposta è Sofia Loren, che
sposò un produttore per proteggersi – afferma ancora Aspesi – da attenzioni
indesiderate”. A chi le chiede se stia giustificando Weinstein, risponde inoltre
“Non giustifico niente. Il femminismo è ancora una delle missioni più
importanti per le donne di tutto il mondo, forse la più importante in assoluto.

https://www.leggo.it/gossip/news/asia_argento_stuprata_da_weinstein_selvagg
ia_lucarelli_frigni_dopo_20_anni_foto_video_11_ottobre_2017-3295503.html (ultimo
accesso 10/01/2018).
6https://www.leggo.it/spettacoli/cinema/asia_argento_weinstein_sfogo_twitter_1
2_ottobre_2017-3297028.html (ultimo accesso 09/01/2018).
7 https://www.vanityfair.it/news/approfondimenti/2017/10/11/weinsteincommento-natalia-aspesi (ultimo accesso 11/01/2018).
5

230

JADT’ 18

È qualcosa in cui ho creduto e credo ancora ciecamente. Ma non mi pare che
con queste denunce possa fare un salto decisivo. Magari sbaglio, ma ho i miei
dubbi”. Il dubbio “Che sia una vendetta fratricida, per togliere di mezzo
Weinstein. Era un produttore potente come pochi e sporcaccione come
moltissimi altri. Che la storia, risaputa da decenni, sia venuta fuori con
questa virulenza soltanto adesso, accompagnata da decine di testimonianze,
non può essere casuale”. A completare la rassegna, un articolo, senza firma,
battuto da ADN Kronos (13/10/2017), che già col solo titolo riesce a
sintetizzare lo stato della polemica Donne che odiano le donne, gogna social per
Asia Argento: “[…] E nel marasma dei commenti social che la accusano di
volta in volta di opportunismo, di prostituzione, di sensazionalismo, a
colpire duro incredibilmente sono soprattutto le donne. Man mano che si
scorrono i commenti agli articoli dedicati al caso in questi giorni dai
principali quotidiani, non è infatti difficile incappare – anzi, è impossibile –
nei tanti insulti lanciati contro l'attrice: a scriverli sono mamme, nonne,
ragazze, studentesse, tutte convinte della colpevolezza di Asia Argento, rea
nel migliore dei casi per chi commenta di aver aspettato troppo a parlare o,
nel peggiore, di essersi prostituita in cambio di un posto al sole di
Hollywood”.8
1.4 La decisione di lasciare l’Italia
“Newspapers ‘slut-shamed’ Asia Argento so badly over the Weinstein saga
that she’s leaving Italy”,9 riporta spesso la stampa straniera nel dar conto
dell’evoluzione della saga di Asia Argento, giudicata coraggiosa e ispiratrice
di altre donne. Fuor di patria. “Part of the criticism from some Italian
newspapers and social media users revolves around the counter-argument
that these celebrities should have come forward years ago (we debunked this
argument here). While these newspapers and internet users are hardly the
only ones engaging in this form of victim-blaming, the violent tone used by
some is alarming and astonishing […].”. Cita quindi il caso di Renato Farina.
La reazione sorprende ancor più la stampa straniera che ha un mezzo di
facile paragone nella solidarietà riservata alle attrici americane protagoniste
di analoghe denunce nei confronti di Weinstein. Giunta a Laura Boldrini la
notizia dell’espatrio volontario, la Presidente della Camera indirizza il
proprio appello all’attrice chiedendole di desistere dai suoi propositi: «Resta

http://www.adnkronos.com/fatti/cronaca/2017/10/13/donne-che-odiano-donnegogna-social-per-asia-argento_4KNSPMO49OoLtVvox04GWN.html
9 http://mashable.com/2017/10/18/asia-argento-harvey-weinstein-sexualharassment-slut-shaming/#YIIO
i.0cNaql
8

JADT’ 18

231

in Italia, non mollare».10 Da sempre impegnata in attività contro la violenza
sulle donne, da New York ha commentato al Corriere della Sera: “Non ho
avuto modo di chiamare Asia Argento perché sono in missione a New York e
in Canada. Le mando, però, questo messaggio: bisogna rimanere in Italia per
rafforzare la solidarietà tra donne. Asia non mollare”. Ha poi aggiunto
“Detesto il fatto che Asia Argento debba arrivare a giustificarsi […]. Questo è
il mondo alla rovescia, non è importante se e quando una donna decide di
denunciare un abuso. Queste sono sue scelte. Lo scandalo è che un uomo di
potere, questo Weinstein, si sentiva libero di saltare addosso alle ragazze che
volevano lavorare. Questo è il sistema marcio che va sradicato”. La stessa
presidente della Camera non è del resto estranea all’azione denigratrice del
web, che ne ha spesso fatto la destinataria di valanghe di insulti e parole
violente. Riporta, tra gli altri, l’intervento di Boldrini il quotidiano Libero,
che,11 il 19 ottobre 2017, titola Laura Boldrini: “Cara Asia Argento resta in Italia,
le donne sono con te” un articolo parco di commento ma nel quale la lingua
non rispettosa del genere e della morfologia della lingua italiana – su tutti la
presidenta – comunica ben più di quanto avrebbero fatto molte parole: “‘Per
quanto riguarda le molestie e gli stupri’, ha sottolinea[to n.d.r.] la presidenta,
‘il problema sono gli uomini e il loro comportamento […]’”.
2. Considerazioni finali
In attesa di uno scandalo a ruoli capovolti, che, da stereotipi culturali e
linguistici dominanti, ad oggi lascerebbe prefigurare tutt’altro genere di
commenti, ci si limiterà a una rosa di citazioni che se anche ampliata
notevolmente non riuscirebbe a spostare di una virgola – chi scrive ne è
convinta – lo stato di polarizzazione che si è venuto a prefigurare in Italia fin
dai primi giorni di diffusione della vicenda. Una polarizzazione oppositiva
che richiama quella tipica del tifo e più di recente della fede politica – che
sembra rendere incapaci di acquisire, anche solo provvisoriamente, una
prospettiva diversa, anche solo in parte, da quella originaria, – alla quale
nessun commento sembra potersi sottrarre. Ragion per cui, per evitare che
anche l’approccio descrittivo tipico dell’analisi del testo possa essere accusato
di faziosità da una o dall’altra parte, occorrerebbe ampliare il corpus di
riferimento di questo lavoro almeno con la disamina quantitativa e
qualitativa di tutti i tweet presenti nell’account di Asia Argento con
riferimento ai profili che li hanno generati; con la disamina almeno

10 https://www.vanityfair.it/news/cronache/2017/10/19/caso-weinstein-lauraboldrini-asia-argento
11 http://www.liberoquotidiano.it/news/politica/13266009/laura-boldrini-caraasia-resta-in-italia-donne-sono-con-te-minigonna-uomini.html

232

JADT’ 18

quantitativa dei segmenti e dei contesti in cui il termine vittima compare
esplicitamente o è richiamato in altro modo; con la disamina dei contesti e
delle forme cui si ricorre per parlare di chi ha offeso, con l’attività social
scaturita dalle cronache relative a momenti clou dell’anno in materia di
violenza o di rivendicazione di genere, nello specifico nei confronti delle
donne, quali la giornata contro la violenza sulle donne o l’8 marzo. Già
attuata a campione, la raccolta e la successiva analisi di messaggi mostra una
pervicace azione a ripetere impermeabilmente le proprie azioni
comunicative, tanto nei contenuti tanto nella forma e nelle costellazioni di
termini che accompagnano il focus di volta in volta oggetto di discussione.
Segno inequivocabile della posizione che gli elementi da cui si irradia la
costellazione stessa hanno nell’enciclopedia e nella coscienza e sensibilità
della comunità linguistica italofona.

JADT’ 18

233

Il cosa e il come del processo narrativo. L’uso
combinato della Text Analysis e Network Text
Analysis al servizio della precarietà lavorativa
Cristiano Felaco1, Anna Parola2
Università degli Studi di Napoli Federico II – cristiano.felaco@unina.it; anna.parola@unina.it

Abstract
This paper shows the analytic procedures in order to use jointly Text
Analysis and Network Text Analysis. Text Analysis allows to detect the main
themes subjects in the narrations and hence the processes of signification,
Network Text Analysis permits to track down the relations between
linguistic expressions of text, identifying therefore the path of flow of
thoughts. Using jointly the two methods is possible not only to explore the
content of narrations, but, starting from the words and concepts with higher
semantic strength, also to identify the processes of signification. To this
purpose, we will present a research aiming to understand high school
students’ perception of employment precariousness in Italy. The lexical
corpus was built by narrations collected from 2013 to 2016 in blog of
Repubblica “Microfono Aperto”.
Riassunto
Il lavoro presenta le procedure analitiche per un uso congiunto delle tecniche
di Text Analysis e Network Text Analysis. La prima permette di cogliere i
temi principali affrontati nelle narrazioni e quindi i processi di significazione,
la seconda di rintracciare le relazioni tra le espressioni linguistiche di un
testo, individuando i percorsi dei flussi di pensiero. L’uso combinato delle
due tecniche permette, dunque, non solo di esplorare i contenuti delle
narrazioni, ma, lavorando su parole e concetti con una maggiore carica
semantica, anche di ricostruire i percorsi attraverso i quali si costruisce il
significato. A tale scopo sarà presentata una ricerca volta a comprendere la
percezione degli studenti delle scuole secondarie superiori sulla precarietà
lavorativa in Italia. Il corpus testuale è stato creato a partire dalle narrazioni
raccolte dal 2013 al 2016 nel blog di Repubblica “Microfono Aperto”.
Keywords: Thematic Analysis of Elementary Contexts; Network Text
Analysis; Employment Precariousness; Students.

234

JADT’ 18

1. Introduzione
La narrazione, e più nello specifico il narrare, è un processo di costituzione di
una tessitura testuale dotata di senso e veicolante significati. Analizzare i
testi permette di cogliere da un lato la percezione di chi narra su un dato
argomento e il processo di significazione attribuita all’esperienza narrata, ma
dall’altro di comprendere i flussi di pensiero, entrando nello specifico delle
parole utilizzate e della loro sequenzialità. L’uso della statistica testuale al
servizio delle narrazioni permette, perciò, il riconoscimento in profondità del
significato delle parole e del senso ivi presente (Bolasco, 2005). Tra le tecniche
di analisi del contenuto, l’uso combinato della Text Analysis (TA) e Network
Text Analysis (NTA) si presta bene a questi scopi. Se la TA permette di
cogliere i temi affrontati, le parole scelte e utilizzate e le dimensioni di senso
attribuite (Lebart et al., 1998), il cosa si narra, l’uso della TNA offre un
ulteriore approfondimento sul come si narra. Analizzando, infatti, la
posizione delle parole all’interno della rete testuale è possibile rintracciare le
parole con una maggiore carica semantica, individuando in questo modo i
diversi percorsi e contesti di significato (Hunter, 2014) mediante lo studio
della natura delle relazioni tra i vari termini. Partendo dall’assunto che la
struttura di relazioni tra le parole di un testo possa corrispondere ai modelli
mentali e alle mappe cognitive messe in atto dagli autori del testo (Carley,
1997; Popping et Roberts, 1997), tale metodo permette di modellizzare il
linguaggio come rete di parole e di relazioni attraverso la creazione di una
mappa cognitiva (Popping, 2000). Il concetto è il nucleo (mentale) che viene
rappresentato attraverso un termine o un’espressione linguistica; i termini
possono essere in relazione tra loro formando un’affermazione. Le
affermazioni che condividono uno stesso concetto formano una struttura
interdipendente creando così una mappa concettuale o rete testuale costituita
da punti (o nodi) che rappresentano le singole parole (o concetti) e da linee,
cioè i legami che li collegano.
2. Metodologia
L’approccio proposto prevede dapprima che i testi prodotti siano sottoposti
ad un’analisi statistica dei dati testuali servendosi del software di analisi
automatica T-lab, e successivamente analizzati in una prospettiva di rete
mediante il software Gephi.
2.1 Pre-trattamento dei testi
Raggruppati all’interno di un unico corpus, la prima fase di lavorazione del
testo si compone di una fase di normalizzazione del corpus e di
personalizzazione del dizionario. La prima ha l’obiettivo di riconoscere le
parole come forme grafiche e ciò comporta una trasformazione del corpus

JADT’ 18

235

(eliminazione di spazi vuoti in eccesso, marcatura degli apostrofi, riduzione
delle maiuscole), e la creazione di stringhe per le locuzioni polirematiche,
insiemi di parole che hanno un significato unitario non desumibile da quello
delle parole che lo compongono, arrivando alla creazione delle multiwords. La
fase di personalizzazione del dizionario è effettuata con le procedure di
lemmatizzazione e disambiguazione del testo che permettono di rinominare
le forme grafiche in lemmi. Lo step della disambiguazione permette di
selezionare le forme omografe per disambiguarle; quello di lemmatizzazione,
partendo dal riconoscimento delle forme con la stessa radice lessicale
(lessema) o appartenenti alla stessa categoria lessicale, di ricondurre ogni
aggettivo e sostantivo al maschile singolare, ogni verbo alla forma di infinito
presente, e così via. Terminata questa fase, si procede al controllo delle
caratteristiche lessicali del corpus per comprenderne la trattabilità a livello
statistico, verificando i valori del type/token ratio, adeguato per un valore
inferiore a 0.2, e gli hapax, adeguato per una percentuale inferiore al 50% per
corpus di grandi dimensioni, e per percentuali leggermente superiori in caso
di corpus di medie o piccole dimensioni. Prima di procedere all’analisi, va,
inoltre, presa visione della lista delle parole chiave, creata con una procedura
automatica dal software, e alla loro occorrenza all’interno del corpus, e si
fissa una soglia di occorrenza minima, escludendo dall’analisi tutte le parole
presenti meno di n. volte. La scelta della soglia di occorrenza dipende dalle
caratteristiche lessicali e dalle dimensioni del corpus in analisi. Le parole
chiave possono dunque essere prese nella loro integrità, ridotte in relazione
alla soglia di occorrenza, o ancora ulteriormente ridotte in base agli scopi
della ricerca.
2.2. Analisi dei testi mediante Analisi Tematica dei Contesti Elementari
L’Analisi Tematica dei Contesti Elementari mediante una Cluster Analysis
permette di costruire ed esplorare i contenuti del corpus in analisi (Lancia,
2004). I cluster sono costituiti da un insieme di contesti elementari definiti
dagli stessi pattern di parole chiave e descritti attraverso le unità lessicali che
maggiormente vanno a caratterizzare i contesti elementari. La cluster
analysis è eseguita mediante un metodo gerarchico-ascendente non
supervisionato (algoritmo bisecting K-means), caratterizzato dalla cooccorrenza dei tratti semantici. Nello specifico, la procedura d'analisi è
costituita da: analisi delle co-occorrenze mediante la creazione di una tabella
dati unità di contesto*unità lessicali con valori di presenza/assenza; pretrattamento dei dati tramite TF-IDF e trasformazione di ogni vettore riga a
lunghezza 1 (norma euclidea); uso del coseno e clusterizzazione tramite
algoritmo bisecting K-means; analisi comparativa con creazione della tabella
di contingenza unità lessicali*cluster; test del chi-quadrato agli incroci

236

JADT’ 18

cluster*unità lessicali. Rispetto al criterio di partizione che determina il
numero dei cluster, viene utilizzato un algoritmo che utilizza il rapporto tra
varianza intercluster e varianza totale assumendo come partizione ottimale
quella in cui questo rapporto supera la soglia del 50%. L’interpretazione della
posizione occupata dai cluster nello spazio fattoriale e delle parole che li
caratterizzano permettono di individuare le relazioni implicite che
organizzano il pensiero dei soggetti, consentendo di cogliere il punto di vista
del narratore nei confronti dell’evento narrato. Quest’ultimo comprende
anche una serie di elementi valutativi, riflessioni, significati, giudizi di valore,
ma anche proiezioni affettive.
2.3. Analisi delle reti
Il secondo step d’analisi prevede l’inserimento del corpus all’interno del
software Gephi. Tale software organizza i vari lemmi in una matrice di
adiacenza (lemma*lemma) consentendo la creazione di una rete 1-mode, uno
strumento utile per visualizzare la struttura di relazioni tra i vari lemmi,
rappresentati da cerchi o nodi, e collegati tramite legami rappresentati da
linee direzionate. Tale tecnica permette di cogliere il modo con cui i nodi
sono connessi tra loro, identificando così le zone di vicinato (neighbourhood), e
individuando quei nodi che occupano una posizione di rilevanza in differenti
set o nell’intero network. A tale scopo, vengono calcolate differenti misure
basate sulla centralità e, tra queste, la degree centrality che indica le parole
usate con maggiore frequenza in connessione ad altre parole all’interno delle
narrazioni e nei vari contesti di significato. Più nel dettaglio, l’incidenza di
ogni nodo può essere espressa sia come in-degree, numero di archi entranti in
un punto, individuando in questo modo i cosiddetti “predecessori” di ogni
unità lessicale, sia come out-degree, numero di archi uscenti dal punto,
mostrando invece i “successori”. Tale relazione tra predecessori e successori
all’interno della rete testuale aiuta a comprendere la varietà semantica
generata dai nodi. Altro indice utilizzato è la betweennes centrality, misura di
centralità globale basata sulla vicinanza, che esprime il grado con cui un
nodo sta “fra” gli altri nodi del grafo. I nodi collocati in queste zone del
network eserciterebbero una funzione di controllo sui flussi informativi e di
“passaggio” permettendo il collegamento tra due o più set del network
(Freeman, 1979). Nell’ottica dell’analisi testuale, questi lemmi, infatti, giocano
un ruolo centrale nella circolazione dei significati all’interno della rete,
fungendo da punto di giunzione da cui si connettono zone diverse di testo e
si snodano specifici percorsi di significato, andando a definire in questo
modo la varietà semantica delle narrazioni.

JADT’ 18

237

3. Caso studio
Presentiamo uno studio condotto attraverso l’uso combinato delle tecniche
allo scopo di comprendere la percezione degli studenti del mondo del lavoro
nel contesto italiano. Gli ultimi dati disponibili mostrano che l’Italia è tra i
paesi europei con il più alto tasso di disoccupazione giovanile (Eurostat,
2017). L’instabilità, la precarietà e la discontinuità delle entrate rendono i
giovani vulnerabili ai cicli economici, modificando natura e tempi della
transizione al mondo del lavoro e riducendo le opportunità di sviluppare
soddisfacenti piani di vita (Leccardi, 2006). La sfiducia incide sui propulsori
della transizione, cioè sul mantenimento di aspirazioni elevate, sulla
cristallizzazione degli obiettivi di carriera e sul comportamento intensivo
della ricerca di un lavoro (Vuolo et al., 2012). Per lo studio abbiamo utilizzato
una fonte di dati testuali provenienti dal blog di Repubblica “Microfono
Aperto” in cui studenti delle scuole superiori, nel periodo dal 2013 al 2016,
hanno risposto al promt “Quattro giovani su dieci senza lavoro. E tu che
pensi? Di chi sono le colpe? Cosa vorresti che venisse fatto al più presto per
garantirti un dignitoso futuro?”. Raccontarsi attraverso la Rete agevola il
processo di riflessione su di sé, sul proprio ruolo e sul rapporto con ciò che
accade nel contesto in cui il giovane è inscritto. In una situazione di
malessere per la precarietà lavorativa, il web può essere un utile contenitore
per la condivisione dell'esperienza di precarietà, costituendo un ambiente di
condivisione e socializzazione delle proprie esperienze (Di Fraia, 2007).
3.1 Risultati
Il corpus conta 130 narrazioni (10110 occorrenze, 2484 forme grafiche, 1590
hapax), utilizzando come variabili descrittive la provenienza territoriale
(nord, centro, sud) e il tipo di istituto frequentato (istituto tecnicoprofessionale e liceo) e soddisfa i criteri statistici di trattabilità. L’analisi
tematica dei contesti elementari ha prodotto quattro cluster (Fig. 1; Tab. 1),
rinominati CL1 “Guardare le opportunità” (14,6%); CL2 “E il governo?”
(19,8%); CL3 “Dai sogni alla crisi” (38,5%); CL4 “La ricerca del lavoro, dove?”
(27,1%). Le narrazioni del cluster “Guardare le opportunità” rimandano
all’analisi di sacrifici e opportunità; emerge in modo marcato la necessità di
una “attività”, di una messa in pratica di azioni nel presente in vista di un
futuro migliore. Per questo motivo, la crisi è al tempo stesso un’opportunità
che i giovani devono cogliere per dimostrare le proprie capacità: Ormai, per
ciò che si sente, chiunque si chiede del proprio futuro. Per garantire che un giorno ci
sia più lavoro, si deve agire ORA. […]. Anche chi cerca lavoro, però, deve volare
basso e accontentarsi, per il momento, di poco, invece di restare a casa arreso.
Secondo me i giovani devono avere l'opportunità di dimostrare ciò che valgono,
dimostrare al mondo ciò che sanno essere e far capire a tutti che sono capaci "se si

238

JADT’ 18

impegnano" di fare qualsiasi lavoro, dal più semplice al più complesso. I testi del
secondo cluster sono maggiormente orientati alla ricerca della “colpa” e ad
una richiesta di soluzioni principalmente dallo Stato: Penso che lo Stato
dovrebbe dare più spazio ai giovani assicurando loro protezione e tutela. I
parlamentari devono conservare i diritti e le possibilità di ogni giovane, siamo noi il
futuro di questo stato, e come tali abbiamo bisogno di opportunità.
Il cluster “Dai sogni alla crisi” rimanda alla dimensione più interna
dell’essere immersi in una società che sta attraversando un momento di crisi
economica. Gli studenti rimarcano che la mancanza di lavoro annulla i sogni:
Sono davvero preoccupata, tutti noi sogniamo cosa fare da grandi e sapere che il
38,7% dei giovani non riesce a trovare lavoro mi rende indignata. I giovani sono il
futuro, il progresso, si impegnano […] Sappiamo tutti cosa dice il primo articolo
della nostra splendida costituzione, eppure sembra sia ignorato. Bisogna dare più
occasioni ai giovani, tenere in considerazione la nostra costituzione, per aprire le
porte al futuro e rendere l'Italia migliore. Le narrazioni dell’ultimo cluster
riguardano trasversalmente tutte le difficoltà del cercare lavoro (la ricerca
affannata, le aziende che non assumono a causa delle troppe tasse) e della
necessità di andare all’estero: L'Italia si ritrova in un periodo di profonda crisi e se
non si riprende economicamente ridando la possibilità a noi giovani di far capire a chi
di dovere che abbiamo le capacità e volontà di lavorare, l'Italia perderà tutti quei
giovani ma soprattutto tutte quelle menti che andranno all'estero in cerca di
condizioni di vita più favorevoli ma soprattutto di maggiori possibilità di lavoro.
La posizione delle variabili descrittive mostra una differenza per la variabile
provenienza territoriale e nessuna differenza per istituto frequentato. Se
infatti il frequentare una scuola piuttosto che un’altra sembra non incidere
sulla percezione del mondo del lavoro e sui vissuti di sfiducia, che sono
invece comuni, l’appartenenza territoriale ha un suo peso. La modalità nord
è, in termini di vicinanza, posta in prossimità dei cluster 1 e 4, il centro del 3 e
il sud del cluster 2. Ciò indica come gli studenti del nord tendano
maggiormente a problematizzare il fenomeno del precariato e la difficile
ricerca del lavoro, mettendo anche l’accento sulle opportunità che i giovani
hanno di dimostrare il proprio valore; le tematiche di quelli del sud vanno
maggiormente nella colpevolizzazione del contesto, in linea con una
maggiore risonanza del tema di discussione a causa di un’elevata incidenza
della disoccupazione giovanile; le narrazioni degli studenti del centro,
invece, maggiormente richiamano i propri vissuti interni.

JADT’ 18

239

Figura 1: Cluster Analysis

La rete prodotta è composta da 259 nodi e 414 legami. Una prima
approfondita forma di visualizzazione della struttura di relazioni tra i vari
lemmi mostra i livelli più alti di degree centrality, in cui “lavoro”, “giovani”,
“futuro”, “problema” e “possibilità” rappresentano i nodi con maggiori
connessioni. Inoltre, questi stessi nodi riportano anche i valori più alti di indegree centrality, nodi “assorbenti” che presentano più legami in entrata che
in uscita rispetto a tutti gli altri punti; gli studenti tendono a indirizzare i
propri discorsi e, più in generale, il flusso di pensiero verso le tematiche
relative al lavoro in termini sia di possibilità future sia analizzandone le
problematiche ad esso legate. Dall’altro canto, “impegnare” (inteso come
impegno messo in atto) e “condizioni” rappresentano il fulcro da cui muove
la narrazione verso altre parole, nodi “sorgente” che hanno più legami in
uscita che in entrata rispetto ai restanti nodi della rete. I lemmi che
rimandano ai vissuti degli studenti, ai propri stati d’animo rispetto all’attuale
condizione e ad una prospettiva lavorativa futura incerta sono quelli che
giocano un ruolo centrale nella circolazione dei significati all’interno della
rete, presentando difatti i valori più elevati di betweenness centrality. In
particolare, “disoccupato”, “costringere”, “rimanere” e “scoraggiare” sono i
nodi che fungono da principale punto di giunzione da cui si snodano
specifici percorsi di significato: le diverse zone del network, e quindi diverse
parti della narrazioni sono collegate tra loro da quei lemmi che ruotano
intorno al tema della precarietà del presente, una situazione di costrizione e
di forte scoraggiamento.

240

JADT’ 18

In-degree Centrality

Out-degree Centrality

Betweenness Centrality
Figura 2

4. Conclusioni
L’uso misto della TA e NTA permette di rappresentare un quadro sintetico
della struttura semantica, comprendere di cosa si parla, ma anche in che
modo lo si fa: la scelta delle parole e l’ordine stesso di presentazione di
un’idea o opinione rispetto al tema in oggetto. L’uso congiunto delle due
tecniche fornisce: a) una sintesi delle informazioni contenute nelle narrazioni;
b) l’analisi dei temi affrontati; c) un focus sulla strutturazione delle frasi in
termini di relazioni tra lemmi. Permette così di mettere in relazione categorie
tematiche e di contenuto in quanto struttura latente, ricostruendo a ritroso il
processo discorsivo.
Bibliografia
Bolasco S. (2005). Statistica testuale e text mining: alcuni paradigmi
applicativi. Quaderni di Statistica, vol. 7: 1-37.

JADT’ 18

241

Carley K.M. (1997). Extracting team mental models through textual analysis.
Journal of organizational behavior, 18(1): 533-558.
Di Fraia G., a cura di, (2007). Il fenomeno blog. Blog-grafie: identità narrative in
rete. Milano: Guerini e Associati.
Eurostat (2017). Statistics on young people neither in employment nor in
education or training. Report.
Freeman L.C. (1979). Centrality in Social Networks Conceptual Clarification.
Social Networks, vol. 1: 215-239.
Hunter S. (2014). A novel method of network text analysis. Open Journal of
Modern Linguistics, vol. 4(2): 350–366.
Lancia, F. (2004). Strumenti per l’analisi dei testi. Milano: Franco Angeli.
Lebart L., Salem A. and Berry, L. (1998). Exploring textual Data. Dordrecht:
Kluwer Academic Publishers.
Leccardi C. (2006). Redefining the future: Youthful biographical
constructions in the 21st century. New directions for child and adolescent
development, vol. 113: 37-48.
Popping R. (2000). Computer-assisted Text Analysis. London: Sage.
Popping R. and Roberts C.W. (1997). Network approaches in text analyisis. In
Klar R. and Opitz O., editors, Classification and knowledge organization.
Berlin, New York: Springer.
Vuolo M., Staff J. and Mortimer, J. T. (2012). Weathering the great recession:
Psychological and behavioral trajectories in the transition from school to
work. Developmental psychology, vol. 48(6): 1759.

242

JADT’ 18

Hablando de crisis: las comunicaciones del Fondo
Monetario Internacional
Ana Nora Feldman
Universidad Nacional de Luján – anafeldman@gmail.com

Abstract
The annual reports of the International Monetary Fund issued annually
under the name of “World Economic Outlook" from the years 2005 to 2012,
are analyzed in this Paper by using the techniques of Statistical Analysis of
Textual Data. The scan tool text, allows us to see the way the IMF describes in
their reports the world crisis, highlighting their strengths and weaknesses in
their role of the ultimate guarantor of global economic balance. Much has
been discussed about the foresight of the crisis and what was the position of
the IMF regarding its consequences. The denial of the crisis, only recognized
in 2010, is consistent with the mission that the International Monetary Fund
considers to carry out, lecturing on how governments should correct their
economies (Weisbrot et al., 2009). All this ignoring that "their prescriptions
failed" (Stiglitz, 2002) as their "structural adjustment policies" … "produced
hunger and unrest" benefiting those who had more resources while "the poor
sometimes sank more and more in misery. " In particular what is analyzed
from the processing of textual corpus with Taltac2 software, developed by
Prof. Sergio Bolasco from the Università di Roma "La Sapienza", are the
concepts and language associated as a contribution to "a significant debate on
a variety of exclusions "..." that encompass the political, economic and social
fields"(Sen et Kliksberg, 2007) and considering that the World Economic
Outlook reports may be useful for understanding the behavior of the IMF in
the context of the financial crisis. The texts analyzed are written by
technicians and bureaucrats, who possess a high level of expertise and
skillful management of common codes, and are the product of a clear
intention on how the global economic situation and the role of the Monetary
Fund (and technicians), within this context, must be read. These reports, as
will be demonstrated meet the goal of preaching the hegemonic conception
on markets and policies, seeking to satisfy goals related to communication
and marketing strategies in order to align public opinion, government
officials and government objectives behind this concept. It is along this line
that the contradictions between the more political text (the introduction and
the summary) and the technical text (the body of the publication) are also
shown.

JADT’ 18

243

Resumen
Con la ayuda de técnicas de Análisis Estadístico de Datos Textuales, se
analizan los informes anuales del Fondo Monetario Internacional que se
publican anualmente con el nombre de “Perspectivas de la Economía
Mundial” entre los años 2005 y 2012. Se trata de evidenciar en los textos la
forma en la que describe el FMI a la crisis, poniendo en evidencia sus
fortalezas y debilidades en su rol de último garante del equilibrio económico
mundial. Mucho se ha discutido acerca de la capacidad de previsión de la
crisis y cuál fue la posición del Fondo Monetario respecto de sus
consecuencias. La negación de la crisis, sólo reconocida en el año 2010, es
coherente con la misión que el FMI considera que debe cumplir, aleccionando
sobre la forma en que los gobiernos deben corregir sus economías (Weisbrot
et al., 2009). Todo esto ignorando que “sus recetas fallaron” (Stiglitz, 2002)
pues “las políticas de ajuste estructural”… “produjeron hambre y disturbios”
beneficiando a quienes poseían más recursos mientras que “los pobres en
ocasiones se hundían aún más en la miseria”. En particular se analizan con la
ayuda de Taltac2, desarrollado por el Prof. Sergio Bolasco de la Università di
Roma “La Sapienza”, los conceptos y el lenguaje asociado como aporte a “un
debate significativo acerca de una variedad de exclusiones” … “que abarcan
el campo político, económico y social” (Sen et Kliksberg, 2007) para
comprender el comportamiento del FMI en el contexto de la crisis financiera.
Los textos analizados son escritos por técnicos y burócratas, que poseen un
alto nivel de especialización y un manejo hábil de códigos comunes, y son
producto de una clara intencionalidad acerca de cómo debe leerse la
situación económica mundial y el rol del Fondo Monetario (y sus técnicos) en
dicho contexto. Estos informes, como se demostrará, cumplen con el objetivo
de predicación de la concepción hegemónica, sobre mercados y políticas,
buscando satisfacer objetivos relacionados con estrategias comunicacionales
y de marketing con el objetivo de alinear a la opinión pública, funcionarios y
gobiernos detrás de esa concepción. En esa óptica es que se muestran
también las contradicciones entre el texto más político (la introducción y el
resumen) y el texto técnico (el cuerpo de la publicación).
Keywords: textual data analysis, content analysis, political language,
economic and financial crisis.
1. Introducción
La crisis económico – financiera que comenzó en Estados Unidos en el año
2007, y que luego se extendió a Europa y otros continentes, fue reconocida de
manera tardía por parte del Fondo Monetario Internacional (FMI).
Considerando que la misión del Fondo es la de prever los riesgos originados
en crisis económicas y brindar recomendaciones acerca de los mecanismos de

244

JADT’ 18

mitigación, la pregunta que se impone es ¿por qué, ante la crisis financiera de
mayor envergadura después de la Gran Recesión de 1930, el Fondo ignoró la
crisis, evitando declarar la emergencia de envergadura mundial? Desde el
punto de vista político (y discursivo), al negar la crisis, el FMI impidió la
puesta en marcha los mecanismos previstos para afrontar problemáticas de
semejante envergadura. En este trabajo se analizan, con técnicas de Análisis
de Datos Textuales, los informes anuales (Perspectivas de la Economía
Mundial) publicados durante 8 años (2005-2012). Congruencias y
contradicciones nos permiten analizar, desde un punto de vista diferente, las
estrategias políticas del Fondo Monetario que ha visto muy desgastada su
imagen como recurso válido e idóneo para el salvataje de economías en
peligro.
2. Corpus
El criterio para la elección del período en análisis es el de relevar información
en momentos diferentes de la crisis. Partiendo desde un “momento 0”
(previo a su aparición), pasando por la instancia de reconocimiento del
estado de situación, para finalmente considerar el cambio más importante en
la política llevada adelante hasta ese momento por parte del FMI, es decir el
paso del paradigma neoliberal “no intervencionista” (ninguna acción por
parte del Estado para que el mercado se regule solo) a una política activa de
ayuda por parte de los gobiernos (de Estados Unidos y de la Unión Europea),
para “salvar las principales empresas, compañías y bancos en quiebra”
(Rapoport et Brenta, 2010). Desde una óptica de análisis del contenido
(Krippendorf, 1969), se realiza un análisis comparativo de dichos informes,
buscando conocer cuál ha sido la forma en la que el FMI ha descrito la crisis y
cuáles son las temáticas asociadas a la misma. La hipótesis, es que este
lenguaje y contenido no neutral de criterios técnicos y políticos, responden al
acuerdo de la que hemos llamado comunidad internacional “de peso real”
(Feldman, 1995).
3. Ocho años de discursos del Fondo Monetario Internacional
Ya hemos trabajado y presentado diferentes aspectos relacionados con las
comunicaciones del Fondo Monetario ante la crisis más importante tanto
económica como financiera. Discursos que dependen del Director General de
turno y el uso de la lexicometría como herramienta para la interpretación de
los informes (Feldman, 2015 a y b).
En este trabajo analizaremos las cuestiones relacionadas con la congruencia y
el uso político que se da en estas publicaciones anuales. La ambigüedad del
discurso, la dificultad de previsión y reconocimiento (o negación) de la
misma, sus causas y consecuencias y los reiterados anuncios del fin de la

JADT’ 18

245

crisis (en los años 2012, 2013 y 2014) que han sido objeto de crítica por todos
los bloques de países más o menos cercanos al FMI.
El objetivo entonces es identificar las posiciones del Fondo Monetario
Internacional en el tiempo. Se trata de comprender cómo habla y cómo calla
el FMI sobre este crucial tema, como aporte a “un debate significativo” sobre
exclusiones que “abarcan el campo político, económico y social” (Sen et
Kliksberg, 2007). Subyace a esta propuesta la idea que la exploración y el
análisis de textos, mediante recursos de estadística exploratoria
multidimensional, permite “una concepción ecológica para el tratamiento de
datos cualitativos” (Bolasco, 2007). El software utilizado es TALTAC.
3.1. El Discurso del FMI
El corpus está constituido por un total de 1.056.336 palabras (u ocurrencias).
Se trata de textos largos (más de 300 páginas incluyendo gráficos y tablas)
con un promedio de 132.042 ocurrencias. Si bien la distribución entre años es
aproximadamente similar, el informe del 2008 se distingue pues concentra el
16% del total de ocurrencias.
Tabla 1 – Análisis Lexicométrico

Así como el año 2008 se destaca por su extensión el del 2009 es el que utiliza
una mayor riqueza de vocabulario. Según nuestra experiencia (Feldman,
1995), la utilización de una cantidad elevada de palabras en un informe
podría estar indicando una situación de “malestar” o bien del uso de
lenguaje “desvirtuado”. Es decir, se deben utilizar más palabras para
describir algo que aún no ha sido consensuado entre los técnicos y, por
consiguiente, no ha sido conceptualizado adecuadamente.

246

JADT’ 18
Tabla 2 – Riqueza de Vocabulario

La distribución en los años de la forma “crisis” es lo suficientemente
ilustrativa acerca del uso dado, por parte del FMI, al correr de los años.

Gráfico 1 – Distribución de la forma “crisis” en el tiempo

3.2. Dos niveles de análisis: año por año los informes del Fondo
Si tomamos en cuenta sólo la Introducción y el Resumen Ejecutivo (a los que
llamaremos “textos políticos”), que preceden al cuerpo del informe técnico
(más de 300 páginas de textos y números) de cada Informe (a los que
llamaremos “textos técnico-económicos”), éstos pueden ser considerados
piezas comunicacionales que tienen un alcance público mayor, pues existe

JADT’ 18

247

una amplia gama de públicos que “consumen” los documentos técnicos del
FMI (periodistas económicos, economistas, público en general) pero que
normalmente no leen los informes completos. Muchas veces son justamente
estos escritos sintéticos, los que tienen un efecto mayor en la modelación de
la opinión pública internacional. ¿Existen entonces diferencias y/o
inconsistencias entre los informes considerados integralmente y los
resúmenes ejecutivos e introducciones? A través de la lectura de los mismos
y el análisis de las principales formas estadísticamente significativas
comentaremos diferencias y similitudes entre estos. Sin presagiar ninguna
crisis, tanto en el año 2005 como en el 2006, en sus textos se registra coherencia
económica a partir de la sintonía entre los contenidos de la primera parte con
aquellas formas estadísticamente significativas del documento técnico:
INFLACIÓN, INVERSIÓN, AHORRO (2005), PRODUCTIVIDAD y
SECTORES PRODUCTIVOS (2006). En el año 2007, el del comienzo de la
crisis el FMI comienza a hablar de un “período incierto y difícil” y las
palabras estadísticamente significativas hacen referencia sobre todo a la
VOLATILIDAD, contemporáneamente habla de crecimiento, registrándose
disonancia económica entre ambas partes. El año 2008, como ya señalado más
arriba es el que concentra el 16% del total de ocurrencias del corpus. Nos
encontramos aquí ante una disonancia discursiva / económica con el uso de
muchos términos no habituales del FMI (VIVIENDA y CAMBIO
CLIMÁTICO) para la descripción de la situación económica (disonancia y/o
incongruencia en el uso de términos, cfr. Feldman, 1995). Ya estallada la crisis
en el año 2009, a partir de la presión internacional, el FMI debe comenzar a
explicar aquello que no previó ni anunció (ver gráfico 1 y Tabla 2).
Encontramos mayor disonancia entre texto y contexto y nuevas formas
significativas (DESPLOME, ALARMAS). Intentando retomar el liderazgo
político, luego de haber sufrido numerosas críticas por su falta de previsión
de la crisis, el FMI, durante el año 2010, donde – entre su parte sintética y el
documento técnico – encontramos coherencia política y disonancia
económica. Entre las formas significativas encontramos CRISIS.
A partir del año 2011 en el que encontramos más distancia entre lo que se lee
en la Introducción y el Resumen Ejecutivo y el contenido del Informe
completo, reaparece la política. Una vez recuperado su espacio institucional y
su razón de ser, los textos del 2012 poseen coherencia tanto política como
económica.
5. Conclusiones
El Fondo realiza una lectura de los indicadores económicos contradictoria,
con una visión poco clara acerca de la gravedad y las consecuencias de esta
crisis. El análisis del contenido de los textos (discursos e informes), con el uso

248

JADT’ 18

de herramientas de estadística textual, permite graficar de manera irrefutable
las contradicciones y los silencios en los que incurre el FMI desde los
primeros síntomas de la crisis en el año 2007. Los conceptos entonces
vertidos en los Informes Perspectivas de la Economía Mundial son el
producto “de una curiosa mezcla de ideología y mala economía, un dogma
que en ocasiones parecía apenas velar intereses creados” recomendando
“soluciones viejas, inadecuadas” con brutales efectos “sobre los pueblos de
los países a los que se aconsejaba aplicarlas” (Stiglitz, 2002). Estas recetas
fallaron en muchas oportunidades y produjeron situaciones sumamente
graves en varios países. Un mensaje, un emisor, un objeto y una misión que
falló, pues el FMI no cumplió con su rol de evitar que el mundo caiga
nuevamente en una nueva Gran Depresión. Los textos analizados permiten
establecer algunas pistas acerca de las motivaciones de este fracaso. En las
contradicciones evidenciadas y en los intentos de negación de una realidad
que no dejaba dudas acerca de la magnitud de esta crisis se afianza la idea de
que existe en el Fondo Monetario y otros organismos internacionales un
problema de gobernanza
Tabla 3 – Análisis de coherencia y disonancia de los Informes año por año

JADT’ 18

249

Bibliografía
Bolasco S., D’Avino E. y Pavone P. (2007) Analisi dei diari giornalieri con
strumenti di statistica testuale e text mining, Publicado en I tempi della vita
quotidiana. Un approccio multidisciplinare all'analisi dell'uso del tempo. ISTAT,
Roma
Feldman, A. (1995), Il concetto di sviluppo umano secondo le Nazioni Unite:
analisi del contenuto in Bolasco, S., Lebart, L. e Salem, A. (eds.). JADT
1995 - Analisi statistica dei dati testuali, Roma, CISU, 2 voll.
Feldman, A. (2015a) Análisis del Posicionamiento del Fondo Monetario
Internacional frente a la crisis del año 2007 en Revista Latinoamericana de
Opinión Pública. Año 2016, número 6, EDUNTREF. Buenos Aires
Feldman, A. (2015b) Text Mining Strategies applied on the annual reports of
the International Monetary Fund. A look at the crisis en ISI 2015 World
Statistics Congress, Rio de Janeiro
Krippendorff, K. (1969). Theories and Anlytical Constructs en: G. Gerbner,
O.R. Holsti, K. Krippendorff, W.J. Paisely y P.J. Stone (eds.) The Analysis
on Communication Content, New York, John Wiley & Sons, p. 6 e ss.
Lebart, L y Salem, A. (2008). Statistique Textuelle, Dunod, Paris.
Nemiña, Pablo. (2009) Aportes para un esquema de análisis del
comportamiento del FMI en crisis financieras a partir de su actuación
durante la crisis argentina (2001-2002). Documentos De Investigación Social
Número 8. ISSN 1851-8788. IDAES, UNSAM, Buenos Aires
Rapoport, M. y Brenta, N. (2010). Las grandes crisis del capitalismo
contemporáneo. Capital Intelectual. Buenos Aires.
Sen, A. y Kliksberg, B. (2007). Primero la Gente. Ediciones Deusto. 9na edición
Editorial Temas, Buenos Aires, Argentina.
Weisbrot, M., Cordero, J. y Sandoval, L. (2009). Empowering the IMF: Should
Reform be a Requirement for Increasing the Fund’s Resources? Center for
Economic and Policy Research. Washington, D.C., Estados Unidos
www.cepr.net

250

JADT’ 18

Brexit in the Italian and the British press: a bilingual
corpus-driven analysis
Valeria Fiasco
Università Roma Tre – valeria.fiasco@gmail.com

Abstract 1 (English)
The spread of English as the Lingua Franca of international communication
has given rise to meaningful language contact phenomena in the world’s
languages like loanwords and pseudo-loanwords, namely, words from one
language (the donor language) are adopted by another language (the
recipient language) sometimes becoming naturalized (Gusmani 1973). From
this perspective, it is thus interesting to observe their behaviour in real
language use. In particular, this study investigates Anglicisms and pseudoAnglicisms found in the newspaper discourse of Brexit by way of a bilingual
corpus collected from two Italian newspapers, i.e. La Repubblica and Il Corriere
della Sera and two British newspapers, i.e. The Independent and The Guardian
selected for both their authoritativeness and their extensive readership. The
exit of the United Kingdom from the European Union was chosen because it
is a widely covered topic both in the Italian and in the British press, thus
providing abundant material for comparative analysis, as well as offering
useful data in order to explore linguistic variation. It was useful for building
an electronic corpus which was retrieved from the digital archives of the
newspapers’ websites in order to carry out an automated text analysis.
The corpus includes articles collected during the periods that both preceded
and followed the Brexit referendum. In order to carry out the analysis,
corpus-driven methodology was used, namely an approach that lets
hypotheses emerge from corpus observation (Tognini-Bonelli 2001). The
investigation was carried out by way of the software TalTac2, and the
automated text analysis, as a result, turned out to be invaluable in order to
investigate and monitor the newspapers’ vocabulary which included
technical terms from the fields of politics, economics and finance as well as
general language words. In order to design and sample a representative
corpus, the parameters proposed by Biber (1993) were used to identify
descriptive criteria so as to select and balance the population.
The aim of this study is to get an overview of the Brexit discourse as used in
the two countries' newspapers’ vocabulary and terminology (of the two
countries) by using text mining to compare and categorize the whole corpus
as a collection of texts and, then, to cluster documents on the basis of the

JADT’ 18

251

lexical similarity of the vocabulary to establish semantic fields or conceptual
areas. Furthermore, by way of the lexical and textual analysis, this study also
investigates Anglicisms and pseudo-Anglicisms in the Italian newspapers,
identifying and analyzing a list of English words used in Italian. The two
British newspapers serve as a reference corpus to compare to the list of
Anglicisms extracted from the Italian corpus. The articles retrieved from the
British newspapers serve to find out which words are typical of each corpus
and to identify pseudo-anglicisms, namely new words that seem to be
English forms, even though they do not exist in English, or if they do exist,
they have a clearly different meaning. Lastly, the data gathered from the
bilingual corpus analysis were later compared with other wider corpora
included in SketchEngine and on the Brigham Young University platform in
order to make generalizations about the distribution of Anglicisms and
pseudo-Anglicisms in general language corpora.
Keywords: Bilingual Corpus, Textual Analysis, Anglicism, Linguistic
Interference
Abstract 2 (Italian)
La diffusione e l’affermazione dell’inglese come lingua franca della
comunicazione internazionale ha generato fenomeni significativi di contatto
linguistico come i prestiti e i falsi prestiti, ossia parole originariamente nate in
una lingua modello che entrano a far parte di un’altra lingua (lingua replica)
alla quale vengono talvolta assimilate e adattate (Gusmani 1973). È quindi
interessante osservarne l’uso e l’andamento in testi autentici che presentano
la lingua nel suo uso corrente. Questo studio analizza gli anglicismi e i falsi
anglicismi nel discorso giornalistico della Brexit, attraverso un corpus tratto
dai quotidiani italiani La Repubblica e Il Corriere della Sera e dai quotidiani
britannici The Guardian e The Independent, che sono stati selezionati per la loro
diffusione e la loro autorevolezza. La scelta della tematica dell’uscita del
Regno Unito dall’Unione Europea è stata dettata da diversi fattori, tra i quali
l’ampia diffusione dell’argomento nella stampa italiana e in quella britannica,
dando la possibilità di creare un corpus per realizzare un’analisi comparativa
attraverso l’esplorazione della variazione linguistica. Dal momento che
queste riviste offrono una versione online che mette a disposizione un
archivio digitale consultabile, sono particolarmente adatte per creare un
corpus che può essere esaminato attraverso l’analisi automatica del testo. Il
corpus è composto da articoli raccolti durante il periodo che precede e segue
il referendum della Brexit e la metodologia utilizzata per condurre l’analisi è
di tipo corpus-driven, ossia un approccio esplorativo in cui, partendo
dall’osservazione del corpus, si arriva alla formulazione delle ipotesi

252

JADT’ 18

(Tognini-Bonelli 2001). Il software TalTac2 e l’analisi automatica dei testi
sono stati estremamente preziosi per esaminare e monitorare il lessico della
stampa che include termini tecnici della politica, dell’economia e della
finanza, insieme a parole che fanno parte del lessico comune. Per progettare
il corpus, sono stati utilizzati i parametri proposti da Biber (1993) con lo
scopo di identificare i criteri descrittivi per selezionare e bilanciare la
popolazione all’interno del corpus. L’obiettivo di questa ricerca è offrire
un’analisi del lessico e della terminologia utilizzata nel discorso sulla Brexit
nei quotidiani italiani e inglesi attraverso il text mining per raffrontare i testi
che compongono il corpus, categorizzarli e raggrupparli sulla base di
somiglianze lessicali per individuare i campi semantici e le aree concettuali.
Inoltre, l’analisi lessicale e testuale ha consentito l’identificazione degli
anglicismi e dei falsi anglicismi nei quotidiani italiani, mentre il corpus dei
quotidiani britannici ha svolto la funzione di corpus di riferimento per
paragonare la lista degli anglicismi estratta dal corpus italiano con i dati
raccolti nel corpus britannico, capire quali parole sono tipiche di ogni lingua
e identificare i falsi anglicismi, vale a dire parole che presentano una forma
inglese, che però non esistono nel vocabolario originario o nel caso in cui
esistano, il loro significato è completamente differente. Infine, i dati raccolti
dall’analisi del corpus bilingue sono stati successivamente confrontati con
altri corpora più ampi, consultabili su SketchEngine e sulla piattaforma della
Brigham Young University con lo scopo di fare delle generalizzazioni sulla
distribuzione degli anglicismi e dei falsi anglicismi in corpora non
specialistici.
Parole chiave: Corpus bilingue, analisi testuale, anglicismo, interferenza
linguistica
1. Introduction
The growing influence of English on many languages in the world represents
the linguistic change produced by language contact. English is used in both
academic and professional settings revealing a pervasive presence of
Anglicisms in European languages (Marazzini & Petralli 2015). This situation
can be traced back to economic and trade developments, as well as political
and social circumstances in the past decades. The Anglo-American
globalization also exerts an influence on language with an increasing number
of EFL (English as a Foreign Language) and ESL (English as a Second
Language) learners and the English use as a Lingua Franca (ELF) for
international communication giving rise to the borrowing of an increasing
number of Anglicisms which have thus become the symbol of the American
lifestyle, an expression of symbols, dynamism and progress. Pulcini, Furiassi,

JADT’ 18

253

Rodríguez Gonzàlez (2012:1) use the term Anglicization to stress the growing
extensive research on lexical borrowing which has had a major impact on
vocabulary and phraseology of English origin. Lexical borrowings adapt to
their receiving language in various ways, from occasional coinages to
integrated words, from more restricted circles to broad groups until reaching
the totality of the speakers of the recipient language. Gusmani (1993:28)
states that there are cases of complete acclimatization in which the speakers
of the recipient language become so used to the foreign word that it is
perceived to be part of the recipient language, i.e. film. One of the main
sources of neologisms and borrowings is from newspapers and magazines
which detect the emerging trends in contemporary language and coin new
words in a creative fashion. According to Beccaria (1983:65), newspapers are
one of the main forums of exchange between written and spoken language,
where different varieties coexist, for example, bureaucratic, technical and
literary language. Moreover, in newspapers, the interaction between the
general and specialized language takes place allowing specific terms to
penetrate the popular culture (Cabré 1999:17).
2. Research design
This paper stems from the assumption that the linguistic interference of
English on Italian brings about significant effects giving rise to lexical
borrowing phenomena like Anglicisms and false Anglicisms, especially in
newspaper language. This bilingual corpus-driven analysis describes both
the Italian and the British discourse of Brexit with the aim of analyzing its
vocabulary and terminology as used in both the Italian and the British press.
By way of text mining, patterns and trends that allow us to make connections
between the two languages under investigation can be discovered. We can
identify Brexit’s main themes, get a picture of how corpus data are shaped
and subdivided into text fragments that correspond to the newspaper
article’s sections (title, subtitle, summary, text). We can investigate the
linguistic interference of English on Italian and the markedness between the
Anglicisms/pseudo-Anglicisms retrieved in the Italian newspapers and their
Italian equivalent words.
The exit of the United Kingdom from the European Union was chosen
because it is a historic and momentous event which has been the focus of
attention of numerous newspapers, thus, providing abundant material to
collect in the corpus. The reason behind the choice of the two languages lies
in the linguistic interference phenomena they are closely involved in: English
performs the role of a highly productive donor language, while Italian is a
recipient language which is under the influence of English.
The bilingual corpus is made up of articles retrieved from two Italian

254

JADT’ 18

newspapers, i.e. La Repubblica and Il Corriere della Sera and two British
newspapers, i.e. The Independent and The Guardian. They were selected for
their authoritativeness, their extensive readership and the possibility to
access their on-line archives with a free subscription. Moreover, they all dealt
with the Brexit issue thoroughly. The corpus was compiled by downloading
and storing all the articles about Brexit published in the on-line versions of
these newspapers from June to October 2016, that is, the period that preceded
and followed the Brexit referendum. The selected articles provide a brief, but
detailed overview of the Brexit, even though they are not representative of all
of the Italian and the British press. The corpus is composed of two corpora,
the Italian and the British one. The Italian corpus includes 42 articles from La
Repubblica and 42 articles from Il Corriere della Sera for a total amount of
51,158 tokens, whereas the British corpus includes 31 articles from The
Guardian and 31 articles from The Independent for a total amount of 49,995
tokens. However, a difference can be observed in the number of articles that
make up the overall corpus, because the average length of the British articles
was shorter than that of the Italian ones. On the whole, the corpus includes
146 articles and 101,153 tokens. The corpus was designed and sampled
according to the parameters proposed by Biber (1993) in order to build up a
representative corpus and to identify descriptive criteria so as to select and
balance the population. The issue of whether a corpus is representative and
reliable is essential, because the information included in the corpus and the
way it is constructed is central in the corpus-driven approach, namely a
method that lets hypotheses emerge from corpus observation (TogniniBonelli 2001). The automated text analysis on the corpus was carried out by
way of the software TalTac2, to investigate the newspapers’ vocabulary, to
observe the behaviour of Anglicisms’, as well as to make a detailed bilingual
analysis. In order to make generalizations about the distribution of
Anglicisms and pseudo-Anglicisms in general language and to retrace their
routes from/into the donor and the recipient language, other general
language corpora were consulted: Sketch Engine (British National Corpus,
itTenTen16 and enTenTen13 corpus) and the online corpora available on the
Brigham Young University website (News on the Web – NOW, Global WebBased English – GloWbE, TIME Magazine Corpus). Furthermore, the software
Iramuteq was used to carry out the cluster analysis of both corpora, to map
them and extract the semantic associations of words according to their
similarity.
3. Results
In order to identify the main themes and semantic fields of the corpus, the
cluster analysis grouped its lexical content so as to maximize the similarity or

JADT’ 18

255

the dissimilarity of different groups of words. The analysis divided the
Italian and the English corpus into 4 homogeneous clusters whose topics are
economics and finance or European and British politics. The output graph
was a dendrogram showing the association of all the words included in the
two corpora according to their similarity. It grouped the words into two
clusters: the first one concerns economics/finance and the second one is
related to politics. The percentage of words included in the Italian economics
cluster equals 31% compared to 23% in the English economics cluster. In both
corpora, the words from the semantic field of economics are homogeneously
distributed, i.e. bank/banca, market/mercato, growth/crescita, fund/fondo,
investor/investimento, rate/tasso. As for the politics cluster, both corpora
subdivide the lexical content into three clusters. In the Italian corpus, the
cluster of politics generates cluster 4 (23%) grouping the words concerning
the British politics and the sub-clusters 1 and 3. Sub-cluster 1 (22%) regards
the European politics and the Brexit referendum, i.e. Unione, europeo, UE,
negoziati, uscire, trattativa, while sub-cluster 3 (23%) is related to European
policies linked to political integration and post-Brexit immigration policies,
i.e. difesa, migrare, integrazione, emergenza. In the English corpus, the cluster of
politics generates cluster 1 (26%) that corresponds to Italian cluster 3, i.e.
movement, immigration, person, European and two sub-clusters (2 and 3) about
the British politics. In particular, sub-cluster 3 is about the Leave campaign,
i.e. Ukip, independence, break, Farage whereas sub-cluster 2 is about the Remain
campaign of the United Kingdom in the European Union, i.e. Cameron,
conservative, labour, tory. Moreover, the dendrogram also shows who the main
actors of this event are: the European Union, David Cameron, Nigel Farage,
Theresa May, Boris Johnson, and Jeremy Corbyn.
By way of its textual analysis, the software TalTac2 also identified the words
occurring within specific text fragments in which the corpus has been
subdivided and labelled, i.e. headline, sub-heading, lead, body. This analysis
particularly focused on the headlines. On the whole, the most frequent lexical
word in both corpora, Brexit, is mainly found in the headlines and in the
body of Italian newspapers, while it can only be observed in the body of the
British press. The concept of “exit, leaving the European Union” mainly
appears in the body of the articles of the British press, while in Italian
newspapers it is predominantly found in headlines. The brief exploration of
the headlines starts with the key topics expressed by the nouns in both the
Italian and the English corpus. The topics refer to the domain of politics, the
governance of the UK, the debate and the negotiations between the two
parties and the problems arising from the exit of the United Kingdom from
the European Union (i.e. referendum, European Union, leader, government,
campaign, support/negoziato, collasso, rischio, leader, rischio, referendum). In

256

JADT’ 18

particular, the most recurrent nouns in both the English and Italian headlines
mirror the themes addressed in the two corpora, i.e. politics: Brexit, EU
referendum, Remain, vote/Brexit, premier, uscita; economics and finance: borsa,
sterlina/pound. As for verbs describing the actions, conditions or experiences
linked to the Brexit, they outline a delicate and unstable situation in both
corpora, i.e. to vote, to fail, to resign, to face, to divide/uscire, crollare, affrontare,
rischiare, intervenire.
As far as the analysis of the linguistic interference is concerned, the Italian
corpus includes 174 Anglicisms (types) for a total amount of 1.096
occurrences (tokens) whose percentage in the corpus is about 2.1%. As to
types, their sum includes a lot of hapax legomena 91 out of 174 Anglicisms to
be exact (approximately 52.3% of types). The 174 Anglicisms belong to the
semantic fields of politics (22.5%), economics (27.5%), general language
(45.5%), and newspaper language (4.5%). The list of Anglicisms extracted
from the Italian corpus was later compared with the British one to check
whether they were actually used in English and how: 81 Anglicisms out of
174 were found in the English corpus. The other 93 Anglicisms are real
English words except for neo-premier (58.64 per million words) which can be
defined as a pseudo-Anglicism. It is a loanblend or a hybrid compound
(Furiassi 2010:40) formed by the English word premier and the Greek-derived
suffix neo-. These two lexical elements are individually used in English, but
they are not used together. The suffix neo- can be found in English
compounds referring to political movements like neo-socialist, neo-fascist or
regarding art and philosophy subjects, i.e. neo-baroque, neo-Aristotelian. The
use and frequency of the compound neo-premier was compared with the
Italian itTenTen16 corpus on SketchEngine. This online corpus displays two
variants of the compound: the hyphenated word neo-premier (0.02 per million
words) and neopremier (0.02 per million words). Conversely, the search of the
same word in English corpora like BNC, enTenTen13, or Now corpus didn’t
produce any results.
The most frequent Anglicisms in the Italian corpus are Brexit (309 tokens,
0.6%), referendum (111 tokens, 0.22%), premier (89 tokens, 0.17%), leader (61
tokens, 0.12%). These four words are particularly frequent in the British
corpus as well: Brexit (232 tokens, 0.46%), referendum (157 tokens, 0.31%),
leader (71 tokens, 0.14%). In particular, the word Brexit is productive in both
the English and the Italian corpus with numerous hyphenated compounds
composed of Latin and Greek suffixs or English-derived morphemes. Some
of them are common to both corpora, i.e. post-Brexit (English corpus 140 per
million words, Italian corpus 58.6 per million words), hard-Brexit (English
corpus 80 per million words, Italian corpus 58.6 per million words), pro-Brexit
(English corpus 100 per million words, Italian corpus 39.1 per million words).

JADT’ 18

257

Other Brexit-compounds like pre-Brexit (39.1 per million words) and dopoBrexit (19.5 per million words) are only found in the Italian corpus, while the
compound anti-Brexit (40 per million words) is only included in the English
corpus. As far as the word premier is concerned, in the English corpus, it only
shows 1 token (20 per million words), while its synonym, prime minister, has a
frequency of 119 tokens (2,380 per million words). The occurrence of this
compound was then compared with larger English corpora like the BNC
where Prime Minister is written both in capital letters (85.17 per million
words) and in lowercase letters (8.33 per million words). On the contrary, the
word premier is present in the BNC and occurs with a frequency of 0.23 per
million words, but it mainly occurs in the semantic field of football, i.e. as a
modifier of the noun league in the collocation premier league. However, it is
also found in the domain of politics as a noun co-occurring with the
modifiers deputy, country. Conversely, in the Italian corpus itTenTen16 in
SketchEngine, premier always occurs in the semantic field of politics. Two
different uses of the word premier and Prime Minister can thus be observed in
the two languages.
4. Conclusion
The aim of this paper has been to provide an outline of the Brexit discourse
as used in the vocabulary and terminology used by two Italian and two
important British newspapers. By way of cluster analysis, the Brexit’s main
themes have been identified: economics, finance, European and British
politics, and the Post-Brexit immigration policies. Another characteristic that
has been explored in this paper is the distribution of the words in various
newspaper article sections which was accomplished by focusing on the
headlines. The analysis showed that the nouns included in newspapers’
headlines refer, for the most part, to Brexit’s main political issues, even
though some words from the field of economics can be found as well.
Whereas verbs aim at describing the difficult circumstances that both the
European Union and the United Kingdom will face. As far as Anglicisms are
concerned, the investigation highlighted that even though they are often
used by newspapers, they represent only about 2% of the whole corpus. This
percentage conforms to the most recent studies on Anglicisms in Italian by
Serianni (2015), Cortellazzo (2015) and Scarpa (2015). They mirror the topic
subdivision of the corpus, and in fact they mainly belong to the semantic
fields of economics and politics, whereas almost half of them can be classified
as general language words. In the Italian corpus, only one pseudo-Anglicism
has been identified, i.e. neo-premier, and its status has been confirmed by
numerous general English corpora. The analysis of Brexit-related Anglicisms
provides a small but interesting contribution to the research on Anglicisms;

258

JADT’ 18

therefore, it would be interesting to keep collecting data about this historical
fact so as to expand the two small corpora under investigation, to make them
as comprehensible and comprehensive as possible, and to carry out an even
more detailed contrastive analysis.
References
Biber D. (1993). Representativeness in Corpus Design. In Literary and Linguistic
Computing, vol. 8 (4): 243-257.
Bolasco S. (1999). Analisi multidimensionale dei dati. Carocci.
Bolasco S. (2013). L’analisi automatica dei testi. Carocci.
Cabré Castellví M. T. (1999). Terminology: Theory, methods and applications.
John Benjamins
Publishing Company.
Cortellazzo M.A. (2015). Per un monitoraggio degli anglicismi incipienti. In
Marazzini C., Petralli A. La lingua italiana e le lingue romanze di fronte agli
anglicismi. Accademia della Crusca.
Furiassi C. (2010). False Anglicisms in Italian. Polimetrica.
Görlach M. (2001). A dictionary of European Anglicisms. Oxford University
Press.
Gusmani R. (1973). Analisi del prestito linguistico. Libreria scientifica editrice.
Gusmani R. (1993). Saggi sull’interferenza linguistica. Le lettere.
Hunston S. (2002). Corpora in Applied Linguistics. Cambridge University Press.
Lenci A., Montemagni S. and Pirrelli V. (2007). Testo e computer. Elementi di
linguistica
computazionale. Carocci.
Marazzini C., Petralli A. (2015). La lingua italiana e le lingue romanze di fronte
agli anglicismi.
Accademia della Crusca.
Pulcini V., Furiassi C. and Rodríguez González F. (2012). The Anglicization of
European lexis. John
Benjamins.
Scarpa F. (2015). L’influsso dell’inglese sulle lingue speciali dell’italiano. Edizioni
Università Trieste.
Serianni L. (2015) Per una neologia consapevole. In Marazzini C., Petralli A.
La lingua italiana e le lingue romanze di fronte agli anglicismi. Accademia
della Crusca.
Sinclair J. (1991). Corpus Concordance Collocation. Oxford University Press.
Tognini-Bonelli E. (2001). Corpus Linguistics at work. John Benjamins
Publishing Company.

JADT’ 18

259

Textual analysis to promote innovation
within public policy evaluation
Viviana Fini1, Giuseppe Lucio Gaeta2, Sergio Salvatore3
2

1 Ospedale Apuane, Massa – vivianafini@gmail.com
Università di Napoli L’Orientale - glgaeta@gmail.com
3Università del Salento - sergio.salvatore65@icloud.com

Abstract
This paper illustrates the contribution by textual analysis in carrying out the
research activities promoted by FORMEZ PA through the REVES (Reverse
Evaluation to Enhance local Strategies) pilot project1 that aims to innovate
public policy evaluation. While evaluation usually embraces a policy/project
viewpoint and adopts a sort of a top-down approach consistent with the flow
of rules/resources from policy makers to citizens’, REVES reverses this
perspective. Indeed, it aims to assess public policies’ performance in
intercepting and supporting development strategies promoted by
citizens/local actors. One of the three case studies carried out by the REVES
project focuses on Melpignano, a small municipality in the Puglia Region of
Southern Italy. Semi-structured interviews were carried out with a sample of
twenty policy actors (national, regional and local policy designer and policy
implementers as well as policy beneficiaries) linked with this municipality.
By using the TLab software, textual analyses of responses were performed in
order to identify their symbolic and latent components and to understand the
actors’ point of view about the world and specifically about local
development. This allowed to assess how similar concepts - such as civic
participation, innovation, community - are used with profoundly different
cultural meanings by the actors. This contributes to understanding public
policies’ difficulties in enhancing local strategies.
Keywords: Local cultures, textual analysis, innovation within evaluation.

The evaluative research was carried out within the framework of the NUVAL
Project, "Actions to support the activities of the National Evaluation System and
Evaluation Units" implemented by Formez PA. The case study was accomplished by
Viviana Fini and Vito Belladonna, under the scientific coordination of Laura Tagle,
Serafino Celano, Antonella Bonaduce, Giuseppe Lucio Gaeta. Viviana Fini carried out
the cultural analysis under the supervision of Sergio Salvatore and thanks with the
contribution of Giuseppe Lucio Gaeta.
1

260

JADT’ 18

Abstract
L'articolo descrive il contributo della ricerca culturale condotta attraverso lo
strumento dell’analisi testuale nella realizzazione del progetto di ricerca
pilota REVES (Reverse Evaluation to Enhance local Strategies) promosso da
FORMEZ PA con l’intento di innovare la valutazione delle politiche
pubbliche. Mentre il processo valutativo tradizionalmente segue il flusso
delle risorse finanziarie e l’attuazione di norme/provvedimenti da parte dei
soggetti locali, REVES propone un capovolgimento di prospettiva,
intendendo valutare le performance delle politiche pubbliche nell’intercettare
e valorizzare le strategie di sviluppo autonomamente elaborate dai territori.
Uno dei casi studio del progetto si incentra sulla città pugliese di
Melpignano. Sono state condotte interviste semi-strutturate con un campione
di 20 attori di policy (policy maker e attuatori di politiche attivi sul piano
nazionale, regionale e locale oltre a potenziali beneficiari delle politiche) a
vario titolo connessi con la città. Con l’ausilio del software TLab sono state
condotte analisi testuali aventi l’obiettivo di evidenziare le componenti
latenti che orientano le visioni del mondo e dello sviluppo proprie degli
attori intervistati. Ciò ha consentito di valutare come concetti simili, ad
esempio partecipazione civica, innovazione, comunità – siano impiegati dagli
attori con significati culturali diversi. Ciò contribuisce alla comprensione del
motivo delle difficoltà delle politiche pubbliche nel valorizzare strategie
localmente elaborate.
Keywords: Culture locali, Analisi testuale, innovazione nella valutazione.
1. Introduzione
L’articolo dà conto dell’indagine culturale - svolta attraverso analisi testuale realizzata per supportare l’innovazione che il progetto REVES ha apportato
al campo della valutazione delle politiche di sviluppo locale. Con un
approccio reverse accountability, il progetto si è domandato se e come le
politiche sovra-locali siano state in grado di cogliere e valorizzare le istanze
di specifici contesti locali, indagando il caso studio “Melpignano”, Comune
in provincia di Lecce, noto in letteratura per aver elaborato, proposto e
attuato, nel corso degli ultimi 30 anni, una visione e una strategia innovativa
di intervento riguardante lo sviluppo locale (Attanasi et al., 2011;
Parmiggiani, 2013). Si discutono qui i risultati dell’indagine culturale e il
vantaggio che l’analisi testuale ha permesso al progetto di realizzare,
consentendo una lettura che è andata oltre il contenuto delle singole
interviste, permettendo di cogliere come concetti simili fossero utilizzati
talvolta – dagli intervistati – con significati culturalmente profondamente
diversi.

JADT’ 18

261

2. L’indagine culturale come presupposto della ricerca valutativa
Il lavoro di ricerca realizzato mediante analisi testuale ha avuto quale fine la
rilevazione delle dimensioni culturali che in modo latente hanno dato forma
alle visioni e agli interventi sullo sviluppo locale. Questo tipo di indagine si
inscrive in una cornice teorica psicologica ad orientamento psicodinamico e
psico-culturale (Carli et al., 2002; Salvatore et al., 2011), che considera i
comportamenti e i discorsi degli attori sociali come espressione di dinamiche
culturali che solo in parte sono consce, in gran parte sono inconsce, latenti
(Matte Blanco, 1975; Fornari, 1979; Carli et al., 2002). Ciò che gli attori fanno,
dicono, ritengono saliente - secondo tale approccio – è funzione di un campo
di forze latenti, un sistema stabile di significati generalizzati, che chiamiamo
cultura (Carli et al., 2002; Salvatore et al., 2011). L’idea di organizzare le azioni
valutative sui risultati dell’indagine culturale ha risposto all’esigenza del
progetto di “costruire” l’oggetto di indagine a partire da una comprensione
profonda delle motivazioni alla base di certi esiti, in conseguenza della
presenza/assenza di alcune iniziative. L’indagine culturale ha consentito di
fare ipotesi su cosa ha avvicinato/distanziato modelli di azione appartenenti
ad attori di policy diversi, consentendo di classificare i loro discorsi in
relazione alla variabilità culturale che li caratterizza e che definisce lo
scenario entro cui ciascuno di essi, senza la mediazione del pensiero
razionale, si è mosso.
2.1 L’analisi testuale: modalità di analisi
Il metodo utilizzato per l’analisi testuale si fonda sul principio delle cooccorrenze lessicali come fonte di ricostruzione del contesto intratestuale.
Tale principio è stato definito all’interno della linguistica (Reinert, 1986) e
successivamente elaborato in chiave psicologica (Carli & Paniccia, 2002;
Lancia, 2004). In termini generali il metodo, utilizzando il software TLab,
trasforma il corpus lessicale in una matrice digitale di co-occorrenze, la quale
viene a sua volta sottoposta ad una procedura di analisi multidimensionale
che permette di estrapolare i cluster semantici attivi nel testo (cioè i cluster di
parole co-occorrenti entro le stesse frasi, in quanto tali indicative di pattern di
significato) che vengono successivamente sottoposti ad interpretazione. La
procedura adottata segmenta il testo in Unità di Contesto Elementari (ECU),
ossia parti di testo interrotte da punteggiatura, che possono contare da un
minimo di 250 caratteri ad un massimo di 500. Attraverso una serie di
operazioni il corpus testuale viene successivamente trasformato in una
matrice digitale in grado di rappresentare il testo in termini di
presenza/assenza dei lemmi nelle ECU che lo compongono. La matrice che si
viene così a definire è sottoposta ad una procedura di analisi
multidimensionale combinata, che unisce l’Analisi delle Corrispondenze

262

JADT’ 18

Multiple (ACM) e l’Analisi dei Cluster (AC). L’ACM permette di estrapolare
le modalità nei termini delle quali i lemmi si associano all’interno delle ECU
(vale a dire: le loro co-occorrenze intra - ECU). Ciascuna dimensione
fattoriale individuata dalla ACM rappresenta un pattern di co-occorrenze che
si ripropone attraverso il testo, o in una sua porzione sufficientemente ampia.
Le dimensioni fattoriali estrapolate dalla ACM vengono quindi utilizzate
come criteri classificatori dalla successiva CA. In questo modo la CA
permette di raggruppare ECU (e lemmi) in base alla loro somiglianza – ossia
in base alle combinazioni di parole per come si danno nelle frasi di testo. Il
risultato finale della procedura è dunque l’identificazione di cluster di frasi
tra loro simili in quanto caratterizzate dalla compresenza delle stesse parole;
oppure, specularmente, dalla identificazione di cluster di parole simili in
quanto tendenti ad essere utilizzate insieme nelle stesse frasi. Per questa loro
caratteristica computazionale, i cluster individuati si prestano ad essere
interpretati nei termini di nuclei tematici, tali in quanto caratterizzati dal
riferimento ad un aggregato sufficientemente stabile di parole (Lancia, 2005).
L’output dell’analisi può essere considerato come una rappresentazione del
campo culturale caratterizzante lo specifico contesto di policy (Carli et al.,
2002), dove sono visibili le dimensioni latenti che dinamizzano il campo
(Fattori) e la variabilità relativa ai diversi modi di pensare dei soggetti
intervistati (Cluster).
2.2 Popolazione di riferimento e campione
La popolazione di riferimento sono gli attori delle politiche. Il campione è
costituito da 20 soggetti che a vario titolo hanno operato in relazione allo
sviluppo locale, con i quali è stata condotta un’intervista in profondità,
considerati figure chiave del contesto studiato per le seguenti variabili
illustrative: ruolo (politici, cittadini, tecnici); tipo di implicazione nella politica
(policy maker, policy designer, attuatori, destinatari); livello di appartenenza
(locale, sovracomunale, regionale, nazionale). Trattandosi di uno studio
pilota, al campione rappresentativo si è preferito un campione a grappolo per
quote non proporzionali (Blalock jr, 1960), facendo riferimento agli attori
presenti entro i contesti, distribuiti in modo tendenzialmente equivalente in
relazione alle tre variabili. La scelta di un campione di questo tipo ha
consentito di costruire ipotesi, più che di verificarle, enucleando lo spettro di
eterogeneità culturale presente entro la popolazione di riferimento.
3. I principali risultati dell’analisi culturale
3.1 I Fattori: le principali dimensioni latenti del campo culturale
I principali fattori estratti sono tre. Di seguito, una loro interpretazione sul
piano culturale.

JADT’ 18

263

Primo Fattore - Simbolizzazione del processo di regolazione sociale: operatività
proceduralizzata vs appartenenza valorizzata
Invitati a parlare della propria visione dello sviluppo, del proprio ruolo in
relazione ad esso, delle politiche in grado di promuoverlo, i soggetti
incontrati parlano, in prima istanza, del modo in cui regolano il processo
relazionale con i propri interlocutori. Da un lato (operatività proceduralizzata)
lo sviluppo del territorio viene visto come esito dell’adesione, da parte degli
attori locali, al frame valoriale e alle azioni proposte dalle politiche di
sviluppo. Dall’altro il riferimento è al costruire un comune sentire
(appartenenza valorizzata), governando e amministrando fatti concreti
riguardanti la vita delle persone, avvalorando le valenze affettive dei legami
di appartenenza. Due differenti modelli di regolazione sociale, che implicano
due visioni alternative di sviluppo: tecnicalità come modello di relazione che
funziona a supposto contesto dato (Carli et al., 1999) - lo sviluppo qui è
realizzabile per decreto - vs modello di regolazione sociale che funziona in
modo esperienziale, - lo sviluppo è qui concepito come sviluppo endogeno
del sistema (Fini et al., 2015).
Secondo Fattore - Forme del desiderio: salvaguardia vs riuscita.
In seconda istanza, i soggetti intervistati parlano della spinta che muove la
loro azione, ossia della forma del loro desiderio. Da un lato (salvaguardia) la
trasformazione in mito della comunità di appartenenza sembra rispondere al
desiderio di sottrarre la propria storia alla contingenza. Operazione che offre
“sicurezza” in cambio di “dipendenza”. Dall’altro (riuscita) viene messa al
centro una dialettica tra identità ed estraneità, con “speranza” e “avvenire”
che prendono il posto di “sicurezza”. In entrambe i casi “comunità” è lemma
centrale, ma mentre nella polarizzazione salvaguardia le parole con cui cooccorre la fanno sembrare valore e scopo dell’azione, nel secondo caso
appare più come un prodotto da costruire, dialogicamente, tra dentro e fuori,
vecchio e nuovo. Due diverse modalità di entrare in rapporto con l’estraneità:
nel primo caso si adatta ciò che è sconosciuto a ciò che già si sa; nel secondo
caso si utilizza il noto per esplorare l’ignoto.
Terzo Fattore - Simbolizzazione della domanda di sviluppo: funzione sostitutiva vs
funzione integrativa
I soggetti intervistati, in terza istanza parlano della domanda di sviluppo. Da
un lato, laddove ci si propone di adeguare i destinatari alle regole della
pianificazione, le regole diventano ordini invalicabili e gli operatori sentono
svilito il proprio ruolo ad un mero adempimento e si sentono impotenti.
Dall’altro i destinatari delle policy si propongono imprenditivamente,
avendo a mente ciò che è rilevante per sé e chiedendo regole che consentano
di muoversi all’interno di aspettative condivise. Emergono, polarizzate, due
domande di sviluppo: la prima soggiacente ad un modello che potremmo

264

JADT’ 18

definire “sostitutivo” (Carli, Paniccia, 1999), che attribuisce alla policy un
potere elevato, valutabile a prodotto finito, che mette l’impotenza al posto
del desiderio. La seconda, relativa ad un modello che potremmo chiamare
“integrativo” (Carli, Paniccia, 1999) che esprime il desiderio di contribuire al
raggiungimento degli obiettivi dei destinatari, in compenetrazione di
funzioni e scelte e che pensa per processi.
3.2 I principali Cluster
La Cluster Analysis ha individuato 4 Cluster principali.

CL_1
Elementary
Context:
407 di 2504
(16,25%)

Tab 1. Contesti Elementari
CL_2
CL_3
Elementary
Elementary
Context:
Context:
840 di 2504
593 di 2504
(33,55%)
(23,68%)

CL_4
Elementary
Context:
664 di 2504
(26,52%)

C1. Le parole con un χ2 maggiormente significativo (che riportiamo tra
parentesi) per questo cluster sono: tema (102,4); amministrazione (100,6); aspetto
(83,9); processo (68,5); economico (66,4); contesto (64,4); imprenditoriale (62,3);
azione (50,6); amianto (52); costruire (49,6); impresa (48); innovazione (43,5).
Abbiamo denominato C1 “Governo imprenditivo dell’innovazione”, per
l’accento posto sull’innovazione, considerata come processo da governare
proattivamente.
C2. Le parole maggiormente rappresentative sono: io (277,2); tu (154,7);
sindaco (80,5); parlare (63,4); trovare (62,1); sentire (56,7); persona (51,7); giorno
(45,5); figlio (41,6); paese (34,9); riuscire (32,6). Abbiamo denominato C2
“Implicazione nella gestione della cosa pubblica”, per l’accento posto sulla
partecipazione diretta e personale, ognuno con il proprio ruolo e la propria
soggettività, al governo del bene comune.
C3. Le parole maggiormente rappresentative sono: cooperativo (224,7);
comunità (182); notte (105,3); anno (103,1); Melpignano (91,2); fare (87,8);
cittadino (83,5); acqua (83); bello (78); casa (75,7); pagare (68,9); euro (63,7);
Taranta (60,7). Abbiamo denominato C3 “Comunità come identità” per
l’accento posto su tutto ciò che ha reso possibile la costruzione di Melpignano
come comunità che si riconosce nella gestione della cosa pubblica e nella
valorizzazione della tradizione popolare.
C4. Le parole maggiormente rappresentative sono: territorio (442,4);
programmazione (191,7); sviluppo (179,4); area (173,9); regione (171,1); GAL
(118,6); attività (104,3); intervento (102,6); livello (90,3); vasto (86,9); Puglia
(77,8); governance (75,2). Abbiamo denominato C4 “Pianificazione come

JADT’ 18

265

sviluppo” per l’identificazione del territorio con i confini amministrativi e la
sovrapposizione tra sviluppo e varie forme di pianificazione, come se
definire confini e pianificare azioni fosse di per sé garanzia di produzione di
sviluppo.
3.3 Discussione
La Tabella 2 mostra il rapporto Cluster-Fattori.

Cluster
CL_01
CL_02
CL_03
CL_04

Tab 2. Rapporto Cluster-Fattori
Fattore1
Fattore2
- 22,2374
14,7017
37,0788
59,5785
60,9616
- 52,9475
- 81,5382
- 11,9437

Fattore3
63,7361
- 22,7426
0
- 30,8565

La proiezione dei Cluster sullo spazio fattoriale ha consentito di comprendere
come concetti simili fossero utilizzati dagli intervistati con significati
culturalmente molto diversi.
È il caso, ad esempio, di C2 (quadrante riuscita - appartenenza valorizzata e
quadrante funzione sostitutiva - appartenenza valorizzata). I discorsi di C2
concernono l’essere attivi nella gestione della cosa pubblica. Ma il loro
differente posizionamento sullo spazio fattoriale ci ha fatto ipotizzare una
differente visione e, di conseguenza, un diverso utilizzo del tema della
partecipazione civica, argomento strategico per il contesto locale e per le
politiche di sviluppo e strettamente connesso con l’attivazione dei cittadini.
Questa ipotesi ha orientato in modo mirato le successive esplorazioni che
hanno evidenziato, sotto lo stesso cappello, micro-processi socioorganizzativi molto diversi: da un lato il destinatario di policy visto come
soggetto da implicare nella produzione del bene, esplorando e valorizzando
il suo desiderio (in coerenza con il quadrante riuscita-appartenenza valorizzata).
Qui la partecipazione è considerata esito di una costruzione dialogica.
Dall’altro (quadrante funzione sostitutiva-appartenenza valorizzata) i destinatari
alternativamente visti come fruitori passivi di un bene prodotto da altri o
soggetti ai quali delegare sovranità e la partecipazione trattata come
strumento di rafforzamento dei sistemi di appartenenza. Questa evidenza ha
consentito di superare la classica distinzione presente in letteratura tra
processi top down/bottom up (Bens, 2005; Sclavi, 2002) e, in una restituzione
ai soggetti locali, di discutere con loro su come lo scarto esistente stesse
piuttosto nelle diverse modalità di presa in carico dell’estraneità relativa al
desiderio del destinatario delle policy. Grazie al tipo di indagine è stato
possibile anche cogliere come temi quali innovazione e comunità, che nelle

266

JADT’ 18

interviste emergevano in modo contiguo come due miti locali per certi versi
sovrapponibili, evidenziassero invece posizionamenti culturali differenti:
quando a prevalere è C1-innovazione (ad esempio: inventare una tradizione
come il Festival di musica popolare La Notte della Taranta; introdurre la
raccolta differenziata; promuovere presso la cittadinanza l’uso dei pannelli
fotovoltaici) le pratiche raccontate sono maggiormente orientate
dall’importanza attribuita al raggiungimento di obiettivi (quadrante
operatività proceduralizzata – riuscita) e dalla necessità di capire come rendere
le innovazioni appetibili per la cittadinanza (quadrante operatività
proceduralizzata – funzione integrativa). Quando invece a prevalere è il tema
C3-comunità (ad esempio promuovere lo sviluppo di una Cooperativa di
Comunità) ciò che sembra essere motore dell’azione è l’idea di rafforzare il
proprio sistema di appartenenza (quadrante appartenenza valorizzata –
salvaguardia; e appartenenza valorizzata-funzione sostitutiva). Infine la proiezione
di C4 sullo spazio fattoriale nel quadrante operatività proceduralizzata –
salvaguardia e operatività proceduralizzata – funzione sostitutiva ha consentito di
cogliere quanto, entro questo assetto culturale, la pianificazione si muova in
modo avulso dai contesti anche laddove la retorica dei programmi preveda
strumenti per l’ascolto e la partecipazione dei destinatari delle policies. Da
sottolineare, poi, come le variabili illustrative si siano polarizzate
maggiormente sul primo fattore: operatività proceduralizzata vs appartenenza
valorizzata. Tecnici da un lato e cittadini/politici dall’altro; policy designer da
un lato e policy maker/destinatari dall’altro. Queste polarizzazione ci hanno
fatto pensare ad una vicinanza culturale tra policy maker/politici e
destinatari/cittadini, evidenziando come la politica locale, a differenza di
quella centrale, sia in una posizione privilegiata per comprendere domande e
interpretare esigenze, limiti, potenzialità di sviluppo dei contesti reali. Gli
attuatori, invece, si posizionano in opposizione a policy maker, destinatari e
policy designer. Questo ci ha interrogati sul loro difficile ruolo di cuscinetto,
tra le domande dei diretti interlocutori della politica (destinatari, policy
maker) e le esigenze intrinseche ai programmi.
4. Conclusioni
L’indagine culturale realizzata mediante analisi testuale ha consentito al team
di ricerca di costruire l’oggetto di indagine a partire da elementi altrimenti
difficilmente individuabili, dal momento che i contenuti proposti dagli
intervistati si presentavano pressoché identici. Poter cogliere tali differenze
sostanziali dal punto di vista culturale ci ha permesso di realizzare
osservazioni, interviste, discussioni con gli attori locali in merito a quando
andavamo capendo ben più mirate e interessanti, anche per i soggetti locali
stessi. In ciò riposa la vera innovazione che l’indagine culturale ha consentito

JADT’ 18

267

al Progetto REVES di apportare nel campo della valutazione delle politiche di
sviluppo locale.
Riferimenti bibliografici
Attanasi, G., Giordano, G. (2011). Eventi, cultura e sviluppo. L’esperienza de “La
Notte della Taranta". Milano: Egea
Bens, I., (2005). Facilitating with ease! Core skills for facilitators, team leaders and
members, managers, consultants and trainers. San Francisco: Josey-Bass.
Blalock, Jr., H. M. (1960). Social Statistics. New York: McGraw-Hill Book
Company.
Carli R., Paniccia, R.M (1999). Psicologia della formazione. Bologna: Il Mulino.
Carli, R., Paniccia, R.M. (2002). L’Analisi Emozionale del Testo. Milano: Franco
Angeli.
Fini, V., Belladonna, V., Tagle, L., Celano, S., Bonaduce, A., & Gaeta, L.G.
(2016), Progetto Pilota di Valutazione Locale, Studio di Caso: Comune di
Melpignano. Come Stato centrale, fondazioni e Regioni possono sollecitare la
progettualità locale retrieved at
http://valutazioneinvestimenti.formez.it/sites/all/files/2_reves_rapporto_cas
o_melpignano.pdf
Fini, V., Salvatore. S. (in press). The fuel and the engine. A general semiocultural psychological framework for social intervention. In S. Schliewe, N.
Chaudhary & P. Marsico (Eds.), Cultural Psychology of Intervention in the
Globalized World. Charlotte (NC): Information Age Publishing.
Fornari, F. (1979). I fondamenti di una teoria psicoanalitica del linguaggio. Torino:
Boringhieri.
Lancia F. (2004). Strumenti per l’analisi dei testi. Introduzione all’uso di T-LAB.
Milano: Franco Angeli.
Matte Blanco, I. (1975). L'inconscio come insiemi infiniti. Saggio sulla bi-logica.
Torino: Einaudi.
Parmiggiani, P. (2013). Pratiche di consumo, civic engagement, creazione di
comunità, in Sociologia del lavoro, 132, 97 – 112.
Reinert, M. (1986). Un logiciel d’analyse textuelle: ALCESTE, in Cahiers de
l’Analyse des Données, 3.
Salvatore, S., & Zittoun, T. (2011). Outlines of a psychoanalytically informed
cultural psychology. In S. Salvatore, & T. Zittoun (Eds). Cultural Psychology
and Psychoanalysis in Dialogue. Issues for Constructive Theoretical and
Methodological Synergies (pp. 3-46). Charlotte, NC: Information Age.
Sclavi, M. (2002). Avventure Urbane. Progettare la città con gli abitanti. Milano:
Euleuthera.

268

JADT’ 18

A proposal for Cross-Language Analysis:
violence against women and the Web
Alessia Forciniti, Simona Balbi
University of Naples Federico II - alessia.forc@libero.it

Abstract
Aim of the paper is investigating the mood on the Web with respect to one of
the most relevant Human Rights violation, without any geographic
distinction: the violence against women. While the literature that studies the
phenomenon is rapidly growing, the action field is still fragile and the
question marks are about the relationship between the public opinion and
the contextual factors. In a first look at the phenomenon, we aim at mapping
gender violence on the Web, in a Big Data perspective. The peculiar problem
we deal with consists in analysing short documents (tweets) written in six
European different languages, in the occasion of a common event: the
International Day for the Elimination of Violence against Women, 25
November 2017. For our statistical analysis, we choose a multi-linguistic,
cross-national perspective. The basic idea is that there are some common
structures, language independent ("concepts"), which are declined in the
different national natural language expressions ("terms"). Investigating those
structure (e.g. factors of lexical correspondence analyses separately
performed on the different collections), enables a double level analysis trying
to understand and visualise national peculiarities and communalities. The
statistical tool is given by Procrustes rotations.
Keywords: Big Data, Text Mining, Cross-national study, Procrustes rotations
1. Introduction
This paper proposes a statistical-linguistic analysis about the mood on Web
in relation to a social issue of universal relevance: the violence against
women (European Union Agency for Fundamental Rights (FRA), 2014; ONU
and United Nations Population Fund, 2016, 2017). The social media, today,
are becoming an important platform of the collective thought of the society
and therefore, they represent an interesting container of context to study. The
constant growth in unstructured information on Web makes the Text mining
applications increasingly important in achieving to knowledge extraction of
the phenomena. This work faces the problem of the public opinion on the
phenomenon of gender-based violence, in Europe, as reply to a common

JADT’ 18

269

event: the International Day for the Elimination of Violence against Women
(United Nations, General Assembly, 1999), 25 November 2017.The proposed
method of analysis is a multi-linguistic, cross-national study of the
multimedia contents extracted from Twitter through Web scraping
techniques. The features of data (Wu X., Wu G-Q.,Zhu et al.,2014) propose an
analysis in terms of Big Data (Zielinski et al., 2012). Considering the aspects
of the comparative research (Finer, 1954; Lijphart, 1975) the choice of number
of cases study does not excess the six European countries; three west
countries, as United Kingdom (Uk), Italy, France and three east countries, as
Bulgaria, Czech Republic and Romania.
The research takes on several methodological issues; it requires the treatment
of multilingual corpora (tweets are written in six different languages) and not
all the treated languages in this study are typical of the Textual Data mining
application. The implications are relative to: a careful pre-processing step
(corpora cleaning from URL and emoticons), it does not exists a package or
software that includes a list of stop words for all investigated languages in
this research and in addition the appropriate system of weights for the
analysis unit in relation to the nature of data (short messages of up 140
characters). The accuracy of these choices is very important for the good
result of the investigation. Therefore, this work has not only a simple
cognitive function of the phenomenon but it represents an opportunity to test
the scientific method. The cross-linguistic perspective is given by projection
on factorial plan of the most frequent terms for couples of countries. In order
to visualize the national peculiarities and communalities, the factors are
projected in the two different natural languages on a common reference
space, per pairwise through the Procrustes rotations.
2.Theoretical Framework
In order to visualize the relationships between document and between terms,
in textual data analysis, is commonly performed a factorial approach. The
starting point is a lexical table, cross-tabulating terms and documents (in this
case terms and tweets).
This study in question intends propose a Procrustes analysis, such as efficient
geometric technique to align lexical matrices. Our research proposes six
lexical tables (X1, ..., X6,) as many as there are the case studies. There is an
extremely wide multivariate analysis literature devoted to the problem of
comparing and synthesising information contained in two or more matrices.
An interesting way of approaching the problem consists in comparing
geometrical configurations in some Euclidean space (Gordon, 1981). In our
case, Correspondence Analysis (CA) is performed on the six tables and
visualises the major themes and suggests similarities and peculiarities

270

JADT’ 18

between countries. In order to have a measure of this similarity for couple of
countries, we can compute the sum of the square distances between
corresponding points in the two configurations:

The data structure consists of two matrices, X (n,p) and Y (n,p). X is the lexical
table having in row the n tweets in which the corpus is organized, and in
columns some content bearing words selected among the most frequent
terms in the corpus for a country. Y is the lexical table having in row the n
tweets, and in columns the content bearing words selected in the natural
language of the other country. Through the CA performed on each corpus,
we compute the principal coordinates and create two matrices: X1 and Y1;
which represent coordinate matrices of each language. The coordinates
matrices have been standardized and normalized so that is not necessary “rescaling” factor”.
3. Data extraction: the Web Scraping
The Social media are a potentially infinite source of user data, and Twitter is
one of the worldwide used Social network. Twitter is a micro-blogging
service which messages (called tweets) of up to 140 characters. Web scraping
is the process of automatically extract data from the Web by an Application
Programming Interface (API) supported by software (or by packages
connected to software). For our research, data extraction has been conducted
with API Twitter and R, respecting specific parameters, common for each
country: a keyword translated in the different 6 languages, the specification
of the language, the geocode (in order to exclude urban semantic deriving
from dialects or territorial slang which change the common sense of words)
and finally the sample size (with technical limits; it is possible to extract until
to n=3200 tweets per day). The monitoring period is a week around the
International Day for the Elimination of Violence against Women, from 23
November to 30 November 2017.
4. Knowledge extraction of the phenomenon
Considering but at same time overlooking the detailed description of the
methodological issues aimed at pre-processing procedures of multi-linguistic
and multimedia content, the argumentation focuses on the results. The
results represent one of the most interesting developments of our proposal.

JADT’ 18

271

However, a note deserves the attention: given the structure and the length of
each document (tweet), the system of weights of elementary unit is tf (term.
frequency) where: wij =
The canonical tools used for Textual Data Analysis, such as occurrence values
of the most frequent terms does not represent, in this case, a useful tool to
comparing relation between countries. There are other statistical tools can
enable us to go deeper in understanding of the phenomenon, such as the
factorial approach.
4.1. Procrustes analysis for a cross-language study
The scientific method that this research intends to test is the Procrustes
Analysis by performing the overlapping of two different configurations. The
configurations to comparing are two normalized CA coordinates matrices.

vita
insinuano

aiutare
exploitation
rape
abuse
domestic
fight
men

approvato
consiglio

internationalday
government
women
rights
activism
elimination
world
aggression
reflect
issue
donnevittime
report
campaign
violence
race
gender
fenomeno
contrastare
giornatainternazionale
genere
stanziato
violenza
maschile
mnl
dirittifondamentali
casadonnepisa
femminista

libera

novembre
legislation
riformatore

-0.5

0.0

0.5

abusi

-1.0

Dimension 2

1.0

1.5

Procrustes errors

-2

-1

0

Dimension 1

Figure 1. Procrustes errors: comparison between Italy-Uk

1

272

JADT’ 18

The graphic representation allows to observe the Procrustes errors between
the two dimensions: points of Italy's normalized principal coordinates matrix
and United Kingdom's points of normalized principal coordinates matrix,
where Uk is the rotated matrix. Beyond the descriptive statistics about the
residual scores, the graph shows how around the axes origin there is a
concentration of points both X1 and Y1 and so we can affirm that there is not a
wide distance between X1 e Y1.Procrustean approach confirms the similarity
estimated by CA maps between Uk and Italy (Figure 2 and Figure 3); where
despite, third quadrant of Italy's and United Kingdom's suffer a dense
overlapping of statistics entities, it is possible note similar topics, which are
collocated nearly in same position on the multidimensional space.

Figure 2. Correspondence Analysis Maps for Uk

Furthermore, through the CA, is possible to investigate structures, language
independent (“concepts”), which is declined in the different national natural

JADT’ 18

273

1.5

language expressions ("terms"). In other words, even though there are terms
that they are not the exact translation from a language to another and so from
Italian to English o conversely, does not changes the conceptual aspect.
Studying the vocabulary of the country we can consider the conceptual
aspect and we can create thematic-groupings and to label the clusters.
Procrustes errors and t Correspondence Analysis permits to observe the
collocation of the statistic entity "abuse". In Procrustes errors plot (Figure 1)
the "term" is distant from others statistics units; therefore it represent a
Procrustes residual. Same consideration is given by observing CA maps
(Figure 2 and 3). Despite, the word "abuse" is the relative translation of natural
language from Italian to English the collocation on the multidimensional
space is different. The "joint terms space" (Figure 4) of the comparison
between Italy and Uk, allows to affirm that the terms that are the exact
translation, are almost close in the projected factorial space; e.g. "women",
"violence", "international day" and "rights".

domestic

1.0

approvato

libera
consiglio

0.5

giornatainternazionale

0.0

government
dirittifondamentaliviolence
novembre
aggression
rights
violenza
women
casadonnepisa
femminista
legislation
genere
contrastare
maschile
stanziato
mnl
reflect
issue
activism
gender
campaign
world donne
community
race
report
riformatore
fenomeno

abuse

fight

abusi
men

vittime

-0.5

Dimension 2 (13.6%)

internationalday

aiutare
rape

-1.0

insinuano vitaexploitation

-0.5

0.0

0.5

1.0

1.5

2.0

Dimension 1 (21.1%)

Figure 3. Correspondence Analysis Maps for Italy

Finally, by confirming the Procrustes errors plot (Figure 1) and the CA maps
(Figures 2 and 3), it is possible to see the unit "abuse" (despite the exact

274

JADT’ 18

translation) is more distant compared to the relative translation of natural
language of the other investigated context. The visualizations of Procrustes
Correspondence Analysis and “Joint terms space”, test the similarity between
Italy and United Kingdom in a cross-linguistic perspective. The graphic
intelligibility allows confirming the concordance between the two profiles in
relation to public opinion on violence against women.

Figure 4. Joint terms space Italy-Uk

In the complex, the visualizations lead us to assert what above mentioned,
while singularly they permit to investigate specific aspects of the linguistic
peculiarities. The "Joint terms space" confirms the overlapping of statistics
units (between countries) around the axis origin, so like the Procrustes errors
graph. Therefore, it does not exist a big difference between Italy and Uk. The
closeness between the "terms" of different languages collocated on the same
reference space recall the thematic-groupings brought out by CA.
5. Conclusion and perspectives
In this paper we faced the problem of comparing corpora when, one is not
the translation of the other. Some investigations (e.g. comparison between Uk
and Italy) indicate that the Procrustes approach is a valid tool for crosslanguage study. However, the cross-national investigations, carried out for
all case studies, bring out some limits relative to semantic of the natural
language expressions of the countries. It is possible that some terms, which

JADT’ 18

275

are natural language expressions of a country does not coincide with the
translation of the language expressions of another country. For example, in
the same case Italy-Uk, we can consider that "reformer" can indicate the
political aspect that Uk shows through terms such as "legislation" or
"government". Different terms (in natural language expressions) could be
ascribable to common conceptual labels since actually are belonging to same
semantic category. The future perspective is addressed to resolve the
semantic problems between countries by performing an analysis that focuses
on study of thematic-axes.
References
Balbi and Misuraca (2006). Procrustes Techniques for Text Minig, in Zani et
al., (Eds.), Data Analysis, Classification and the Forward Search, pp.227-234
Berlin, Heidelberg: Springer.
Bolasco (1999), Analisi multidimensionale dei dati. Metodi, strategie e criteri
d’interpretazione, Carocci, Roma.
Bolasco (2005), Statistica testuale e text mining: alcuni paradigmi applicativi.
Quaderni di Statistica, Vol.7, pp. 1-37.
European Union (2017). Report on equality between women and men in the
EU.
Feldman et al. (1998). Mining text using keyword distributions. Journal of
Intelligent Information Systems. Vol. 10, Issue 3, pp. 281–300.
Finer (1954). Metodo, ambito e fini dello studio comparato dei sistemi
politici, in Studi politici, III, 1, pp. 26-43.
FRA, European Union Agency for fundamental Rights (2014). Report
summary: Violence against women: an EU-wide survey. Results at a
glance. Publications Office of the European Union.
Gower (1975). Generalised Procrustes Analysis. Psychometrika, vol.(40):33-51.
Lijphart (1975). The comparable-cases strategy in comparative research, in
Comparative political studies, VIII, pp. 161-174.
Wu X., Wu G-Q., Zhu et al. (2014). Data mining with big data. IEEE
Transactions on Knowledge and Data Engineering. Vol. 26, Issue: 1.
Zielinski et al. (2012). Multilingual Analysis of Twitter News in Support of
Mass Emergency Events, Multilingual Twitter Analysis for Crisis
Management.

276

JADT’ 18

La verbalisation des émotions
Béatrice Fracchiolla, Olinka Solène De Roger
University of Lorraine in Metz
beatrice.fracchiolla@univ-lorraine.fr; olinka-solene.de-roger8@etu.univ-lorraine.fr

Abstract
Our study concerns the correlation between the perception of negative
emotions and discursive productions to express them. Our study is based on
26 transcribed oral interviews to be analyzed with Lexico3 (13 men and 13
women). We study the way in which healthy volunteers react verbally to the
conditioned production of negative emotions after viewing the government
realized video stop jihad, broad casted on television after the 2015 attacks.
Interviews were collected between November 2016 and February 2017
through out the COREV1project framework (understanding verbal violence
in reception). At the same time, following an identical protocol, we showed
another "neutral" video to the same people in order to have a control group.
All the subjects saw both videos, but in different orders, after 11hours of
intervals. According to our methodology of analysis with Lexico3 we were
able to extract the linguistic data allowing to have an over view of the
emotional feelings perceived by the volunteers after viewing each neutral or
violent video and to propose a synthetic card of them. The analysis was
conducted with three tools for statistic alanalysis of textual data proposed by
Lexico3:search for specificity according to the partitions using the PCLC tool
(Main Lexicometric Characteristics of the Corpus), the concordances, the
graphs of ventilation by partition. The over all analysis of the results shows
firstly that the emotions are distributed according to the nature of the videos
(neutral video: positive emotions and /or neutral - violent video: negative
emotions) and that the violent video provokes a quantity of speech longer
than the neutral. Then, if the intensity of perceived emotions seems to differ
according to the person wehere show it also is globally correlated to the
order of diffusion of the videos. We can see in the responses and the
construction of the speeches a correlation of positive or negative intensity of
the emotions according to the video which is seen first Like wise, the analysis

The Corev project (2016-2017) which allowed us to constitute the corpus
studied is an association of the CNRS, the University of Lorraine and the hospital of
Pitié Salpêtrière in order to make a comparative analysis of the neurophysiological
responses, emotional and discursive to exposure to (verbal) violence before / after
sleep and before / after waking.
1

JADT’ 18

277

seems to show that the reception of the violence invites volunteers and urges
them to express them selves more about their feelings: can we see here a
correlation also between discursive productivity and negative emotions - a
form of verification to the French proverb that "happy people have nothing
to say " ?
Résumé
Notre étude porte sur la corrélation qui existe entre la perception d’émotions
négatives et les productions discursives pour les exprimer. Elle est réalisée à
partir de 26 entretiens individuels oraux retranscrits pour être analysés via
Lexico3 (13 hommes et 13 femmes). Nous étudions la manière dont des
volontaires sains réagissent verbalement à la production conditionnée
d’émotions négatives après avoir visionné la vidéo stop-djihad du
gouvernement, diffusée à la télévision après les attentats de 2015. Les
entretiens ont été recueillis entre novembre 2016 et février 2017 dans le cadre
du projet COREV2 (comprendre la violence verbale en réception).
Parallèlement, suivant un protocole identique, nous avons montré une autre
vidéo « neutre » aux mêmes personnes afin d’avoir un groupe contrôle. Tous
les sujets ont vu les 2 vidéos, mais dans des ordres différents, à 11h
d’intervalles. Suivant notre méthodologie d’analyse via Lexico3 nous avons
pu extraire les données linguistiques permettant d’avoir un aperçu des
ressentis émotionnels perçus par les volontaires après le visionnage de
chaque vidéo neutre ou violente et d’en proposer une carte synthétique.
L’analyse par Lexico 3 a été menée via trois outils d’analyse statistiques des
données textuelles proposés par Lexico3: la recherche de particularité selon
les partitions à l’aide de l’outil PCLC (Principales Caractéristiques
Lexicométriques du Corpus), les concordances, les graphiques de ventilation
par partition. L’analyse globale des résultats montre tout d’abord que les
émotions sont réparties selon la nature des vidéos (vidéo neutre : émotion
positive et ou neutre – vidéo violente : émotion négative) et que la vidéo
violente suscite un temps de prises de parole plus long que la neutre. Si
l’intensité des émotions perçues semble différer selon la personne nous
montrons ici qu’elle est également relative à l’ordre de diffusion des vidéos.
Des indices lexicaux ou discursifs nous permettent de vérifier que les sujets
qui ont vu d’abord la vidéo djihad réagissent avec plus d’émotions positives

Le projet Corev (2016-2017) qui nous a permis de constituer le corpus étudié est
issu d’une association entre le CNRS, l’Université de Lorraine et l’hôpital de la Pitié
Salpetrière dans le but de faire une analyse comparée des réponses
neurophysiologiques, émotionnelles et discursives à une exposition à de la violence
(verbale) avant / après sommeil et avant /après réveil.
2

278

JADT’ 18

à la vidéo « neutre »et , inversement, que celles et ceux qui ont vu la vidéo
neutre en premier réagissent avec plus d’émotions négatives lors de la
projection de la vidéo stop-djihad. Autrement dit : nous constatons dans les
réponses et la construction des discours une corrélation d’intensité positive
ou négative des émotions en fonction de la vidéo qui est vue en premier. De
même, l’analyse semble montrer que la réception de la violence interpelle les
volontaires et les pousse à plus s’exprimer sur leur ressenti : peut-on voir ici
une corrélation également entre productivité discursive et émotions
négatives – soit une forme de vérification du proverbe selon lequel « les gens
heureux n’ont rien à dire ».
Keywords: verbal violence, discourse analysis, emotions, textual statistical
analisis, Lexico3
1. Introduction
Dans cette étude, nous nous intéressons à la manière dont des sujets
confrontés à des éléments violents extériorisent verbalement leurs émotions.
Dans l’expérimentation que nous avons conçue pour y arriver, nous avons
travaillé sur différents types de réponses émotionnelles obtenues sur 26
sujets ayant visionné une vidéo « violente » (la vidéo « stop-djihad » diffusée
par le gouvernement français suite aux attentats de 2015 – désormais notée
vidéo V) et une vidéo « neutre » (sur la nouvelle région Languedoc
Roussillon midi Pyrénées – désormais notée N). Le protocole multimodal
suivi pour récupérer nos données a été réalisé en milieu hospitalier3. Nous
avons recueilli plusieurs entretiens individuels semi-directifs portant sur le
ressenti émotionnel avant et après la vision des différentes vidéos, ainsi que
de nombreuses données neurovégétatives. Cette recherche soutenue par la
mission à l’interdisciplinarité du CNRS entre novembre 2016 et décembre
2017 visait plus particulièrement la compréhension et la perception de la
violence verbale chez des sujets sains (Fracchiolla et al., 2013).
L’expérimentation ainsi menée nous permet à la fois de mettre en évidence
certains des éléments marqueurs d’extériorisation émotionnelle verbale et de
comparer les types de réponses aux vidéos V et N. La présente publication
porte exclusivement sur la dimension verbale de l’extériorisation des
émotions, une fois le corpus des entretiens menés avec nos sujets retranscrit
et étudié à l’aide du logiciel Lexico3. Notre approche sera ici plus

3 Dans le service de et en collaboration avec la Professeure Isabelle Arnulf,
Neurologue, directrice de l'unité des pathologies du sommeil de l'hôpital de la PitiéSalpêtrière, professeure de neurologie à l’Université Pierre et Marie Curie (UPMC),
laboratoire : ICM UMR 7225.

JADT’ 18

279

spécifiquement de nous demander si les mots que nous utilisons pour nous
exprimer sont en adéquation avec ce que nous pensons et surtout avec les
émotions ressenties. Notre corpus est ainsi constitué de 26 entretiens répartis
en deux groupes comme suit : le Groupe 1 a vu les vidéos dans l’ordre : 1/
Vidéo N – 2/ Vidéo V. Le Groupe 2 : a vu les vidéos dans l’ordre inverse 1/
Vidéo V – 2/ Vidéo N4.
2. Manifestations d’un discours « émotionné »
2.1. Analyse des PCLP
La répartition du corpus selon la partition « vidéo » avec l’outil PCLC
(Principales caractéristiques lexicométriques du corpus), montre les
spécificités de cette première partition par vidéo et par groupe. Les
interventions des enquêtrices n’y sont pas inclues.
Tableau 1 : Principales caractéristiques de la partition « vidéo »
Partie
V1 N1
V1 N2
V1 Neutre
V2 Dj1
V2 Dj2
V2 Djihad
Groupe 1
V1 Dj1
V1 Dj2
V1 Djihad
V2 N1
V2 N2
V2 Neutre
Groupe 2

Occurrences
8295
33359
41654
7872
40191
48063
89717
12794
35405
48199
5790
36002
41792
89991

Formes
1227
2926
4153
1224
3325
4549
8702
1677
2966
4643
961
3013
3974
8617

Hapax
689
1538
2227
685
1679
2364
4591
906
1492
2398
517
1561
2078
4476

Fréquence Max
300
1049
1349
260
1225
1485
2834
368
1096
1464
168
1205
1373
2837

Forme
de
de
de
de
de
de
de
Et
Je
Je
La
Je
Je
Je

Pour le groupe 1 (N en 1 et V en 2) la forme la plus fréquente est « de » alors
que pour le groupe 2, c’est « je ». Les caractéristiques sont à peu près
équivalentes quelle que soit la vidéo projetée en 1. Quelle que soit la vidéo
projetée, et quel que soit l’ordre, pour les deux groupes on remarque que la
première exposition à la vidéo provoque moins de réactions (paroles=
nombre de formes) que la seconde, ce qui est a priori dû au fait que les
entretiens 2 (soir) et 3 (lendemain matin) contiennent un entretien de

L’un des principaux critères de recherche était de voir si les émotions étaient
plus ou moins mieux intégrées à 11h d’intervalle de jour ou de nuit. Tous les sujets
ont donc vu les 2 vidéos deux fois, à 11h d’intervalle entre chaque projection. 13 sujets
dans l’ordre vidéo V matin et soir et N soir et matin, 13 sujets au contraire dans
l’ordre vidéo N matin et soir et V soir et matin.
4

280

JADT’ 18

mémoire de la vidéo, avant la seconde projection sont plus longs. Cependant,
quel que soit l’ordre de passage, l’ensemble des sujets, tout groupes
confondus, parlent plus (environ 7000 occurrences de plus), à propos de la
vidéo V (stop djihad), qu’à propos de la N. Une tendance se dessine ainsi
selon laquelle la confrontation à la violence provoquerait une prise de parole
en « je » et un besoin de parler plus important.
2.2. Analyse du lexique « émotionné »
Reconnues comme des « moments » spécifiques instantanés, les émotions
sont définies comme « une réaction physique et/ou psychologique due à une
situation. », dont l’effet peut parfois se prolonger plus ou moins dans le
temps en fonction de leur intensité (Coletta & Tcherkassof, 2016; voir aussi
Bourbon, 2009 ; Feldman et al,. 2016 ou Fiehler, 2002). Pour étudier le lexique
des émotions, nous avons regroupé sous formes de listes des mots identifiés
dans le corpus et en fonction des concordances comme se rapportant à
l’expression de 4 des 6 émotions de base selon Ekman (1972) à savoir : la joie,
la colère, la tristesse et la peur (ici nommée inquiétude). Ce choix de 4
émotions et du terme « inquiétude » au lieu de « peur » a été fait en
adéquation avec les tests BMIS (échelles d’auto-évaluation de l’état
émotionnel par les sujets) demandés aux volontaires avant et après chaque
projection de vidéo. Les termes du lexique « émotionné » sont rassemblés cidessous par « groupes de formes ». Ainsi par exemple agréable+ contient
agréable(s)(ment) :
Bonheur/ Joie : Adoucit ; agréable+ ; allégresse ; ambiance+ ; amusé+ ;
apaisant+ ; bon+ ; calme+ ; content+ : désir+ ; emballer+ ; émerveillé ;
émouvoir+ ; excitant+ ; fière ; gai+ ; heureux+ ; jaloux* ; joie,+ ; marrant+ ;
paisible ; ravi ; serein+ ; surpris+
Colère : aberrant+ ; agacée+ ; agressé+ ; blasé+ ; chiffonne ; choc/choquer+ ;
colère ; énerver+ ; fâcher ; frappant+ ; furieux ; haine ; hard ; heurté+ ;
horreur+ ; horripile+ ; hostile+ ; irriter+ ; révolter+ ; saoulé
Inquiétude/ Peur :agitation+ ; angoissant+ ; anxiété+ ; apeuré+ ; crainte ;
effraiement*, effrayant+ ; flippant+ ; gêne+ ; incompréhensible+ ; nerveux+ ;
perdre+ ; peur+ ; stressant+ ; terreur
Tristesse : affecter+ ; affreux+ ; attristé+ ; bouleversé+ ; déception/déçu+ ;
dégoût+ ; déprimant+ ; dérange+ ;désolant+ ; impuissance ;
malheureusement, malheureux ; mélancolique; navrée ; peine+ ; triste+
Nous avons ici fusionné les émotions positives et neutres dans un même
groupe, ce qui explique que sous « joie » soient listés les termes « apaisante,
calme, serein » qui ne signifient pas éprouver de la joie, mais dont l’axiologie
est évaluée comme positive car exprimant une certaine neutralité
émotionnelle (Kerbrat-Orrechioni, 1980). De même, le terme « jaloux » dans
la colonne « joie » prête à interrogation : la jalousie est normalement associée

JADT’ 18

281

à l’expression d’un désir négatif, de l’ordre de l’inquiétude et de la colère ;
mais elle traduit ici du désir, comme le montre le contexte : «… ça faisait, ça
faisait très envie et ça rendait un peu jaloux». Ici, « jaloux », comme « envie »,
exprime un désir positif, qui va dans le sens d’un bien-être, contrairement à
son axiologie sémantique intrinsèque. De même, le terme « chiffonne »
(préoccuper, contrarier) est également une émotion négative qui devrait
trouver sa place plutôt dans la colonne de l’inquiétude. Mais en contexte, il
correspond ici à de la colère (« énerve » serait ici un synonyme) : « … ça me,
ça me chiffonne un peu de voir ce genre de, de, de vidéo à chaque fois ».
Enfin, le néologisme « éffraiement* », substantif masculin construit sur le
verbe effrayer, est ici associé à la peur, nous permettant de le classer dans la
colonne inquiétude : « un petit peu de peur et, et d’effraiement5 ». D’une
manière générale, pour une étude fine, tous les termes ici listés nécessiterait
une analyse développée, en contexte ; ce qui est l’objet d’une autre
publication.
3. Evaluation des émotions en contexte
L'analyse en concordance du lexique émotionné relevé ci-dessus révèle des
éléments significatifs avec le tri « avant », synthétisés dans le tableau cidessous. Ces résultats ont été doublés par des graphiques de ventilation :
Tableau 2 : synthèse des locutions adverbiales ou adverbes accompagnant
les expressions des émotions
Joie
un (petit) peu
un (peu) plus
(encore/beaucoup) plus
aussi
assez
plutôt
moins
pas très
pas
très
vraiment
autant
surtout

10
8
20
0
5
8
7
8
12
13
0
0
0

Colère
37
0
27
2
9
8
5
0
0
0
3
0
0

Inquiétude
37
4
8
2
2
1
0
0
0
1
4
3
0

Tristesse
36
0
9
0
0
2
0
0
7
0
0
0
4

On peut ici interroger à un niveau plus large le principe même de la création
néologique en rapport avec le contexte de l’émotion, qui peut se traduire au niveau de
la production verbale comme au niveau du corps, par différentes perturbations
(bégaiement, intonation, respiration changée, ne plus trouver ses mots…) (voir
Plantin, 2016) ; perturbations dont la création de néologismes serait l’une des
manifestations sur le plan lexical.
5

282

JADT’ 18

Figure 1 : Histogramme représentant les locutions adverbiales présentes à proximité des
expressions d’émotion (fréquences relatives)

Le contexte interactionnel de l’étude où l’on demande aux interviewés
d’évaluer les émotions ressenties, génère comme on le voit des réponses
presque systématiquement accompagnées d’adverbes ou locutions
adverbiales exprimant une intensité positive, équivalente, ou négative. De
manière significative, on relève ensuite une accentuation de l’intensité
positive lorsqu’il s’agit d’exprimer la joie (« encore/beaucoup/plus » 20 fois,
« très » 13 fois) alors que « un (petit) peu » est hyper présent pour atténuer
significativement les émotions négatives ressenties (colère, inquiétude,
tristesse). La seconde projection graphique permet de voir que, lorsque la joie
est exprimée, elle l’est de manière plus diverse, comparativement aux
émotions négatives. Ces résultats indiquent que pour le corpus étudié, qui
s’intéresse à la réception d’un discours violent, l’expression de l’intensité
correspond à celle d’une atténuation. On peut voir par exemple que
l’inquiétude et la tristesse sont les émotions qui attirent le plus la locution
d’intensité « un peu » qui tend à restreindre l’intensité de l’émotion perçue
par le locuteur (Coupin, 1995). Il est possible également que cela soit dû au
fait que ce sont des émotions plus diffuses et plus difficiles à caractériser de
manière tranchée que la joie et la colère, que l'on identifie assez facilement
lorsqu’on les ressent. Cela est confirmé par le fait que les émotions positives
sont accompagnées de locutions adverbiales marquant une forte intensité
(encore/beaucoup ; plus et très) : les locuteur.trice.s expriment leur joie avec
certitude et n’ont pas peur de la dire. De manière significative, c'est
également le cas pour l'expression de la colère, qui semble être l'émotion la
plus caractérisée adverbialement, à la fois par des éléments atténuateurs et
par des éléments intensificateurs («un (petit) peu» 37 occ. et
« encore/beaucoup/plus » 27 occ.), ce que l’on peut interpréter comme
l’expression du fait que les volontaires ne sont pas particulièrement

JADT’ 18

283

heureux.ses de se trouver exposé.e.s deux fois à la vidéo V et le manifestent
de cette manière. Le contexte apparaît ici fondamental : la colère est liée
d’une manière ou d’une autre ici à une forme d’impuissance face à la fois aux
attentats terroristes, aux images montrées qui sont en lien plus ou moins
direct selon les sujets, avec les attentats et l’état d’urgence et avec la situation
des civils syriens.

Figure 2 : Graphiques de ventilation par partition : V en N

Les graphiques de ventilation par partition vidéo V et N montrent les
émotions exprimées par les volontaires selon les vidéos visualisées. Les
émotions négatives (colère, inquiétude, tristesse) sont élevées en V ; à
l’inverse la joie est assez élevée en N. On remarque une variation des
émotions entre le premier et le second visionnage des vidéos : en effet, la
verbalisation des émotions négatives tend à baisser lors du second
visionnage (V1 à V2) alors que les émotions positives augmentent de V1 à V2.
Le même phénomène s’observe à l’inverse :les émotions positives baissent de
N1 à N2, et les négatives augmentent de N1 à N2, ce que montre le tableau cidessous :
Tableau 3: tableau récapitulatif des graphiques de partition v1 et v2

Groupe 1

Joie
Colère
Inquiétude
Tristesse

V1=N

V2=DJ

159
153
145
84

154
215
202
134

Groupe 2
V1 –
V2
5
62
57
50

V1=DJ

V2=N

245
167
100
124

259
105
43
74

V1 –
V2
14
62
57
50

284

JADT’ 18

Conclusion
Les réactions des sujets montrent de manière attendue, que la vidéo V génère
des émotions négatives et N, des émotions positives. En revanche, l'intensité
des émotions exprimées tend à être influencée par l'ordre dans lequel sont
vues les vidéos :dans le groupe 1, 1’expression de la joie est exprimée 159
fois ; elle est exprimée 259 fois en N dans le groupe 2. Lorsque les volontaires
voient d'abord la vidéo V, il semble que leurs réactions émotionnelles tendent
statistiquement à l'inverse de ce à quoi elles tendent dans l'ordre contraire :
ainsi l’expression verbale d’une émotion de bonheur tend à être supérieure
lorsqu'ils voient la vidéo N après la V, et l'expression de la colère,
l’inquiétude et la tristesse sont nettement inférieures. L’étude du lexique
émotionné tend à montrer que les sujets ressentent plus de bien être lorsqu'ils
voient la vidéo N après la V, comme un soulagement, un apaisement qui
arrive après une scène violente. Lorsque la vidéo N est vue en premier,
néanmoins, un certain facteur de stress émotionnel demeure, dû
probablement au fait que les sujets découvrent l'expérimentation et ne savent
pas ce qu'ils vont voir.
References
Bourbon B., (2009). L’expression des émotions & des tendances dans le langage,
University of Michigan Library.
Colletta J.-M. et Tcherkassof A. (2003). Les émotions. Cognition, langage et
développement. (P. Mardaga, Éd.) Belgique:Mardaga.
Coupin C. (1995). La quantification de faible degré : le couple peu/un peu et la
classe des petits opérateurs, thèse de doctorat, dir. Oswald Ducrot, EHESS.
Feldman B. L., Lewis M., Haviland-J. et Jeanette M. (2016). Handbook of
Emotions, Fourth Edition, The Guildford Press.
Fiehler R. (2002). « How to Do Emotions withWords : Emotionality in
Conversations », in Fussel, Susan (ed.) The Verbal Communication of
Emotions, London, Lawrence Erlbaum,pp.87-107.
Fracchiolla B., Moïse C., Romain C. et Auger N. (2013). Violences verbales
Analyses, enjeux et perspectives. Rennes: Presses Universitaires de Rennes.
Kerbrat-Orecchioni C. (1980) L’énonciation. La subjectivité dans le langage, Paris,
A. Colin.
Perrin L. (2016). « La subjectivité de l’esprit dans le langage », in Rabatel A. et
al. (éds) Sciences du langage et neurosciences (Acte du colloque de l’ASL
2015), Lambert-Lucas, 189-209.
Plantin Ch. (2011). Les bonnes raisons des émotions. Principes et méthode pour
l’étude du discours émotionné. Berne, Peter Lang.

JADT’ 18

285

Improving Collection Process for Social Media
Intelligence: A Case Study
Luisa Franchina1, Francesca Greco2, Andrea Lucariello3,
Angelo Socal4, Laura Teodonno5
1

AIIC (Associazione Italiana esperti in Infrastrutture Critiche) President –
blustarcacina@gmail.com
2Sapienza University of Rome – francesca.greco@uniroma1.it
3Hermes Bay Srl – a.lucariello@hermesbay.com
4Hermes Bay Srl – a.socal@hermesbay.com
5Hermes Bay Srl – l.teodonno@hermesbay.com

Abstract
Social Media Intelligence (SOCMINT) is a specific section of Open Source
Intelligence. Open Source Intelligence (OSINT) consists in the collection and
analysis of information that is gathered from public, or open sources. Social
Media Intelligence allows to collect data gathering from Social Media web
sites (such as Facebook, Twitter, YouTube etc…). Both OSINT and SOCMINT
are based on the Intelligence Cycle. This Paper aims to illustrate advantages
gained by applying text mining to collection phase of the intelligence cycle,
in order to perform threat analysis. The first step for detecting information
related to a specific target is to define a consistent set of keywords. Web
sources are various and characterized by different writing styles. Repeating
this process manually for each source could be very inefficient and time
consuming. Text mining specific software have been used in order to
automatize the process and to reach more reliable results. A partially
automatized procedure has been developed in order to gather information on
specific topic using the Social Media Twitter. The procedure consists in
searching manually a set of few keywords to be used for a specific threat
analysis. Then TwitteR of R Statistics was used to gather tweets that were
collected in a corpus and processed with T-Lab software in order to identify a
new list of keywords according to their occurrence and association. Finally,
an analysis of advantages and drawbacks of the developed method.
Abstract
La Social Media Intelligence (SOCMINT) è una sezione specifica di Open
Source Intelligence. L’Open Source Intelligence (OSINT) consiste nella
raccolta e analisi di informazioni da fonti pubbliche o aperte. La Social Media
Intelligence consente di raccogliere dati da siti Web di social media (come
Facebook, Twitter, YouTube ecc.). Sia l’OSINT che la SOCMINT sono basate

286

JADT’ 18

sul ciclo di Intelligence. Il presente documento intende illustrare i vantaggi
ottenuti applicando tecniche di text mining alla fase di raccolta del ciclo di
intelligence, al fine di eseguire analisi delle minacce. Il primo passo per
individuare le informazioni relative ad un obiettivo specifico è definire un
insieme coerente di parole chiave. Le fonti Web sono varie e caratterizzate da
diversi stili di scrittura. La ripetizione manuale di questo processo per
ciascuna fonte potrebbe essere molto inefficiente e dispendiosa in termini di
tempo. Sono stati utilizzati software specifici di text mining per
automatizzare il processo e ottenere risultati più affidabili. È stata sviluppata
una procedura parzialmente automatizzata al fine di raccogliere
informazioni su argomenti specifici utilizzando il Social Media Twitter. La
procedura consiste nella ricerca manuale di un gruppo di poche parole
chiave da utilizzare per un'analisi specifica delle minacce. Quindi il pacchetto
TwitteR di R Statistics è stato utilizzato per raccogliere i tweet che sono stati
raccolti in un corpus ed elaborati con il software T-Lab al fine di identificare
un nuovo elenco di parole chiave in base al loro verificarsi e associazione.
Infine viene fornita un'analisi dei vantaggi e degli svantaggi della procedura
sviluppata.
Keywords: Social Media Intelligence, Twitter, text mining, data collection
1. Introduction
“Open Source Intelligence [OSINT] is the discipline that pertains to
intelligence produced from publicly available information that is collected,
exploited, and disseminated in a timely manner to an appropriate audience
for the purpose of addressing a specific intelligence requirement”
(Headquarters Department of the Army, 2010, p. 11-1). OSINT is mainly used
in the framework of national security, by law enforcement to conduct
investigations, and in business field to gather important information. Social
Media Intelligence (SOCMINT) is a specific section of OSINT which focuses
on Social Media.
In recent years, with the spread of Internet, and the high amount of readily
accessible data, which give a picture of the actual state of things, the
importance of OSINT and SOCMINT has grown, becoming a key enabler of
decision and policy making. To bring the best out of such flow of data, the
intelligence process must take place as a systematic approach structured
around clear steps: planning and direction; collection; processing; analysis
and production; dissemination. These stages, each of which is vital, create the
Intelligence Cycle (CIA - Central Intelligence Agency, 2013). In order to
automatically collect data from both the web and the Social Media, OSINT
dashboards are being developed (Brignoli et Franchina, 2017).

JADT’ 18

287

This paper describes the contribution provided by automated support tools
in the collection phase of the Intelligence Cycle from a Social Media (Twitter)
on the phenomenon of interest. To capture the real essence of text available
and turn data publicly collected into valuable and reliable knowledge, text
mining techniques were implemented. To this aim, text mining plays a
relevant role as it enables the detection of meaningful patterns to explore
knowledge from textual data. As stated by Feldman and Sanger: “Text
mining can be broadly defined as a knowledge-intensive process in which a
user interacts with a document collection over time by using a suite of
analysis tools. In a manner analogous to data mining, text mining seeks to
extract useful information from data sources through the identification and
exploration of interesting patterns” (Feldman et Sanger, 2007, p. 1).
2. The use of Twitter
Twitter is a common Social Media, a microblog mainly for real time
information and communication. With Social Media becoming the main tool
for informational exchange, in October 2017, Twitter reached about 330
million users (Statista, 2018).
Twitter’s specific characteristics makes such a social particularly suitable for
SOCMINT purposes. Contents can be accessed by anyone, with no need to
create an account. Its users interact with short messages called “tweet”,
whose length is limited to 280 characters and can be embedded, replied to,
liked and unliked. Tweet quick nature, which can then be easily compared to
SMS (Short Messaging Service) messaging, fosters the use of acronyms and
slang, providing a real-time feel as they bring the first reaction to an event.
Phrasing can be simple in structure or imply a large amount of hapax.
With Twitter becoming one of the most important web application, it
provides a big amount of data and therefore it constitutes a vital source for
Social Media Intelligence. Thanks to its characteristics (potential reach, oneon-one conversation, promotional impact), Tweeter gained importance over
years in different social fields, from policy, to media communication and
terrorism. As a result, it is commonly considered a valuable source to
monitor social phenomena and their changing pattern.
3. Case Study
This paragraph illustrates how text mining tools can be integrated into the
SOCMINT data collection phase. The aim of the procedure is to select a
suitable and limited list of keywords allowing for an effective and efficient
information retrieval in order to support the analyst work.
In this case study the analyst was interested in collecting tweets on the
criminal and antagonist threat macro thematic that is related to many specific

288

JADT’ 18

topics as, for example, critical infrastructures or telecommunications. The
collection process has to identify a list of keyword able to collect the
messages concerning, for example, "the criminal and antagonist threat in
relation to critical infrastructures". The process can be illustrated by a cycle of
four different steps: selection of keywords related with the specific tropic
performed by the analyst; tweets collection; text mining; and verification and
list of keywords definition (figure 1).

Figure 1: illustration of automatic process for Twitter’s data collection four steps cycle

3.2. Keywords selection
The first step is performed by the analyst and consists in defining a suitable
list of words which could be used in order to collect tweets related to a
specific thematic, which in our example could be Critical Infrastructures. To
each X topic there is a set of keywords defining it (X1, X2, … Xn), e.g., railway,
station, airport. The same topic is made by all possible sets, given by the
formula:

3.1. Tweets collection
Once the keywords are selected, the second step consists collect data from
Twitter repository, e.g. using the twitteR package of R statistics (Gentry,
2016), in order to identify the keywords allowing for the collection of a
certain amount of tweets, that in our example was more than one hundred in
a day. That is, a word could perfectly represent the topic but could be rarely
used in the messages, resulting in a collection of a small sample of tweets.
The aim of this step is to find these words that allows for an effective data
collection (n ≥ 100), eliminating those words that are rarely used in the

JADT’ 18

289

messages (n < 100). That makes information retrieval more effective as the
number of keywords that can be used is limited.
3.3. Text Mining
After the keywords’ data collection efficacy was checked, a ten day messages
collection was performed including the retweets (49,3%), which is the data
retrieval maximum limit of the Twitter repository. The large size corpus
(token = 284.253) of 19.491 tweets was cleaned and pre-processed by the
software T-Lab (Lancia, 2017) in order to build a vocabulary (type = 19.765;
hapax = 8.947) and a list of content words (nouns, verbs, adverbs, adjectives)
(table 1). Then the list of content words was checked in order to identify the
new keywords and to implement the list.
Table 1: List of the first 20 lemmas of the list
Word
stazione

n
6066

Word
elettrico

n

Word

n

Word

2226

treno

1198

via

n
825

Word
ferrovia

n
659

aeroporto

4734

nuovo

1581

regione

1025

Milano

731

repubblica

632

impianti

3605

rifiuti

1536

Zingaretti

1022

autorizzare

720

giorni

627

Roma

3337

comune

1317

aiutare

Italia

679

centrale

605

896

In order to perform a content analysis, keywords were selected. In particular,
we used lemmas as keywords filtering out the lemmas below ten
occurrences. Then, on the tweets per keywords matrix, we performed a
cluster analysis with a bisecting k-means algorithm (Savaresi et Boley, 2004)
limited to twenty partitions, excluding all the tweets that did not have at
least two keywords co-occurrence. The eta squared value was used to
evaluate and choose the optimal solution.
The results of the cluster analysis show that the keywords selection criteria
allow the classification of 98.53% of the tweets. The eta squared value was
calculated on partitions from 3 to 19, and it shows that the optimal solution is
13 clusters (η2 = 0,19) (figure 2). Then, the analyst controlled for the lexical
profile of each cluster in order to detect the words useful to focus data
collection by means of the Boolean operators.
This procedure allows for the identification of a short list of most used words
(about 20) with regard to both the macro thematic and the related topic. The
list of keyword was then further reduced and it was reached a set off five
meaningful words for each intersection of the macro thematic with a specific
topic. Such a reduction stems from the fact that the use of a bigger amount of
words led to an exponential increase of false - positive production rate.

290

JADT’ 18

Figure 2: Eta squared difference per partition

As abovementioned, though such a work methodology effectively enables to
extract more often used words, with regard to Twitter it is still necessary to
test keywords to delete “noise” they produce, which however will not be
eliminated entirely. In other words, this methodology affects keywords’
amount on the basis of redundancies used by users. However, keywords’
quality should be tested in Twitter search engine in order to reach a level of
acceptance which includes both false and positive negative. Such words
made up the vocabulary to be used to identify intersection between the
macro thematic and a specific topic, i.e in the first case “criminal and
antagonist's threat with regard to critical infrastructure”, in the second case
“criminal and antagonist’s threat with regard to telecommunication” etc.
Between words identified there is an OR relationship. Example: terrorism OR
attack OR attack at station OR airport OR railway. Intersection between
cluster “criminal and antagonist’s threat” and “critical infrastructure is
synthetized by the following formula:

Where A is the cluster “criminal and antagonist’s threat”, B is “critical
infrastructure” and C is the intersection, which is “criminal and antagonist’s
threat with regard to “critical infrastructures”. The following image shows an
example.

JADT’ 18

291

Figure 3: an example of a possible set of words defining the intersection of the cluster
“criminal and antagonist’s threat”, with the topic “critical infrastructure”

3.4. Verification test
Finally, the list of keywords was tested on the Open Source Intelligence
dashboard. Collected Tweets were analyzed in order to identify the level of
its reliability to monitor the desired phenomena.
4. Conclusion
The developed process reflects the reliability of text mining software in
supporting information gathering process for Social Media Intelligence
purposes. The vocabulary identified for four different clusters, each of one
covering a specific topic, is being tested at this very moment on an advanced
dashboard in order to evaluate reliability. However, the role of the analyst is
still fundamental. The relationship between OSINT dashboard and analysts
must be complementary: dashboard plays a key role in gathering a big
amount of tweet, but it is still necessary the analyst support in choosing the
suitable keywords to be upload in the database, in order to render
information collection more effective. Indeed, OSINT dashboard can’t
understand Twitter users’ use of metaphors and similarities: keywords
choice must be made in accordance with monitoring targets. It should be
recalled that Italian language is really complex and it might occur that users’
language don’t refer to chosen target. Let’s see a practical example: some
keywords which usually refer to criminal threats (bomba - bomb or furto theft) can be used in Italian language also to refer to synthetic concepts with
regard to football or business offers (“bomba” might be used to mean a goal
scored through a powerful strike; “furto” might be used to mean that a
particular business offer is uneconomical). Another very important issue,
which can’t be solved without analysts, regard ironic tweets: dashboard

292

JADT’ 18

collects all information uploaded into database but it can’t subdivide tweets
into ironic and non-ironic by means of interpretation. To conclude, as
dashboards don’t understand textual meaning of words, analysts are
required to support dashboards’ capabilities, being the only ones to interpret
the specific meaning of words.
References
Brignoli M. A., and Franchina L. (2017). Progetto di Piattaforma di
Intelligence con strumenti OSINT e tecnologie Open Source. Proceedings
of the First Italian Conference on Cybersecurity (ITASEC17), Venice, Italy,
pp. 232-241.
CIA,
Central
Intelligence
Agency
(2013).
Kids'
Zone.
CIA,
https://www.cia.gov/kids-page/6-12th-grade/who-we-are-what-wedo/the-intelligence-cycle.html
Feldman R. and Sanger J. (2006), The Text Mining Handbook: Advanced
Approaches in Analyzing Unstructured Data. Cambridge University Press.
Gentry J. (2016). R Based Twitter Client. R package version 1.1.9.
Headquarters Department of the Army (2010). FM 2-0 Intelligence: Field
Manual. USA Army, https://fas.org/irp/doddir/army/atp2-22-9.pdf
Lancia F. (2017). User’s Manual: Tools for text analysis. T-Lab version Plus 2017.
Savaresi S.M. and Boley D.L. (2004). A comparative analysis on the bisecting
K-means and the PDDP clustering algorithms. Intelligent Data Analysis,
8(4): 345-362.
Statista (2018). Twitter: number of monthly active users 2010-2017. Statista,
https://www.statista.com/statistics/282087/number-of-monthly-activetwitter-users/

JADT’ 18

293

The impact of language homophily and similarity of
social position on employees’ digital communication
Andrea Fronzetti Colladon, Johanne Saint-Charles, Pierre Mongeau

1. Introduction
Knowledge creation and organizational communication are fundamental
assets to obtain strategic competitive advantage (Tucker, Meyer, &
Westerman, 1996) and in modern organization most of these happen through
digital communication. We know that the way employees use digital
communication can predict their engagement level (Gloor, Fronzetti
Colladon, Giacomelli, Saran, & Grippa, 2017) as well as future business
performance (Fronzetti Colladon & Scettri, 2017). Hence there is a need to
better understand what is affecting employees’ participation in internal
communication in order to foster the efficacy of internal communication and
to deliver effective messages and campaigns in the most strategic way. Based
on the idea of homophily, this paper examines if employees’ participation in
their organization intranet is linked with their similarity in discourse and in
network positions. Communication, digital or not, encompasses both the
language people are using to communicate and the interactions and
relationships they have (Tietze, Cohen, & Musson, 2003; White, 2011). In the
last two decades scholars have explored how people’s discourse1 and
relationships are intertwined notably through the lenses of social network
analysis. Among others, those studies have shown that social relationships or
interactions between people are linked to the similarity of the words and
expressions they use (Basov & Brennecke, 2018; Nerghes, Lee, Groenewegen,
& Hellsten, 2015; Roth & Cointet, 2010; Saint-Charles & Mongeau, 2018).
Also, Gloor and colleagues have proposed a framework to study online social
dynamics in which language plays an important role, especially with regards
to the dimensions of sentiment, emotionality and complexity (Gloor et al.,
2017). Such results align with the notion of homophily that corresponds to
the tendency to relate to others on the basis of similarities (Lazarsfeld &
Merton, 1954). A tendency now acknowledged as an important factors for the
constitution of social networks (Mcpherson, Smith-Lovin, & Cook, 2001). It is
assumed that this similarity leads to the development of relationships since
similarity is linked to attraction towards the other (Montoya & Horton, 2013).

1 Discourse is define here as “a general term that applies to either written or
spoken language that is used for some communicative purpose” (Ellis, 1999, p. 81).

294

JADT’ 18

Considering digital communication Brown, Broderick, & Lee (2007) and
Yuan & Gay (2006) showed that ties strength and computer-mediated
interaction increases with homophily. Most of the studies have explored
similarities with regards to sociodemographic variables but several authors
have expanded this to a wide range of variables including attitudes,
psychological traits, values, etc. as latent homophily factors (Lawrence &
Shah, 2007; Shalizi & Thomas, 2011). Hence, given that interaction in digital
communication happens through written text, we assume that discourse
similarity of employees’ messages is a key homophilic determinant for
employees’ interactions in the network of internal digital communication.
Similarity can also be observed with regard to network position. Indeed,
occupying an equivalent position in a network was shown to lead to similar
outcomes (attitudes, points of view, roles, etc.) (Borgatti & Foster, 2003; Burt,
1987). In the study of large on-line networks, actors’ similarity in centrality
has proven useful for identifying role-similarity of actors in the network
(Roy, Schmid, & Tredan, 2014). According to Gloor et al. (2017), it is also
important to investigate the dynamic evolution of social positions. Rotating
leaders, for example, proved to play a very important role in online
communities, supporting their growth and participation (Antonacci,
Fronzetti Colladon, Stefanini, & Gloor, 2017). In sum, the “homophily
phenomenon” has been largely demonstrated through the study of various
types of similarities. This paper seeks to explore this phenomenon in the
context of the use of internal digital communication system in an
organization and we propose to use discourse and network position
similarity measures to this avail, our overall hypothesis being that the two
are correlated and that they are correlated with interactions.
2. Research Design and Methodology
We analyzed the digital communications of about 1,600 employees working
for a large multinational company, mainly operating in Italy. This company
has a largely popular intranet social network, structured as an online forum,
where only employees can interact, exchanging opinions and ideas through
the sharing of news and comments. We could extract and analyze more than
23,000 posts (news and comments), written in Italian over a period of one
and a half year. Users were mostly males (68%) and a small part of them also
played the role of content managers (7%).
The first step in our analysis was to build the social network which
represents the forum interactions. This network is made of N nodes, one for
each forum user, and M edges. In general, there is an edge between two
nodes if the corresponding employees had at least one interaction – for
example, they exchanged knowledge or opinion through subsequent

JADT’ 18

295

comments, or one answered a question of the other. We then proceeded to
calculate the similarity measures for both discourse and network position.
Based on what was presented above, we looked at five aspects of discourse
similarity: words use, sentiment, emotionality, complexity and length.
Additionally, we studied employees’ connectivity and interactivity, as
suggested by Gloor and colleagues (2017). We further explored employees’
use of language by looking at the sentiment, emotionality, complexity and
length of their forum posts. Length is simply calculated as the average
number of characters used in forum posts by an employee – after having
removed stop-words and punctuation, via a script written using the Python
programming language and the package NLTK (Perkins, 2014). Sentiment
expresses the positivity or negativity of forum posts and is calculated thanks
to the machine learning algorithm included in the social network and
semantic analysis software Condor (Gloor, 2017). Sentiment varies between 0
and 1, where 0 represents a totally negative post and 1 a totally positive one.
Emotionality expresses the variation from neutral sentiment and is computed
by Condor using the formula presented by Brönnimann (2014). Posts that
convey less neutral expressions, either positive or negative, are considered
more emotional. Lastly, complexity represents the deviation from common
language and is calculated as the probability of each word of a dictionary to
appear in the forum posts (Brönnimann, 2014); when rare terms appear in
forum posts more often, complexity is higher. Even this last measure was
obtained from Condor. Concerning the study of employees’ positions in the
social structure, we referred to network centrality measures (Freeman, 1979).
To measure centrality, we used the two well-known metrics of degree and
betweenness centrality. Degree centrality measures the number of direct
links of a node, i.e. the number of people an employee interacted with, in the
online forum. Betweenness centrality, on the other hand, takes into account
the indirect links of a node and counts how many times a social actor lies inbetween the paths that interconnect his/her peers. Betweenness centrality is
calculated by considering the shortest network paths that interconnect every
possible pair of nodes and counting how many times these paths include a
specific employee (i.e. the node for which the betweenness centrality is
calculated). Employees’ interactivity was operationalized by calculating
rotating leadership. This variable counts the oscillations in betweenness
centrality of a social actor, i.e. the number of times betweenness centrality
changed reaching local maxima or minima. If an employee maintains a static
position, his/her rotating leadership is zero. On the other hand, we have
rotating leaders when people oscillate more between central and peripheral
positions, activating or taking the lead of some conversations and then
leaving space to other people in the network. As control variables, we could

296

JADT’ 18

access to employees’ gender and forum role (content manager or not). Even if
gender homophily is not always supported by social networks studies, it is
very often used as a control variable, as it has been shown that gender can
influence online social communication and behavior (Thelwall, 2008, 2009).
Similarly, we control for content manager role, as we expect different
behaviors when employees have the assignment of informally moderating
the intranet social network. All the variables presented above were first
calculated at the node level and subsequently transformed into similarity
matrices. Like a network adjacency matrix, a similarity matrix is made of N
row and columns, where each row and column represents a specific
employee. For categorical attributes (gender and being a content manager or
not) we have a value of 1 in a cell of the matrix if the two corresponding
employees share the same attribute (for example they are both females), and
0 otherwise. For continuous variables, we populated the matrices with the
absolute value of the differences in individual actor scores.
3. Results
In general, we notice a prevalence of male employees, even if more forum
content managers are females (most of them working in the internal
communication department, which is mostly populated by females). Being a
content manager is also associated with more central and dynamic network
positions: content managers have on average higher scores of degree and
betweenness centrality and they rotate more. To put it in other words, they
have interactions with more people, often act as brokers of information and
in general do not keep a static dominant position after having fostered a
conversation.
As described in the previous section, we measured similarity with respect to
several characteristics of employees: their gender, content manager role, use
of language, centrality and interactivity. Text similarity shows the strongest
association with digital communication (ρ = 0.48). Employees who more
frequently use the same vocabulary communicate more between themselves.
Apart from gender and sentiment, homophily effects seem to be significant
for all the other variables included in our study. Employees that are more
similar with respect to their use of language, degree of interactivity and
network position tend to interact more between themselves.
As per agreed privacy arrangements, we are prohibited from revealing the
company name or other details that could help in its identification. It might
be useful to replicate our research to see if our findings are confirmed in
different business contexts. Future studies could include more control
variables, particularly those which are supposed to produce homophily
effects – such as employees’ age (Kossinets & Watts, 2009). Having more

JADT’ 18

297

accurate timestamps could also help in the assessment of average response
time, to see if more reactive users tend to cluster. As our was mainly an
association study, we advocate further research to carry out a longitudinal
analysis which could tell us which actor similarity effects can be considered
as significant antecedents of digital communication.
Our findings have practical implications both for company managers and
administrators of online communities. For example, if a company wants to
attract the attention of employees on a strategic topic, in the light of our
results, it appears vital to choose a language close to that of the target people.
Employees’ participation in conversations can be fostered by online messages
aligned with the general use of language and by choosing social ambassadors
who have network positions similar to the target.
References
Antonacci, G., Fronzetti Colladon, A., Stefanini, A., & Gloor, P. A. (2017). It is
Rotating Leaders Who Build the Swarm: Social Network Determinants
of Growth for Healthcare Virtual Communities of Practice. Journal of
Knowledge Management, 21(5), 1218–1239. https://doi.org/10.1108/JKM11-2016-0504
Basov, N., & Brennecke, J. (2018). Duality beyond Dyads: Multiplex
patterning of social ties and cultural meanings. Research in the Sociology
of Organizations, in press.
Borgatti, S. P., & Foster, P. C. (2003). The network paradigm in organizational
research: A review and typology. Journal of Management.
https://doi.org/10.1016/S0149-2063(03)00087-4
Brönnimann, L. (2014). Analyse der Verbreitung von Innovationen in sozialen
Netzwerken. University of Applied Sciences Northwestern Switzerland.
Retrieved from http://www.twitterpolitiker.ch/documents/Master_
Thesis_Lucas_Broennimann.pdf
Brown, J., Broderick, A. J., & Lee, N. (2007). Word of mouth communication
within online communities: Conceptualizing the online social network.
Journal
of
Interactive
Marketing,
21(3),
2–20.
https://doi.org/10.1002/dir.20082
Burt, R. S. (1987). Social Contagion and Innovation: Cohesion versus
Structural Equivalence. American Journal of Sociology, 92(6), 1287–1335.
https://doi.org/10.1086/228667
Ellis, D. G. (1999). From Language To Communication. New York, NY:
Routledge.
Freeman, L. C. (1979). Centrality in social networks conceptual clarification.
Social Networks, 1, 215–239.
Fronzetti Colladon, A., & Scettri, G. (2017). Look Inside. Predicting Stock

298

JADT’ 18

Prices by Analysing an Enterprise Intranet Social Network and Using
Word Co-Occurrence Networks. International Journal of Entrepreneurship
and Small Business, in press. https://doi.org/10.1504/IJESB.2019.10007839
Gloor, P. A. (2017). Sociometrics and Human Relationships: Analyzing Social
Networks to Manage Brands, Predict Trends, and Improve Organizational
Performance. London, UK: Emerald Publishing Limited.
Gloor, P. A., Fronzetti Colladon, A., Giacomelli, G., Saran, T., & Grippa, F.
(2017). The Impact of Virtual Mirroring on Customer Satisfaction.
Journal of Business Research, 75, 67–76.
https://doi.org/10.1016/j.jbusres.2017.02.010
Huang, A. (2008). Similarity measures for text document clustering. In
Proceedings of the sixth new zealand computer science research student
conference (NZCSRSC2008) (pp. 49–56). Christchurch, New Zealand.
Jivani, A. G. (2011). A Comparative Study of Stemming Algorithms.
International Journal of Computer Technology and Applications, 2(6), 1930–
1938. https://doi.org/10.1.1.642.7100
Kossinets, G., & Watts, D. J. J. (2009). Origins of Homophily in an Evolving
Social Network. American Journal of Sociology, 115(2), 405–450.
https://doi.org/10.1086/599247
Krackhardt, D. (1988). Predicting with networks: Nonparametric multiple
regression analysis of dyadic data. Social Networks, 10(4), 359–381.
Lawrence, T. B., & Shah, N. (2007). Homophily: Meaning and Measures. In
Paper presented at the International Network for Social Network Analysis
(INSNA). Corfu, Greece.
Lazarsfeld, P. F., & Merton, R. K. (1954). Friendship as a Social Process: A
Substantive and Methodological analysis. Freedom and Control in Modern
Society, 18, 18–66. https://doi.org/10.1111/j.1467-8705.2012.02056_3.x
Mcpherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a feather:
Homophily in social networks. Annual Review of Sociology, 27(1), 415–
444. https://doi.org/10.1146/annurev.soc.27.1.415
Montoya, R. M., & Horton, R. S. (2013). A meta-analytic investigation of the
processes underlying the similarity-attraction effect. Journal of Social and
Personal Relationships, 30(1), 64–94.
https://doi.org/10.1177/0265407512452989
Nerghes, A., Lee, J.-S., Groenewegen, P., & Hellsten, I. (2015). Mapping
discursive dynamics of the financial crisis: a structural perspective of
concept roles in semantic networks. Computational Social Networks, 2(16),
1–29. https://doi.org/10.1186/s40649-015-0021-8
Perkins, J. (2014). Python 3 Text Processing With NLTK 3 Cookbook. Python 3
Text Processing With NLTK 3 Cookbook. Birmingham, UK: Packt
Publishing.

JADT’ 18

299

Roth, C., & Cointet, J. P. (2010). Social and semantic coevolution in
knowledge networks. Social Networks, 32(1), 16–29.
https://doi.org/10.1016/j.socnet.2009.04.005
Roy, M., Schmid, S., & Tredan, G. (2014). Modeling and measuring graph
similarity: The case for centrality distance. In Proceedings of the 10th
ACM international workshop on Foundations of mobile computing, FOMC
2014 (pp. 47–52). New York, NY: ACM.
https://doi.org/10.1145/2634274.2634277
Saint-Charles, J., & Mongeau, P. (2018). Social influence and discourse
similarity networks in workgroups. Social Networks, 52, 228–237.
https://doi.org/10.1016/j.socnet.2017.09.001
Shalizi, C. R., & Thomas, A. C. (2011). Homophily and contagion are
generically confounded in observational social network studies.
Sociological Methods and Research, 40(2), 211–239.
https://doi.org/10.1177/0049124111404820
Tata, S., & Patel, J. M. (2007). Estimating the selectivity of tf-idf based cosine
similarity predicates. ACM SIGMOD Record, 36(2), 7–12.
https://doi.org/10.1145/1328854.1328855
Thelwall, M. (2008). Social networks, gender, and friending: An analysis of
mySpace member profiles. Journal of the American Society for Information
Science and Technology, 59(8), 1321–1330.
https://doi.org/10.1002/asi.20835
Thelwall, M. (2009). Homophily in MySpace. Journal of the American Society for
Information Science and Technology, 60(2), 219–231.
https://doi.org/10.1002/asi.20978
Tietze, S., Cohen, L., & Musson, G. (2003). Understanding organizations through
language. Understanding Organizations Through Language.
https://doi.org/10.4135/9781446219997
Tucker, M. L., Meyer, G. D., & Westerman, J. W. (1996). Organizational
communication: Development of internal strategic competitive
advantage. Journal of Business Communication, 33(1), 51–69.
https://doi.org/10.1177/002194369603300106
White, H. C. (2011). Identité et contrôle. Une théorie de l’émergence des formations
sociales. Paris: Éditions de l’École des hautes études en sciences sociales.
Yuan, Y. C., & Gay, G. (2006). Homophily of network ties and bonding and
bridging social capital in computer-mediated distributed teams. Journal
of Computer-Mediated Communication, 11(4), 1062–1084.
https://doi.org/10.1111/j.1083-6101.2006.00308.x

300

JADT’ 18

Looking Through the Lens of Social Sciences:
The European Union in the EU-Funded Research
Projects Reporting
Matteo Gerli
University for Foreigners of Perugia – matteogerli81@gmail.com

Abstract
In the last decades, European integration and scientific production have
come to be deeply intertwined as a result of the Europeanization of many
research activities. On one side, European institutions promote the
realization of research projects aiming at developing a type of knowledge
“close” to the end users’ interests; on the other side, the resulting knowledge
contributes to conditioning the practices that take place in the European and
national institutions, according to a circular process that brings the
innovations to feed back into the system that expresses them. The purpose of
this paper is to explore this relationship by examining two peculiar scientific
products realized by researchers operating within the broad domain of the
Socio-economic Sciences and Humanities (SSH), as a part of the research
projects financed by the Seventh Framework Programme (2007-2013) of the
European Union: final reports and policy briefs. In other words, it aims to
analyse all reports as a whole using some automatic text analysis tools, while
incorporating some supplementary variables which help to define the
broader context of scientific production.
Keywords: European Union, International Research Projects, Socio-economic
Sciences and Humanities, Textual Data Exploration, Quantitative Discourse
Analysis, IRaMuTeQ.
1. Introduction
The European Research Policy plays a strategic role for thousand of
researchers and research institutions which operate within the EU borders.
Thanks to the concomitant decrease in national public funds for scientific
activities (see for instance, Vincent-Lacrin, 2006; 2009), the European research
agenda has dramatically increased its appeal among scholars and
consequently its ability to have an impact on the directions and processes of
scientific knowledge production. Indeed, starting from the 90s, the European
Commission has equipped itself with new means to combine and manage, on
the basis of medium to long-term planning cycles, the whole set of scientific
and technological initiatives financed by the European budget: the framework

JADT’ 18

301

programme (Ippolito, 1989; Ruberti and André, 1995; Guzzetti, 1995;
Menéndez and Borras, 2000; Borras, 2000; Banchoff, 2002; Cerroni and
Giuffredi, 2015). In short, the underlying logic is that of the programmatic
intersection between research activities and other European policies, so that
the promotion of scientific excellence complements the need to foster the
creation of cross-border and interdisciplinary collaborations intended for
producing a type of knowledge “close” to the end users’ interests.
As it was observed in previous studies (Adler-Nissen and Kropp, 2015),
European integration and scientific production have come to be deeply
intertwined: on one side, the progress of integration process influenced (and
still influences) research activities through the promotion of particular forms
of knowledge and research questions (as far as we are concerned, mainly
through the realization of cross-national and cross-disciplinary research
projects); on the other side, the resulting knowledge contributes to
conditioning the practices that take place in the European and national
institutions, according to a circular process that brings the innovations to
feed back into the system that expresses them. Social Sciences and
Humanities, which are less directly involved in the production of knowledge
with a clear practical usability, are by no means unconcerned with this kind
of phenomenon. At this regard, the Journal of European Integration has recently
published a special issue on the relationship between social sciences and
European integration, hosting some important articles that have highlighted
the existence of several “crossroads” between the European Union and the
scientific community’s “itineraries”1: Rosamond (2015), for instance,
observed how certain theories on the political and economic integration (in
particular that of the Hungarian Béla Balassa, from the economics side, and
the neofunctionalism, from the political science side) had been informing the
“strategic narrative” adopted by the European Commission during the 60s
and 70s to legitimize its newly-formed institutional role and its economic
policy position, according to a quite peculiar two-ways traffic of influences
process, being the economic integration theorized while it was happening;
Deem (2015) pointed out the existence of a relationship between the birth of a
new field on higher education studies, the simultaneous evolution of national
university systems and the launch of the so-called Bologna process at
European level; Vauchez analysed, through a sociogenetic approach, the
historical process through which the acquis communautaire «has been
formulated, stretched, criticized, revised and finally naturalized as the most
rigorous and objective measure of Europe against other possible methods»
(2015: 196) thanks to the work of those who have been defined

1

Journal of European Integration, 37 (2015).

302

JADT’ 18

“methodological entrepreneurs”, that is European officials who have
politically invested and succeeded in establishing Europe’s cognitive and
technical equipment.
Looking beyond such individual cases, what is really relevant to our purpose
is the underlying idea about the possibility of studying science production
from a sociological point of view, basically by rejecting what was
traditionally regarded as an internal/external division (Adler-Nissen and
Kropp, 2015: 161-163), and thus admitting that even scientific and academic
concepts can be formulated in conjunction with political-economic ambitions
and practical problems (see Bohme et al., 1983; Funtowicz and Ravez 1993;
Slaughter and Leslie 1997; Gibbons et al., 1994; Ziman, 2000; Albert and
Mcguire, 2014), such as those above mentioned. This does not mean that
science is equal to politics or economics (Breslau, 1998); what it does mean is
that, in order to understand science production, one needs to recognize that
“non-academic” resources (such as, for instance, financial or material
resources, ideas and beliefs, symbolic resources, political or normative
resources, people, etc.) may overstep scientific boundaries and be used for
the production of new knowledge. Bourdieu (1975, 1984, 1990, 1992, 1994,
1995, 2001) described this phenomenon through the concept of “fields
interrelations”. In few words, the social word is composed of multiple semiautonomous fields, basically microcosms characterized by different stakes,
rules of the game and particular resources which one needs to possess to get
access to the game itself and its specific advantages. He conceptualized these
sphere as partially independent, by which he means that, even though each
field develops its own institutions, hierarchies, problems, tacit or explicit
rules, they necessarily interact and affect each other. This is particularly true
for cultural fields (art, cinema, religion, science, journalism, etc.), since they
are structurally dependent and subordinated to political and economic fields.
Going straight to the point, this is to say that, if one is dealing with a
sociological analysis of a cultural product (e.g. a text), thus one neither can
just consider its formal characteristics, nor be limited to its context of
production. Instead, one should use a “relational approach”, taking into
account both the internal features of the product and its external determinants.
In engaging with this broad issue, this paper will try to further contribute to
the understanding of the topic by examining two peculiar scientific products
realized by researchers operating within the broad domain of the Socioeconomic Sciences and Humanities (SSH), as a part of the research projects
financed by the Seventh Framework Programme (2007-2013) of the European
Union: final reports and policy briefs. By using some automatic text analysis
tools, it will thus statistically explore the contents of such documents not per
se, but in connection with some variables, which help to define the broader

JADT’ 18

303

context of production. In its exploratory character, this study does not have
strong hypothesis to be tested. Nevertheless, following Bourdieu’s approach,
it aims to give an original perspective through which observing the
relationship between the field of social sciences and the public policy field of
the European Union (Gerli, 2017).
2. The corpus and methodology
Unlike the studies discussed earlier, which are mainly based on microsociological observation, our investigation covers a macro-sociological
analysis of a quit large corpus made of 46.513 graphic forms, equal to
3.025.960 occurrences. It is an ad-hoc constructed corpus: it contains 360 texts,
of which 205 belonging to final reports and 155 to policy briefs, which were
collected from the digital database CORDIS2, the main institutional source of
information related to the research projects financed by the European Union.
The choice to focus on these documents is not accidental, but depends on
their strict relevance to our research objectives. In fact, both include a
summary of the project results and conclusions, with a description of their
potential socio-economic impact (EC 2010), even though policy brief is strictly
designed for policy makers (both European and national ones), while final
report is addressed to a wider audience, which may include (at least
potentially) lay people as well. In this perspective, they represent an effective
“shortcut” through which empirically observe the way in which the research
groups awarded a grant “actualized” the inputs they received from the
Commission. This is, to resume the previous discussion, to analyse how
European institutions and social scientists contribute together to the
definition and resolution of some EU-related issues.
With regard to the methodology, both simple and multivariate analyses were
performed with the IRaMuTeQ software (Lebart et al., 1998; Bolasco, 2013). In
particular, the lexicographical analysis was used for a first exploration of the
corpus, that is to identify and format texts units, turn texts into text segments
(TS) and classify words by their frequency. The multivariate analysis,
instead, was performed to detect the associations between textual data and
the following supplementary variables related to what in the 7FP was
defined as macro-activity (MA) and financing scheme (FS)3. Going into more
details, the 7FP included eight macro-activities: Growth, employment and
competitiveness in a knowledge society (MA1); Combining economic, social and
environmental goals in Europe: towards sustainable development (MA2); Major

http://cordis.europa.eu/projects/home_it.html.
For more details: Decision No 1982/2006/EC of the European Parliament and of
the Council of 18 December 2006.
2
3

304

JADT’ 18

trends in society and their implications (MA3); Europe in the world (MA4); The
citizen in the European Union (MA5); Socio-economic and scientific indicators
(MA6); Foresight studies (MA7); Strategic activities (MA8). As for the financing
schemes, the 7FP included five main different types, which differed from
each other by the research team size and the type of purposes to be achieved
(the first three mainly focused on the development of new knowledge, while
the last two were mainly thought for the coordination and support of
research activities and policies): Small or medium-scale focused project (FS1);
Small or medium-scale focused research project aimed at international cooperation
(FS2); Large-scale integrating project (FS3); Coordination action (FS4); Support
action (FS5). Additionally, we also took into account the starting year of the
project and the geographic area in which the coordinating institution was
located.
As a whole, our sample (of non-probabilistic type) involves 223 research
projects out of 251 realized in 2007-2013 (equal to 88.8%) and broadly covers
all macro-activities and financing schemes above mentioned. In Tab. 1, a
description of the corpus and its main subsets is provided.
Tab. 1: Description of the corpus
Type
Final report
Policy Brief
Corpus

Number of texts
205
155
360

Graphic forms
42.047
19.795
46.513

Occurrences
2.441.168
584.792
3.025.960

3. The main findings
At first glance, the most frequent “full” words used in the SSH research
reports do not provide particularly relevant insights. The first ten (social,
policy, research, European, project, EU, countries, public, national, Europe)
concerns the “general context of meaning” where discourses on Europe and
related issues took shape. Ten words that, without having a clear disciplinary
connotation, define some “semantic coordinates” common to all research
projects carried out. Interesting enough, it is the wide use of the words
country/es (freq.=10.531) and national (freq.=5.527) which, compared with the
words European (freq.=9.190), EU (freq.= 8.563) and Europe (freq.=5.408), prove
the great importance of the “national” level of analysis, mainly in a
comparative way. Scrolling down the list, we can also recognise some typical
words of the socio-economic lexicon (economic, market, growth, employment,
financial), the socio-political lexicon (people, education, State, young, groups,
cultural, society, governance), and the methodological one, namely related to
the operative context of the research activities (date, case, results, impact,
analysis, study). Yet these are terms that, at this early stage of the analysis, do
not provide any clear “message”.

JADT’ 18

305

At a closer look, however, we can identify some specific words which are, in
a broad sense, linked to the political macro-orientations defined by the
Lisbon Strategy (European Council 2000), demonstrating the “osmosis”
existing between European institutions and social sciences. Here some
examples: innovation (freq.=5.793), cornerstone of industrial competitiveness
and economic growth (EC 2003, 2006); development (freq.=5.176), to be
understood, among the various meaning, mainly as sustainable development
(EC 2005, 2009); education (freq.= 3.490) and knowledge (freq.=3.221), which,
together with the already mentioned “innovation”, represent the “three
sides” of the so-called “knowledge triangle”, from the European
Commission’s perspective, the ground for a greater economic and social
dynamism.
For the aim of this study, what is of particular interest is also the
geographical scope of the research activities. Indeed, the most frequent
toponyms refer to EU based countries. Among these, the five main sponsors
and recipients of the framework programs (Germany, UK, France, Italy and
Spain) are placed at the top of the ranking. As for the extra-European
countries, several of them are placed in Asia (e.g. China, Japan, India, Vietnam
and Thailand), North Africa (Morocco, Tunisia, Egypt and Libya) and South
America (Brazil, Argentina, Colombia, Peru and Chile). This is indicative of a
globalization process, which is affecting both European institutions and
researchers by expanding their interests (“political”, with regard to the first
ones, and “scientific”, for the second ones) beyond the European borders.
What matters is that they are moving together insofar we can suppose the
existence of a clear synergy between the emergence of a new multipolar area
of political, commercial and cultural influence, in which the European Union
is now required to act, and the production of knowledge on topics with a
potential “global” added value.
3.1 The main semantic groups and their connections with the “context”
To go deeper in the analysis, and to explore the relationship between the
selected texts and some variables related to their context of production, we
performed a Descending Hierarchical Analysis (DHA). Indeed, this method
allowed us, first, to identify clusters with similar vocabulary within text
segments and, then, to visualize them in conjunction with the supplementary
variables (Camargo and Justo 2013; Curbelo, 2017). In Fig. 1, the output of the
DHA is summarised.

306

JADT’ 18

Fig. 1: Dendrogram of top-down hierarchical classification (Reinert method) of the corpus

As it can be easily seen in Fig. 1, the DHA algorithm allowed the
identification of five clusters, each with its own specific semantic content.
Following Reinert (1987), they can be interpreted as “lexical words”, namely
specific semantic structures which, in our case, refer to different and even
competing scientific representations of the European Union and related
issues. The second cluster has the greater representation (26,8% of the SSH
discourses) and identifies a semantic sphere characterized by a language
mainly oriented towards political and social issues. Indeed, the most central
word in this cluster is political, followed by cultural, identity, citizenship, border,
conflict, citizen, State and so on. Immigration (migrant) and related issues
appear to be particularly relevant as well. The fifth cluster (24,1%) delineates
a quite peculiar semantic sphere based on a set of words (such as project,
conference, research, university, workshop, dissemination, website, etc.) strictly
linked with the management and realization of European research projects
and, more in general, with scientific research and related activities. The first
cluster, third in terms of representativeness (19%), refers to the relationship
between economic development and environmental protection, being the
most central word innovation, followed by development, economic, sustainable,
environmental, change, rural and so on. This interpretation seems to be
supported by the presence of several words that refer to the need for a
change with respect to a situation that is perceived as not desirable (change,

JADT’ 18

307

impact, strategy, challenge, need, solution, improve, step, etc.). The third cluster
(16,2%), instead, covers a semantic area mainly related to the economy and
the market. It is a language that involves two main branches, the one of the
real economy (income, price, household, wage, firm, energy, poverty, etc.), and the
one of the finance (financial, bank, risk, monetary, credit), but above all it is
characterized by the large presence of technical terms and acronyms (gdp,
estimate, asset, inflation, emu, Eurozone, insurance, macroeconomic, etc.). Finally,
the fourth linguistic cluster (13,9%) includes words essentially associated to
the relationship between education, training and employment, as shown by
the presence of terms such as young, person, child, school, education, aspiration,
background, vocational and compulsory. It is a cluster that differs from the
others due to the greater concreteness of the language, as proved by the
recurring use of words referring to “concrete” social actors (child, parent,
student, teacher, mother, friend, volunteer, etc.).
Fig. 2, resulting from a Lexical Correspondences Analysis (LCA), shows the
relationship between clusters (left side) and between clusters and the
supplementary variables (right side). The main aim here was to verify
whether or not SSH discourse exhibits clear evidence of “adaptability” with
regard to the macro-activities and the financing schemes, as defined by the
European Commission.

Fig. 2: Association between clusters and supplementary variables

The first two factors summarize together 67,5% of the total inertia: the first
one (39,97%) marks a clear opposition between cluster 5 (positive half-plane)
and the other four clusters (negative half-plane); the second factor (27,47%),
instead, highlights a significant opposition between clusters 1 and 3 (positive
half-plane) and clusters 2 and 4 (negative half-plane). As a whole, we can
distinguish three different (partially autonomous) semantic contexts, arising
from the association between the “cultural” and “socio-political” discourses

308

JADT’ 18

(third quarter), the “economic” discourse and that on “innovation” and
“sustainable development” (forth quarter), and finally the discourse on
“research activities” (in-between the first and the second quarters).
As far as the relationship between discourses (clusters) and supplementary
variables, Figg. 3 and 4 show the most significant categories (those with a
larger chi-square and a lower p-value), referring to the “macro-activity” and
“financing scheme” variables. As shown in the first figure, MA1 and MA2
categories are only significant in the definition of clusters 1 (innovation) and
3 (economics); MA5 is the most relevant for cluster 2 (politics); similarly,
MA3 category is the only significant for cluster 4 (culture); and finally, MA4
and MA8 categories predominate on cluster 5 (research activities). In short,
these results strongly support the thesis of adaptability, insofar the different
scientific representations of the European Union emerged from the analysis
resulted strongly associated with the macro-activities defined by the
European Commission.
Cluster
1
2
3
4
5

Category

Chi2

%

p-value

MA2

1226.7

25,7

<0.0001

MA7

762.9

36.5

<0.0001

MA5

5220.0

54.8

<0.0001

MA1

1282.4

28.9

<0.0001

MA2

1414.2

27.0

<0.0001

MA3

5238.5

33.0

<0.0001

MA4

839.9

33.6

<0.0001

MA8

534.9

43.7

<0.0001

Fig. 3: Chi2 significance of variable “macro-activity” by cluster

On the other hand, the role of the “financing scheme” variable resulted much
less significant in discriminating the five clusters, except for categories FS4
and FS5, which are the most significant for cluster 5, and category FS1, which
instead clearly prevail on cluster 4. Nothing relevant emerged in relation to
the variables “geographic area” and “starting year”.
Cluster
1

Category

Chi2

%

p-value

FS2

186.3

25,7

<0.0001

FS3

145.1

24.7

<0.0001

2

FS1

487.6

29.0

<0.0001

3

FS1

286.5

17.6

<0.0001

4

FS1

1245.0

16.7

<0.0001

FS4

2195.0

51.5

<0.0001

FS5

1583.2

58.5

<0.0001

5

Fig. 4

JADT’ 18

309

4. Conclusions
The findings presented herein indicate a close relationship between the
programmatic framework, defined by the Commission, and the contents of
the final reports and policy briefs, supporting the thesis of a co-construction of
the European integration (Adler-Nissen, Kropp 2015). The scientific
discourse has come to be structured around few semantic macro-aggregates
arisen from DHA, which in turn resulted associated with the variables
performed in LCA. Furthermore, the SSH linguistic space shows a clear
cleavage between the economic discourse and the cultural discourse, which
points out the existence of a lack of interaction between these two spheres.
From a more “general” point of view, all this means that, in connecting the
social sciences field with the policy field, the European research projects
produced a scientific discourse that, on the whole, is structurally homologous
with the “space of possibilities” inherent to the 7PQ.
References
Adler-Nissen R., Kropp K. (2015). A Sociology of Knowledge Approach to
European Integration: Four Analytical Principles. Journal of European
Integration, 37(2): 155-173.
Albert M., Mcguire W. L. (2014). Understanding Changes in Academic
Knowledge Production in a Neoliberal Era. Political Power and Social
Theory, 27: 33-57.
Banchoff T. (2002). The Politics of the European Research Area. ACES
Working Paper 3, Paul H. Nitze School for Advanced International Studies.
Böheme G., Van den Daele W., Hohlfeld R., Krohn W., Shafër W. (1983).
Finalization in Science. The Social Orientation of Scientific Progress.
Dordrecht: Riedel.
Bolasco S. (2013). L’analisi automatica dei testi. Fare ricerca con il text mining.
Roma: Carocci.
Borras S. (2000). Science, Technology and Innovation in European Politics.
Research Paper n. 5, Roskilde University.
Bourdieu P. (1975). The Specificity of Scientific Field and the Social Condition
of the Progress of Reason. Social Sciences Informations, 6: 19-47.
Bourdieu P. (1984). Homo academicus, trad. it. (2013) Homo academicus. Bari:
Edizioni Dedalo.
Bourdieu P. (1992). Les règles de l’art, trad. it. (2013) Le regole dell’arte. Milano:
Il Saggiatore.
Bourdieu P. (1994). Raisons pratiques. Sur la théorie de l’action, trad. it. (2009)
Ragioni pratiche. Bologna: Il Mulino.
Bourdieu P. (1995). Champ politique, champ des sciences sociales, champ

310

JADT’ 18

journalistique, trad. it. (2010) Campo politico, campo delle scienze sociali, campo
giornalistico. In Cerulo M. (a cura di). Sul concetto di campo in sociologia.
Roma: Armando.
Bourdieu P. (2001). Science de la science et réflexivité, trad. it. (2003) Il mestiere di
scienziato. Milano: Mondolibri.
Breslau D. (1998). In Search of the Unequivocal: The Political Economy of
Measurement in U.S. Labor Market Policy. London: Praeger.
Camargo B. V., Justo A. M. (2013). R Interface for Multidimentional Analysis
of Texts and Questionnaires, IraMuTeQ tutorial, available on:
http://www.iramuteq.org.
Cerroni A., Giuffredi R. (2015). L’orizzonte di Horizon 2020: il futuro europeo
nelle politiche della ricerca. Futuri, 6: 29-39.
Curbelo A. A. (2017). Analysing the (Ab)use of Language in Politics: the Case
of Donald Trump. Working Paper n. 2. University of Bristol: SPAIS.
Deem R. (2015). What is the Nature of the Relationship between Changes in
European Higher Education and Social Science Research on Higher
Education and (Why) Does It Matter?. Journal of European Integration.
37(2): 263-279.
European Commission (2010). Communicating research for evidence-based
policymaking. Bruxelles: Directorate-General for Research.
European Commission (2003). Politica dell’innovazione: aggiornare l’approccio
dell’Unione Europea nel contesto della Strategia di Lisbona. COM(2003) 112
definitivo, 11.03.2003.
European Commission (2005). Comunicazione della Commissione al Consiglio e al
Parlamento europeo sul riesame della strategia per lo sviluppo sostenibile. Una
piattaforma d’azione. COM(2005) 658 definitivo, 13.12.2005.
European Commission (2006). Mettere in pratica la conoscenza: un’ampia
strategia per l’innovazione per l’UE. COM(2006) 502 definitivo, 10.05.2006.
European Commission (2009). Integrare lo sviluppo sostenibile nelle politiche
dell’UE: riesame 2009 della strategia dell’Unione Europea per lo sviluppo
sostenibile. COM(2009) 400 definitivo, 24.07.2009.
Funtowicz S., Ravez J. (1993). Science for the Post-Normal Age. Future, 25:
735-755.
Gerli M. (2017). Il campo sociale dei progetti di ricerca europei. Il caso delle
SSH. Studi Culturali, 1: 127-150.
Gibbons M., Limoges C., Nowotny H., Schwartzman S., Scott P. e Trow M.
(1994). The New Production of Knowledge. London: Sage.
Guzzetti L. (1995). A Brief History of European Union Research Policy.
Luxembourg: Publications Office of the European Communities.
Ippolito F. (1989). Un progetto incompiuto. La ricerca comune europea: 1958-88.
Bari: Edizioni Dedalo.

JADT’ 18

311

Lebart L., Salem A., Berry L. (1998). Exploring Textual Data. New York:
Kluwer Academic.
Menéndez L. S., Borrás S. (2000). Explainig Changes and Continuity in EU
Technology Policy: The Politics of Ideas. In Dresner S. e Gilbert N. (eds),
Changing European Research System. Aldershot: Ashgate.
Reinert M. (1987). Classification descendante hiérarchique et analyse lexicale
par contexte: application au corpus des poésies d’Arthur Rimbaud.
Bulletin de Méthodologie Sociologique, 13: 53-90.
Rosamond B. (2015). Performing Theory/Theorizing Performance in
Emergent Supranational Governance: The Live Knowledge Archive of
European Integration and the Early European Commission. Journal of
European Integration, 37(2): 175-191.
Ruberti A., André G. (1995). Uno spazio europeo della scienza. Riflessioni sulla
politica europea della ricerca. Firenze: Giunti.
Slaughter S., Leslie L.L. (1997). Academic Capitalism: Politics, Policies and the
Entrepreneurial University. Baltimore: The John Hopkins University Press.
Vauchez A. (2015). Methodological Europeanism at the Cradle: Eur-lex, the
Acquis and the Making of Europe’s Cognitive Equipement. Journal of
European Integration, 37(2): 193-210.
Vincent-Lacrin S. (2006). What is Changing in Academic Research? Trends
and Futures Scenarios. European Journal of Education, 41(2): 169-202.
Vincent-Lacrin S. (2009). Finance and Provision in Higher Education: A Shift
from Public to Private?. Higher Education to 2030 (vol. 2), Centre for
Education Research and Innovation: OECD.
Ziman J. (2000). Real Science: What It Is, and What It Means. Cambridge-New
York: Cambridge University Press.

312

JADT’ 18

Spécialisation générique et discursive d’une unité
lexical L’exemple de joggeuse dans la presse
quotidienne régionale
Lucie Gianola1, Mathieu Valette2
Université de Cergy-Pontoise – lucie.gianola@u-cergy.fr
2Institut National des Langues et Civilisations Orientales– mvalette@inalco.fr
1

Abstract
In this paper, we study the distribution of lexical items designating outdoor
sport practitioners (joggeur/joggeuse, randonneur/randonneuse, unneur/runneuse,
promeneur/promeneuse), in order to identify links between gender, semantic
themes and genre across press discourse in French. The corpus is sampled
from newspaper articles from regional newspapers. In press discourse, we
observe a convergence between gender and genre through the actualized
semantic classes.
Résumé
Nous étudions dans cet article la distribution d’unités lexicales désignant les
pratiquant·e·s de sport de plein air (joggeur/joggeuse, randonneur/randonneuse,
runneur/runneuse, promeneur/promeneuse) afin d’identifier les corrélations
entre genres sexuels, thèmes sémantiques et genres textuels dans le discours
journalistique en français. Le corpus est constitué à partir d’un
échantillonnage d’articles de la presse quotidienne régionale. Il apparaît que
dans le discours journalistique, on observe une convergence entre genres
sexuels et genres textuels par le biais des classes sémantiques instanciées.
Keywords: Press discourse, textometrics, semantic class, genre, gender
1. Introduction
Nous proposons une étude de lexicologie textuelle sur la distribution
d’unités lexicales choisies dans un corpus de textes de presse. L’étude n’a pas
été réalisée dans une perspective corpus-driven, comme c’est souvent le cas en
textométrie, mais avec une approche corpus-based (Biber, 2009) où les
observables ont été prédéfinis. Notre objectif est en effet de nous focaliser sur
les désignations des pratiquant·e·s de sport de plein air suivant une
opposition en genres sexuels : joggeur vs joggeuse, randonneur vs randonneuse,
runneur vs runneuse, promeneur vs promeneuse. Il s’agit d’identifier les
corrélations entre genres sexuels, isotopies et genres textuels dans le discours
journalistique de la presse quotidienne régionale française.

JADT’ 18

313

2. Problématique
2.1. Sommation des isotopies de genres et de discours en signifiés
La lexicologie textuelle consiste en l’analyse du lexique à partir des
conditions textuelles de sa production. Elle repose sur l’hypothèse selon
laquelle les unités lexicales subissent un ensemble de contraintes
intertextuelles et infratextuelles de la même nature que les formes
sémantiques diffuses et non lexicalisées et qui en conditionnent les régimes
de production et d’interprétation. Dans de précédents travaux, ont été
proposées les conditions théoriques d’une analyse textuelle du lexique,
principalement focalisées sur l’étude de la néologie sémantique – ou néosémie
(Rastier et Valette, 2009) et des formes sémantiques diffuses en voie de
lexicalisation synthétique ou protosémie (Valette, 2010ab). Il s’agit ici d’étudier
l’utilisation systématique d’une unité lexicale donnée dans un genre textuel
précis et l’incidence de cette utilisation sur son sémantisme. En effet, tout mot
placé dans un texte en reçoit des déterminations sémantiques, qui sont
susceptibles de modifier son signifié (afférence de sèmes). Posant l’hypothèse
selon laquelle le signifié est une forme sémantique lexicalisée (Valette 2010b),
on considérera que les sèmes des isotopies du texte peuvent se propager au
signifié d’une unité lexicale par le processus de sommation décrit par (Rastier,
2006). L’observation a pu être faite concernant les isotopies de domaine
(redomanialisation d’une unité lexicale dans le cas de la néosémie par
exemple) mais les isotopies génériques (relatives au genre textuel) ou
discursives (relatives au discours) peuvent-elles transformer le signifié d’un
mot de la même façon que les isotopies domaniales ? C’est à cette question
que nous allons tâcher de répondre ici.
2.2. Présentation du corpus
Le corpus est donc constitué suivant deux axes, lexical et discursif : nous
avons utilisé 8 formes considérées comme des mots-clés pour collecter des
textes exclusivement issus du discours journalistique et, plus précisément, de
la presse quotidienne régionale, sans considération de genre textuel. Le
corpus a été collecté de manière semi-automatique à l’aide d’un script
d’aspiration de pages web puis nettoyé et dédoublonné manuellement, afin
d’écarter des articles constitués de reprises de dépêches AFP qui se
retrouvent d’un titre à un autre. Le script, basé sur la commande Linux cURL,
est alimenté pas une liste d'URL collectées sur les sites des titres de presse à
l’aide de requêtes effectuées sur le moteur de recherche Google
(site:nomdusite forme, modulée par un inhibiteur -blade dans le cas de
« runner » afin d'écarter les articles à propos du film Blade Runner). Entre 100
et 130 URL ont été collectées pour chaque forme. La phase de nettoyage a
permis de supprimer les en-têtes, sommaires, liens annexes, légendes

314

JADT’ 18

d’images, etc., pour ne conserver que le titre et le corps de l’article. Le corpus
est organisé en huit sous-corpus correspondant aux 8 formes étudiées :
Joggeur, Joggeuse, Promeneur, Promeneuse, Randonneur, Randonneuse, Runner,
Runneuse, dont les statistiques sont présentées dans le tableau suivant.
Table 1 : Analyse factorielle des correspondances sur les parties du discours
Sous-corpus

Nombre de mots

Joggeur

40 671

Joggeuse

48 285

Randonneur

35 162

Randonneuse

31 931

Promeneur

44 497

Promeneuse

31 009

Runner

22 212

Runneuse

31 367

Total

285 134

Les articles sont issus principalement de titres de la presse quotidienne
régionale comme Nice Matin, Ouest-France, L’Est Républicain, La Dépêche du
Midi, La Montagne, Corse-Matin, La Provence. La collecte n’a pas été orientée
sur une rubrique en particulier mais sur l’ensemble des titres, et nous
n’avons pas défini de limite temporelle.
3. Analyses1
3.1. Observations générales
Une analyse factorielle préliminaire (figure 1) portant sur les seules parties
du discours montre une opposition marquée sur l’axe 1 entre les sous-corpus
Runner et Runneuse et les autres sous-corpus. Cet écart s’explique par les
genres textuels des sous-corpus considérés. En effet, comme l’ont montré les
travaux pionniers de (Biber, 1988) et, à leur suite, ceux de (Malrieu et Rastier,
2001), les variables locales que constituent les parties du discours sont des
marqueurs de genre particulièrement stables. Ici, il apparaît que Runner et
Runneuse relèvent du genre du compte rendu d’événements sportifs tandis
que les 6 autres sous-corpus sont composés en grande majorité de faits
divers. Autrement dit, la plupart des unités lexicales choisies pour nos
requêtes, qui correspondent à des pratiques sportives de plein air,

1 Le corpus a été analysé au moyen du logiciel de textométrie TXM
(http://textometrie.ens-lyon.fr/) (Heiden et al. 2010).

JADT’ 18

315

n’appartiendraient pas – ou alors à la marge – au vocabulaire des genres
sportifs du discours journalistique.
L’analyse factorielle des correspondances sur les formes, dont la fréquence
est au moins égale à 10 occurrences, offre à voir une distribution très
différente. Runner et Runneuse sont toujours très proches mais il en est
désormais de même de Randonneur et Randonneuse (désormais Randonneur·se)
(figure 2). Les sous-corpus Joggeur, Promeneur et Promeneuse se situent à la
croisée des axes et seront étudiés individuellement, mais Joggeuse se
singularise.
3.2. Analyses des classes sémantiques constituantes
L’analyse des spécificités (formes) des regroupements ainsi constitués nous
indique les contextes d’instanciation des différentes formes.
Le regroupement a priori très homogène Randonneur·se offre à voir un
vocabulaire associé aux accidents de montagne. Le corpus est structuré en 3
classes sémantiques principales,
- celle des accidents : « chute », « mortelle », « mètre »,
« avalanche », « fracture », « cheville », « hôpital », « blessée »,
« trauma », « glisser » etc.
- celle des disparitions : « disparu », « alerte », « retrouvé »,
« emporté », « inquiet », etc.
- celle des secours : « PGHM » (pour Peloton de gendarmerie de
haute montagne), « hélicoptère », « Dragon » (un modèle
d’hélicoptère) « évacué·e », « pompiers », « CRS », « secouriste »,
« secteur », « équipe », « sauveteur », « secourir », etc.
Le sous-corpus Promeneur et le sous-corpus Promeneuse relatent
essentiellement 3 types d’événements :
- la promenade : « sentier », « phare », « littoral », « patrimoine »,
« chemin », etc.
- les accidents : accident de chasse essentiellement : « chasseurs »,
« chasse », etc.
- les découvertes : « macabres », « corps », « cadavre », « tronc »,
« jambe », « squelette », « ossement », « obus », « pépite », etc.
Le sous-corpus Joggeur ne comporte quant à lui qu’une classe sémantique
principale, celle des accidents n’incluant pas de tiers humain : « arrêt,
malaise, crise cardiaque », « algues vertes », attaques d’animaux (« rapace »,
« aigle », « buse »), sulfure d’hydrogène, H2S, intoxication, toxique, gaz. Il est
à noter que cette classe ne s’actualise pas dans le sous-corpus Joggeuse.

316

JADT’ 18

Les deux sous-corpus restants, le regroupement très homogène Runner et
Runneuse (désormais Runneur·se) et Joggeuse méritent toute notre attention.
D’un point vue ontologique, le jogging comme le running sont des formes
similaires de course à pied relevant du domaine du sport. Mais leur usage
dans le discours journalistique diffère très sensiblement. Dans le
regroupement Runneur·se, qui comporte, comme nous l’avons vu,
essentiellement des articles relatant des événements sportifs, le vocabulaire
est structuré autour des classes sémantiques suivantes :
- définitoire : hyperonyme « sport », synonyme « coureur », etc.
Ainsi, le sous-corpus Runneur·se est le seul dont le sens
correspond à la signification.
- classe de la compétition : « course » « marathon », « semimarathon »,
« trail »,
«triathlon »,
« championnat »,
« inscription », « départ », « épreuve », « km », « victoire »,
« podium », « médaille », « sponsors », etc.
-

classe des blessures : « blessure », « foulure », « ampoule »,
« contracture », etc.
Il comporte également deux classes sémantiques liées aux techniques
associées à la pratique :
- classe
des
équipements :
« équipement »,
« baskets »,
« chaussures », « brassière », « connectés », « GPS » ou « montre
GPS », etc.
- classe des entrainements : « entrainement », « préparation »,
« fractionné », « cardio », « conseils », « performances », « yoga »
(comme activité complémentaire destinée à éviter les blessures),
etc.
Il est à noter que le sous-corpus Runneuse se singularise par la mention
d’événements sportifs caritatifs liés à la lutte contre le cancer du sein :
« octobre rose », « prévention ».
A l’inverse, la joggeuse dans le sous-corpus éponyme n’est nullement une
sportive, mais sa caractérisation textuelle est remarquablement précise : elle
est une femme agressée pendant son jogging et les classes sémantiques
actualisées dans ce sous-corps relèvent du crime, du droit et de l’enquête
judiciaire :
- classe des agressions : « meurtre », « tentative », « agressée »,
« agression sexuelle », « viol », « enlèvement », « tuée »,
- classe des agresseurs : « homme », « suspect », « meurtrier »,
« présumé », « portrait-robot », « violeur », « exhibitionniste »
- classe des procédures judiciaires : « enquêteurs », « avocats »,

JADT’ 18

317

« cour », « procureur », « réquisition », « réclusion », « prison »,
« accusé »,
« interpellé »,
« agresseur »,
« condamné »,
« procédure », « instruction », « ADN », etc.
3.3. Synthèse
A l’issue de cette analyse, on choisit de se concentrer sur la définition en
miroir de la joggeuse et de la runneuse, laissant de côté les autres unités
lexicales détaillées ci-dessus. Les isotopies génériques et discursives qui
constituent la trame sémantique des articles dans lesquels occurrent ces deux
formes donnent lieu à la construction de deux signifiés antagonistes, par
sommation :
La joggeuse apparaît :
1. /isolée/ (elle court seule),
2. /vulnérable/ (elle est sans défense face à un agresseur) et, quoi
qu’il arrive, puisque le genre du fait divers l’exige,
3. /victime/ (elle est agressée, violée, tuée).
A l’inverse la runneuse est :
1. /entourée/ (elle court dans le cadre d’événement sportifs
collectifs),
2. /sécurisée/ (par la technologie, notamment les montres GPS qui
permettent de gérer l’effort et d’optimiser ses performances, par
l’entraînement suivi. Les blessures subies apparaissent par
ailleurs bénignes par rapport aux risques encourus par la
joggeuse),
3. /compétitrice/ (elle participe à des compétitions).
4. Conclusion
Dans cet article, nous avons tenté de montrer comment les fonds sémantiques
issus des genres et des discours pouvaient modifier, par sommation, les
signifiés des unités lexicales qui sont utilisées. Pour deux unités lexicales
partageant a priori un référent identique, celui d'une femme pratiquant la
course à pied, l'actualisation en corpus journalistique fait émerger des
contenus sémantiques très différents. Il ne s’agit pas de considérer que les
joggeuses sont nécessairement des femmes en danger mais la régularité avec
laquelle le mot joggeuse est actualisé dans la presse comme une /victime/,
/vulnérable/ et /isolée/ pourrait avoir, à terme, une incidence sur la
perception d’une pratique dont la réalité médiatique est exclusivement
macabre. En d’autres termes, dans le discours de presse, pour les femmes, le
jogging est une pratique dangereuse, la joggeuse une victime d'agression,
alors que la runneuse une sportive impliquée dans des événements sociaux
et le running une pratique sûre et valorisante.

318

JADT’ 18

Références
Biber, D. (1988). Variation across Speech and Writing . Cambridge, Cambridge
University Press.
Biber, D. (2009). Corpus-Based and Corpus-driven Analyses of Language
Variation and Use. In B. Heine and H. Narrog (editors) The Oxford
Handbook of Linguistic Analysis, 159–191. Oxford.
Heiden S., Magué J.-P., et Pincemin B. (2010). TXM : Une plateforme logicielle
open-source pour la textométrie – conception et développement, S.
Bolasco. editors., Journées internationales d’Analyses statistiques des Données
Textuelles, vol(2), 1021-1032.
Malrieu, D. et Rastier, F. (2001). Genres et variations morphosyntaxiques, In
Traitements automatiques du langage, 42, 2, 547-577.
Rastier, F. (2006). Passages. In Corpus, 6, 125-152.
Rastier, F., Valette, M. (2009). De la polysémie à la néosémie. In Le français
moderne, vol. (77), 97-116.
Valette, M. (2010a). Propositions pour une lexicologie textuelle. In Zeitschrift
für Französische Sprache und Literatur, vol. (37): 171-188.
Valette, M. (2010b). Méthodes pour la veille lexicale, In L. Messaoudi, et al.
editors Sur les dictionnaires, Publication du laboratoire Langage et société,
Université Ibn Tofail, Kénitra: 251-272.

JADT’ 18

319

The Transparency Engine – A Better Way to Deal
with Fake News
Peter A. Gloor1, Joao Marcos de Oliveira2, Detlef Schoder3
1

MIT Center for Collective Intelligence, Cambridge MA – pgloor@mit.edu
2Galaxyadvisors, Aarau Switzerland – jmarcos@galaxyadvisors.com
3University of Cologne, Germany – schoder@wim.uni-koeln.de

Abstract
We introduce the “Transparency Engine”, a social network search engine to
separate fact from fiction by exposing (1) the hidden “influencers” and (2)
their “tribes”. Our goals are to quantify the influence and relevancy of
persons, concepts, or companies on institutions, issues or industries by
tracking the dynamics and changes in the observed environment. In
particular we visualize the networks of influence for a given social or
economical ecosystem, thus providing a tool to both the scientific and general
public (including journalists, or anyone interested to check news) to track the
diffusion of new ideas, both good and bad. In particular, the Transparency
Engine exposes the hidden influencers behind fake news. We propose a
unique solution, which combines three subsystems we have been developing
over the last five years: (I) Powergraph, (II) Tribefinder, and (III)
Swarmpulse, The powergraph displays the degree and power of the
spreader’s position by re-constructing her/his (social) network via Web sites
and social position in the Twitter-universe. The tribefinder exposes the tribal
echo chambers on Twitter nurturing fake news items through social media
mining, thus allowing the news consumer to develop an informed opinion
for identifying the motivation of the spreaders of fake news. This is done
through mining Twitter word usage of tribe members with neural networks
using tensorflow. The swarmpulse system finds the most relevant fake and
non-fake news on Wikipedia and Twitter by combining their emergent
patterns.
Keywords: Fake News, Transparency Engine, News, Truth, Belief System,
Machine Learning, Big Data
1. Introduction
According to independent investigations, Russian misinformation and fake
news by Western conspiracy theorists on social media may have contributed

320

JADT’ 18

to the outcome of the Brexit vote1 and the election of Donald Trump2.
Misinforming news has become a significant threat to societal discourse and
opinion formation. Mechanisms to deal with this type of fake news by
making them transparent are urgently needed. The goal of this project is to
understand the concept of “fake news” in the context of forming collective
awareness through social media. The concept of truth is dependent on a
personal belief system. On the other hand, conspiracy theories and satire is
nothing new, and people who WANT to believe these have always embraced
them. Categorizing news as “Fake news” happens when they are against
one's innermost and most passionate beliefs. The more somebody is
embedded into a predefined belief system, the more likely they are to believe
fake news. For instance, people who use Facebook as their major news
source, are more likely to believe fake news (Silverman & Singer-Vine, 2016).
What mental processes are happening when we embrace fake news? When
embedded in a particular belief system, individuals recognize fake news
immediately when they read them, because they do not want to believe
them, similarly they also immediately categorize news as true news when
they read them, because they perfectly fit into their belief system. For
instance, Trump followers label mainstream news as “fake news”, while
mainstream news labels news from Trump followers as “fake news”.
2. Related Work
There are many approaches to creating more transparency in societal
discourses. In fact, this may be seen as the core task of quality journalism.
Most if not all of these approaches, however, are not well supported by IT
tools, do not scale well, and many do not reveal the applied algorithms. Fact
checking Websites such as Wikitribune, Snopes.com, PolitFact, and
FactCheck.org, and corporate/proprietary initiatives like Facebooks’s fake
news detection tools mostly rely on human volunteers and/or paid staff to do
fact checking, which has major disadvantages:
- human bias: fact checkers might have a “leftist” or “right-wing” bias
- non-scalable: the human pool of fact checkers is by definition restricted
- deferred access: the machine can check any news item immediately, 24/7, and it
does not take the expensive detective work of the human fact checker
- non-replicable: as the fact checking is done by different users, the reader will not be
able to understand why a certain fact has been categorized in a particular way

1 Londongrad - Russian Twitter trolls meddled in the Brexit vote. Did they
swing it?. Economist, Nov. 23rd 2017
2 https://en.wikipedia.org/wiki/Russian_interference_in_the_2016_United_States
_elections

JADT’ 18

321

Among the automated approaches, Kloutscore (www.klout.com) gives a
metric for the social media influence of a person. However the kloutscore has
to be requested manually by a user who wants a kloutscore, so it is heavily
skewed towards self-promoters. Another solution for finding the social
media profiles of users is to leverage the Google Knowledge Graph
(https://en.wikipedia.org/wiki/Knowledge_Graph),
which
has
been
employed in theoretical work by Ciampaglia et al. (2015) for fact checking by
measuring the shortest path distance between related concept nodes.
Another approach consists of using machine learning to identify fake news,
for instance it has been shown by Ott et al. (2011) that machine learning
based on word usage beats humans by wide margins to identify fake reviews
in tripadvisor by computing feature vectors from the text of the reviews.
More generally, (Youyou et al. 2015) have shown that to identify (tribal)
attributes of people, having the computer look at their Facebook likes
through machine learning will be more reliable than human judgment. A
similar research question is addressed when identifying Twitter bots based
on their networking pattern and word usage. For instance, Botcheck
(botcheck.me) and Botometer (https://botometer.iuni.iu.edu/#!/) (Varol et al.
2017) check the likelihood of any Twitter id to be a bot, based on number of
followers and friends, tweeting dynamics, and content of tweets.
3. Motivation – How Influencers Spread Fake News
Today’s online social media consumers are exposed to a cacophony of fact
and fiction as never before. “It is true, I read it on the Internet” is
unfortunately a prominent way for information to spread. For example,
immediately after the 2016 US Presidential elections, in early November 2016,
Hillary Clinton was accused of running a pedophile ring out of a pizza
restaurant in Washington. Called “pizzagate”, this news item became a
favorite call to arms among right-wing extremists and Donald Trump
supporters, leading one incensed fanatic to drive a few hundred miles from
Salisbury, North Carolina to Washington DC, and firing his automatic gun
into the pizza restaurant. The origin of this fake news story has been well
documented, starting from a white supremacist Twitter account, then picked
up by the conspiracy News Web site of Sean Adl-Tabatabai, where it fell on
the willing ears of the American right. Just like Google has revolutionized the
way we access information, our proposed Transparency Engine intends to
change the way how we look at such information, by exposing the hidden
influencers like “Sean Adl-Tabatabai” who inject new information into the
public discourse.

322

JADT’ 18

3.1 The concept of tribes and how they perceive information
Besides knowing the sources of rumors, it is essential to also know the
(political) orientation of these influencers. Quantum physics suggests that
there are many different universes, with our current world being embedded
into just one out of infinitely many other universes. Looking at radically
different interpretations of the same news item, it seems we are indeed living
in different quantum universes. These different universes can be grouped
into “tribes” (Sloterdijk 2011). Each of these tribes has its own reality,
defining fact or fiction for the members of the tribe. Previous research (De
Oliveira et al. 2017) has exemplified this idea. What is fact for one tribe is
fiction for another tribe. It all depends on the tribe, and what the members of
the tribe WANT to believe. Examples are the denial of human-influenced
global warming, the explanation of evolution through “intelligent design”, or
the causal relationship between vaccination and autism where some tribes
perceives related issues as “fact” and “truth” whereas other tribes perceive
the objectively same issues as “fiction”, “lie” or “fake news”, thus creating an
“alternate reality”. In contrast to the power of states and corporations, the
growing power and dynamics of networks is mostly invisible. Unlike
hierarchical structures, the central influencers in networks are hard to
identify by the “naked eye”. What matters to spread any news – fact or fake –
is the influence of the spreader. The main way to quantify the influence of the
spreader is her/his position in a given network and with it the power to
“multiply” the word to larger audiences. More specifically, the degree and
power of the spreaders’ position can be measured by re-constructing their
(social) network via their Web sites and their social position for example in
the Twitter-universe (and other social networking platforms) thus measuring
the influence of Web sites and the influence of Twitter (accounts) on a
specific topic.

Figure 1 Twitter retweet network “pizzagate” (left), and Twitter influence network (right)

JADT’ 18

323

Pizzagate only spread because a moderately influential spreader, Sean AdlTabatabai, discovered the original tweet and posted it on his conspiracy
News Web site. Figure 1 illustrates how social media analysis can increase
trust and transparency by visualizing the echo chambers of fake news about
pizzagate using our social media analysis system Condor (Gloor 2017). The
picture at left shows the Twitter network about pizzagate, each node is a
person tweeting, a link between two people means either that one person is
retweeting a tweet sent by the other person, or is mentioning the other
person in a tweet. There is a large cluster in the center of the network, made
up of believers in the fake news. They are reinforcing each other, and
increasing the traffic in their echo chamber. The few supporters of Hillary
Clinton, trying to debunk the fake news, are pushed aside; their tweets are
ignored by the large echo chamber of conspiracy theory believers. The people
in the periphery (the “asteroid belt”) are tweeting into the void, as their
tweets are ignored by friends and foes alike.
Using an influencer algorithm (Gloor 2017) shows that the discourse about
pizzagate on Twitter is dominated by Trump followers (the picture at right
above). Our algorithm makes somebody an influencer, if the words she or he
is using, are picked up by others and spread quickly through the network. As
the picture at right in figure 1 shows, there is just one voice of reason left,
while the proponents of pizzagate reinforce each other much more, with a
cluster of influential spreaders of wild ideas in the center, and other
conspiratorialists in the periphery of the cluster, being retweeted by
hundreds of likeminded others (shown as “parachutes” in the graph).
4. Our Solution – Transparency Engine
We introduce the “Transparency Engine”, a social network search engine to
separate fact from fiction by exposing the hidden influencers and their
“tribes” behind fake news. Just like Google has revolutionized the way we
access information, Transparency Engine changes the way we look at such
information, by exposing the hidden influencers. Our goals are fourfold: (1)
Quantify the influence and relevancy of persons, concepts, companies on
institutions, issues or industries. (2) Qualify the dynamics and changes in the
observed environment. (3) Visualize the networks of influence for a given
social or economical ecosystem. (4) Provide a tool to track the diffusion of
new ideas, both good and bad.
4.1. Powergraph
Our solution combines three subsystems we have been developing over the
last five years (Fuehres et al. 2012, de Oliveira et al. 2016, de Oliveira et al
2017): Power graph, tribe finder, and swarmpulse. Power graph measures the

324

JADT’ 18

importance of “notable” people as defined by Wikipedia through calculating
the number of other Wikipedia people pages than can be reached within two
degrees of separation from a particular people page on Wikipedia. This is a
proxy for social capital, as it basically measures the influence of the people a
person is connected to. The system also identifies those people with Twitter
accounts by matching them with sources of information like Wikidata and
Google knowledge graph.

Figure 2. Sample Powergraph for “global warming”

Figure 2 illustrates our prototype version of the Powergraph, showing the
social network of the most influential people about “global warming”, based
on their Wikipedia and Twitter presence. We find, not surprisingly, that
Donald Trump and the former US presidents are most influential. We
measure the importance of people through calculating the number of other
Wikipedia people pages and Twitter friendship networks than can be
reached within two degrees of separation from a particular people page. This
is a proxy for social capital, as it basically measures the influence of the
people a person is connected to (Fuehres et al. 2012).
4.2 Tribefinder
The second component of our system, tribefinder (de Oliveira et al. 2017),
identifies the tribal affiliations of the opinion leaders about any news item.
To assign a tribe to an influencer, our system analyzes their word usage,
using deep learning. An integral component of the tribefinder system is
“TribeCreator", this subsystem automatically helps the user to find people
that belong to a newly defined tribe by looking at profile self-descriptions,

JADT’ 18

325

the content of tweets, and at followers, and Twitter friends. For example, if
users wants to create a tribe for Treehuggers (people who like nature), they
can search for people with profile descriptions that match the idea of this
tribe: “nature lover”, “I love nature”, “nature”, etc., for people who follow
pages about nature, or tweet about nature. In the second step we calculate
the vocabulary that these influentials are using in their tweets. This
vocabulary is then used to match the vocabulary against the vocabulary of
any Twitter user, calculating their tribal affiliates. Knowing the tribal
affiliations of the thoughtleaders for a news item allows readers to correctly
position the news item, deciding for themselves if they want to trust the
news coming from a particular influencer.
4.3 Swarmpulse
The third component of our system is Swarmpulse (de Oliveira et al 2016).
Swarmpulse finds the most recently edited Wikipedia pages and uses Twitter
to see which people are talking about those subjects. This system helps users
to serendipitously spot most recent news items they were not aware of, and
then check their influencer network on the power graph and calculate their
tribal affiliations with tribefinder.
5. Conclusion
The best approach for fact-checking is a critical, well-informed mind. Our
world needs more powerful ways and tools to support the critical mind.
Transparency is a key enabler for this. The Transparency Engine thus
provides the foundation for informing the critical mind: The global
Powergraph will display the power network of the one million globally most
influential people on Wikipedia people pages and the most popular Twitter
users. It will allow all other Twitter users to position themselves within the
context of the Powergraph. The Tribefinder will show the “truth of tribes” by
creating tribes through their use of language on social media and assigning
each influencer to one or more tribes and showing the tribal affiliations in the
Powergraph. Swarmpulse will build an index of most recent significant news
by combining new edits on Wikipedia with the most popular tweets from
influential twitterers and show the actors involved through Powergraph. The
landscape of transparency generating approaches calls for a scientific, open
approach such as the Transparency Engine proposes. Our aim is to
substantially contribute to popularizing and democratizing fact checking for
the whole world. Everyone should be enabled to do this easily and simply by
themselves!

326

JADT’ 18

References
Ciampaglia, G. L., Shiralkar, P., Rocha, L. M., Bollen, J., Menczer, F., &
Flammini, A. (2015). Computational fact checking from knowledge
networks. PloS one, 10(6), e0128193.
de Oliveira, J. Gloor, P. (2016) The Citizen IS the Journalist - Automatically
Extracting News from the Swarm. Rome, Italy June 9-11, 2016, Designing
Networks for Innovation and Improvisation: Proceedings of the 6th
International COINs Conference (Springer Proceedings in Complexity)
de Oliveira, J. Gloor, P. (2017) GalaxyScope – Finding the "Truth of Tribes" on
Social Media. Detroit September 11-14, 2017. Proceedings of the 7th
International COINs Conference (Springer Proceedings in Complexity)
Fuehres, H. Gloor, P. Henninger, M. Kleeb, R. Nemoto, K. (2012)
Galaxysearch: Discovering the Knowledge of Many by Using Wikipedia
as a Meta-Search Index. Proceedings Collective Intelligence 2012, April 1820, Cambridge, MA
Gloor, P. (2017) Sociometrics and Human Relationships: Analyzing Social
Networks to Manage Brands, Predict Trends, and Improve Organizational
Performance , Emerald Publishing, London 2017
Ott, M., Choi, Y., Cardie, C., & Hancock, J. T. (2011, June). Finding deceptive
opinion spam by any stretch of the imagination. In Proceedings of the
49th Annual Meeting of the Association for Computational Linguistics:
Human Language Technologies-Volume 1 (pp. 309-319).
Silverman, C. Singer-Vine, J. (2016) "Most Americans who see fake news
believe
it,
new
survey
says."
BuzzFeed
News;
https://www.buzzfeed.com/craigsilverman/fake-news-survey
Sloterdijk, P. (2011). Bubbles: microspherology. MIT Press
Varol, O., Ferrara, E., Davis, C. A., Menczer, F., & Flammini, A. (2017). Online
human-bot interactions: Detection, estimation, and characterization. arXiv
preprint arXiv:1703.03107.
Youyou, W. Kosinski, M. Stillwell, D. (2015) Computer-based personality
judgments are more accurate than those made by humans. Proceedings of
the National Academy of Sciences (PNAS)

JADT’ 18

327

Brexit and Twitter: The voice of people
Francesca Greco, Leonardo Alaimo, Livia Celardo
Sapienza University of Rome – francesca.greco@uniroma1.it;
leonardo.alaimo@uniroma1.it; livia.celardo@uniroma1.it

Abstract 1
There is an increase in Euroscepticism among EU citizens nowadays, as
shown by the development of the ultra-nationalist parties among the
European states. Regarding the European Union membership, public opinion
is divided in two. British referendum in 2016, where citizens chose to “exit”
shaking the public opinion, and the following general election in June 2017,
where the British Europeanist parties won the election according to the 1975
British referendum where 72% of citizens chose to “Remain”, are clear
examples of this fracture. There are still few studies concerning the
investigation of Brexit discourses within the social media and most of them
focus on the 2016 British referendum. Due to that, this exploratory research
aims to identify how Brexit and the EU are nowadays discussed on Twitter,
through a text mining approach. We collected all the tweets containing the
terms “Brexit” and “EU”, for a period of 10 days. Data collection has been
performed with TwitteR package, resulting in a large corpus to which we
applied multivariate techniques in order to identify the contents and the
sentiments behind the shared comments.
Abstract 2
Negli ultimi anni c'è stato un aumento dell'euroscetticismo tra i cittadini
dell'UE, come testimoniato dallo sviluppo di partiti ultra nazionalisti in
diversi stati europei. Sul tema "Europa", l'opinione pubblica è divisa fra
europeisti e euroscettici. Un chiaro esempio di questa divisione è dato dalle
recenti vicende britanniche: infatti, nel referendum del 2016 i cittadini
britannici hanno scelto di "uscire" dall’UE scuotendo l'opinione pubblica,
mentre le successive elezioni politiche di giugno 2017 hanno visto
l'affermazione dei principali partiti filo-europeisti. Vi sono ancora pochi studi
in letteratura che indagano come nei social media venga affrontato il tema
della Brexit in relazione all’UE, dato che la maggior parte di essi si focalizza
su cause e potenziali effetti del voto di giugno 2016. In tal senso, questa
ricerca esplorativa ha lo scopo di identificare in che modo Brexit e l'Unione
Europea vengano discusse su Twitter in questo momento storico attraverso
l’analisi automatica del testo. A questo scopo sono stati raccolti tutti i
messaggi contenenti i termini "Brexit" e "EU" per 10 giorni attraverso

328

JADT’ 18

l'utilizzo del pacchetto TwitteR, ottenendo un corpus di grandi dimensioni a
cui sono state applicate delle tecniche multivariate, al fine di individuare i
contenuti e i sentimenti relativi al tema in esame.
Keywords: Brexit, Twitter, Emotional text mining.
1. Introduction
There is a growing increase in Euroscepticism among EU citizens nowadays,
as shown by the development of the ultra-nationalist parties among the
European states. Regarding to European Union membership, public opinion
is divided between Eurosceptics and pro-Europeans, as shown by the 2016
British referendum ("Brexit"), where 52% of citizens chose to “Leave”. For
further evidence of this division, the following general election of June 2017
saw the affirmation of the main Europeanist parties (especially the Labour
Party) and the results led to a hung Parliament. Brexit has shaken the
European public opinion as it revealed the relevance of the anti-Europeanist
trend. During the 60th Anniversary of the Treaties of Rome in 2017, millions
of citizens expressed their support to the EU participating to Europeanist
demonstrations in many European cities.
One useful starting point for explaining the results of Brexit is to focus on the
electoral issue: the relationship between the UK and Europe. This has always
been a central and rather controversial issue in the British public debate. The
media, public opinion and the political class have always been deeply critical
and sceptical about the European integration. This position influences
citizens' attitudes towards the Union, which is not only considered distant
and inadequate to resolve everyday issues (immigration, unemployment,
and so on), but it is often perceived as their major cause, by limiting the
political and economic power of United Kingdom. The electoral outcome
created disbelief all over the world. Britain is the home of the term
Euroscepticism (Spiering 2004, p.127). But, while it is clear that a large
proportion of UK residents are sceptical about Europe, it is not clear enough
that this position coincides with the wish to leave the EU. However,
Euroscepticism should not be confused with this wish. Szczerbiak and
Taggart (2008) have distinguished two different types of Euroscepticism: the
Hard Euroscepticism that is a principled opposition to the EU and European
integration and Soft Euroscepticism that concerns on one (or a number) of policy
areas lead to the expression of qualified opposition to the EU.
Although there are several studies exploring British Euroscepticism, only few
of them investigate the Brexit discourses within the social media. Due to that,
we decided to perform a quantitative study, where the online discourses
regarding Brexit and EU are analysed using two different approaches,

JADT’ 18

329

Content Analysis and Emotional Text Mining. The aim is to explore not only
the contents but also the sentiments shared by users on Twitter. For this
paper, we used one of the most important and known blog tools, Twitter. It is
an online platform for sharing real-time, character limited communication
with people partaking of similar interests that, in 2017, counted over than 300
million users and an average of about 500 million of tweets sent per day.
2. Data collection and analysis
In order to explore the sentiments and the contents on Brexit and EU in
twitter communications during ten days, we scraped all the messengers in
English language produced from September 22nd to October 2nd, 2017,
containing together the words Brexit and EU. The data extraction was carried
out with the TwitteR package of R Statistics (Gentry, 2016). We started
collecting 221,069 messengers, including 83% of retweets, from which two
samples of tweets were extracted. The first we used for the sentiment
analysis is composed of 99,812 messengers, where the retweets were limited
to the threshold of 31, resulting in a large corpus of 1,601,985 of tokens; the
second one we used for content analysis, where we excluded all the retweets,
resulted in a large corpus of 37,318 tweets and 618,255 tokens. In order to
check whether it was possible to statistically process data, two lexical
indicators were calculated: the type-token ratio and the hapax percentage
(TTRcorpus 1 = 0.02; Hapaxcorpus 1 = 39.8%; TTRcorpus 2 = 0.04; Hapaxcorpus 2 =
52.31%). According to the large size of the corpus, both lexical indicators
highlighted its richness and indicated the possibility to proceed with the
analysis.
2.1. Emotional text mining
We know that people sentiments depend not only on their rational thinking
but also, and sometimes most of all, on the emotional and social way of
functioning of people’s mind. If the conscious process set the manifest
content of the narration, that is what is narrated, the unconscious process can
be inferred through how it is narrated, that is, the words chosen to narrate
and their association within the text. According to this, it is possible to detect
the associative links between the words to infer the symbolic matrix
determining the coexistence of these terms in the text (Greco, 2016). To this
aim we perform a multivariate analysis based on a bisecting k-means
algorithm to classify the text (Savaresi et Boley, 2004), and a correspondence
analysis to detect the latent dimensions setting the cluster per keywords
matrix (Lebart et Salem, 1994) by means of T-Lab software. The interpretation
of the cluster analysis results allows to identify the elements characterizing
the emotional representation of Brexit, while the results of correspondence

330

JADT’ 18

analysis reflect its emotional symbolization. By the clusters interpretation, we
classify the emotional representations in positive, neutral and negative
sentiments, determining the percentage of messages for each sentiment
modality. To this aim, first corpus was cleaned and pre-processed with the
software T-Lab (T-Lab Plus version, 2017) and keywords selected. In
particular, we used lemmas as keywords instead of types, filtering out the
lemma Brexit and EU and those of the low rank of frequency (Greco, 2016).
Then, on the tweets per keywords matrix, we performed a cluster analysis
with a bisecting k-means algorithm limited to twenty partitions, excluding all
the tweets that do not have at least two keywords co-occurrence. The
percentage of explained variance (η) was used to evaluate and choose the
optimal partition. To finalize the analysis, a correspondence analysis on the
keywords per clusters matrix was made in order to explore the relationship
between clusters and to identify the emotional categories setting Brexit
representations.
2.2. Content analysis
Content analysis is a technique used to investigate the content of a text; in
text mining, many methods exist to analyse it automatically. One of these is
Text Clustering, where the corpus is splits in different subgroups based on
words/documents similarities (Iezzi, 2012). In this paper, a text co-clustering
approach (Celardo et al., 2016) is used. The objective is to simultaneously
classify rows and columns, in order to identify groups of texts characterized
by specific contents. To do that, data were pre-processed with Iramuteq
software lemmatizing the texts, removing stop words and terms with
frequency lower than 10. The weighted term-document matrix was then coclustered through the double k-means algorithm (Vichi, 2001); the number of
clusters for both rows and columns was fixed using the Calinski-Harabasz
index.
3. Emotional text mining main results and discussion
The results of the cluster analysis for ETM show that the 655 keywords
selected allow the classification of 88,6% of the tweets. The percentage of
explained variance was calculated on partitions from 3 to 19, and it shows
that the optimal solution is six clusters (η= 0.057). The correspondence
analysis detected six latent dimensions. In table 1, we can appreciate the
emotional map of Brexit and the EU emerging from the English tweets. It
shows how the clusters are placed in the factorial space produced by five
factors. The first factor represents the political and economic domain where
Brexit seems to have its main impact; the second factor reproduces the
possible solutions of Brexit: a separation or a new agreement; the third factor

JADT’ 18

331

represents the national or European level of reaction to Brexit; the fourth
factor is the blame, distinguishing the blame of politicians from the one of the
willingness to be independent; and the fifth factor is the political leadership,
differing old and new policies.
Table 1  Correspondence analysis results (the percentage of explained inertia
is reported between brackets beside each factor).
Factor 1 (27.5%)
NP

NP

PP

try

Macron

war

pro

support

chance

Brussel

deliver

Europe
an
good

Florenc
e
Delay

remaine
r
concern

zero

divorce

better off

save

laureate

debate

union

finger

proposa
l
fight

leaving

progres
s
negotiat
or
pay

brexiteer miracle
s
help
market

allow

single

finish

chief

event

Merkel

row

0.070.02 ac

economi
st
6.49-4.40
ac

NP

4.721.50 ac

PP

Factor 3 (19.8%)

negotiatio bill
n
Briton
Barnier

future

PP

Factor 2 (24.3%)

0.350.12
ac

1.5-1.61
ac

Factor 4 (15.6%)
NP
blame
march

withdraw stay
al
blast
speech
states

0.350.05
ac

PP

Factor 5 (12.9%)
NP

referendum leader
Johnson

remai
n
Verhofstadt walk

PP
people
Tory
hard

independen urge
voter
t
conservativ destroy
May T. party
e
anti
migrant
hope
happe
n
Blair
vow
Cataloni
call
a
reverse adopt
die
time
0.550.29
ac

5.220.94
ac

3.651.24
ac

10.281.49 ac

NP =negative pole; PP = positive pole; ac = absolute contribution (10-3)

The six clusters are of different sizes and reflect the representations of Brexit
(table 2), that correspond to three different sentiments: positive, negative for
domestic reasons, and negative for foreign ones (table 1). The first cluster
represents the choice to leave EU as a good option, underlining the need to
proceed; the second cluster focuses on the EU political reaction fixing divorce
conditions, perceiving EU political representatives as unfavourable and
therefore threatening; the third cluster represents Britons’ hope to improve
their economic condition leaving EU as naive; the fourth cluster represents
the old British political leadership as incompetent, being unable to protect
and adequately inform Britons in order to support them in remaining in the
EU; the fifth cluster reflects the negotiation of the divorce conditions,
perceiving the negotiation as unfair and the costs of leaving EU as a
punishment; and the sixth cluster represents Brexit as a Britons informed
choice, highlighting that its consequences belong to the policy domain who
should respect the citizens’ choice.

332

JADT’ 18
Table 2  Clusters (the percentage of context units classified
in the cluster is reported between brackets).

Cluster 1
(10.0% CU)

Cluster 2
(14.9% CU)

Cluster 3
(20.9% CU)

Cluster 4
(13.4% CU)

Cluster 5
(19.2% CU)

Cluster 6
(21.7% CU)

Good Choice

EU Reaction Uncertain Future British Leadership Divorce Conditions Informed Choice
people

bill

referendum

leaving

Tory

Barnier

Corbyn

Briton

hard

brussel

Johnson

Theresa May market

chance

voter

progress

think

urge

warn

zero

party

divorce

independent

call

single

better off

happen

negotiator

Boris

walk

business

Nobel

Florence

pay

Verhofstadt

UKIP

minister

economist

stay

chief

Florence

government

Europe

laureate

Catalonia

demand

destroy

hope

move

tell

believe

national

try

look

Merkel

rating

Spain

Davis

policy

mean

miracle

law

Rees Mogg

offer

issue

leader

Macron

negotiation

remain

European

time

good

From 1611
to 620 CU

From 2004
to 951

From 1844 to
668 CU

From 2506 to
461 CU

From 2705 to 843
CU

From 2098 to
512 CU

CU = context units classified in the cluster.

By the clusters interpretation, we detected six different representations of
Brexit that correspond to three different sentiments (table 1). We have
considered as positive (21,7%) the representation of Brexit as a Good Choice
or an Informed Choice, and negatives all the other representations (78,3%).
Among the negative clusters, we distinguished negativity according to the
origin of the problem: Uncertain Future and British Leadership are negative
for domestic reasons (34,2%), that is, the lack of UK political leadership’s
competences; and EU Reaction and Divorce Condition are negatives due to
foreign factors (34,1%) as the EU after Brexit seems to be perceived as
vindictive and, therefore, threatening.
4. Content analysis main results and discussion
The pre-processing phase, implemented on the second corpus, allowed us to
identify a set of 1.957 keywords, representing the 97% of the tweets; so, on
the term-document matrix of dimension (1.957 × 36.383) we calculated the
Calinski-Harabasz Index in order to define the number of clusters for rows
and columns. After calculating the index values for partitions from 2 to 10 for
each dimension, the Calinski-Harabasz Index suggested to classify the words
in three groups and the tweets in five groups. In table 3, the centroids of the
clusters are exposed.

JADT’ 18

333
Table 3  Centroids matrix (Terms × Documents).

Cluster 1

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Cluster 5

(55%)

(20%)

(12%)

(11%)

(2%)

0,005

0,003

0,004

0,000

0,000

Cluster 2

0,002

0,063

0,003

0,149

0,012

Cluster 3

-0,002

0,000

0,090

-0,003

0,309

Table 4  Words groups (first 10 words listed below by frequency of occurrence).
Cluster 1
Negotiation
stay

Cluster 2
Economic Transformation
leave

Cluster 3
British Identity
home

Junker

move

sound

ambassador

transition

cake

cry

late

plan

track

deal

datum

surge

trade

live

peer

retain

finish

shape

post

Id

turmoil

Macron

idea

survive

urge

national

As shown in the table 3, the algorithm has identified five blocks of
specificities; in fact; the first cluster of words is connected to the first group of
tweets; the second is specific of the second and the fourth cluster of tweets
and the third is related to the third and the fifth group of tweets. In table 4,
the groups of words are presented.
The first group of words is related to the need of defining new rules and
settlements within the negotiation and it represents more than half of the
tweets; it has no strong specificities related to the texts, but in comparison to
all the documents clusters, it seems to be more connected to those words. On
the other hand, for the other two groups of words, there are more effective
specificities; the second cluster of words is about the definition of new
economic agreements, and it is connected to the 31% of the tweets, while the
third one, related to the requirement in specifying a new identity after Brexit,
is representative of the 14% of the corpus documents.
5. Conclusions
The results of the two analyses showed a strong relationship between the
terms “Brexit” and “EU”, not only in terms of sentiment, but also in terms of

334

JADT’ 18

contents. According to the literature, the sentiment analysis revealed the
presence of both positive and negative opinions in respect to the exit of
United Kingdom from the EU. On the other hand, starting from the analysis
of the contents we found that the Twitter communications on Brexit focuses
primarily on the concept of negotiation. The remaining part of the messages
take into account both the Brexit economic features and the need of the
national identity redefinition. To conclude, the results of the two analyses
revealed that Brexit is a theme with a strong emotional charge, mostly
negative. British people seem to focus their attention basically toward three
issues: the new asset, the economic consequences, and the national identity.
These subjects are treated positively and negatively from the users, probably
because of the lack of cohesion within the country.
References
Celardo L., Iezzi D.F. and Vichi M. (2016). Multi-mode partitioning for text
clustering to reduce dimensionality and noises. In Proceedings of the 13th
International Conference on Statistical Analysis of Textual Data.
Gentry J. (2016). R Based Twitter Client. R package version 1.1.9.
Greco F. (2016). Integrare la disabilità. Una metodologia interdisciplinare per
leggere il cambiamento culturale. Franco Angeli.
Hobolt S. (2016). The Brexit vote: a divided nation, a divided continent.
Journal of European public policy, 23(9): 1259–1277.
Iezzi D. F. (2012). Centrality measures for text clustering. Communications in
Statistics-Theory and Methods, 41(16-17), 3179-3197.
Lebart L. and Salem A. (1994). Statistique Textuelle. Dunod
Savaresi S. M. and Boley D. L. (2004). A comparative analysis on the bisecting
K-means and the PDDP clustering algorithms. Intelligent Data Analysis,
8(4): 345-362.
Spiering M. (2004). British Euroscepticism. In Harmsen R. and Spiering M.,
editors, Euroscepticism: Party Politics, National Identity and European
Integration. Editions Rodopi B.V.
Szczerbiak A. and Taggart P. (2008). Opposing Europe? The Comparative Party
Politics of Euroscepticism. Volume 1: Case Studies and Country Surveys.
Oxford University Press.
Vichi M. (2001). Double k-means clustering for simultaneous classification of
objects and variables. Advances in classification and data analysis, 43-52.

JADT’ 18

335

A text mining on clinical transcripts of good and poor
outcome psychotherapies
Francesca Greco1, Giulio de Felice2, Omar Gelo3
Sapienza University of Rome & Prisma S.r.l. – francesca.greco@uniroma1.it
2 Sapienza University of Rome & NCU University – giulio.defelice@uniroma1.it
3 University of Salento & Sigmund Freud University – omar.gelo@unisalento.it
1

Abstract
The text mining of clinical transcripts is broadly used in psychotherapy
research, but is limited to top-down approaches, with a-priori vocabularies
that code the transcripts according to a theoretical predetermined
framework. Nevertheless, the semantic level that a word or clinical
intervention can assume depends on the relational field in which the
discourse is produced. Thus, bottom-up approaches seem to be particularly
meaningful in addressing such a relevant issue. With the aim of investigating
possible similarities and differences between good outcome and poor
outcome psychotherapies, we applied a multivariate analysis on the
transcripts of eight single cases of brief experiential psychotherapy (four
good outcome vs four poor outcome cases), in order to identify the general
core themes, and their difference according to therapy outcome. The results
showed a significant difference in the number of context units classified in
two of the six core themes (clusters) between good and poor outcome cases
(χ2, df=5, p<0,01). These findings show how the bottom-up technique of text
analysis on clinical transcripts turned out to be an enlightening tool to let
their latent dimensions emerge, setting the clinical process and outcome, and
therefore, providing a very useful tool for clinical purposes.
Abstract
L’analisi delle trascrizioni cliniche è stata ampiamente utilizzata nella ricerca
in psicoterapia, sebbene prevalentemente si basi sull’utilizzo di un dizionario
che consente la codifica del testo in funzione di criteri predeterminati.
Tuttavia, la polisemia che una parola, o un intervento clinico,” può assumere
dipende dal campo relazionale in cui il discorso è prodotto. Pertanto, gli
approcci bottom-up sembrano essere particolarmente utili nell'affrontare tale
questione. Allo scopo di indagare gli elementi caratterizzati le trascrizioni
cliniche con esito positivo e negativo, è stata effettuata un’analisi multivariata
di un corpus composto da otto trascrizioni di psicoterapia breve (quattro con
esito positivo e quattro con esito negativo) al fine di identificare i temi
centrali generali e la distribuzione delle unità di contesto nei diversi temi in

336

JADT’ 18

funzione dell’esito della terapia. I risultati hanno evidenziato una differenza
significativa tra i casi con esito positivo e quelli con esito sfavorevole (χ2, df =
5, p <0,01), mettendo in evidenza come l'analisi automatica del testo delle
trascrizioni dei colloqui clinici possa essere uno strumento utile a far
emergere le dimensioni latenti organizzatrici del processo e del risultato,
configurandosi così come un utile strumento ai fini clinici.
Keywords: Emotional Text Mining, clinical transcripts, psychotherapy
outcome.
1. Introduction
The text mining of clinical transcripts is very broadly used in psychotherapy
research, but is limited to top-down approaches where a-priori vocabularies
code them according to a theoretical predetermined framework.
Nevertheless, the semantic level that a word, or clinical intervention, can
assume, depends on the relational field in which the discourse is produced.
Thus, bottom-up approaches seem to be particularly meaningful in
addressing such relevant issue. Psychotherapy can be considered a dynamic
communicative exchange between the client and the therapist (e.g., Gelo et
Salvatore, 2016). Within such an exchange, the content (i.e., the semantic) of
what is said plays a primary role. Thus, the textual analysis of therapy
transcripts may represent a very useful tool for psychotherapy process
researchers as well as for clinicians (Gelo et al., 2013; Salvatore et al. 2017). In
the field of psychotherapy research, some methods of text mining have been
developed and applied, such as the Therapeutic Cycle Model (Mergenthaler,
2008) and Referential Activity (Bucci et al., 1992). Following a top-down
approach, these methods use predefined content categories to semantically
classify units of text. Each of these categories corresponds to a thematic
dictionary containing all the words indicative of the content represented by
that category. Even though these top-down methods of text mining allow for
a reliable and valid investigation of the therapeutic process, they present a
major limitation, disregarding the contextual nature of the linguistic meaning
(Carli et al., 2004; Salvatore et al., 2012). In fact, the meaning of a word is
polysemic and depends on the way it combines with other words in the
communicative interaction, i.e., it depends on its association with other
words. Grounded on these considerations, there has recently been a
development of text mining approaches which, by means of their bottom-up
logic, allow for a context-sensitive textual analysis (e.g., Salvatore et al., 2012;
2017; Cordella et al., 2014; Greco, 2016). The aim of this study is to investigate
possible similarities and differences between good outcome and poor
outcome psychotherapy cases by applying the Emotional Text Mining
(Cordella et al., 2014; Greco, 2016). Our assumption is that it is possible to

JADT’ 18

337

detect the associative links between the words in order to infer the symbolic
matrix determining the coexistence of the terms in the text. To this aim, we
perform a multivariate analysis based on a bisecting k-means algorithm
(Savaresi et Boley, 2004) to classify the text, and a correspondence analysis
(Lebart et Salem, 1994) to detect the latent dimensions setting the cluster per
keywords matrix. The interpretation of the cluster analysis allows for the
identification of the elements characterizing the core themes of the treatment,
while the results of the correspondence analysis reflect the emotional
symbolisation characterising the therapeutic exchange. The advantage of
such an approach is to interpret the factorial space according to word
polarization, thus identifying the emotional categories that generate the core
themes, and to facilitate the interpretation of clusters, exploring their
relationship within the symbolic space (Greco et al., 2017).
2. Data collection and analysis
2.1. Data collection
The sample of the present study was drawn from the York Depression Study
I, a randomized clinical trial to assess the efficacy of brief experiential
therapy for depression (Greenberg et Watson, 1998; Watson et al., 1998).1
From the original sample, we initially selected the six best outcome cases and
the six cases worst outcome cases based on the Reliable Change Index of the
Beck Depression Inventory (BDI; Beck et al., 1988). We then excluded four
cases due to missing session transcripts. Our final sample was thus
comprised of a total of eight cases, with four good outcomes and four poor
outcomes. The treatment length was between 15 and 20 sessions (M = 17.62;
SD = 1.38), for a total of 141 sessions. Patients (one man and seven women;
M=37.1 years old) met the criteria for major depressive disorder assessed by
means of the Structured Clinical Interview for DSM-III-R (SCID; Spitzer et al.,
1989). Therapists (seven women and one man; M= 5.5 years of therapeutic
experience) had six months of training in experiential psychotherapy
(Greenberg et al., 1993). The transcripts were collected in a large size corpus
of 1090234 tokens. In order to check whether it was possible to statistically
process data, two lexical indicators were calculated: the type-token ratio and
the percentage of hapax (TTR = 0.01; hapax = 35.3%). They highlighted the
richness of the corpus indicating the possibility of proceeding with the
analysis.

1 We are grateful to Dr. Les Greenberg for having us provided with files of the
transcripts for these cases.

338

JADT’ 18

2.2. Data analysis
First, data were cleaned and pre-processed with the software T-Lab and
keywords selected. In particular, we used lemmas as keywords instead of
type. We selected all the lemmas in the medium rank of frequency (upper
frequency threshold = 933), and those of the low rank of frequency until the
threshold of 17 occurrences, that is, equal to the average number of sessions
made on average by the patients (Greco, 2016). Then, in order to identify the
core themes common to all the psychotherapies, we performed a cluster
analysis on the keywords per context units (CU) matrix, by means of a
bisecting k-means algorithm (Savaresi et Boley, 2004), limited to ten
partitions, excluding all the CU that did not have at least two keywords cooccurrences. The eta squared value was used to evaluate and choose the
optimal solution. To finalize the text mining, we performed a correspondence
analysis on the keywords per clusters matrix (Lebart et Salem, 1994) in order
to explore the relationship between clusters, and to identify the emotional
categories setting the psychotherapeutic process. The interpretation of the
factorial space was performed according to the procedure proposed by
Cordella and colleagues (2014) in which each keyword is considered only in
the factor with the greatest absolute value. To finalise the analysis, we
performed a chi squared test on the contingency table cluster per therapy
outcome, calculating the standard residual in order to identify the differences
between good outcome and poor outcome clinical transcripts in terms of core
themes.
3. Main results and discussion
The results of the cluster analysis show that the 1351 keywords selected
allow for the classification of 56.6% of context units. The high proportion of
unclassified context units is due to the transcripts richness in terms of paraverbal interactions (i.e. mhm, yeah, etc). The eta squared value was calculated
on partitions from 3 to 9, and it showed six clusters as the optimal solution
(η2 = 0.034). In table 1, we can appreciate the emotional map emerging from
the clinical transcripts representing the clusters location in the factorial space
produced by the interpretation of the five factors. The first factor reflects
patient positioning, which can be passive or active; the second factor refers to
the relationship that could be familiar or unfamiliar, i.e., a person facing
something new and unpredictable; the third factor represents the
communication content that can be emotional or concrete; the fourth factor
reflects the outcome of the therapeutic work, that is, the patient’s
empowerment or making sense of the patient’s experiences; and the fifth
factor distinguishes the issues within the daily ones, concerning everyday life,

JADT’ 18

339

from the relational ones, concerning the loved ones.2
Table 1  Factorial space representation (the percentage of explained inertia
is reported between brackets under each factor).
Cluster

1

Label
(CU%)

Family Structure
(11.6%)
Transformative Process
(12.1%)
Concrete thinking
(16.1%)
Therapeutic Relationship
(22.4%)
Relational Issues
(14.6%)
Feelings
(23.1%)

2
3
4
5
6

Factor 1
(26.7%)
Positioning
Passive
0.20
Active
-0.46
Passive
0.84
Active
-0.25
0.04
0.06

Factor 2
(25.8%)
Relationship
Familiar
-0.56
Unfamiliar
0.29
Unfamiliar
0.34
Familiar
-0.18
Familiar
-0.14
Unfamiliar
0.58

Factor 3
(21.5%)
Content
Emotional
-0.16
0.06
Concrete
0.42
Concrete
0.41
Emotional
-0.47
Emotional
-0.43

Factor 4
(14.5%)
Outcome
-0.01
To empower
-0.35
To empower
-0.19
To understand
0.28
To empower
-0.18
To understand
0.49

Factor 5
(11.5%)
Issues
Daily
-0.32
Daily
-0.16
0.05
Relational
0.16
Relational
0.45
Daily
-0.14

CU = context units classified in the cluster.
Table 2  Psychotherapy core themes.
Cluster 1
Family
Structure

Cluster 2
Transformative
Process

Cluster 3
Concrete
Thinking

Cluster 4
Therapeutic
Relationship

Cluster 5
Relational
Issues

Cluster 6
Feelings

keyword CU keyword

CU keyword CU keyword CU keyword

CU keyword

home

525 start

507 hear

455 week

699 mother

399 understand 416

kid

371 able to

504 money

326 sense

675 life

335 hurt

300

house

290 change

438 dollar

267 day

438 problem

333 important

298

father

241 different

396 accept

205 bad

432 hard

292 person

231

husband

213 situation

288 pay

196 angry

381 care

268 hard

213

child

205 point

237 listen

175 call

253 deal

252 support

185

parent

194 go on

216 believe

135 night

189 family

237 inside

170

stay

190 mind

213 matter

130 morning

169 relationship 233 strong

168

live

179 trying

183 sell

126 set

162 Father

153

195 pain

CU = context units classified in the cluster.

The six clusters are of different sizes (table 1) and reflect the core themes of
the brief psychotherapy (table 2). The first cluster describes the family
structure with its role and places; the second cluster reflects the transformative

2 In the negative pole of the fifth factor (Daily Issues) we find the following
words: house, stay, TV, rule, street, teacher, move out, neighbour, pounds, and in the
positive pole we find words as mother, life, problem, sister, relationship.

CU

340

JADT’ 18

process characterising a psychotherapy; the third cluster highlights the
concrete thinking process, a way to think that could be defined as concrete
thinking, which is often rational and frequently concerning economic issues;
the fourth cluster represents the therapeutic relationship that is made of
concrete limits, and the process of making sense of personal experiences; the
fifth cluster reflects the relational issues of the patient’s private life; and the
sixth cluster refers to the process of detecting, recognizing, and
understanding feelings, characterizing internal emotional experiences.
There is a significant difference in the number of content units classified in
each cluster among the good and poor outcome therapies (χ2, df = 5, p < 0.01).
In particular, the differences lay on the relevance of two of the six core
themes: the concrete thinking and the feelings. While the good outcome brief
psychotherapies are characterized by a high number of context units
classified in the cluster feelings (SE = 6.8) and a low number of context units
classified in the cluster concrete thinking (SE = -5.8); the poor outcomes
psychotherapies are characterized by a high number of context units
classified in the cluster concrete thinking (SE = 6.8) and a low number of
context units classified in the cluster feelings (SE = -7.0). Namely, it would
seem that patients tend to dwell upon their emotional experiences in the
good outcome psychotherapy, while they tend to dwell upon facts in the
poor outcome psychotherapy, probably not connecting them to their
emotional experiences. Given that we classified the interactions among the
patients and the therapists in this analysis, the therapy outcome could derive
both from the patient’s ability in dealing with feelings or the therapist’s
ability to support the patient in doing so.
The above-mentioned differences between good and poor outcome cases are
coherent with findings obtained on the same sample by means of a principal
component analysis made on the transcripts coded according to three
dictionaries: abstract language, emotional positive language, and emotional
negative language (de Felice et al., 2018). In this study, differences in the
correlation matrices between good outcome and poor outcome cases were
evident. The most obvious one concerned the dynamic in which the patient
made use of abstract/concrete language, interpreted very positively in poor
outcome cases and very negatively in good outcome cases. In the latter, it
was probably and correctly considered as a patient’s defense mechanism to
address. This was confirmed by the use of positive and negative emotional
language, inversely proportional to abstraction, only in poor outcome cases.
4. Conclusion
Talking about concrete events without any sort of emotional involvement in
the clinical literature is a defence mechanism that goes under the name of

JADT’ 18

341

rationalisation, and it represents a way to protect the mind from painful
feelings using an abstract, intellectual and often concrete attitude in dealing
with them. While the good outcome psychotherapeutic relationships seem to
be capable of addressing the emotional content laying under the surface of
the psychotherapeutic field (i.e. use of the therapist’s negative emotional
language), the poor outcome dynamics seem to be completely wrapped up in
a process of avoiding it. Both the PCA (de Felice et al 2018) and text analysis
on clinical transcripts confirmed the difficulty in poor outcome
psychotherapies to work on the patient’s emotional aspects. This bottom-up
technique of text analysis on clinical transcripts turned out to be an
enlightening tool to let their latent dimensions emerge, arranging the clinical
process and outcome, therefore, providing a very useful tool for clinical
purposes.
References
Beck A.T., Steer R.A. and Garbin M. G. (1988). Psychometric properties of the
Beck Depression Inventory: Twenty-five years of evaluation. Clinical
Psychology Review, 8: 77 100.
Bucci W., Kabasakalian-McKay R. and RA Research Group (1992). Scoring
referential activity. Ulm, Germany: Ulmer Textbank.
Carli R., Dolcetti F. and Dolcetti (2004). L’Analisi Emozionale del Testo
(AET): un caso di verifica nella formazione professionale. In Purnelle G.,
Fairon C. and Dister A., editors, Actes JADT 2004: 7es Journées
internationales d’Analyse statistique des Données Textuelles, pp. 250-261.
Cordella B., Greco F. and Raso A. (2014). Lavorare con Corpus di Piccole
Dimensioni in Psicologia Clinica: Una Proposta per la Preparazione e
l’Analisi dei Dati. In Nee E., Daube M., Valette M. and Fleury S., editors,
Actes JADT 2014 (12es Journées internationales d’Analyse Statistque des
Données Textuelles, Paris, France, Juin 3-6, 2014), pp. 173-184.
de Felice G., Orsucci F., Mergenthaler E., Gelo O., Paoloni G., Scozzari A.,
Serafini G., Andreassi S., Vegni N. and Giuliani A. (2018). What
differentiates good and poor outcome psychotherapies? A statistical
mechanics approach to psychotherapy research. Nonlinear Dynamics,
Psychology and Life Sciences. Submitted.
Gelo O.C.G. and Salvatore S. (2016). A dynamic systems approach to
psychotherapy: A meta-theoretical framework for explaining
psychotherapy change processes. Journal of Counseling Psychology, 63(4):
379-395.
Gelo O.C.G., Salcuni S. and Colli A. (2013). Text analysis within quantitative
and qualitative psychotherapy process research: introduction to special
issue. Res. Psychother. Psychopathol. Process Outcome 15: 45–53.

342

JADT’ 18

Greco F. (2016). Integrare la disabilità. Una metodologia interdisciplinare per
leggere il cambiamento culturale. Franco Angeli.
Greco F., Maschietti D. and Polli A. (2017). Emotional text mining of social
networks: The French pre-electoral sentiment on migration. Rivista Italiana
di Economia Demografia e Statistica, 71(2): 125:36.
Greenberg L., Rice L. and Elliott R. (1993). Facilitating emotional change. The
moment by moment process. Guilford Press.
Greenberg LS, Watson JC (1998). Experiential therapy of depression:
differential effects of client-centered relationship conditions and process
experiential interventions. Psychotherapy-Research 8: 210–224.
Lebart L. and Salem A. (1994). Statistique Textuelle. Dunod
Mergenthaler E. (2008). Resonating minds: A school-independent theoretical
conception and its empirical application to psychotherapeutic processes.
Psychotherapy Research, 18(2): 109-126.
Salvatore S., Gelo O., Gennaro A., Metrangolo R., Terrone G., Pace V.,
Venuleo C., Venezia A. and Ciavolino E. (2017). An automated method of
content
analysis
for
psychotherapy
research:
A
further
validation. Psychotherapy Research, 27(1): 38-50.
Salvatore S., Gennaro A., Auletta A.F., Tonti M. and Nitti M. (2012).
Automated method of content analysis: A device for psychotherapy
process research. Psychotherapy Research, 22(3): 256-273.
Savaresi S.M. and Boley D.L. (2004). A comparative analysis on the bisecting
K-means and the PDDP clustering algorithms. Intelligent Data Analysis,
8(4): 345-362.
Spitzer R., Williams J., Gibbons M. and Firs M. (1989). Structured Clinical
Interview for DSM-III-R. American Psychiatric Association
Watson J.C., Greenberg L. S. and Lietaer G. (1998). The experiential paradigm
unfolding: Relationship & experiencing in therapy. In Greenberg L.S.,
Watson J.C. and Lietaer G., editors, Handbook of experiential psychotherapy,
Guilford Press.

JADT’ 18

343

DOMINIO: A Modular and
Scalable Tool for the Open Source Intelligence
Francesca Greco1, Dario Maschietti2, Alessandro Polli3
1

La Sapienza University of Rome, Prisma S.r.l. – francesca.greco@uniroma1.it
2 Prisma S.r.l – d.maschietti@prismaprogetti.it
3 La Sapienza University of Roma – alessandro.polli@uniroma1.it

Abstract
Prisma has developed an innovative technology for the Open Source
Intelligence (OSINT) which aims to provide a solution for those processes of
knowledge management, which require the intervention of a human
operator, unaided by information technology (IT) support, in one or more
stages of the procedure. Such intervention involves a considerable waste of
time and resources that could be reduced through the use of an IT tool,
partially or totally automating entire stages of the procedure. DOMINIO is a
platform that implements tools for automatic online information aggregation,
its analysis, the possible alignment with traditional databases and the
representation through infographic and georeferencing tools, in order to
generate a report. This paper describes the platform architecture, the main
algorithms used in the analysis stage of the contents and possible directions
of development.
Abstract
Prisma ha sviluppato una tecnologia innovativa finalizzata all’Open Source
Intelligence (OSINT) che intende fornire risposta alle necessità di knowledge
management, che richiedono l’intervento di un operatore umano, non
assistito da supporti di information technology (IT), in una o più fasi della
procedura. Tale intervento comporta un notevole dispendio di tempo e
risorse che potrebbe essere ridotto attraverso l’utilizzo di uno strumento di
IT, automatizzando parzialmente o totalmente intere fasi della procedura.
DOMINIO è una piattaforma che implementa strumenti per l’aggregazione
automatica di informazioni on line, la loro analisi, l’eventuale allineamento
con banche dati di tipo tradizionale, la rappresentazione attraverso tool di
infografica e georeferenziazione, allo scopo di generare una reportistica. Il
presente contributo descrive l’architettura della piattaforma, i principali
algoritmi adottati nella fase di analisi dei contenuti e le possibili direzioni di
sviluppo.
Keywords: knowledge management, Open Source Intelligence tool,
Information Technology,

344

JADT’ 18

1. Introduction
There is a close link between data management and knowledge on the one
hand, and knowledge and innovation on the other. The growing mass of
unstructured information from disparate channels (search engines, RSS
feeds, social networks) and traditional databases entails the need to
drastically simplify the preparation, analysis and reporting stages required to
structure the information. In fact, only a structured information translates
into knowledge. Knowledge, in turn, is a major driver of innovation and,
properly managed, it translates into a competitive advantage. The idea at the
basis of the tool OSINT (Open Source Intelligence) stems from the needs
expressed by analysts – mainly involved in the field of sentiment analysis
and opinion mining industry. However, this idea is enough comprehensive
to encompass all those activities of knowledge management, similar to the
former, which require intervention by a human operator, unaided by IT
support (Information Technology), in one or more stages of the procedure,
the intervention of which involves a great deal of time and resources.
Although in high-end solutions machine learning systems are starting to
spread, the available technology is still characterized by significant
limitations, especially in the presence of unstructured information. In
particular, with regard to supervised machine learning systems, intervention
is required by an operator in the initial stages of the procedure and, in
general, with reference to any automated system applied to the analysis of a
text, it is still impossible to identify complex cognitive functions (for example,
irony). Of course, these problems are immanent in many fields of OSINT,
and they also affect the stage of reporting, which requires a direct
involvement of the analyst, unaided by IT. So, the availability of an IT tool
that minimizes human operator intervention − partially or totally automates
entire stages of the procedure − would result in substantial advantages, like
time savings, increased productivity and the resulting increased efficiency in
the allocation of human and financial resources.
Prisma has developed an innovative technology of OSINT, which aims to fix
the problems briefly described above. The platform implements tools for
automatic aggregation of the online information, their analysis, the alignment
with traditional databases, the representation through infographic and
georeferencing tools, aimed to automate also the phase of elaboration of the
final report.
This paper will describe the architecture of the platform, the main analysis
modules and the possible directions of development.

JADT’ 18

345

2. Platform Architecture
DOMINIO is an OSINT (Open Source Intelligence) platform that
automatically aggregates information from online and traditional databases,
analyses it and generates reports on a user-defined subject. The platform
collects information by querying several channels: search engines (Google,
Yahoo, Bing), social networks (Facebook, Twitter, Google+), RSS feeds, blogs
(Blogger, Wordpress, Tumblr), traditional databases. The goal of DOMINIO
is to build a structured set of contents, as broad as possible, and to carry out a
wide range of qualitative and quantitative analysis. DOMINIO stores these
contents within a non-relational database (DB) (MongoDB, 2018; Morphia,
2018), classifying the various documents by channel of origin (Twitter,
Facebook, RSS, etc.) to ensure the homogeneity of the collections.
Among the options, the DOMINIO user can make queries on-demand or in a
continuous mode. The on-demand option carries out an asynchronous
search, while the continuous mode option enables to aggregate periodically
data and to track a subject over an extended time span. The DOMINIO’s
architecture allows the user to switch from one mode to another; the
availability of two searching modes allows overcoming the trade-off between
accuracy of analysis and speed of processing.
With regard to one or more subjects selected by the operator, DOMINIO
performs synchronous or asynchronous research on a set of Internet
channels, such as search engines (Google, Yahoo, Bing), social networks
(Facebook, Twitter, Google+), RSS feeds, blogs (Blogger, Wordpress, Tumblr).
The user can also extend the search to the Deep Web, through specific search
engines, such as Torch or Grams.
Moreover, to meet specific information needs, DOMINION can match these
search results with the information achievable from the traditional databases
to support many types of analysis (brand reputation, country risk
assessment, opinion polls, cyber security, etc.), considerably increasing the
operability and flexibility of the tool.
Among the traditional databases already available, DOMINIO includes:
 IHS Jane's (2018), which provides updates on military and political
situation, terrorist acts, civil wars, transportation system, for most of
the countries in the world;
 Bureau Van Dijk (2018), which collects firms data on ratings,
shareholdings, equity investments and M&A;
 MIG (a geographic information database drawn up by one of the
authors).
In addition, for specific information purposes, DOMINIO is open for
interfacing with Enterprise Resource Planning databases (like SAP, Oracle,
etc.) through market tools (Business Object, Quick View).

346

JADT’ 18

The search results are recalled by the analyst, who operates from a CMS
(Content Management System) application to manage the structured set of
content and conduct a wide range of qualitative and quantitative analyses
(from simple summary statistics to sophisticated multivariate analyses and
text and opinion mining techniques).
The statistical methods implemented on DOMINIO are chosen by the Prisma
research team according to a set of criteria that privileges the suitability of
one algorithm to automate entire stages of the procedure, in accordance with
the original design idea. Moreover, the modular architecture of DOMINIO,
described briefly below, allows a quick integration of the latest analysis tools
and innovative methodologies produced in the academic field.
Once the stage of content analysis is completed, the CMS application
generates a micro-site containing the results (geo-referenced maps, summary
statistics, multivariate analysis results, textual and semantic analysis of
sentiment analysis, etc.). After selecting a graphic layout for the final report,
the analyst has only to write notes and final remarks.
The possibility of including features generating automatic and/or autocompletion comments, customizable by the user, is also being studied. Once
the last stage is completed, the report is ready for online publication or
traditional diffusion in pdf format, or linked to external services.
From an architectural point of view, DOMINIO is designed following the
most modern criteria of modular software design, with the parallel
development of the platform’s modules. In short, in order to ensure a greater
fault tolerance and high safety standards, the system is divided into three
independent logical units (cfr. Figure 1):
• DOMINIO Engine Unit (MEU), which implements the features of
1) scraping information from the sources mentioned above (web,
social networks, RSS feeds, traditional databases); 2) storage of
results on MEDB database; 3) qualitative and quantitative
analysis;
• DOMINIO RESTurl Unit (MRESTU), which receives requests from
the MCMS unit, verifies the consistency and forwards the request
to the unit ME. Upon receiving the response, it implements the
request by adding additional fields (username, token, etc.) and
returns them to the MCMS client. The MRESTU unit contains the
database (MRESTDB) for user profiling;
• DOMINIO Content Management System Unit (MCMSU), which
manages the stage concerning the reporting and archiving of
reports according to pre-logical criteria (organization by topic,
chronologically, for templates, etc.).

JADT’ 18

347

Figure 1 - DOMINIO General Overview

3. Main analysis modules
3.1. Country Threat Assessment
The Country Threat Assessment module supports the Company Intelligence
and Security analyst in the country's risk assessment process. Through a
responsive type interface, it aggregates information from major global
industry databases (eg, IHS Jane's) giving an assessment of external and
internal risk and that due to political and socio-economic factors and
potential outbreaks or revolutionary movements for 192 different countries.
Country Threat Assessment is integrated with intelligence information
updated weekly on each country. Through an automatic report, data is
aggregated into a single file by optimizing timing of risk assessment and
providing a solid foundation for any further detailed analysis. DOMINIO
offers the possibility of making a full or partial information download, and
the generation of an automatic report, thus optimizing any drafting
processes.
3.2. Due Diligence
The Due Diligence module supports the Economic Intelligence analyst in the
process of business valuation in relation to suppliers, partners and
customers. Among the sectors analysed in the module are included

348

JADT’ 18

assessments of profitability and financial performance as well as
creditworthiness. Through a simple and intuitive interface, the module
aggregates information from leading industry databases and returns an
economic, financial and credit risk profile on hundreds of millions of
businesses around the world. The Due Diligence Module also allows an
assessment of individuals, through the analysis of individuals exposed
politically, returning an automatic report that integrates the main aspects of
each business and its economic risk analysis.
3.3. Open Source Intelligence
On completion of the aggregation of large amounts of data from major social
networks (Facebook, Twitter, Youtube) and the main Italian newspapers
based on predetermined keywords analyst, a statistical representation of the
main trending topic is returned and an output of structured data for
subsequent multivariate analysis is generated. Furthermore, the module
allows the geo-referencing of content, highlighting even at geographic levels
useful signs for the analyst. As for each of DOMINIO’s modules, it is possible
to generate automatic reporting.
3.4. Geographic Information Module
This is a module that analyses the information inferable from a dataset of
basic statistical information and related indicators, with reference to a
multitude of subjects, 9 of which are in a current stage of development. The
basic statistical information, refers to the division of the Italian territory into
provinces, covering a time period between 1995 and the latest available year,
which for some subject areas is ongoing or, more frequently, the previous
year to the current one. The dataset will be supportive to a wide range of
applications - from forecasting and scenario analysis, counterfactual analysis
to spatial analysis.
3.5. Text Mining Module
On completion of the automatic analysis of textual data using statistical
methods (Lebart et Salem, 1994; Feldman et Sanger, 2006; Bolasco, 2013), in
order to extract structured information, the main statistical methods of
analysis of textual data implemented in DOMINIO are: factor analysis
(correspondence analysis, multiple correspondence analysis); cluster analysis
(k-mean, bisetting k-mean, fuzzy clustering, etc.); network analysis; Markov
analysis; pattern recognition.
For example, during the French presidential campaign of 2017 we analysed
the sentiment about migration, that was one of the most debated theme. We
performed an Emotional Text Mining (Greco et al., 2017) in order to explore

JADT’ 18

349

the emotional content of the Twitter messages concerning migration written
in French in the last two weeks before the first round of the presidential
election in 2017. The aim was to analyse the opinions, feelings and shared
comments, classifying the contents and the sentiments. We retrieved the
messanges from the Twitter repository collecting a sample of over une
hundred thousand tweets The large size corpus of 2.154.194 tokens (TTR =
0,01; Hapax percentage = 40,4) underwent a multivariate analysis based on a
bisecting k-means algorithm (Savaresi et Boley, 2004) to classify the text, and
a correspondence analysis (Lebart et Salem, 1994) to detect the latent
dimensions setting the cluster per keywords matrix. The advantage
connected with this approach is to interpret the factorial space according to
words polarization, thus identifying the emotional categories that generate
migration representations, and to facilitate the interpretation of clusters,
exploring their relationship within the symbolic space (Greco, 2016).
The results interpretation allowed for the detection of seven representations
of migrants that corresponded to three different sentiments: positive (42%),
negative for the community (45%), and negative for migrants (13%). We
considered as negative the representation of migrants as squatters, invaders,
terrorists, trafficking slaves and migration victims, and positive the sport
heroes and the EU solidarity target. Among the negative clusters, we
distinguished negativity according to the direction of the action: squatters,
terrorists and invaders are negative for the community and trafficking slaves
and migration victims are negatives for migrants themselves (see Greco et al.,
2017). Moreover, It was possible to highlight the connection between the real
life events and the tweets production. While the terrorist attack three days
before the first round of voting in the centre of Paris had slightly modified
the production of messages, the candidates’ interviews had a higher impact.
This suggests that the medialization was more important than the terrorist
attack in the production of messages (see Greco et al., 2017).
4. Conclusion
The innovative aspect that characterizes DOMINIO is the ability to aggregate
data of different types and from different channels of information,
automatically, simply and transparently. Moreover, its structure allows for
the integration of the latest analytical tools and innovative methodologies
produced in academia. By means of an automated reporting system, the
analyst is supported in the assessment of risk and the collection of
information in the geopolitical and economic field and from open sources.
The set of modules allows the analyst to generate knowledge from an evergrowing amount of data by optimizing the processes of assessment and risk
reduction.

350

JADT’ 18

References
Bolasco S. (2013). L’analisi automatica dei testi: Fare ricerca con il text mining.
Carocci.
Bureau von Dijk (2018). A Moody’s Analytics Company. Bureau von Dijk,
https://www.bvdinfo.com/it-it/home
Feldman R. and Sanger J. (2006). The Text Mining Handbook: Advanced
Approaches in Analyzing Unstructured Data. Cambridge University Press.
Greco F. (2016). Integrare la disabilità. Una metodologia interdisciplinare per
leggere il cambiamento culturale. Franco Angeli.
Greco F., Maschietti D. and Polli A. (2017). Emotional text mining of social
networks: The French pre-electoral sentiment on migration. RIEDS, 71(2):
125:36.
IHS Jane’s (2018). Jane’s Information Group. IHS Jane’s, http://www.janes.com
Lebart L. and Salem A. (1994). Statistique Textuelle. Dunod
MongoDB
(2018).
MongoDB
for
GIANT
ideas.
MongoDB,
https://www.mongodb.com
Morphia (2018). The Java Object Document Mapper for MongoDB. MongoDB,
https://mongodb.github.io/morphia/
Savaresi S.M. and Boley D.L. (2004). A comparative analysis on the bisecting
K-means and the PDDP clustering algorithms. Intelligent Data Analysis,
8(4): 345-362.

JADT’ 18

351

Is training worth the trouble?
A PoS tagging experiment with Dutch clinical records
Leonie Grön, Ann Bertels, Kris Heylen
KU Leuven – leonie.gron@kuleuven.be; ann.bertels@kuleuven.be; kris.heylen@kuleuven.be

Abstract
Part-of-speech (PoS) tagging is a core task of Natural Language Processing
(NLP), which crucially influences the output of advanced applications. For
the tagging of specialized language, such as that used in Electronic Health
Records (EHRs), the domain adaptation of taggers is generally considered
necessary, since the linguistic properties of such sublanguages may differ
considerably from those of general language. Previous research suggests,
though, that the net benefit of domain adaptation varies across languages.
Therefore, in this paper, we present a case study to evaluate the effect of
training with in-domain data on the tagging of Dutch EHRs.
Keywords: Electronic Health Records; Part-of-Speech tagging; medical
sublanguage; Dutch
1. Background
EHRs are valuable resources for data-driven knowledge-making. To unlock
the relevant information from free text, domain-specific NLP systems are
required. Such systems must deal with a text genre that can be characterized
by a high density of specialized terms, including non-canonical variants, and
non-standard syntactic constructions. These properties affect all further steps
in a processing pipeline, starting from core tasks such as PoS tagging. Since
PoS values are important features for further processing, the output of many
systems, such as tools for term extraction and term-to-concept mapping (e.g.
Doing-Harris et al., 2015; Scheurwegs et al., 2017), crucially depends on the
accuracy of the PoS tags assigned in the first place. Processing suites such as
cTAKES (e.g. Savova et al., 2010), which have been developed specifically for
the medical domain, are known to boost tagging performance. As most tools
are only available for English, though, systems dealing with other languages,
such as Dutch, are required to start the domain adaptation from scratch.
Typically, this process involves the re-training of an existing tool on handcoded data, which is time- and labor-intensive. Besides, evidence from
German challenges the wide-held belief that domain training is a prerequisite
to achieve good tagging performance (Wermter et Hahn, 2004).

352

JADT’ 18

Given these considerations, we conduct a pilot study to investigate the
potential benefit of domain adaptation for the PoS tagging of Dutch EHRs.
Firstly, we assess the impact of training with a hand-coded clinical dataset on
the accuracy of an off-the-shelf tagger. Secondly, we evaluate how the
difference in accuracy affects the output of a term extraction method based on
PoS patterns.
2. Related Work
For the PoS tagging of clinical writing, the main challenges reside in the
particular linguistic properties of the genre, both at the lexical and the
syntactic level: On the one hand, EHRs contain a high proportion of
specialized terminology and idiosyncracies, including misspellings and noncanonical abbreviations; a tagger developed for general language will thus
encounter a high number of out-of-vocabulary words (Knoll et al., 2016). To
complicate this matter, the PoS distributions in clinical corpora differ from
those found in general language, which may be detrimental to the statistical
classification of unknown or ambiguous tokens (Pakhomov et al., 2006). On
the other hand, EHRs are typically composed in a telegraphic style, which can
be characterized by the omission of functional syntactic elements; the lack of
linguistically informative context may prevent the accurate prediction of PoS
transitions within n-grams (Coden et al., 2005). At the same time, the average
sentence length in EHRs is relatively short; the high number of inter-sentential
transitions may pose additional pitfalls for an out-of-domain tagger
(Pakhomov et al., 2006). Most previous research thus agrees that the use of offthe-shelf taggers on clinical writing is highly prone to errors, which are likely
to be propagated through the different levels of an application (Ferraro et al.,
2013).Therefore, many state-of-the-art systems use an annotated set of EHRs
for training. The creation of training materials comes at a cost, though, and
entails a range of methodological challenges in itself, such as the creation of
suitable guidelines and tagsets (Albright et al., 2013). To circumvent these
issues, alternative ways of domain adaptation have been explored, including
the integration of a domain-specific vocabulary, and the exploitation of
morphological features to classify unknown words (Knoll et al., 2016).
However, other languages than English may present a different case: In an
early study, Wermter & Hahn (2005) come to the conclusion that in German,
taggers trained on newswire perform very well on EHRs. This surprising
finding can be partly attributed to the rich inflectional system of the language,
which lends itself to the prediction of PoS categories. On the other hand, the
low complexity of the medical sublanguage may be a factor: In their study, the
general training data subsumed all PoS transitions found in the clinical test
data, so that the tagger was sufficiently equipped to handle the latter.

JADT’ 18

353

3. Methods
3.1. Corpus and manual tagging
Our study is based on the analysis of a mixed sample of EHRs, containing a
total of 375 documents. As detailed in Table 1, the subsets of this sample
differ with regard to their medical subdomain, institutional origin and
document structure: The EN and RD sets cover only one medical specialty,
whereas the DL, SP and GP sets are less homogeneous; the DL, EN and RD
sets were composed at a single institution, while the documents in the GP
and SP sets are drawn from a multi-source database, Integrated Primary Care
Information (ICPI), which contains EHRs from medical practices all across
the Netherlands. Finally, the EHRs in four subsets (DL, GP, RD, SP) had been
split into shorter fragments to comply with privacy standards; therefore,
these documents are much shorter than those in the EN set, which count
204.2 tokens on average. All EHRs are tokenized with the NLTK tokenizer1
and manually labelled by the authors, using the Universal Tagset (Petrov et
al., 2012). Finally, for each subset, the EHRs are split into a training and test
set, containing 67% vs. 33% of the files respectively.
Table 1: Overview of the subsets of our file sample. The first three columns specify the name of
the subset, the document types, the origin and the number of institutions involved in their
creation. The remaining columns give the number of documents, the absolute length in tokens,
and the average document length in tokens.
Subset

Document
types

Origin

Nr. of
sources

Nr. of documents

Subset
length

DL

Clinical
discharge
letters
EHRs from
endocrinology
EHRs from
general
practitioners
EHRs from
radiology

EMC
Rotterda
m
UZ
Leuven
IPCI
(Vlug et
al., 1999)
EMC
Rotterda
m
IPCI
(Vlug et
al., 1999)

One

88

3597

Average
documen
t length
40.88

One

80

16337

204.2

Multipl
e

60

1431

23.85

One

60

1441

24.02

Multipl
e

87

4784

54.99

Σ

375

27590

73.57

EN
GP

RD

SP

Specialist
letters from
various fields
(e.g.
cardiology)

1

http://www.nltk.org/_modules/nltk/tokenize.html

354

JADT’ 18

3.2. Evaluation
3.2.1. Effect of domain training on tagging performance
Firstly, we assess the impact of using in-domain data for training on tagging
accuracy. For evaluation, we use the state-of-the-art Perceptron Tagger.2 This
tagger uses context tokens as well as suffix features for classification. As
Knoll et al. (2016) show, this configuration outperforms a primarily
sequential tagger, as used by Wermter & Hahn (2005), on clinical data. The
pre-compiled model for Dutch is trained on the Alpino Treebank (van
Noord, 2006). In addition, we build a domain-specific model based on the
manually labelled training set. Then, we feed both models into the tagger to
classify the test set. To measure the accuracy of each model, we calculate the
precision, i.e. the proportion of tags that match those in the manually labelled
gold standard.3 To compare the effect across the different subsets, we
calculate the gain in precision achieved with the domain model relative to the
precision achieved with the Alpino baseline.
3.2.2. Effect of tagging performance on term recognition and extraction
Secondly, we quantify the effect of tagging performance on pattern-based
term recognition. For the identification of term candidates, we use a set of
PoS sequences that are characteristic for termhood in the domain. Similar to
Scheurwegs et al. (2017), we focus on complex nominals, i.e. nouns
surrounded by one or more modifiers; Table 2 provides some examples of
such patterns.
Table 2: Examples of PoS patterns used for term retrieval. The left column lists the target tag
sequence, the middle and right column provide Dutch examples and English translations of
term candidates.

PoS pattern
adjective noun
noun adposition noun
noun noun

Dutch example
‘diabetische
retinopathie’
‘syndroom van Apert’
‘zwelling enkel’

English translation
diabetic retinopathy
syndrom of Apert
swelling ankle

Using a sliding-window approach, we iterate through the three tagged
versions of the test set, i.e. the manually tagged gold standard, the version

http://www.nltk.org/_modules/nltk/tag/perceptron.html
The Alpino model uses a more fine-grained tagset than the Universal Tagset
used for the manual tagging. To enable the comparison across models, the redundant
labels from Alpino are mapped to the respective categories of the Universal Tagset
(e.g. adj <adjective>, comparative <comparative> → ADJ <adjective>).
2
3

JADT’ 18

355

tagged with the Alpino model and the version tagged with the domain
model. We identify all PoS sequences that match the pre-specified patterns,
and extract the respective tokens for manual validation. For each version, we
calculate the precision as the proportion of true positives, i.e. domain-specific
phrases, relative to the total list of matches.4 To assess the individual effect
size, we also calculate the relative gain in precision for each subset.
3.3. Results
3.3.1. Effect on tagging performance
For PoS tagging, training on domain data has a sizeable effect on precision:
The domain model reaches 85.8% accuracy on the test set of held-out EHRs,
compared to 66.9% with the Alpino baseline. Regardless of the model, the
best results are achieved for DL, followed by RD and EN; for SP and GP,
precision stays at the lowest levels. To evaluate the improvement across the
different subsets, we compare the increase in precision relative to the value
achieved with the baseline. The comparison of these values reveals
considerable differences of the individual effect sizes: In SP, the training
effect is most striking, followed by GP and EN; in RD and DL, the
improvement is less evident.
3.3.2. Effect on term recognition and extraction
The increase in accuracy has a strong effect on the term retrieval task: When
using the tags assigned by the Alpino model, only 3.42 of the retrieved
candidates are correct; with the domain model, precision jumps to 9.3%.
Again, the results vary substantially across the different datasets: Overall, the
best results are obtained for EN, followed by RD and DL. In SP and GP,
precision remains at the lowest levels. Judging from the relative gain in
precision, though, we find the strongest increase in GP, followed by DL. In
RD, EN and SP, we only find weaker effects. Table 3 provides the full results
for both tasks.
For error analysis, we label all false positives with the nature of
misclassification, whereby we distinguish between three types of errors:
Firstly, errors based on erroneous PoS tags (e.g. ‘merkt hypoglycemie’ notices
hypoglycemia, whereby the verb is tagged as an adjective); secondly,
segmentation errors, whereby one token is associated with an unrelated one
(e.g. ‘oedeem Lipitor’ edema Lipitor, whereby two unrelated nouns are

To qualify as domain-specific, a phrase must contain at least one noun that has
a concept entry in the clinical terminology SNOMED-CT (International Release July
2017; http://browser.ihtsdotools.org/). For instance, ‘echografie rechterschouder’
echography right shoulder, which refers to a clinical procedure, would count as a true
positive; the general expression ‘pak koekjes’ bag of biscuits would not.
4

356

JADT’ 18

mistaken for a compound); thirdly, term candidates that match a target PoS
pattern, but are not domain-specific (e.g. ‘kleine boterhammen’ small
sandwiches). Then, we calculate the proportion of error types among the false
positives provided by both models. With the Alpino model, the vast majority
of errors (74.4%) is based on false PoS tags. About 18.2% of the proposed
term candidates are out-of-domain, while only a small portion (7.3%) of
errors is caused by mistakes in segmentation. Conversely, with the domain
model, most false positives (49.7%) are out-of-domain terms; errors in
tagging and segmentation account for 30.1% and 20.2% respectively.
Table 3 : Precision of PoS tagging and term extraction across subsets. The first
column specifies the subset. The second and third column provide the percentage of
correct tags assigned by the domain model and the Alpino model respectively; the
fourth column contains the relative increase in precision. The remaining three
columns provide the corresponding values for the extraction task.
Term extraction

PoS tagging

%
% Prec
Prec
%
domain
subset % Prec domain model % Prec Alpino % increase model Alpino increase
DL

89.62

EN
GP

76.61

16.99

7.33

2.64

177.87

86.82

67.5

28.62

21.48

8.04

167.1

79.81

61.76

29.23

3.28

0.84

291.31

RD

88.98

74.1

20.08

8.89

3.31

168.52

SP

83.68

54.5

53.53

5.52

2.26

144.09

Σ

85.78

66.9

29.69

9.3

3.42

189.78

4. Discussion
Overall, the positive effect of domain adaptation is evident: Using clinical
data for training improved the accuracy of PoS assignments and, as a
consequence, the output of the term extraction method. Based on our results,
we do not see a clear relation between the amount of training data and the
global level of precision: For PoS tagging, DL and RD, which are among the
smaller subsets, score highest; on the other hand, for the term extraction task,
EN, which is the largest subset, produces the best results by far. This
indicates that the benefit of training hinges on linguistic and semantic
qualities, rather than the mere quantity of the data.
In particular, tagging performance correlates with the homogeneity and wellformedness of the data. The homogeneity depends, on the one hand, on the
medical field: A dataset such as RD, which is confined to one clinical

JADT’ 18

357

specialty, only makes reference to a fairly limited number of medical
concepts; by contrast, a more heterogeneous set, such as SP, covers a wider
range. Besides, the number of institutions involved in data creation plays a
role: In an EHR sample provided by a single hospital, such as EN, it is likely
that preferred terms and phrases are perpetuated throughout the dataset. By
contrast, in a set drawn from a multi-source database, such as GP, the
potential for variation is higher. Both these factors affect the overall size of
the vocabulary, which, in turn, determines the complexity of the tagging
task. The well-formedness, on the other hand, depends mainly on the EHR
type. The GP set, for instance, contains mostly notes intended for internal
documentation; these notes are written in an informal style, whereby
function words and suffixes may be left out or truncated. As these features
usually serve as predictors for PoS classification, their omission may cause a
drop in tagging performance. While the global level of precision is thus
lowest in conceptually and lexically EHR samples, such as GP and SP, the
relative benefit of domain adaptation is the greatest here.
5. Conclusion
We conclude that the training with in-domain data benefits the output of PoS
taggers for clinical Dutch. Especially if the file sample covers different
subdomains, or if the language used deviates strongly from the standard, the
potential gain in performance is great. At the same time, considerable
training efforts are required to achieve only marginal improvements.
Depending on the scope of the project and the composition of the sample, it
may thus be preferable to implement a cheaper alternative, for instance by
integrating a domain dictionary into the tagger.
Acknowledgements
This work was supported by Internal Funds KU Leuven.
References
Albright D., Lanfranchi A., Fredriksen A., Styler W.F., Warner C., Hwang
J.D., Choi J.D. et al. (2013). Towards Comprehensive Syntactic and
Semantic Annotations of the Clinical Narrative. J Am Med Inform Assoc
vol. 20: 922–30.
Coden A.R., Pakhomov S.V., Ando R.K., Duffy P.H. and Chute C.G. (2005).
Domain-Specific Language Models and Lexicons for Tagging. J Biomed
Inform vol. 38: 422–30.
Doing-Harris K., Livnat Y. and Meystre S. (2015). Automated Concept and
Relationship Extraction for the Semi-Automated Ontology Management
(SEAM) System. J Biomed Semantics vol. 6 (15): 1–15.

358

JADT’ 18

Fan J.-W., Prasad R., Yabut R.M., Loomis R.M., Zisook D.S., Mattison J.E. and
Huang Y. (2011). Part-of-Speech Tagging for Clinical Text: Wall or Bridge
between Institutions? In AMIA Annu Symp Proc, pp. 382–91.
Ferraro J.P., Daumé H.I., DuVall S.L., Chapman W.W., Harkema H. and
Haug P.J. (2013). Improving Performance of Natural Language Processing
Part-of-Speech Tagging on Clinical Narratives through Domain
Adaptation. J Am Med Inform Assoc vol. 20: 931–39.
Knoll B.C., Melton G.B., Liu H., Xu H. and Pakhomov S.V.S. (2016). Using
Synthetic Clinical Data to Train an HMM-Based POS Tagger. In 2016
IEEE-EMBS (International Conference on Biomedical and Health Informatics),
pp. 252–55.
van Noord, G. (2006). At Last Parsing Is Now Operational. In Proceedings of
TALN 2006, pp.20–42.
Pakhomov S.V., Coden A. and Chute C.G. (2006). Developing a Corpus of
Clinical Notes Manually Annotated for Part-of-Speech. Int J Med Inform
vol. 75: 418–29.
Petrov S., Das D. and McDonald, R. (2012). A Universal Part-of-Speech
Tagset. In Piperidis N.C., Choukri K., Declerck T., Doğan M.U., Maegaard
B., Mariani J., Moreno A., Odijk J., and Piperidis S. Proceedings of the Eight
International Conference on Language Resources and Evaluation (LREC’12), pp.
2089–96.
Savova G.K., Masanz J.J., Ogren P.V., Zheng J., Sohn S., Kipper-Schuler K.C.
and Chute C.G. (2010). Mayo Clinical Text Analysis and Knowledge
Extraction System (cTAKES): Architecture, Component Evaluation and
Applications. J Am Med Inform

JADT’ 18

359

Les outils de la statistique textuelle pour analyser
les corpus de données d’enquêtes
de la statistique publique
France Guérin-Pace, Elodie Baril
Institut national d’études démographiques

Abstract
For more than 20 years, textual statistic methods have been allowing us to
explore and analyze data from official statistics survey and the different
corpora it contains: answers to an open question, associated words,
significant life events. Based on three corpora of data: Population-Lived
Spaces-Environments survey (Ined, 1992), EuroBroadMap survey on
representations of Europe in the world (2009), and more recently the
Information and Daily Life survey on adult reading skills (INSEE, 2011), we
have demonstrated the diverse use cases of these methods and the richness
that helps identify the corpus content in relation to the individual
characteristics of respondents as well as to the survey questions. In recent
years, we have mobilized these methods to post-codify the events collected in
the IVQ survey. Today we will present to you the results of this work: the
benefits and limitations of textual statistic method.
Résumé
Réponses à une question ouverte, mots associés, évènements marquants de la
biographie, constituent autant de corpus issus de données d’enquêtes de la
statistique publique que nous avons explorés et analysés avec les méthodes
de la statistique textuelle, depuis plus de 20 ans. A partir de trois corpus de
données : enquête Populations-Espaces de vie-Environnements (Ined, 1992),
enquête EuroBroadMap sur les représentations de l’Europe dans le monde
(2009), et plus récemment l’enquête Information et Vie quotidienne sur les
compétences en lecture des adultes (Insee, 2011), nous montrons la diversité
d’applications de ces méthodes, leur richesse pour cerner le contenu des
corpus en lien avec les caractéristiques individuelles des répondants mais
aussi d’autres questions d’enquête. Plus récemment nous avons mobilisés ces
méthodes pour post-codifier les évènements recueillis dans l’enquête IVQ.
Nous présenterons les apports et les limites de cette démarche.
Keywords: textual statistics, open-ended questions, associated words corpus,
post-coding.

360

JADT’ 18

1. Des corpus de nature variée
Introduire un questionnement ouvert dans une enquête en population
générale est toujours un défi pour les concepteurs même si les méthodes de
la statistique textuelle ont prouvé depuis longtemps leur intérêt et leur
efficacité pour leur traitement. Cerner les contours et l’acception d’un mot
valise était l’objectif de l’introduction de la question ouverte «Si je vous dis
environnement, qu'est-ce que cela évoque pour vous? » dans l’enquête «
Populations-Espace de vie-Environnements » réalisée en 1992 (INED) auprès
d'un échantillon de 6 000 personnes, représentatif de la population française.
Un des objectifs consistait à examiner quelles représentations les populations
construisent sur la notion même d'environnement.
Une technique un peu différente de recueil est celle adoptée, par exemple,
dans l’enquête EuroBroadMap conduite en 2009 dans 18 pays. Enquêter près
de 10 000 étudiants à travers le monde sur leurs représentations de l’Europe
est l’un des objectifs de ce projet européen. Une pièce centrale de ce dispositif
est de recueillir les mots associés à l’Europe par les étudiants1 après leur
avoir demandé de délimiter, selon leur perception, ses contours sur une carte
du monde. A la différence du corpus précédent, les mots ne sont pas
proposés sous forme de liste et ce sont les représentations spontanées qui
sont recueillies. Cette technique des mots associés a pour intérêt de
contraindre davantage le format des réponses et d’obtenir un corpus plus
homogène. Une des principales difficultés de ce corpus est celle de la langue
de recueil des mots associés. Pour résoudre en partie ce problème, nous
avons choisi de traduire les réponses en anglais pour chacun des pays au
moment de la saisie, selon des consignes précises2.
Une autre forme de matériau qualitatif intéressant à recueillir dans les
enquêtes concerne les événements de vie. Pour les démographes, le recueil
d’éléments des parcours individuels possède une dimension explicative très
pertinente, qu’ils s’agissent de points d’inflexion, de ruptures au sein des
parcours biographiques ou d’éléments ponctuels sans conséquence à long
terme (Laborde et al., 2007). C’est ce que nous avons mis en place dans
l’enquête Information et Vie quotidienne (Guérin-Pace, 2009). Les
évènements marquants peuvent être recueillis de manière ouverte ou fermée.
L’intérêt de les recueillir, sous forme de question fermée, est de pouvoir

La question posée était « Quels sont les mots que vous associez le plus à l’ «
Europe » ? Choisissez 5 mots au maximum. »
2 Pour des raisons de coût et de délai, l’instruction donnée aux partenaires était
de traduire, eux-mêmes, en anglais les mots associés au moment de la saisie des
questionnaires. Les premiers traitements textuels ont permis de repérer des
incohérences et nécessité un retour vers les questionnaires dans leur langue d’origine.
1

JADT’ 18

361

effectuer des comparaisons systématiques dans la mesure où tous les
enquêtés répondent à une même question. Nous avons introduit dans
l’enquête sous forme de question fermée les évènements les plus
fréquemment cités (divorce ou séparation des parents, décès d’un proche,
problème de santé, etc.). Les événements recueillis de manière « fermée » ne
permettent pas d’aborder tous les thèmes notamment ceux portant sur des
sujets sensibles (cas de violence par exemple). Le recueil sous forme
d’énumération devient en effet vite intrusif, parfois déplacé, si les personnes
ne sont pas concernées. Par ailleurs, par cette démarche, on fait l’hypothèse
de la nature a priori traumatisante d’un événement sans savoir si Ego l’a
vécu comme tel durant son enfance (Laborde et al., 2007). Nous avons ainsi
fait le choix de compléter ce questionnement par la question ouverte suivante
« Avez-vous connu un autre événement marquant durant votre enfance ? Si
oui, lequel ? ». Près d’un quart des répondants déclarent un « autre
événement marquant » de leur enfance en réponse à cette question. Parmi
eux, un sur deux évoque un décès, un sur dix un événement lié à un
problème de santé, et dans les mêmes proportions une situation de violence
vécue durant l’enfance (Baril, Guérin-Pace, 2016).
Tableau 1 : Description des corpus analysés
Enquêtes
PopulationsEspaces de VieEnvironnement
(1992)
EuroBroadMap
(2009)
Information et Vie
Quotidienne (2011)

Corpus

Nombre de
réponses

Nombre
d’occurrences

Nombre de
mots
distincts

Environnement

4596

28716

2130

9343

40800

5111

3167

15993

2161

Mots associés à
l’Europe
Evènements
marquants de
l’enfance

2. Une étape sous-estimée : lecture des mots du corpus et les statistiques
lexicales
Une première étape essentielle d’analyse est la lecture du lexique des mots
les plus fréquents associé à un corpus d’enquête. Ce lexique donne à lui seul
un aperçu de la tonalité du vocabulaire (positive ou négative) et des registres
abordés. Par exemple, dans le corpus de mots associés à l’Europe, le premier
mot à connotation péjorative n'apparait qu'en 26ème position (colonialism). La
lecture des événements les plus fréquents indique quant à elle le caractère
individuel ou collectif, le plus souvent historique, des événements perçus.
Pour les enquêtes internationales ou à passage répété, le recours aux

362

JADT’ 18

statistiques lexicales permet de comparer la richesse du vocabulaire de
manière pertinente. Ainsi, dans le corpus « Europe », la comparaison des
proportions de mots distincts (Figure 1) apporte des informations
intéressantes. Il apparaît ainsi que les étudiants interrogés dans des pays les
plus éloignés de l’Union européenne (Cameroun, Chine, Russie, Brésil, Inde)
ont une vision plus consensuelle ou partagée de l’Europe que ceux des pays
qui en sont membres, ou à la marge.

Figure 1 : Diversité des mots associés à l’Europe selon les pays d’enquête
Source : Enquête EuroBroadMap (2009)

3. Faire émerger le contenu d’une question ouverte à partir du TLE
Une autre application des méthodes d’analyse textuelle à un corpus de
réponses à une question ouverte consiste à extraire les mondes lexicaux selon
la méthodologie Alceste. Une CDH effectuée sur le tableau croisant les
réponses à la question ouverte avec le lexique associé au mot
« environnement » met en évidence deux approches fondamentalement
différentes de la notion d’environnement (Figure 2). La première aborde
l’environnement selon une approche cognitive concernant un espace
physique et social (qualité de vie, univers local, etc.), tandis que la deuxième
approche est plus symbolique ou imaginaire (iconographie de la nature,
sensation de bien-être.).

JADT’ 18

363

Figure 2 : Les mondes lexicaux du corpus « environnement » (Alceste)
In Guérin-Pace F., 1997

4. Croiser les réponses spontanées avec un questionnement fermé
Les limites d’interprétation d’une question ouverte résident dans
l’impossibilité d’interpréter ce qui n’a pas été évoqué par les répondants.
Compléter ce dispositif par un questionnement fermé permet d’y remédier.
Nous avons ainsi, à la suite de la question ouverte, introduit deux questions
fermées qui proposaient une liste de mots et d’adjectifs pouvant être associés
ou non, par le répondant, au mot « environnement »3. L’observation conjointe

3 Les questions étaient libellées de la manière suivante : « Voici une liste de noms
(adjectifs). Lesquels vous semblent liés à la notion d’environnement ? (Pour chacun,
précisez oui ou non).

364

JADT’ 18

des réponses à ces deux modes de questionnement par une ACM sur le TLA
permet d’enrichir l’analyse du contenu « spontané » au regard des
représentations fermées.
On observe ainsi (Figure 3) que l’opposition entre un environnement fait de
« relations » et un environnement fait de « nature » (axe horizontal)
s’accompagne, par exemple, du choix ou du refus de mots et d’adjectifs qui
décrivent les nuisances urbaines. Sur l’axe vertical, à l’opposition entre un
environnement conçu comme une proximité immédiate et un environnement
basé sur les relations entre « l’homme et son milieu » correspond un
vocabulaire associé qui renforce cette perception. Proche de la première
perception, on relève les mots « maison-oui », « amical-oui », « sécurité-oui »
et « planète-Non ».

Figure 3 : Proximité entre formes du corpus « environnement » et associations proposées
Guérin-Pace F., Garnier B, 1995 Lecture : à proximité des mots « santé » ou « liberté » cités en
réponse à la question ouverte, on relève les réponses « non » à l’association du mot
environnement aux mots « ville » ou « violence ».

5. Post-coder les événements marquants de l’enfance par la statistique
textuelle
Une autre application plus récente de ces méthodes pour post-codifier des
réponses à une question ouverte peut sembler contradictoire avec l’esprit
même de la statistique textuelle. Il s’agit plus précisément de post-codifier les
évènements recueillis dans l’enquête Information et Vie quotidienne (IVQ).
Pour cela, nous avons effectué une classification (CDH) sur le tableau lexical

JADT’ 18

365

entier croisant les réponses à la question « Avez-vous vécu d’autres
événements marquants ? » avec le lexique du corpus. On retient une partition
en cinq classes au sein de laquelle on observe une première dichotomie entre
des événements de nature collective (guerre d’Algérie, Mai 1968, etc.) et un
ensemble de classes qui évoquent des événements de nature individuelle :
décès, maladie, accident et violence (Figure 4). Nous avons ajouté à ces cinq
classes deux classes supplémentaires : une classe intitulée « Refus »
regroupant toutes les réponses qui marquent une volonté de l’enquêté de ne
pas détailler l’événement marquant à l’enquêteur (tout en ayant donné une
réponse affirmative à la question « Avez-vous connu un autre événement
marquant ? ») ; une classe « Autre » au sein de laquelle nous avons regroupé
les réponses non classées4. Nous avons ensuite cherché à affiner cette
typologie en précisant les acteurs éventuels impliqués dans les événements.
Par exemple, au sein de la classe « Maladie » (classe 2), nous avons filtré au
moyen d’un vocabulaire familial (père, mère, frère, sœur, tante, ami, etc.) et
constitué 4 sous-modalités distinctes selon les personnes concernées.

Figure 4 : Typologie des événements marquants de l’enfance
Source : Enquête IVQ, Iramuteq (classification Méthode Reinert)

Nous avons procédé de la même manière pour la classe « violence » en
distinguant cette fois les personnes concernées par l’événement et son auteur
éventuel. Nous obtenons finalement une typologie construite sur les
questionnements ouverts et fermés, composée de 43 items (Baril, GuérinPace, 2016), qui pourrait être réutilisée pour d’autres enquêtes nationales.
En conclusion, ces différentes applications sur des corpus variés d’enquêtes

4

Près de 90 % des 3167 réponses à cette question sont classées.

366

JADT’ 18

de la statistique publique permettent de mettre en évidence la diversité des
apports des méthodes de la statistique textuelle. Aujourd’hui, de plus en plus
d’enquêtes nationales abordent des thématiques sensibles (violences,
précarité, illettrisme, etc.). Le recours à un questionnement ouvert s’avère
ainsi indispensable en permettant au chercheur d’objectiver sa démarche. Les
méthodes de la statistique textuelle se révèlent incontournables dans cette
perspective.
Références
Baril E., Guérin-Pace F. (2016). Compétences à l’écrit des adultes et
événements marquants de l’enfance : le traitement de l’enquête
Information et vie quotidienne à l’aide des méthodes de la statistique
textuelle, Economie et statistique, n°490, pp. 17-36.
Guérin-Pace F. (2009). Illettrismes et parcours individuels, Economie et
statistique, n°424-425.
Brennetot A., Emsellem K, Guérin-Pace F., Garnier B. (2013). Dire l’Europe à
travers le monde. Les mots des étudiants à travers l’enquête
EuroBraodMap, Cybergeo : European Journal of Geography.
Guérin-Pace F., Collomb P. (1998). Les contours du mot environnement :
Enseignements de la statistique textuelle, L’Espace Géographique, n°1, pp.
41-52.
Guérin-Pace F. (1997). La statistique textuelle : un outil exploratoire en
sciences sociales, Population, n°4, pp. 865-888.
Laborde, C., Lelièvre, E., Vivier, G. (2007). Trajectoires et événements
marquants, comment dire sa vie : Une analyse des faits et des perceptions
biographiques. Population, vol. 62,(3), pp. 567-585.
ssoc vol. 17: 507–13.
Scheurwegs E., Luyckx K., Luyten L., Goethals B. and Daelemans W. (2017).
Assigning Clinical Codes with Data-Driven Concept Representation on
Dutch Clinical Free Text. J Biomed Inform vol. 69: 118–27.
Vlug A. E., van der Lei J., Mosseveld B.M., van Wijk M.A., van der Linden
P.D., Sturkenboom M.C., and van Bemmel J.H. (1999). Postmarketing
Surveillance Based on Electronic Patient Records: The IPCI Project.
Methods Inf Med 38 (4/5): 339–44.
Wermter J. and Hahn U. (2004). Really, Is Medical Sublanguage That
Different? Experimental Counter-Evidence from Tagging Medical and
Newspaper Corpora. In Fieschi M., Coiera E. and Li Y.-C.L. Proc. of the
11th World Congress on Medical Informatics (MEDINFO 2004), pp. 560–64.

JADT’ 18

367

Annotation-based Digital Text Corpora Analysis
within the TXM Platform
Serge Heiden
Université de Lyon, ENS de Lyon, IHRIM – UMR5317, CNRS – slh@ens-lyon.fr

Abstract
This paper presents new developments in the TXM textual corpora analysis
platform
(http://textometrie.org)
towards
direct
text
annotation
functionalities. Some annotations are related to a web based external historic
ontology called SyMoGIH and others to co-reference information between
words or to word properties like part of speech or lemma.
The paper discusses the methodological stakes of unifying in a single
framework the production and the analysis those annotations with the
traditional ones already available in TXM corresponding to the XML markup
of the text sources and to the linguistic annotations automatically added to
texts by NLP tools.
Keywords: textometry, TXM, digital text representation, XML, TEI,
annotation, ontology, co-reference, part of speech, digital hermeneutic circle.
1. Introduction
TXM (Heiden, 2010) is a software platform offering textual corpora analysis
tools. It is delivered as a standard desktop application for Windows, Mac and
Linux and as a web portal server application (http://textometrie.org).
Its analysis tools combine qualitative types of tools like word lists,
concordancing or text edition navigation (close reading) with synthetic
quantitative types of tools like factorial analysis, clustering, keywords or
statistical co-occurrence analysis (distant reading).
To be able to work on texts, the platform imports first the corpus sources to
build a rich internal representation of texts through the following general
workflow:
a) first the “base text” of each text is established: this operation
implements “digital philology” principles and consists of decoding
information in the various formats of the source documents5 to

5
TXM can analyze three main types of corpora : corpora of written texts,
possibly including paginated editions including images of facsimiles ; record
transcriptions corpora, possibly time synchronized with the audio or video source ;

368

JADT’ 18

decide primarily where are the text limits, internal structures
boundaries and words and punctuations of the text. Its result is
represented in a pivot XML format especially designed for TXM
called “XML-TEI TXM” and extending the standard encoding
recommendations of the Text Encoding Initiative consortium (TEI
Consortium, 2017) ;
b) then, natural language processing (NLP) tools are optionally applied
to the base text to automatically add linguistic information like
sentence boundaries, grammatical category (pos = part of speech)
and lemma of words by eg TreeTagger (Schmid, 1994), etc. As NLP
tools generally don’t take XML format as input, the pivot
representation is first converted to raw text for NLP processing and
results are added back into the XML-TEI TXM representation ;
c) finally a specialized representation of texts is built into TXM for
efficient execution of its tools (by indexing for search engines and
text edition rendering).
From the point of view of TXM, NLP tools results in b) are seen as automatic
annotations added to the initial XML-TEI TXM representation of texts built in
a), and the XML tags of the initial XML-TEI TXM representation in a) can be
seen as manual annotations added to the base text (or raw text), typically
philologically edited with the help of specialized XML editors (like Oxygen
XML Editor6) outside of TXM when the source is in XML format, or as
automatic annotations added by TXM when converting from some other
format into XML-TEI TXM. All TXM tools apply indiscriminately to all types
of annotation regardless of their origin (automatic or manual).
Thus, TXM implements a traditional workflow combining a “text source
encoding and annotation” step to an “application of analysis tools to
annotated texts” step. The text analysis tools use text annotations (for
example word pos) to offer their services and produce their results (for
example the concordance of all infinitive verbs). The workflow is
unidirectional and the whole of it must be passed through again completely
if any annotation needs to be corrected. To add or correct annotations, the
user has to edit the sources or the annotations outside of TXM. For example
word properties can be exported from the XML-TEI TXM representation,
edited in a spreadsheet and inserted back into the texts before re-import7.

and parallel multilingual corpora aligned at the level of a textual structure such as the
sentence or the paragraph.
6
https://www.oxygenxml.com
7
see
for
example
this
tutorial
based
on
TXM
macros:
https://groupes.renater.fr/wiki/txm-users/public/tutoriel_correction_mots.

JADT’ 18

369

This paper introduces new services developed in TXM to annotate directly
texts from within the results view of specific tools for a better integration of
philological and analytic work.
2. Annotation services in TXM
The new annotation services concern both adding and correcting information
and all the annotations edited are meant for further exploitation by usual
TXM tools.
2.1. SyMoGIH annotation by concordance
The first new service, developed in partnership with the LARHRA research
laboratory in history8, is based on the annotation of concordance pivots: any
sequence of words composing the pivots can be annotated with any semantic
category9 coming from the SyMoGIH10 historical ontology framework
(Beretta, 2015). In this architecture, the SyMoGIH web platform hosts the
ontology of historic facts and knowledge, and concordances provide the user
interface to link identifiers of those data to text spans for further analysis. As
an illustration, see figure 1 the annotation of the “Faculté de droit d’Aix”
entity (of id CoAc13562) in unverified OCRed texts of the “Bulletin
administratif de l'Instruction publique" corpus11.
TXM internal management of those annotations is equivalent to a re-import
of the current pivot representation of the annotated texts. After re-import
(after saving annotations) the new annotations are available for all TXM tools
to work on like any original “annotation” of the texts (internal structures and
their properties, word properties, etc.).
2.2. URS annotation in text edition
The second new service is based on manual annotation of word sequences
inside text editions with elements of a Unit-Relation-Schema (URS)
annotation model. URS type annotations are designed to encode discourse
entities like co-reference chains in texts (Schnedecker, Glikman, & Landragin,
2017). In a URS model, Units or entities have any number of properties and
can be linked together by the two other annotation types: Relations, having
any number of properties (1-to-1 relation type), and Schemas, having any

http://larhra.ish-lyon.cnrs.fr
pivots can also optionally be annotated with simple keywords or with keyvalue pairs, managed by TXM in a local repository.
10
http://symogih.org/?lang=en
11
see the Bibliothèque historique de l'éducation (BHE) project:
http://www.persee.fr/collection/bhe
8
9

370

JADT’ 18

number of properties (1-to-n relation type). Any types and properties of
units, schemas, and relationships are definable in the annotation model
before and during annotation. The types and properties are chosen by the
user, they are not limited to co-reference chains.

Figure 1: TXM screenshot of a Concordance of a “Faculté de droit d’Aix” word sequence
pattern to annotate (top) and of browsing SyMoGIH semantic categories to use for the
annotation (bottom).

The original URS model has been designed and developed in the Glozz
(Widlöcher & Mathet, 2009) and Analec (Landragin, Poibeau, & Victorri,
2012) software. It is being integrated into TXM through the text edition
reading tool for a project funded by the French National Research Agency
(ANR) called DEMOCRAT12.
As an illustration, see figure 2 the annotation of the “ses loix” word sequence

12
http://www.agence-nationale-recherche.fr/en/anr-fundedproject/?tx_lwmsuivibilan_pi2%5BCODE%5D=ANR-15-CE38-0008

JADT’ 18

371

with a unit of type MENTION, of “GN.POS” grammatical category and “les
lois de la divinité” referent, in the first chapter of the 1755 edition of De
l'esprit des lois by Montesquieu. TXM internal management of those
annotations can be represented as new XML-TEI stand-off annotations
anchored to the word elements of the XML-TEI TXM representation of texts
(Grobol, Landragin, & Heiden, 2017).

Figure 2: TXM screenshot of the edition of the first page of De l'esprit des lois with units of
type MENTION highlighted in yellow and the selected unit in bold (top) and the current
values of the properties of the selected unit (bottom).

2.3. Word properties annotation by concordance
The third service will be based on the annotation of concordance pivot
words: a word present in the pivots of a concordance will be able to be
annotated with properties. The primary goal of that service is to annotate and
correct grammatical properties and lemma of word elements of the XML-TEI
TXM representation of texts. This development is done for a project cofunded by the ANR and Deutsche Forschungsgemeinschaft (DFG) called
PaLaFra13 <http://palafra.org>.
2.4. Editing XML sources
Finally we are developing the possibility to directly edit the XML sources

13
http://www.agence-nationale-recherche.fr/en/anr-fundedproject/?tx_lwmsuivibilan_pi2%5BCODE%5D=ANR-14-FRAL-0006

372

JADT’ 18

from within TXM through an internal XML editor. This editor will eventually
be accessed through TXM tools as a “back to source” operation similar to the
current “back to text” operation (for example from a concordance line to a
text edition page).
3. Discussion
By using a common XML-TEI pivot representation for internal management
of corpora for all the annotation services, TXM unifies transcription and
annotation activities in a single framework. In this framework, annotations
represent manual (user), semi-automatic (machine+user) or automatic
(machine) interpretation results used further for analysis and interpretation
work. The reflexive nature of the resulting text analysis workflow is
schematized in figure 3. Texts are first digitized by OCR, transcribed or
converted from digital formats. They are then philologically corrected and
established through XML-TEI manual encoding. Then automatically
processed by NLP tools while being imported into TXM to produce the TXM
internal corpus model. Corpus analysis is then assisted by TXM tools applied
to the corpus model. The pivot representation that gathers all annotations
produced by annotation tools is figured as the node labeled « Pivot rep. » and
the interpretation workflow itself is figured as a digital hermeneutic circle.

Figure 3: Digital hermeneutic circle integration into TXM.

JADT’ 18

373

Legend:
- red box = automatic annotation activity - black box = tool
- blue box = manual annotation activity - green box = TXM corpus
data model
- purple disk = data representation
- black arrow =
activity
- green arrow = annotation equivalence
4. Conclusion
All the new annotation services integrated into TXM are building a
comprehensive annotation-based digital text corpora analysis platform. From
an epistemological point of view, the integration of different annotation
models and tools into the platform should help its users to better define what
comes from the source corpus they analyze and what comes from their own
or from others interpretation work.
This work was funded by the ANR and the DFG under grant numbers ANR15-CE38-0008 (DEMOCRAT project) and ANR-14-FRAL-0006 (PaLaFra
project).
References
Beretta, F. (2015). Publishing and sharing historical data on the semantic
web : the SyMoGIH project – symogih.org. Presented at the Workshop:
Semantic Web Applications in the Humanities. Retrieved from
https://halshs.archives-ouvertes.fr/halshs-01136533
Grobol, L., Landragin, F., & Heiden, S. (2017). Interoperable annotation of
(co)references in the Democrat project. Presented at the Thirteenth Joint
ISO-ACL Workshop on Interoperable Semantic Annotation. Retrieved
from https://hal.archives-ouvertes.fr/hal-01583527/document
Heiden, S. (2010). The TXM Platform: Building Open-Source Textual Analysis
Software Compatible with the TEI Encoding Scheme. In K. I. Ryo Otoguro
(Ed.), 24th Pacific Asia Conference on Language, Information and Computation
(pp. 389–398). Institute for Digital Enhancement of Cognitive
Development, Waseda University. Retrieved from http://halshs.archivesouvertes.fr/halshs-00549764/en/
Landragin, F., Poibeau, T., & Victorri, B. (2012). ANALEC: a New Tool for the
Dynamic Annotation of Textual Data (pp. 357–362). Presented at the
International Conference on Language Resources and Evaluation (LREC
2012).
Retrieved
from
https://halshs.archives-ouvertes.fr/halshs00698971/document
Schmid, H. (1994). Probabilistic Part-Of-Speech Tagging Using Decision

374

JADT’ 18

Trees. In Proceedings of the International Conference on New Methods in
Language Processing (Vol. 12).
Schnedecker, C., Glikman, J., & Landragin, F. (2017). Les chaînes de
référence : annotation, application et questions théoriques. Langue
française, (195), 5–16. https://doi.org/10.3917/lf.195.0005
TEI Consortium. (2017). TEI P5: Guidelines for Electronic Text Encoding and
Interchange. TEI Consortium. Retrieved from http://www.teic.org/Guidelines/P5
Widlöcher, A., & Mathet, Y. (2009). La plate-forme Glozz: environnement
d’annotation et d’exploration de corpus. In Actes de la 16e Conférence
Traitement Automatique des Langues Naturelles (TALN’09), session posters (p.
10). Senlis, France, France. Retrieved from https://hal.archivesouvertes.fr/hal-01011969

JADT’ 18

375

Quantifying Translation : an analysis of the
conditional perfect in English-French comparableparallel corpus
Daniel Henkel
Université Paris 8 Vincennes St-Denis – dhenkel@univ-paris8.fr

Abstract
The frequency of the conditional perfect in English and French was observed
in an 8-million-word corpus consisting of four 2-million-word comparable
and parallel subcorpora, tagged by POS and lemma, and analyzed using
regular expressions Intra-linguistically the Wilcoxon-Mann-Whitney test was
used to compare authors and translators. Frequencies in source and target
texts were evaluated using Spearman's correlation test to identify interlinguistic influences. Overall, the past conditional in English was found to
have a stronger influence in the translation process.
Résumé
La fréquence du conditionnel parfait en anglais et en français a été observée
dans un corpus de 8 millions de mots comprenant quatre sous-corpus
comparables et parallèles de 2 millions de mots chacun, étiquetés par
catégorie grammaticale et par lemme, et analysés par expressions rationnelles
(regex). Le test de Wilcoxon-Mann-Whitney a servi pour comparer les
auteurs et traducteurs, tandis que la corrélation entre textes-sources et -cibles
a été évaluée au moyen du coefficient de corrélation de Spearman.
Globalement, l'influence du conditionnel parfait en anglais sur le processus
traductionnel paraît plus sensible.
Keywords: corpus, translation, regular expressions, statistical analysis,
Wilcoxon-Mann-Whitney, Spearman, conditional perfect
1. Introduction
Since Corpus-based Translation Studies (CBTS) first began to gain
momentum around the turn of the 21st century, differences have consistently
been shown between corpora of translated English, French and other
languages in comparison with untranslated reference corpora in the same
languages. The hybrid nature of translated texts is now thus widely

376

JADT’ 18

acknowledged as an established fact among specialists1 in the field so much
so that any further proof might seem superfluous. These studies have
focused on phenomena such as the use of 'that' to introduce subordinate
clauses (Olohan & Baker, 2000), contractions (Olohan, 2003), manner-ofmotion verbs (Cappelle, 2012), existential predications (Loock & Cappelle,
2013) most often in terms of their overall frequency2. Such comparisons have
provided valuable insights about the languages involved and the translation
process. Little consideration has been given so far, however, to the fact that
each language-system consists of many individual styles or idiolects which
gravitate around a common center, but individually exhibit widely differing
characteristics. In other words, while the variation from one author or
translator to another is inherent in the very nature of corpus linguistics, this
dimension remains absent from the equation in many, if not most, corpusbased translation analyses.
2. Methods
Two important terminological distinctions must be made at the outset. The
first is between ex nihilo, a.k.a. 'original', English (En0) and French (Fr0), i.e.
discourse in each language produced independently of any known prior
influence, as opposed to English-translated-from-French (EtrF) and Frenchtranslated-from-English (FtrE), which will be used to refer to translations into
each language, based on a pre-existing work in the other language, and
therefore potentially subject to inter-linguisitic influences. The second
distinction is between two sorts of bilingual corpora, 'comparable' and
'parallel'. In keeping with the clarification offered by McEnery & Xiao (2007),
the term 'comparable corpus' will hereafter refer to a bilingual corpus
consisting of two subcorpora of ex nihilo English and French texts, which are
therefore not translations of one another, but which share a certain number of
common characteristics, whereas the term 'parallel corpus' will designate a

Albeit with some divergence of opinion as to whether such differences are
best interpreted as evidence of source-language interference or as consequences of the
translation process regardless of the source-language, i.e. characteristics inherent in
the 'third code' or 'translationese' (cf. Koppel & Ordan, 2011).
2
Olohan (2002) apparently subscribes to Stubbs' (2001) view that “corpus
linguistics […] investigates relations between frequency and typicality, and instance and
norm. It aims at a theory of the typical,” (while nonetheless encouraging investigation of
individual translators' styles in her conclusion), and the predominance of this
approach is confirmed again over a decade later by Loock (2013) who observes that
“many studies within the CBTS framework still solely rely on overall quantitative analyses to
establish differences between original and translated languages.”
1

JADT’ 18

377

corpus made up of one sub-corpus of ex nihilo works in a source-language
and another sub-corpus consisting of the translations of those same works
into the target-language.
The corpora used in this study were compiled from public domain works
available in electronic format (.epub, .mobi, .html or .txt), the translations of
which were also available in electronic format via publicly available sources
(primarily Project Gutenberg). Common criteria3 based on size and date were
then used to select 20 works by 20 different authors in En0 and the same
number in Fr0, so as to obtain, first of all, two reference sub-corpora
comparable in terms of date, size, discourse type and diversity:
Table 1 Summary of characteristics for comparable En0 and Fr0 subcorpora.
Subcorpus 1 En0 (n=20)

Subcorpus 2 Fr0 (n=20)

Wordcounts4

Max. 199,976
(Collins,
The
Moonstone)
Min. 59,771
(Mansfield,
The
Garden-party)
Median 99,558 (Wells, The War in
the Air)
Total 2,114,517

Max. 192,521 (Zola, Les trois villes
Paris)
Min. 62,539
(Rolland,
Les
précurseurs)
Median 90,873 (Leroux, La chambre
jaune)
Total 2,083,787

Dates

Max. 1928 (Woolf, Orlando)
Min. 1868 (Collins, The Moonstone)
Median 1901 (Kipling, Kim)

Max. 1921 (Leblanc, Les dents du
tigre)
Min. 1866 (Gaboriau, L'affaire
Lerouge)
Median 1901 (Bazin, Les Oberlé)

The translations of these works were then compiled into two sub-corpora of
EtrF and FtrE, so as to produce an 8m-word 'super-corpus' consisting of four
2m-word sub-corpora, designed to be both comparable and parallel and
thereby provide a basis for three types of comparisons:
– between En0 and Fr0, in order to establish benchmark data for each
language,
– between EtrF and En0, so as to ascertain whether the linguistic indicator

Whenever several works by the same author were available, preference was
given either to the most recent or the one with the highest word-count. In general date
was given precedence over size, except in cases where a major difference in wordcount was found between works published within a relatively close interval.
4
Word-counts were estimated using the text editor Geany, after replacing
punctuation with whitespaces, given that punctuation has been found to artificially
inflate word-counts in French as compared to English.
3

378

JADT’ 18

under investigation, i.e. the conditional perfect, has a similar distribution in
EtrF compared to En0, and likewise for FtrE in comparison with Fr0,
– between source- and target-texts, to determine whether correlations exist
between the parallel subcorpora (i.e. EtrF~Fr0 and FtrE~En0) which could be
taken as evidence of interlinguistic interference.
All of the texts were cleaned of metatext, tagged for POS and Lemma in
TreeTagger, and interrogated in TextSTAT using the following regular
expressions to target the conditional perfect.
English (all verbs):
d) (((w|c|sh)ould)|('d)|(might)|(ought))(e?st)?/\S+(
\S+/RB[RS]?/\S+)*(
to/\S+)?( ((ha|')ve|of)/\S+)( \S+/RB[RS]?/\S+)* \S+/V[BHV][ND]/
French (verbs taking AVOIR as an auxiliary, verbs taking ÊTRE, reflexive
constructions):
e) \S+/VER:cond/avoir( \S+/ADV/\S+)* \S+/VER:pper
f) \S+/VER:cond/être(
\S+/ADV/\S+)*
\S+/VER:pper/(r[eé])?(aller|(ad|de|inter|par|pro|sur)?venir|rester|deme
urer|(ap|dis)?paraître|naître|mourir|décéder|arriver|partir|tomber|mo
nter|descendre|passer|rentrer|retourner|sortir)
g) ((je/\S+(
\S+/ADV/\S+)*
m[e']/\S+)|(tu/\S+(
\S+/ADV/\S+)*
t[e']/\S+)|(nous/\S+(
\S+/ADV/\S+)*
nous/\S+)|(vous/\S+(
\S+/ADV/\S+)* vous/\S+)|(s[e']/\S+))( en|y/\S+)* \S+/VER:cond/être(
\S+/ADV/\S+)* \S+/VER:pper/
The results obtained from these queries were converted into frequencies per
1000 words (freq./1k) for each author or translator and analyzed using the
Wilcoxon-Mann-Whitney and Spearman tests as described in the following
section.
3. Results and analysis
The data collected from each of the subcorpora are presented in the following
tables and summarized in Fig. 1.
Table 2a Conditional perfect frequencies in En0
Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Buchan

139

102022

1.36

Lewis

58

83799

0.69

Burnett

78

84093

0.93

London

57

100816

0.57

Collins

326

199976

1.63

Mansfield

67

59771

1.12

ConanDoyle

108

105040

1.03

Reid

200

94254

2.12

Cox

142

114352

1.24

Stevenson

81

70366

1.15

Eliot

319

164456

1.94

Stoker

127

161255

0.79

JADT’ 18

379

Hardy

254

153076

1.66

Wallace

135

101948

1.32

Hope

115

83189

1.38

Wells

54

99558

0.54

Joyce

26

69225

0.38

Wilde

76

79412

0.96

Kipling

109

107601

1.01

Woolf

76

80308

0.95

max: 2.12, min: 0.38, median: 1.03
Table 2b Conditional perfect frequencies in EtrF
Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Tr.Barbusse

48

116179

0.41

Tr.Leroux

127

74920

1.7

Tr.Bazin

74

76312

0.97

Tr.Loti

15

65837

0.23

Tr.Benoît

41

64301

0.64

Tr.Massenet

42

57736

0.73

Tr.Flaubert

125

175678

0.71

Tr.Maupassant

45

76070

0.59

Tr.France

66

76830

0.86

Tr.Mirbeau

76

101959

0.75

Tr.Gaboriau

335

170870

1.96

Tr.Proust

408

198721

2.05

Tr.Gourmont

76

69399

1.1

Tr.Rolland

27

65872

0.41

Tr.Hugo

104

125428

0.83

Tr.Vanderem

80

95884

0.83

Tr.Huysmans

46

130181

0.35

Tr.Verne

89

63760

1.4

Tr.Leblanc

112

128493

0.87

Tr.Zola

179

205503

0.87

max: 2.05, min: 0.23, median: 0.83
Table 2c Conditional perfect frequencies in Fr0
Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Barbusse

47

114877

0.41

Leroux

78

90873

0.86

Bazin

41

78395

0.52

Loti

15

72386

0.21

Benoît

33

67915

0.49

Massenet

45

76711

0.59

Flaubert

108

149808

0.72

Maupassant

46

75598

0.61

France

20

71998

0.28

Mirbeau

59

117035

0.5

Gaboriau

53

120464

0.44

Proust

296

170105

1.74

Gourmont

60

73000

0.82

Rolland

11

62539

0.18

Hugo

18

118095

0.15

Vanderem

44

91476

0.48

Huysmans

22

132824

0.17

Verne

50

76890

0.65

Leblanc

47

130277

0.36

Zola

141

192521

0.73

max: 1.74, min: 0.15, median: 0.5

380

JADT’ 18
Table 2d Conditional perfect frequencies in FtrE
Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Tr.Buchan

69

105082

0.66

Tr.Burnett

74

80743

Tr.Collins

138

Tr.ConanDoyle

Cond.Pf.
(n=)

Words
(n=)

Freq./1k

Tr.Lewis

80

96211

0.83

0.83

Tr.London

49

86378

0.57

198988

0.69

Tr.Mansfield

82

68674

1.19

119

117280

1.01

Tr.Reid

120

93025

1.29

Tr.Cox

194

130967

1.48

Tr.Stevenson

64

76757

0.83

Tr.Eliot

120

168125

0.71

Tr.Stoker

167

176623

0.95

Tr.Hardy

217

151435

1.43

Tr.Wallace

97

87316

1.11

Tr.Hope

99

82966

1.19

Tr.Wells

74

108529

0.68

Tr.Joyce

49

72739

0.67

Tr.Wilde

63

82430

0.76

Tr.Kipling

68

124885

0.54

Tr.Woolf

56

87475

0.64

max: 1.48, min: 0.54, median: 0.83

Fig. 1 Distributions of conditional perfect frequencies in En0, EtrF, FtrE and Fr0.

As is readily apparent from Fig. 1, the conditional perfect is used more
frequently in En0 than in Fr0, which, aside from one extreme outlier (Proust),
is situated below the 1st quartile of En0. EtrF and FtrE (as usual) occupy an

JADT’ 18

381

intermediate zone, with practically identical medians (0.83) which are both
inferior to Q1 in En0 and superior to Q3 in Fr0. The most striking difference
is between authors in Fr0 and translators, who use the conditional perfect
almost twice as often in FtrE. As a result, the entire distribution in FtrE is
superior to the median for Fr0, with 75% of FtrE (Q2-Q4) in the same range as
the top quartile (Q4) of Fr0. Wilcoxon-Mann-Whitney confirms that a similar
disparity could hardly occur by chance (U=337, n1=n2=20, p=0.0002) and that
it is therefore reasonable to infer that – notwithstanding the considerable
amount of variation that can be observed from one author or translator to
another – FtrE and Fr0 are clearly different with respect to their use of the
conditional perfect. Between EtrF and En0, however, the difference is less
obvious. Although the interquartile range for EtrF (0.63-1) is noticeably lower
than in En0 (0.9-1.37), there is nonetheless a great deal of overlap between the
two distributions, and Wilcoxon-Mann-Whitney (U=135, n1=n2=20, p=0.08)
indicates that the risk of error is too great to say with confidence whether any
substantial difference exists between EtrF and En0 in their use of the
conditional perfect.
To what extent such differences may be attributed to the influence of the
analogous forms in the source-texts can be assessed statistically as illustrated
in Fig. 2a and 2b:

Fig. 2a Frequency of conditional perfect forms
in FtrE vs. En0. (ρ=0.47, p=0.036)

Fig. 2b Frequency of conditional perfect forms in
EtrF vs. Fr0. (ρ=0.57, p=0.009)

In both cases, Spearman's5 correlation test reveals a statistically significant
(p<0.05) positive correlation (ρ=0.57 for EtrF/Fr0, ρ=0.47 for FtrE/En0) of
moderate strength, which somewhat unexpectedly obtains a higher score for

5
Spearman's was preferred due to the presence of outliers. Pearson's R yields
an almost identical result for FtrE/En0, and a somewhat stronger coefficient (r=0.67)
for EtrF/Fr0, with similar p-values in both cases.

382

JADT’ 18

EtrF/Fr0. These correlations of similar strength suggest an intuitively
plausible tendency to translate individual instances of the conditional perfect
in one language by the analogous form in the other language in both
directions and in roughly similar proportions (although this remains to be
verified by manual examination of translation segments). Such a hypothesis
would help to explain why the medians and interquartile ranges observed in
EtrF and FtrE occupy a middle zone between En0 and Fr0, but it does little to
account for the greater disparity between FtrE and Fr0 as opposed to EtrF
and En0. Other contextual parameters may well be involved, or perhaps the
higher frequency of the conditional past in En0 exerts a sort of subliminal
effect on translators, who then use it more freely in FtrE with or without a
syntactic counterpart in the corresponding En0 segment.
4. Conclusion
These findings demonstrate how quantitative analysis of translated parallel
corpora in comparison with untranslated comparable corpora, can be used
both to identify disparities between target-texts and the target-language as
represented in an ex nihilo corpus, and to assess the influence of the sourcetexts on the target-texts. Such relationships are often asymmetrical: in this
case the correlation between the original French conditional perfect and the
translations into EtrF is stronger, while the higher frequency of conditional
perfect forms in English, though less strongly correlated on a text-to-text
basis, nonetheless fosters a style of French-translated-from-English which is
markedly different from ex nihilo French. While the exact mechanisms
involved will require further investigation, the conditional perfect in English
appears to exert a stronger influence in the translation process than the
corresponding form in French.
References
Hu K. (2016). Introducing corpus-based translation studies. Springer.
Koppel M and Ordan N. (2011). Translationese and Its Dialects Proceedings of
the 49th Annual Meeting of the Association for Computational Linguistics,
pp. 1318–1326, June 19-24, 2011
Kruger A., Wallmach, K. and Munday J. (Eds.). (2011). Corpus-based
translation studies: Research and applications. Bloomsbury Publishing.
Loock R. (2013). Close encounters of the third code. In Lefer M.A. and
Vogeleer S., eds, Interference and normalization in genre-controlled
multilingual corpora, Belgian Journal of Linguistics 27: 61-86
Olohan M. (2002). Comparable corpora in translation research. In LREC
Language Resources in Translation Work and Research Workshop Proceedings
pp. 5-9.

JADT’ 18

383

Zanettin F. (2013). Corpus methods for descriptive translation studies.
Procedia-Social and Behavioral Sciences, 95, 20-32.
Hüning Matthias. TextSTAT 2.9c © 2000/2014 Niederländische Philologie,
Freie
Universität
Berlin,
http://neon.niederlandistik.fuberlin.de/en/textstat/
R Core Team (2017). R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria URL
https://www.R-project.org/.
Schmid H.
TreeTagger,
Universitaet
Stuttgart,
http://www.cis.unimuenchen.de/~schmid/tools/TreeTagger/

384

JADT’ 18

Extraction of lexical repetitive expressions from
complete works of William Shakespeare
Daniel Devatman Hromada
Univesität der Künste, Berlin, Germany – daniel at udk dash berlin dot de

Abstract
Rhetoric tradition has canonized dozens of repetition-involving figures of
speech. Our article shows a way how hitherto ignored repetition-involving
schemata can be identified by means of translation of so-called “entangled
numbers” into backreferencing regular expressions. Each regex is
subsequently exposed to all utterances in all works of William Shakespeare,
allowing us to pinpoint 3367 instances of 172 distinct repetitive schemata.
Keywords: rhetoric stylometry, figures of speech, repetition, chiasm,
entangled numbers, regular expressions, William Shakespeare, non-zipfian
distribution
Résumé
On montre, comment peut-on identifier les figures de styles jusqu'ici
inconnues. Le but en question est atteint grâce au fait qu'on peut concevoir
un certain groupe de figures de style tel un nombre ayant quelques
propriétés particulières. Une fois découverte et énumérés, on peut transcrire
ces nombres en expressions régulières qui peuvent ensuite être éxposé à un
corpus textuel. Dans le cas de notre étude préliminaire, il s'agissait du corpus
de William Shakespeare.
Mots clés: stylométrie rhétorique, rfigures de style, répétition, chiasme,
expressions régulières, répetition, William Shakespeare
1. Introduction
Masterpieces of litterature and drama abound with repetitions. Rhetorics
abounds with repetitions, succesful oratories abound with repetitions. Many
a schema and a figure exists which exploits repetition : e.g. a polysyndeton
and an anaphore, an anadiplose and an epistrophe, a symploche and an
antanaclasis, paranomasis and an antimetabole. And alliterations and
paregmenons, and polypoptons, epizeuxiae or even a good old psittacism ?
Many are such schemata, many are such figures. Woe to the one who
thinking he knows them all !

JADT’ 18

385

Our article presents a way of enumerating of many a new schemata
involving one or more repetition of one or more lexical signifiants. The
procedure starts with a theoretical insight, that at least certain subset of the
set of all such schemata, is easily enumerable. This insight is subsequently
transcribed into an algorithm enumerating natural numbers which satisfy
following properties. These numbers once identified, they are to be translated
into Perl Compatible Regular Expressions exploiting some back-references
and negative lookaheads.
1.1. Computational rhetorics and its roots
In literature studies it is fairly common to speak about so-called "rhyme
schemes" like AAAA for monorhymes, ABAB for alternate rhyme, ABBA for
enclosed rhymes etc.
It is therefore barely surprising that analogic formalisms - that is, formalisms
that involve alphabetic indices - have been adopted by scholars aiming to
formalize a subgroup of rhetoric figures, known as the group of schemes. For
example (Harris et DiMarco, 2009) use a following formalism:
[W ]a::: [W ]b::: [W ]b::: [W ]a
to denote the rhetoric figure known as antimetabole. Subsequent studies in
automatized chiasm identification pursue a similiar route and often use
formulae like ABXBA, ABCBA, ABCXCBA to denote schemata
corresponding to utterances such as: "Drake love loons. Loons love Drake.", "All
as one. One as all." (Hromada, 2011) or "In prehistoric times women resembled
men, and men resembled women." (Dubremetz & Nivre, 2015) .
Table 1: 14 lowest E-numbers, their corresponding alphabetic representations and some
corresponding Shakespearean expressions .

E-number

Alphabetic

Example expression

11

AA

"we split we split " 1

111

AAA

"we split we split we split "

1111

AAAA

"justice justice justice justice "

1122

AABB

"gross gross fat fat "

1212

ABAB

"to prayers to prayers "

1
Note that sometimes one single word is attributed the role of a distinct
« brick » , sometimes a concatenation of two or even more words assumes such a role.
As will be indicated in sections two and three, this behaviour is not a bug, but an
anticipated property of our method.

386

JADT’ 18

1221

ABBA

"my hearts cheerly cheerly my hearts "

11111

AAAAA

"so so so so so "

11122

AAABB

"great great great pompey pompey "

11212

AABAB

"come come buy come buy "

11221

AABBA

"high day high day freedom freedom high day "

11222

AABBB

"o night o night alack alack alack "

12112

ABAAB

"too vain too too vain "

12121

ABABA

"come hither come hither come "

12122

ABABB

"come buy come buy buy "

1.2. Entangled numbers
The set of entangled numbers (or E-numbers) is a subset of a set of natural
numbers (i.e. integers). Entangled numbers are defined as “words of length n
over an alphabet of size 9 that are in standard order and which have the property that
every letter that appears in the word is repeated. “ (OEIS, 2016)
Note that the term word, as used in the preceding, as well as in following
citations, is used in mathematician's sense, meaning something as « sequence
of symbols » : “A word is in "standard order" if it has the property that whenever a
letter i appears, the letter i-1 has already appeared in the word. This implies that all
words begin with the letter 1.” (Arndt et Sloane, 2016). Hence, numbers like 22
or 33 are not entangled numbers because they are not in “standard order”
and numbers like “12” or “121 ”are not entangled because some (or all) of
their digits are not repeated. Fourteen smallest (i.e. with the lowest numeric
value) entangled numbers and their corresponding alphabetic transcriptions
are enumerated in Table 1.
Given that entangled numbers are natural numbers, they can be easily
enumerated by an incremental algorithm starting at one and iterating
towards infinity. Once enumerated (OEIS, 2016), we can bridge the realm of
numbers with the realm of text and apply our method.
2. Method
The core idea behind our method can be stated as follows:
Any E-number can be "translated" into a backreference-endowed regular
expression.
Concretely speaking, every digit of an E- number can be interpreted as an
element or a "brick". In this article, we work only with one type of bricks,
those corresponding to sequences which are between two to twenty-three

JADT’ 18

387

characters long 2. Such sequences can correspond to one or multiple lexical
units. A first occurence of a novel brick can be represented as a PERLcompatible regular expression (Friedl, 2002 ; Aho, 2014):
(.{2,23})
However, any subsequent repeated occurence of a digit in an E- number is
interpreted not as an occurence of the new brick, but rather as a
backreference to the brick which was already denoted by the same digit. The
very first E- number 11 is therefore NOT to be translated into regex /(.{2,23})
(.{2,23})/. For this would imply existence of two distinct bricks. Rather, the Enumber 11 is to be translated into regex:
(.{2,23}) \1
wherein the expression \1 denotes the backreference to the content matched
by the regex-brick specified in first parentheses, i.e. brick no.1 .
Hence, the E-number 111 can be easily translated into a regex /(.{2,23}) \1 \1/,
1111 into a regex /(.{2,23}) \1 \1 \1/ etc.
What's more, when we combine the backreference with a negative lookahead
operator – traditionally expressed by the formula (?!) - we can make sure that
a so-called non-identity principle is also satisfied. That is :
"Each distinct digit corresponds to distinct content"
For example, by translating the E-number 121 into the regex
(.{2,23}) (?!\1)(.{2,23}) \1
we can make sure that the content matched by the brick denoted by digit 2
shall be different from the content matched by the brick denoted by digit 1.
Thus, a phrase "no no no" shall not be matched by such a regex while an
expression "no yes no" shall.
Going somewhat further, an E-number 12321 - which could be understood as
an instance of chiasm or antimetabole ABXBA - is to be translated into regex
(.{2,23}) (?!\1)(.{2,23}) (?!\1\2)(.{2,23}) \2 \1
whereby the disjunctive backreference contained in the negative lookahead

2

These are the only variable parameters of our method.

388

JADT’ 18

(?!\1\2) assures that the content matched brick no.3 - corresponing to filler X
- shall be different from content matched by the brick representing digit 1 as
well as the brick representing digit 2.
3. Corpus & Processing
A digital, unicode-encoded version of Craig's edition of "Complete works of
William Shakespeare" has been downloaded from a publicly available
Internet source3 . This corpus contains 17 txt files stored in the sub-folder
"comedies", 10 txt files stored in the sub-folder "tragedies" and 10 txt files
stored in the sub-folder "historical".
Texts were subsequently split into utterances by interpreting closing tags
(e.g. </PERSONA>, </MIRANDA> etc.) as utterance separator. Even more
concretely, one can simply consider the slash symbol / to be the utterance
separator.
Only two further text-processing steps have been executed during the
initialization phase of the experiment hereby presented. Primo, content of
each utterance has been put into lowercase. Secundo, non-alphabetic symbols
(e.g. dot, comma, exclamation mark etc.) have been replaced by blank spaces.
We are aware that such replacement could potentially lead to certain amount
of loss of prosody- or pathos- encoding information. However, we consider
this step as legitimate because the objective of our experiment was to focus
on repetition of lexical units4.
Pre-processing code once executed, identification of expressions containing
diverse types of lexical repetitions is as simple as matching each
Shakespearean utterance with each regex.
4. Results
All in all, 3667 instances of a repetitive expressions have been detected in
Shakespeare's complete works. These were contained in 2295 distinct
utterances and corresponded to 172 distinct schemata. Among these, 71
matched more than one instance: these schemata could thus potentially
correspond to a certain cognitive pattern or a habitus in Shakespeare's mind.
Table 2 contains summary information concerning 23 schemata matching at
least five distinct utterances.

3

http://www.lexically.net/downloads/corpus_linguistics/ShakespearePlaysPlus.zip
4
Regexes matching repetitions of phonotactic clusters, syllables, or phrases,
are also possible. We prefer, however, not to focus on this topic within the limited
scope of this conference proposal.

JADT’ 18

389

Table 2: Repetitive schemata matching at least 23 distinct utterances present in collected
works of William Shakespeare.
Instances
2332
525
170
100
48
35
32
32
30
23

E-number
11
1212
111
123123
12121
1221
12341234
1122
1111
121212

Example
"bestir bestir "
"to prayers to prayers "
"ha ha ha "
"cover thy head cover thy head "
"come hither come hither come "
"fond done done fond"
"let him roar again let him roar again "
"with her with her hook on hook on "
"great great great great "
"come on come on come on "

Another phenomenon may be found noteworthy by a reader interested in
purely quantitative aspects of our research. It concerns the relation between
the length of the E-number (i.e. the amount of corresponding bricks) and the
number of utterances matched by such numbers. In case of trivial repetitions,
this relation seems to be plainly Zipfian. For example : Shakespeare's dramas
seem contain 2332 duplications (e.g. E=11), 170 triplications (E=111), 30
tetraplications (E=1111), 8 pentaplications (E=11111) two hexaplications
(E=111111), one heptaplication (E=1111111) and zero octaplications.
Table 3: Comparison of frequencies of occurrence of schemata of certain length and amount
Digits
2
3
4
5
6
7
8
9

Theoretical
1
1
4
11
41
162
715
3425

Matched
2332
170
622
91
211
56
86
67

It is worth mentioning, however, that generic relation between the length (in
digits) of an and the amount of utterances which matches seems not to be
Zipfian. As indicated by Table 3, an observed preference for repetitive
expressions including two, four, six or eight bricks cannot be explained in
terms of number-theoretical distribution of E-numbers themselves.
For example, there exists eleven E-numbers with five digits and fourty-one Enumbers of length six. However, when exposed to Shakespeare corpus,
regexes generated from six digits long seem to match 211 utterances while
five brick long regexes match only ninety-one of them. Whether this

390

JADT’ 18

observed asymmetry is an artefact of our method or whether it is due to a sort
of cognitive bias, a sort of preference for balanced repetitions within the Poet's
mind poses us in front of an argument which we do not dare to tackle here.
4. Conclusion
Insight that certain class of repetition-based schemata can be enumerated
allows us to generate myriads hitherto unseen Perl Compatible Regular
Expressions5 which involve back-references and negative lookaheads.
In the end, such regexes have been exposed to corpus containing collected
works of William Shakespeare.
Matching all utterances with all regexes generated out of all 4360 E-numbers
with less than 10 digits lasted 9555 seconds in case Shakespearean comedies,
6607 seconds in case of tragedies and 6900 seconds in case of historical
dramata. All this on one single core of an 1.4 GHz CPU.
This approach allowed us to pinpoint 36676 utterances matching at least one
among 172 distinct repetitive schemata. 23 among these schemata matched at
least 5 distinct utterances, 71 among them matched at least two utterances.
This may potentially point to a sort of neurolinguistic habit residing in the
opaque sphere between the syntactic and lexical layers.
We believe that at least some among these «figures » could be of certain
interest not only for scholars trying to understand inner intricacies of
Shakespeare's genius, but also to address more generic topics in fields as
distinct as digital humanities, computational rhetorics, discourse stylometry
or even more general cognitive sciences.
References
Aho, A. V. (2014). Algorithms for finding patterns in strings. Algorithms and
Complexity, 1:255.
Arndt, J., Sloane, N. J. A. (2016). Counting words that are in "standard order".
The on-line encyclopedia of integer sequences.
https://oeis.org/A278984/a278984.txt.
Dubremetz, M., Nivre, J. (2015). Rhetorical figure detection: the case of
chiasmus. On Computational Linguistics for Literature, page 23.

We remind the reader that PCREs are much more powerful than so-called
regular grammars. For example, regular grammars are unable to backreference, while
for PCREs, backreferencing is a completely legal act.
6
See
https://refused.science/rhethorics/shakespeare-regex/matches.csv
(Licenced under CC BY-NC-SA) for list of all matched utterances, including the
information about the respective entangled numbers, theater pieces, genres (comedy /
tragedy / drama) and the dramatis personae.
5

JADT’ 18

391

Friedl, J. E. F. (2002). Mastering regular expressions. O’Reilly Media, Inc.
Harris, R., DiMarco Ch. (2009). Constructing a rhetorical figuration ontology.
In Persuasive Technology and Digital Behaviour Intervention
Symposium, pages 47–52. Citeseer.
Hromada, D. D. (2011). Initial experiments with multilingual extraction of
rhetoric figures by means of PERL-compatible regular expressions. In
RANLP Student Research Workshop, pages 85–90.
OEIS (2016). List of words of length n over an alphabet of size 9 that are in
standard order and which have the property that every letter is repeated
at least once. https://oeis.org/A273978

392

JADT’ 18

Spécificités des expressions spatiales et temporelles
dans quatre sous-genres romanesques (policier,
science-fiction, historique et littérature générale)
Olivier Kraif, Julie Sorba
Univ. Grenoble Alpes, LIDILEM
olivier.kraif@univ-grenoble-alpes.fr; julie.sorba@univ-grenoble-alpes.fr

Abstract
In this paper, we aim to test if the classifications of the phraseological units
based on recurring trees and ngram methods are functional in order to
separate novel genres one from another. Our results confirm that these two
methods are relevant for the expressions relative to space and time into our
corpora.
Résumé
Notre objectif est de tester les classifications des phraséologismes, opérées
par les méthodes des ALR et des SR, dans le but de distinguer des sousgenres romanesques les uns des autres. Dans nos corpus, nos résultats
confirment la pertinence de ces classifications pour les deux champs de
l’espace et du temps.
Keywords: ngram, recurring trees, novel genres, phraseology
1. Introduction
Notre étude, qui s’inscrit dans le cadre de l’analyse exploratoire des données
textuelles, concerne des romans français contemporains rassemblés dans le
cadre du projet ANR-DFG PhraseoRom. Ce corpus (plus de 110 millions de
mots pour le français) est partitionné en plusieurs sous-corpus correspondant
à différents sous-genres littéraires (policier, science-fiction, fantasy, roman
historique, roman sentimental, littérature générale). Notre objectif est de
caractériser ces genres et sous-genres textuels par les unités phraséologiques
spécifiques qu’ils contiennent. À l’instar de Boyer, nous postulons que
« chaque genre comprend un certain nombre de sous-ensembles, des séries
fondées sur la réutilisation de composantes identiques » (1992, p.91). Dans la
mesure où la phraséologie étendue s’intéresse à tout ce qui est
« préfabriqué » dans les séquences lexicales, elle constitue donc un point
d’entrée privilégié pour mettre en évidence ces « séries ».
Pour cette étude, nous retenons spécifiquement 4 sous-genres : les romans de

JADT’ 18

393

science-fiction (SF), les romans policiers (POL), les romans historiques (HIST)
et les romans de littérature dite blanche ou générale (GEN). La fouille des
textes utilise la technique de repérage des Arbres Lexicosyntaxiques
Récurrents (ou ALR, Kraif & Diwersy, 2012 ; Kraif, 2016) dont la validité a
déjà été montrée par le repérage d’unités phraséologiques spécifiques dans
les textes scientifiques (Tutin & Kraif, 2016). Nous proposons en outre de
comparer ici cette technique d’extraction avec celle des segments répétés
(Salem, 1987), les ALR ayant montré une meilleure prise en compte de la
variabilité syntaxique pour le repérage des routines, mais s’avérant parfois
défaillants pour identifier des segments figés en surface, du fait du modèle
dépendanciel employé.
Dans des travaux antérieurs, nous avons montré comment les ALR
permettaient de repérer des motifs récurrents construits autour d’expressions
spécifiques fortement liées à la composante thématique des sous-genres en
question : c’était le cas pour « scène de crime » dans POL (Kraif, Novakova &
Sorba, 2016). Ici, nous nous concentrons sur des expressions moins
directement liées aux univers de référence des sous-genres (le crime, l’amour,
la science, etc.), afin de mettre en évidence des traits moins prévisibles. C’est
pourquoi, nous avons choisi de sélectionner les séquences – bien souvent
adverbiales – liées à l’expression du temps et de l’espace.
Nous allons désormais présenter les résultats obtenus dans des travaux
antérieurs (partie 2), puis décrire notre méthodologie expérimentale (partie
3). Enfin, nous exposerons et discuterons nos observations (partie 4) avant de
proposer des conclusions et perspectives à notre étude (partie 5).
2. Travaux antérieurs
Lefer, Bestgen & Grabar (2016) s’appuient sur une extraction de n-grammes
de 2 à 4 mots pour caractériser 3 genres textuels : des débats parlementaires
européens, des éditoriaux de presse et des articles scientifiques. Ces auteurs
utilisent une méthode d’AFC pour identifier les expressions les plus typiques
et en tirent des observations contrastives concernant l’expression de la
certitude et de l’opinion. De notre côté, nous avons analysé des contrastes
génériques sur un plan qualitatif, en identifiant des ALR dans des corpus de
romans policiers et de science-fiction, en nous fondant sur des mesures de
spécificité (Kraif, Novakova & Sorba, 2016). Nous avons également utilisé
l’extraction des ALR pour classer automatiquement, dans une approche
supervisée, des sous-corpus POL, SF et GEN (Chambre & Kraif, 2017). Ces
travaux préliminaires ont montré que les ALR donnaient de meilleurs
résultats que les autres catégories de traits (ponctuation, morphosyntaxe,
lexique), et permettaient de classer correctement 98% des textes du corpus à
partir d’une sélection de traits discriminants. La plupart de ces traits

394

JADT’ 18

appartenaient à des champs lexicaux précis, liés aux univers de référence
propres à chaque sous-genre, comme ceux du ‘téléphone’ (le numéro de
portable, passer un coup de fil, etc.) ou de la ‘voiture’ (à travers le pare-brise,
démarrer en trombe, etc.) pour POL. De plus, des expressions temporelles (p.ex.
pour POL à huit heures, vingt et une heure, au bout de X minutes) et des
indications spatiales très variées (p.ex. pour SF par la voie, dans le territoire,
dans la sphère, dans l’espace, la zone de) ont été mises en évidence.
Nous proposons ici un prolongement de cette expérimentation, d’une part,
en étudiant les expressions spatiales et temporelles, et d’autre part, en
ajoutant le sous-genre des romans historiques (HIST), afin de déterminer si
ces classes d’expression sont suffisantes pour différencier les quatre sousgenres (POL, SF, GEN, HIST).
3. Méthodologie
Pour chaque sous-genre, notre corpus comporte un échantillon d’environ 8
millions de mots, correspondant à environ 70 œuvres d’une quarantaine
d’auteurs (cf. Tableau 1). Ces œuvres sont toutes postérieures à 1950, et la
majorité d’entre elles ont été publiées pour la première fois après 2000. La
classification des œuvres en genre a été effectuée a priori selon des critères
éditoriaux, en fonction des collections de publication.
Auteurs

Romans

Taille

POL

46

69

8 008 395

SF

36

75

8 001 582

HIST

38

70

8 015 933

46

69

8 008 395

GEN

Tableau 1 : Constitution du corpus

Figure 1 : ALR représentant l’expression en une fraction de seconde

Pour identifier les expressions phraséologiques caractéristiques des différents
sous-genres, nous utilisons deux méthodes de repérage :
- la méthode des ALR : nos corpus étant analysés en dépendances avec XIP

JADT’ 18

395

(Aït-Mokhtar et al., 2002), ces ALR sont des sous-arbres respectant des
critères de fréquence (ici ≥ 10 occurrences), de dispersion (ici ≥ 10 auteurs
différents, appartenant à au moins 3 sous-genres différents) et de taille (ici ≥ 3
nœuds et ≤ 8 nœuds). En outre, lors de la recherche de ces ALR, une mesure
d’association est calculée afin de ne retenir que les nœuds significativement
associés avec le reste de l’arbre. La figure 1 montre un exemple d’ALR
correspondant à l’expression en une fraction de seconde.
- la méthode des segments répétés (ou SR, Salem, 1987) : nous avons appliqué
les mêmes critères de dispersion et de taille (≥ 3 et ≤ 8), afin de comparer les
deux méthodes in fine. Les SR sont constitués de séquences de lemmes
(obtenus avec XIP), et non de formes fléchies. Cette dernière méthode est
plus simple à mettre en œuvre et nécessite peu de ressources linguistiques,
bien qu’elle pose des problèmes d’explosion combinatoire (cf. partie 4).
Dans un second temps, nous appliquons un filtrage par mots-clés afin de ne
retenir que les séquences liées aux deux sous-domaines étudiés, à savoir
l’expression du temps et de l’espace. Les mots-clés pour l’espace sont des
noms de lieux, d’espaces naturels, de description géographique, des mesures
de distance, des adverbes de lieu, sélectionnés après un premier sondage des
ALR extraits :
- Mots-clés ESPACE : cave, salon, hôpital, immeuble, bâtiment, camp, restaurant,
village, route, rue, quai, chaussée, terrasse, ministère, parc, bureau, carlingue,
maison, toit, chambre, hôtel, palais, rez-de-chaussée, entrée, pont, escalier, chemin,
place, salle, jardin, seuil, cour, couloir, colline, sentier, sol, rive, rivage, plage, rivière,
mont, montagne, mer, océan, lac, bois, forêt, espace, endroit, coin, pays, continent,
frontière, direction, cap, sud, est, nord, ouest, confins, mètre, kilomètre, annéelumière, hectare, acre, loin, proche, près de, au bord de, orée, distance.
Les mots-clés pour le temps désignent des moments de la journée et de
l’année, des unités de mesure et des découpages conventionnels de période
(noms, adverbes et locutions adverbiales) :
- Mots-clés TEMPS : matin, soir, soirée, après-midi, nuit, jour, temps, fois, moment,
instant, toujours, jamais, parfois, souvent, autrefois, jadis, tôt, tard, longtemps,
brièvement, immédiatement, subitement, tout à coup, tout de suite, aujourd'hui,
demain, hier, lendemain, maintenant, heure, minute, seconde, journée, semaine,
mois, an, année, décennie, siècle, millénaire, printemps, été, automne, hiver.
Ces listes ne prétendent pas être exhaustives et le filtrage opéré produit à la
fois du silence et du bruit, du fait des ambiguïtés. Celles-ci demeurent
toutefois marginales (d’après un sondage manuel, le bruit est inférieur à
10 %).
Pour identifier les ensembles de traits pertinents du point de vue des sousgenres, nous injectons ces expressions (ALR ou SR) dans un système de
classification automatique. De la sorte, nous visons un double objectif : d’une

396

JADT’ 18

part, vérifier que nos classes constituées a priori sont cohérentes et corrélées à
des critères objectivables ; d’autre part, identifier ces critères sous la forme
d’ensemble de traits discriminants pour la classification.
4. Résultats et discussion
Dans une première étape, nous avons extrait les 6000 ALR les plus fréquents
sur l’ensemble du corpus. En effectuant une classification sur ces traits, avec
un modèle SVM optimisé par SMO (avec la plate-forme Weka, Eide et al.
2016), on obtient, dans une évaluation croisée à 10 plis, une précision de 74 %
(123 sur 166), avec un Kappa de 0,65, ce qui correspond à un très bon accord
avec la classification de référence. La matrice de confusion (cf. Tableau 2)
montre que les deux genres les mieux classés sont SF (93,1 %) et POL
(79,5 %). Le genre GEN obtient la précision la plus faible (64%) avec des
confusions fréquentes avec POL et HIST ; HIST est de son côté fréquemment
confondu avec GEN.
L’examen des ALR les plus discriminants montre, comme on pouvait s’y
attendre, la forte présence de certains thèmes dans POL, HIST et SF (la
voiture, le crime, le téléphone pour POL ; la guerre, la religion pour HIST ;
l’univers spatial et les artefacts technologiques pour SF) et l’absence de traits
saillants dans GEN.
4.1 Sélection des traits TEMPS+ESPACE
Lorsqu’on sélectionne les traits liés à l’expression du temps seul (environ un
millier), on obtient une dégradation par rapport aux résultats précédents,
avec une précision globale de 48,8 % et un Kappa de 0,31 signifiant un accord
faible entre la classification a priori et la classification automatique. Les
expressions spatiales, de leur côté (on en obtient 1560, mais nous avons
retenu les 1000 plus fréquentes afin de disposer de résultats comparables),
obtiennent des résultats un peu meilleurs, toutefois moins bons que les traits
non filtrés : la précision est de 59,6 %, avec un Kappa de 0,46 correspondant à
un accord modéré.
Quand on sélectionne conjointement les ALR de TEMPS et ESPACE, on
obtient une légère amélioration par rapport à la classification avec ESPACE
seul : 61,4 % (102 instances bien classées sur 166), avec un Kappa assez bon
de 0,48. La matrice de confusion (cf. tableau 2) montre que POL obtient la
meilleure précision (69%) et GEN la moins bonne (55,9 %).
Si on sélectionne les traits les plus discriminants (attributs SfcSubsetEval avec
méthode BestFirst dans Weka), on obtient un ensemble de 54 attributs. On
peut évaluer, de manière indicative, le pouvoir classificateur de ces attributs
sur notre corpus en les réinjectant dans une classification par SMO : on
obtient alors une précision globale très légèrement supérieure (62 %), mais il

JADT’ 18

397

est intéressant de noter que les genres marqués POL, SF et HIST sont très
bien classés sur la base de ces traits (précision de 85,7% pour HIST, 84 % pour
SF, 75,7 % pour POL) avec une dégradation forte pour GEN (43,4%), comme
le montre la matrice de confusion ci-dessous (tableau 2).
Tableau 1 : Matrices de confusion pour les classifications avec (1) tous les traits, (2) les ALR
filtrés (TEMPS+ESPACE) et (3) les ALR sélectionnés
(1) Tous les traits
(6000 ALR plus fréquents)

(2) TEMPS+ESPACE
(2571 traits filtrés)

(3) TEMPS+ESPACE
Sélection de 54 traits

SF

POL

GEN

HIST

SF

POL

GEN

HIST

SF

POL

GEN

HIST

SF

27

2

2

5

18

5

6

7

21

2

13

0

POL

1

35

9

1

5

29

12

0

3

28

15

0

GEN

1

5

32

8

3

3

33

7

1

6

36

3

HIST

0

2

7

29

3

5

8

22

0

1

19

18

L’examen détaillé des 54 traits sélectionnés révèle plusieurs points saillants :
- d’une manière générale, les ALR relatifs à l’espace sont très largement
majoritaires avec 33/54 contre 17/54 pour le temps, après élimination du bruit
(4/54).
- si on considère les traits spécifiques à HIST, les expressions spatiales
désignent surtout des lieux de pouvoir (la place forte, de son palais, salle du
palais, salle du château, pénétrer dans la grande salle) et la mer (sur la mer, de la
mer), tandis que les expressions temporelles font référence à une temporalité
longue (au bout de quelques mois, règne de X années, avoir le temps) et à des
datations absolues ou relative (du Ne siècle, venir le lendemain, à trois heures de
l’après-midi).
- pour POL, en revanche, les expressions temporelles indiquent des datations
horaires (à 8 heures, 21 heures) et des durées courtes (une vingtaine de secondes).
Les expressions spatiales, nombreuses, indiquent des pièces et des espaces
intérieurs (de la salle de bain, vers la salle de bain, entrer dans le bureau, vers le
bureau, dans le coin), des lieux urbains (aller à l’hôtel, passer à l’hôpital, à
l’hôpital), et des localisations vagues (dans le coin au sens de « dans les
parages »).
- pour SF, les expressions temporelles sont plus nombreuses (7/18) que dans
les autres sous-genres. Elles font référence à des durées extrêmes par leur
longueur (milliers d’années, de mille ans) ou leur brièveté (une fraction de
seconde, un centième de seconde). Pour l’espace, on trouve des expressions de
distances chiffrées (dizaines de mètres, centaine de mètres, plusieurs centaines de
mètres), des références attendues à l’espace intersidéral (dans l’espace, à travers
l’espace, être dans l’espace, voyager dans l’espace, flotter dans l’espace), à l’espace-

398

JADT’ 18

temps et des expressions avec sol (sur le sol, sous-sol).
- pour GEN : la seule expression spécifique apparaissant dans les traits
sélectionnés est chemin de traverse.
4.2 Comparaison avec les segments répétés
Nous n’avons pas réussi à extraire la totalité des SR de 3 à 8 mots pour
l’ensemble du corpus, du fait des problèmes d’explosion combinatoire
(environ 40 000 000 SR générés pour 100 textes du corpus). Nous avons donc
retenu les SR contenant les mots-clés sélectionnés pour TEMPS et ESPACE,
en conservant les 1000 SR les plus fréquents afin d’avoir des ensembles de
traits comparables aux ALR filtrés. On obtient de meilleurs résultats que
pour les ALR, avec une précision de 66,7 % pour ESPACE et 58,3 % pour
TEMPS contre respectivement 59,6 % et 48,8 %. Pour TEMPS+ESPACE, on
constate une certaine dégradation, avec une précision qui tombe à 64,1 %. À
ce stade de nos observations, il nous est difficile d’interpréter ces résultats
quantitatifs car la sélection du meilleur ensemble de traits pour ESPACE
donne peu ou prou les mêmes expressions qu’avec les ALR :
le chambre de, le cour de, à le cour, dans le espace, le salle de bain, de le espace, dans
son bureau, de le immeuble, le maison et, à le hôtel de, centaine de mètre, sur le
bureau, sur le place de, le palais de, dans le grand salle, de bureau de, de le salle de
bain, sur son bureau, cour de France, en route pour, dans mon bureau, dans tout le
direction, un dizaine de mètre, de son pays, à le rue, dans le sous-sol, quitter le salle,
dans un restaurant, sur le rivage, mètre plus bas, vers le bureau, route vers le,
dizaine de mètre de, un kilomètre de, à ministère de, dans le espace et, de un
montagne, le espace et le.
Les deux méthodes donnent donc des résultats convergents en termes
qualitatifs en extrayant les mêmes expressions. Néanmoins, des
investigations complémentaires seront nécessaires pour interpréter
correctement le fait que les SR obtiennent de meilleurs résultats quantitatifs.
5. Conclusion et perspectives
Cette étude confirme que les expressions phraséologiques constituent de
bons descripteurs pour la classification en sous-genre (Chambre & Kraif,
2017). En effet, même si les résultats obtenus ici à partir du sous-ensemble
constitué des expressions spatiales et temporelles sont sensiblement
inférieurs à ceux obtenus à partir de traits plus directement liés aux univers
de référence de chaque sous-genre (61.4 % /vs/ 98 %), ces expressions moins
riches sur le plan informatif permettent cependant de classer les romans dans
les sous-genres marqués POL, SF et HIST de manière satisfaisante. En
revanche, pour la catégorie des romans généraux (GEN), elles ne sont pas
discriminantes. Notre méthode permet aussi de dégager des spécificités

JADT’ 18

399

génériques propres à ces deux champs ESPACE et TEMPS (lieux de pouvoir
dans HIST /vs/ intérieur et lieux urbains dans POL ; durées et distances
extrêmes dans SF). Enfin, à partir de cette sélection d’expressions spatiotemporelles, la méthode des segments répétés produit une classification en
sous-genres plus précise que celle des ALR. Ce point, difficile à interpréter à
partir de nos premières observations qualitatives, nécessite une étude plus
approfondie. Ces résultats nous incitent à poursuivre l’exploration d’autres
champs lexicaux en marge des univers de référence de chaque sous-genre,
afin, d’une part, d’affiner notre méthodologie et, d’autre part, de cibler les
éléments au cœur de la phraséologie.
Références
Aït-Mokhtar S., Chanod J.-P. and Roux C. (2002). Robustness beyond
Shallowness: Incremental Deep Parsing. Natural Language Engineering,
8:121-144.
Boyer A.-M. (1992). La paralittérature. Presses Universitaires de France.
Chambre J. et Kraif O. (2017). Identification de traits spécifiques du roman
policier et de science fiction. Communication présentée aux Journées
Internationales de la Linguistique de Corpus - JLC2017, Grenoble, 05.07.2017.
Eibe F., Hall M. A. and Witten I. H. (2016). The WEKA Workbench. Online
Appendix for "Data Mining: Practical Machine Learning Tools and
Techniques", Morgan Kaufmann, Fourth Edition.
Kraif O., Novakova I. et Sorba J. (2016). Constructions lexico-syntaxiques
spécifiques dans le roman policier et la science-fiction. Lidil, 53 : 143-159.
Kraif O. et Diwersy S. (2012). Le Lexicoscope : un outil pour l'étude de profils
combinatoires et l’extraction de constructions lexico-syntaxiques. Actes de
la conférence TALN 2012, pp. 399-406.
Lefer M.-A., Bestgen Y. et Grabar N. (2016). Vers une analyse des différences
interlinguistiques entre les genres textuels : étude de cas basée sur les ngrammes et l’analyse factorielle des correspondances. Actes de la conférence
conjointe JEP-TALN-RECITAL 2016, pp. 555-563.
Tutin A. et Kraif O. (2016). Routines sémantico-rhétoriques dans l’écrit
scientifique de sciences humaines : l’apport des arbres lexico-syntaxiques
récurrents. Lidil, 53 : 119-141.
Salem A. (1987). Pratique des segments répétés. Essai de statistique textuelle.
Klincksieck.

400

JADT’ 18

Les phrases de Marcel Proust
Cyril Labbé1, Dominique Labbé2
1 Univ. Grenoble Alpes, CNRS, Grenoble INP*, LIG, F-38000 Grenoble France
(cyril.labbe@imag.fr)
2 Univ. Grenoble Alpes, PACTE (dominique.labbe@umrpacte.fr)

Abstract
Analysis of sentence lengths in Marcel Proust’s A la recherche du temps perdu.
Counting standards and the various available measures are presented. For
most of his reading time, the reader of this novel is confronted with very long
and syntactically-complex sentences. A comparison with other writers shows
that these sentences are atypical but not unique and that some of their
characteristics can be observed in a number of other works, some of which
are cited in the Recherche du temps perdu.
Résumé
Analyse des longueurs de phrases dans A la recherche du temps perdu de
Marcel Proust. Présentation des normes de dépouillement et des différentes
mesures possibles. Durant la majorité de sa lecture, le lecteur se trouve
confronté à des phrases très longues et syntaxiquement complexes. Une
comparaison avec un large panel d’écrivains montre qu’il s’agit d’un
phénomène exceptionnel mais pas unique et que certaines caractéristiques se
retrouvent dans quelques œuvres dont certaines sont citées dans la Recherche
du temps perdu.
Keywords: lexicometry - stylometry - sentence length – French literature Proust
1. Introduction
Les phrases de Marcel Proust (1871-1922) sont-elles exceptionnelles ? La
question a été surtout traitée sous l’angle qualitatif (notamment Curtius
1970). Il existe quelques estimations quantitatives (Bureau 1976, Brunet 1981,
Milly 1986), avec des résultats divergents pour des raisons qui seront
explicitées au début de cette communication. Mais surtout, nous présentons
une comparaison statistique avec d’autres écrivains qui permettra de juger de
l’exceptionnalité de la phrase proustienne.
L’analyse des phrases soulève plusieurs des problèmes auxquels est
confrontée la lexicométrie (statistique appliquée au langage). En premier lieu,
ici, il y a le choix de l’édition de référence. En effet, pour la Recherche du temps

JADT’ 18

401

perdu, ce choix existe et introduit une légère incertitude concernant la
ponctuation de l’oeuvre (discussion dans Ferré 1957 et Serça 2010),
spécialement pour les trois derniers volumes. Nous nous sommes tenus au
principe général selon lequel fait foi l’ultime version révisée par l’auteur ou,
à défaut, la plus proche de sa mort. Il s’agit ici de l’édition originale chez
Gallimard (annexe 1). De plus, cette édition originale s’impose puisqu’elle est
dans le domaine public et peut être communiquée librement aux chercheurs
soucieux de reproduire nos résultats et d’aller plus loin dans cette analyse.
2. Le mot et la phrase
Le mot est défini comme l’occurrence d’un vocable, c’est-à-dire une entrée
dans le lexique de la langue française selon la norme présentée par Muller
1963. Cette norme est fondée notamment sur la nomenclature de Hatzfeld et
al. 1898. Son implémentation est décrite dans Labbé 1990. Par exemple,
"aujourd’hui", "parce que" ou "Saint-Loup" sont des mots uniques et non
deux "formes graphiques". Il y a 1 449 "parce que" dans la Recherche, soit plus
d’un mot pour mille ; et 787 fois "Saint-Loup" (l’un des principaux
personnages du roman). A l’inverse, les formes graphiques "le", "la", "les" ont
deux entrées (pronom ou article) ; "du" ou "des" sont la contraction de deux
entrées du lexique - préposition "de" et article "le". En fonction de la norme
retenue (vocable ou formes graphiques), le nombre de mots dans un texte
peut varier de près 10%. Selon cette "norme Muller", la Recherche compte 1
327 859 mots (N dans la suite) et 21 836 vocables différents.
Quant à la phrase, il y a un accord général pour la définir comme l’empan de
texte dont le premier mot comporte une majuscule initiale et qui se trouve
compris entre deux ponctuations majeures. Les ponctuations majeures sont le
point, les points d’interrogation et d’exclamation, les points de suspension.
Cependant, aucun de ces 4 signes typographiques ne marque
automatiquement une fin de phrase :
- le point dans « M. Verdurin » ne termine pas une phrase même s’il est suivi
d’un mot à majuscule initiale. Il y a dans la Recherche 3 152 « monsieur » écrits
"M.". C’est le deuxième substantif le plus fréquent dans la Recherche (juste
derrière "Mme"), soit 2,4 pour mille mots. Ce point "non-terminal" se retrouve
dans les initiales que Proust utilise pour "anonymiser" certains noms (Mme
X.) ou derrière des abréviations (etc.).
- dans la Recherche, plus de trois points d’interrogation sur 10 sont internes à
la phrase (721).
- il y a 1 201 points d’exclamation internes à la phrase et 190 points de
suspension également dans cette situation. Proust a plusieurs fois déclaré son
hostilité envers ces derniers mais il les utilise parfois. Par exemple : « La
duchesse émit très fort, mais sans articuler : « C’est l’... i Eon l... b... frère à

402

JADT’ 18

Robert. » (la Prisonnière).
Cette rapide discussion permet de comprendre la solution adoptée : un
automate détermine les fins de phrase et, en cas de doute, l’opérateur choisit :
fin de phrase ou ponctuation interne ? A condition que l’opérateur suive
toujours la même norme, le dépouillement est fait sans erreur et, surtout, les
résultats obtenus sur un auteur sont comparables à ceux de tous les autres.
Ce recensement établit le nombre de phrases de la Recherche (voir tableau en
annexe). P = 37 336 phrases. Comment caractériser ces phrases en fonction de
leurs longueurs ?
3. Les indices statistiques usuels.
Les P phrases sont rangées par longueur croissante, dans des classes
d’intervalles égaux (ici 1 mots). Par exemple, la première classe (1 mot,
généralement une exclamation) contient 124 phrases, soit 0,37% du total.
L’effectif de chaque classe est ainsi recensé et son poids relatif est calculé. Ce
recensement fournit les informations suivantes :
- Etendue de la distribution : 1 à 931 mots. La plus longue phrase est celle sur
les homosexuels au début de Sodome et Gomorrhe. Les phrases de la Recherche
ne sont pas réparties uniformément sur cet intervalle. La seconde plus longue
– celle sur les chambres au début de Combray – compte 542 mots ; la troisième
(le salon des Verdurins dans la Prisonnière) : 430 ; la quatrième (l’église de
Combray) : 399. Ensuite, il n’y a plus de "trou" important dans l’étalement
des longueurs.
- Le mode est la classe la plus peuplée, ou longueur de phrase que le lecteur a
le plus de chance de rencontrer : 11 mots. Il y a donc, dans la Recherche, une
prédominance des phrases courtes et syntaxiquement simples. Il en est ainsi
dans la plupart des textes en français.
- La médiane est la valeur de la variable pour l’individu du milieu ou
individu "médian". Dans les P phrases rangées par longueurs, l’individu
médian est celui qui occupe la place (P+1)/2. Lorsque l’effectif total de la
population (P) est pair, la médiane est la moyenne des valeurs de la variable
pour les 2 individus situés de part et d’autre. Dans un texte étendu comme la
Recherche, la médiane se trouve dans une classe dont l’effectif est assez élevé.
Dans ce cas, la valeur est interpolée en divisant l’intervalle de la classe où se
situe l’individu médian par l’effectif de cette classe. Dans la Recherche, ce
calcul aboutit à une médiane de 26,28 mots. Etant donné que la variable
"longueur de phrase" ne prend que des valeurs entières, les décimales
indiquent le sens de l’arrondi et la position de la borne. La longueur médiane
des phrases de la Recherche est donc de 26 mots. Ou encore la moitié des
phrases ont une longueur inférieure ou égale à 26 mots et l’autre moitié une
longueur supérieure à 26.

JADT’ 18

403

- La moyenne (N/P) : 35,57 mots. A cet indice est associée une déviation
"standard" des valeurs de la variable autour de la moyenne (écart-type) :
racine carrée de la variance (moyenne des carrés des écarts de chaque valeur
de la variable à la moyenne arithmétique). L’écart type de la longueur des
phrases de la Recherche est de 31,42 mots.
La dispersion des valeurs autour de la moyenne mesurée par le coefficient
de variation relative : rapport de l’écart-type à la moyenne arithmétique (ici
89%). Etant donné l’effectif considéré (37 336 phrases), si les valeurs de la
variable "longueur de phrase" étaient distribuées normalement autour de la
moyenne (cas d’une population homogène), ce coefficient serait d’environ
4%. Autrement dit, les observations sont extrêmement dispersées. Dans ce
cas, la moyenne n’est pas représentative de la série et, en particulier, il n’est
pas possible de considérer que cette moyenne se situe à peu près "au milieu"
de la population. Dès que la dispersion relative approche les 50% de la
moyenne, celle-ci est située dans la partie basse de l’étendue de la
distribution qui est fortement asymétrique. Le profil de la distribution des
longueurs de phrases dans la Recherche est donné par la figure 1 dans laquelle
l’effectif relatif de chaque classe est représenté par la hauteur du bâton
correspondant (histogramme).

Figure 1. Histogramme de la distribution des longueurs phrases

D’une part, le graphique s’interrompt à la classe 200+ mots et le bâton pour
cette classe – à l’extrême-droite du graphique - correspond aux 96 phrases
longues de 200 mots et plus (0,3% du total des phrases mais 2,1% de la
surface du texte). Le graphique complet est encore plus étalé sur la droite, la
grande masse des phrases apparaissant serrées sur la gauche… D’autre part,
le bâton le plus haut correspond au mode principal (11 mots) mais l’on
observe de nombreux modes secondaires (17, 20, 24, etc.) : plusieurs

404

JADT’ 18

populations sont donc mélangées. La plupart des phénomènes sociaux
présentent des caractéristiques semblables et, en premier lieu, la distribution
des revenus ou des patrimoines. Dans de pareils cas, l’analyse ne se contente
pas des valeurs centrales. Elle se centre sur la distribution du caractère étudié
(ici la surface du texte) au sein de la population (ici les phrases).
4. L’inégal partage de la surface du texte entre les phrases
Ce renversement de perspective présente un avantage : la surface de texte
correspond grosso-modo à la durée de la lecture. Deux méthodes sont
possibles pour l’évaluer.
4.1 Quantile et médiale
Les phrases étant classées par longueurs croissantes, la surface du texte
qu’elles couvrent est découpée en masses égales (tableau 1).
Tableau 1. Partage de la surface du texte en fonction de la longueur des phrases
Surface divisée en quantiles
Premier décile
Deuxième décile
Premier quartile
Troisième décile
Quatrième décile
Deuxième quartile (médiale)
Sixième décile
Septième décile
Dernier quartile
Huitième décile
Neuvième décile

Longueur (mots)
18.58
26.70
29.53
33.30
41.35
49.93
60.20
72.93
81.13
90.57
121.00

% des phrases (cumulé)
33,8
49,6
54,5
60,6
70,1
77,5
84,6
89,7
92,3
94,2
97,8

Dans ce tableau, le premier décile est la borne supérieure de l'intervalle
comprenant les phrases les plus courtes couvrant en tout 10% de la surface
du texte et la borne inférieure du 2e décile. Il indique que les phrases de
longueurs inférieures ou égales à 18 mots couvrent 10% du texte et
représentent plus du tiers du total des phrases (33,8%). Le lecteur n’y passe
au mieux qu’un dixième du temps de la lecture. Or c’est au-dessus de cette
longueur que l’on commence à rencontrer des phrases syntaxiquement
complexes. Autrement dit, au mieux, le lecteur de la Recherche se trouve face
à des phrases simples pendant un dixième de sa lecture (ou il est face à des
phrases plus ou moins complexes pendant les neuf dixièmes !)
A l’opposé, 2,2% des phrases (700) comptent plus de 121 mots (9e décile).
Elles couvrent également 10% du texte, c’est-à-dire la même surface que le
tiers évoqué ci-dessus. Cela signifie que le lecteur de la Recherche passe (au
moins) autant de temps à lire des phrases très longues – dont la construction
est nécessairement complexe -, qu’il n’en consacre à la masse des phrases les

JADT’ 18

405

plus brèves et structurellement simples.
Dans cette perspective, la valeur centrale la plus caractéristique est la
longueur de la phrase qu’il faut atteindre pour avoir lu la moitié du texte.
Pour éviter les confusions, cette seconde médiane est appelée médiale (Ml).
Elle correspond à la borne haute du cinquième décile (ou du deuxième
quartile). Dans la Recherche, elle est égale à 49,93 mots, soit 50 mots. Le
tableau indique que 77,5% des phrases (près de 8 sur 10) sont inférieures à
cette médiale. Autrement dit, le lecteur de la Recherche passe au moins la
moitié de son temps confronté à des phrases de 50 mots et plus, ce dont la
plupart d’entre eux n’ont guère l’habitude. Malgré le talent de l’écrivain, c’est
évidemment cela que les lecteurs retiennent.
4.2 Mesure de l’inégalité
Deuxième méthode, un indice unique mesure l’inégale répartition de la
surface du texte entre les phrases (en fonction de leurs longueurs). Deux
calculs sont proposés :
- le rapport entre la médiane (26,28) et la médiale (49,93) soit 0,90. Autrement
dit la médiale est de 90% supérieure à la médiane (pour des comparaisons
avec d’autres écrivains, voir l’annexe 2). Cet écart considérable suffit à
attester la prédominance des phrases longues dans la Recherche.
- le second calcul est utilisé en science économique pour étudier la
distribution des revenus ou des patrimoines. Il s’agit de l’indice de Gini qui
mesure l’écart entre la situation réelle et celle qui serait observée en cas
d’égale répartition du caractère (ici la surface du texte) entre les individus
(les phrases) composant le livre. En cas d’équirépartition, toutes les phrases
de la Recherche auraient la longueur moyenne (≈ 36 mots). Pour chaque
centile, on calcule la proportion de la surface de texte couverte et l’écart par
rapport à ce que serait cette surface dans l’hypothèse d’équirépartition.
L’indice de Gini est la somme de ces écarts. Ici, il est égal à 55,4%. Autrement
dit, dans la Recherche, les longueurs de phrases s’écartent de plus de 55% de
ce qui serait constaté dans une population homogène.
Le "diagramme de Gini" permet de visualiser cette situation. Les phrases
étant rangées par longueurs croissantes, on compte le nombre qu’il faut lire
pour atteindre 1% de la surface (premier centile), puis 2%, etc. jusqu’à 100%.
Les valeurs observées pour chaque centile sont reportées sur la figure 2 où la
diagonale représente l’hypothèse d’équirépartition. L’indice de Gini est la
surface comprise entre la diagonale et la courbe. Deux auteurs
contemporains, et importants pour M. Proust, sont ajoutés sur le diagramme
afin d’en illustrer les propriétés.

406

JADT’ 18

Figure 2 Diagramme de concentration (Gini) de la surface de la Recherche sur les phrases
longues, comparée à celle de J. Barbey d’Aurevilly et de A. France.

Ce diagramme permet de comprendre pourquoi la médiane ou la moyenne
rendent mal compte des distributions fortement asymétriques comme les
longueurs de phrase. Par exemple, les deux tiers des phrases ont des
longueurs inférieures à la moyenne et pourtant ces phrases ne couvrent qu’à
peine plus d’un tiers du texte (34,5%).
La figure 2 montre également que, si les phrases de la Recherche sont
singulières par rapport à certains écrivains du XIXe - à commencer par A.
France qui aurait fourni le modèle de Bergotte (Levaillant 1952) –, elles
semblent très proches de quelques livres comme Une vieille maîtresse (1851) de
Barbey d’Aurévilly, écrivain que Proust cite à plusieurs reprises (Rogers
2000). C’est la dernière question abordée dans cette communication.
5. Singularité de Proust ?
Pour juger de cette singularité : à qui le comparer ? Et comment décider si les
écarts constatés sont statistiquement significatifs ?
Premièrement, il faut comparer Proust à lui-même. Un de ses ouvrages se
trouve dans le domaine public : Les Plaisirs et les jours (1896) dont les valeurs
centrales sont indiquées en première ligne dans le tableau 2.

JADT’ 18

407

Tableau 2. Caractéristiques des phrases des Plaisirs et les jours comparés à la Recherche

Plaisirs et
jours
Recherche

Etendue
1-250

Mode
7

Médiane
21,30

Moyenne
27,87

Médiale
37,16

Me/Ml
0,754

Gini
0,542

1-931

11

26,28

35,57

49,93

0,900

0,554

Toutes ces valeurs sont significativement inférieures à celles observées dans
la Recherche. Cependant, l’indice de Gini indique que le jeune Proust avait
déjà tendance à concentrer une proportion importante du texte dans les
phrases longues.
Deuxièmement, il faut comparer Proust aux auteurs qu’il cite explicitement
ou par allusion, non seulement dans la Recherche (Nathan 1968) mais aussi
dans ses autres œuvres et dans sa correspondance (Chantal 1967). Dans la
Recherche, Racine et Mme de Sévigné sont les plus cités, puis en seconde
position : Balzac et Saint-Simon ; en troisième : Chateaubriand, Hugo,
Molière, Musset, Sand et Vigny. La singularité des phrases théâtrales (Labbé
& Labbé 2010) ne permet pas de comparer la Recherche (qui est un roman)
avec les pièces produites par Molière, Hugo, Musset, Racine ou Vigny.
Enfin, il faut le comparer aux autres romanciers contemporains : ont été
ajoutés les principaux écrivains du XIXe et du début du XXe - comme
Bourget, Giraudoux, Flaubert, Maupassant, Zola – et quelques auteurs moins
connus mais singulièrement proches de Proust.
L’annexe 2 présente un échantillon des résultats. Chaque écrivain est
singulier et parfois les indices peuvent varier selon ses oeuvres. La Recherche
se situe dans la partie haute pour tous les indices et notamment pour la
propension à concentrer une proportion importante du texte dans les phrases
les plus longues (Gini). Cependant, on observe des caractéristiques
supérieures à celle de Proust dans quelques œuvres - Huysmans (A rebours),
les frères Goncourt (Mme Gervaisais) - ou proches dans Barbey d’Aurevilly,
mais aussi dans les Lettres de Mme de Sévigné ou les Mémoires de SaintSimon.
6. Conclusions
Lorsque, dans une population – ici les phrases d’un texte -, un caractère (la
surface de ce texte) est très inégalement réparti, la moyenne et la dispersion
standard sont de peu d’utilité. L’indice statistique le plus éclairant est la
seconde médiane ou médiale. Pour mesurer le degré de dispersion de la série
autour de cette valeur centrale, de nombreux indices sont concevables,
notamment les rapports entre quantiles extrêmes. Cependant, le rapport
entre médiane et médiale, ou l’indice de Gini paraissent les plus aptes à
donner une indication de la concentration du caractère sur une proportion

408

JADT’ 18

plus ou moins restreinte de la population totale.
Ces indices montrent que, durant la majorité du temps, le lecteur de la
Recherche se trouve confronté à des phrases très longues (50 mots et plus) et
syntaxiquement complexes. Ils confirment que M. Proust a une propension à
concentrer une proportion importante du récit dans les phrases les plus
longues.
Ces conclusions ont été acquises grâce à un dépouillement rigoureux, à des
indices statistiques adaptés et à une vaste base de textes traités selon les
mêmes procédures. A ce prix, la statistique lexicale peut être une auxiliaire
utile de l’analyse littéraire.
Enfin, dans une œuvre littéraire, il n’existe pas un type de phrase unique
mais plusieurs qui ont chacun leurs particularités lexicales et stylistiques
(Monière et al. 2008 ; Labbé & Labbé 2010). Une prochaine publication
présentera ces types de phrases avec leurs singularités lexicales, stylistiques
et thématiques. Elle répondra aussi à une question pendante : comment
déterminer que les écarts entre œuvres et auteurs sont ou non significatifs ?
References
Brunet E. (1981). La phrase de Proust. Longueur et rythme. Travaux du cercle
linguistique de Nice, p. 97-117.
Bureau C. (1976). Marcel Proust ou le temps retrouvé par la phrase.
Linguistique fonctionnelle et stylistique objective. Paris : PUF, p. 178-231.
Curtius E.-R. (1971). Etude de lilas. Le rythme des phrases. In Tadié J.-Y.
(dir.). Lectures de Proust. Paris : A. Colin.
Milly J. (1975). La phrase de Proust. Des phrases de Bergotte aux phrases de
Vinteuil. Paris : Larousse.
Ferré A. (1957). La ponctuation de M. Proust. Bulletin de la Société des Amis de
Marcel Proust, 7, p 171-192.
Hatzfeld A., Darmeisteter A., Thomas A. (1898). Dictionnaire général de la
langue française du commencement du XVIIe siècle jusqu'à nos jours. Paris :
Delagrave.
Labbé C., Labbé D. (2010). Ce que disent leurs phrases. In Bolasco S., Chiari
I., Giuliano L. (Eds). Proceedings of 10th International Conference Statistical
Analysis of Textual Data. Rome : Edizioni Universitarie di Lettere
Economia Diritto. Vol 1, p. 297-307.
Labbé D. (1990). Normes de saisie et de dépouillement des textes politiques.
Grenoble : Cahiers du CERAT.
Levaillant J. (1952). Note sur le personnage de Bergotte. Revue des sciences
humaines. Janvier-Mars 1952, p 33-48.
Milly J. (1986). La longueur des phrases dans "Combray". Paris-Genève :
Champion-Slatkine.

JADT’ 18

409

Monière D., Labbé C. & Labbé D. (2008). Les styles discursifs des premiers
ministres québécois de Jean Lesage à Jean Charest. Canadian Journal of
Political Science / Revue canadienne de science politique. 41:1, p. 43-69.
Muller C. (1963). Le mot, unité de texte et unité de lexique en statistique
lexicologique. Langue française et linguistique quantitative. Genève-Paris:
Slatkine-Champion, 1979, p 125-143.
Nathan J. (1969). Citations, références et allusions de Marcel Proust dans A la
recherche du temps perdu. Paris : Nizet (Première édition : 1953).
Rogers B. (2000). Proust et Barbey d’Aurevilly. Le dessous des cartes. Paris :
Champion.
Serça I. (2010). Les coutures apparentes de la Recherche. Proust et la ponctuation.
Paris : Champion.
Annexe 1 Corpus A la Recherche du temps perdu (Marcel Proust. Paris Gallimard 1919-1927)
Livre

Longueur

Vocabulaire

Combray

79 906

6 502

1 727

Un amour de Swann

84 142

5 859

2 226

Noms de pays : le nom

19 434

2 823

374

Du côté de chez Swann (1919)

183 482

9 347

4 327

Autour de Mme Swann

91 451

6 532

2 511

Noms de pays : le pays

134 192

8 283

3 334

225 643

10 396

5 845

Le côté de Guermantes 1

75 494

6 281

1 903

Le côté de Guermantes 2, chapitre 1

84 354

6 368

2 781

A l'ombre des jeunes filles en fleur (1919)

Le côté de Guermantes 2, chapitre 2

N phrases

89 727

6 707

2 700

249 575

6 707

7 384

Sodome et Gomorrhe

13 512

2 476

271

Sodome et Gomorrhe 2, chapitre 1

30 699

3 779

2 082

Sodome et Gomorrhe 2, chapitre 2

117 774

7 822

3 056

Sodome et Gomorrhe 2, chapitre 3

57 603

5 311

1 811

Sodome et Gomorrhe 2, chapitre 4

8 137

1 373

250

227 725

10 972

7 470

Le côté de Guermantes (1920-21)

Sodome et Gomorrhe (1921-22)
La prisonnière (1923)

173 409

9 062

5 124

La fugitive (1925)

115 866

6 456

3 255

Le temps retrouvé (1927)

152 159

8 708

3 931

Dernier volume (posthume)

441 434

13 518

12 310

1 327 859

21 837

37 336

Total général (A la recherche du temps perdu)

410

JADT’ 18

Annexe 2 Longueur des phrases chez quelques écrivains antérieurs ou contemporains de
Proust
Recherche
Balzac
Barbey
(Chevalier)
Barrès

d’A.

Etendue

Mode

Médiane

Moyenne

Médiale

Me/Ml

Gini

931

11

26,28

35,57

49,93

0,900

0,554

391

10

17,27

21,88

29,00

0,680

0,511

192

7

21,92

29,4

43,00

0,964

0,557
0,497

195

8

17,86

21,94

28,59

0,601

Bourget

201

7

16,62

21,34

29,58

0,780

0,539

Chateaubriand
(Mémoires)
Daudet

195

22

24,46

28,5

34,28

0,401

0,437

203

5

13,14

17,84

25,26

0,923

0,549

Dumas

243

7

14,90

20,28

29,00

0,947

0,567

Flaubert

231

7

13,75

18,37

25,24

0,837

0,528

France

394

8

15,79

19,98

26,06

0,651

0,504

Gautier*

282

18

27,11

33,07

41,90

0,546

0,493

Giraudoux*

466

4

18,60

25,77

37,76

1,031

0,580

Goncourt
(Gervaisais)
Goncourt
(Journal)
Hugo*

670

8

24,17

34,05

51,47

1,130

0,597

373

3

19,80

25,37

37,62

0,900

0,580

828

6

11,39

16,89

23,68

1,079

0,561

Huysmans
(A
rebours)
Maupassant*

254

28

44,24

51,49

65,82

0,488

0,557

168

6

14,44

18,98

26,39

0,828

0,542

Musset*

197

16

19,56

23,82

29,57

0,512

0,485

Nerval*

136

12

19,93

24,21

31,27

0,569

0,499

Saint-Simon

361

18

27,89

34,15

44,14

0,523

0,506

Sand (Champi)

117

21

22,11

26,19

32,56

0,473

0,477

Sévigné (Lettres)

307

11

25,72

31,99

40,96

0,593

0,490

Stendhal

235

18

20,18

23,92

29,79

0,477

0,463

Vigny*

315

17

20,82

27,47

37,41

0,797

0,538

Zola

153

8

15,80

19,91

25,66

0,624

0,491

* Uniquement les romans

JADT’ 18

411

Verso un dizionario corpus-based del lessico dei beni
culturali: procedure di estrazione del lemmario
Ludovica Lanini1, María Carlota Nicolás Martínez 2
1

Università degli Studi di Roma La Sapienza– ludovica.lanini@uniroma1.it
2 Università degli Studi di Firenze – cnicolas@unifi.it

Abstract
The vocabulary of Italian cultural heritage has become a crucial object of
interest for different categories of users from a number of countries.
However, there are no satisfactory multilingual lexical resources available.
The present work moves in that direction. The aim of the paper is twofold: on
the one hand, it describes the LBC database, a resource for developing a
multilingual electronic dictionary of cultural heritage terms, made up of
comparable corpora from nine languages; on the other hand, a corpus-based
method for building a comprehensive headword list is proposed.
Keywords: electronic lexicography, multilingual lexical resources, corpus
linguistics
1. Introduzione
Di fronte a un interesse crescente, a livello internazionale, per il lessico
italiano dei beni culturali, emerge oggi l’esigenza, da parte di diverse
categorie di utenti, di risorse elettroniche multilingui relative al patrimonio
culturale; nonostante ciò, allo stato attuale, non sono disponibili strumenti
multilingui adeguati. Il progetto LBC (Lessico dei Beni Culturali) si propone
di affrontare il problema, sviluppando una banca dati testuale comprendente
corpora specialistici e comparabili per nove lingue (cinese, francese, inglese,
italiano, portoghese, russo, spagnolo, tedesco, turco). Fine ultimo è la
creazione di un dizionario multilingue del lessico dei beni culturali a base
testuale, che abbia come principali utenti studiosi del settore, ma anche
traduttori e operatori turistici. L’approccio corpus-based viene applicato sin
dal processo di definizione del lemmario, focus specifico del contributo.
2. La Banca dati LBC
La Bd-LBC (Banca dati LBC) è un database testuale multilingue progettato per
essere rappresentativo del lessico dei beni culturali: per il suo disegno si è
considerato l’italiano quale punto di partenza, ma si è pensato anche al
valore aggiunto derivante dalla possibilità di stabilire relazioni tra le diverse
lingue. L’italiano viene scelto come punto di riferimento in virtù della sua

412

JADT’ 18

centralità nello sviluppo storico del lessico dei beni culturali; molti testi non
italiani relativi a tale dominio hanno inoltre lo sguardo rivolto proprio verso
le tecniche e i monumenti realizzati in Italia. La prima fase di lavoro, dedicata
alla raccolta dei materiali, è partita dunque dai testi italiani che sono alla base
della storia dell’arte e dalle relative traduzioni, ma anche da opere in altre
lingue, applicando una metodologia di studio che facesse leva sulle
potenziali sinergie plurilingui. Per dare fondamento alla struttura del corpus
(Cresti et Panunzi 2013:57), la rappresentatività della risorsa è stata definita
fin dall’inizio attraverso dei criteri di campionamento dei testi (Billero et
Nicolás 2017: 208): «la rilevanza storico-culturale dell’opera dell’ambito
specifico di studio (ad es. testi di Vitruvio o Leonardo); la diffusione
internazionale di un’opera relazionata con l’ambito di studio (es. libri di
Vasari); il prestigio dato a livello internazionale al patrimonio italiano da
parte di un’opera (es. testi di Stendhal o Ruskin); la specificità dell’argomento
in rapporto alla storia dell’arte italiana ed in particolare della Toscana (es.
Burckhardt) ». Si è in questo modo delimitato un nucleo di testi di base
condivisi tra lingue, tale da rendere il corpus parzialmente parallelo, cui si
sono aggiunti via via testi peculiari per ogni lingua.
La progettazione del database ha previsto inoltre una macrostruttura
omogenea per i diversi corpora, che condividono i metadati associati a ogni
testo, a partire dai quali viene generato automaticamente un nome di file
univoco. Per quanto riguarda la microstruttura, la regola fondamentale è
stata quella di rispettare il testo originale, mantenendo eventuali note,
divisione in capitoli e tratti ortografici arcaici. Seguendo tali regole
strutturali, ogni squadra di lavoro, specificamente rivolta a una delle lingue,
ha avviato lo sviluppo dei singoli corpora (Corpus LBC-francese, Corpus
LBC-inglese, etc.), sottoposti a un’operazione di validazione della
digitalizzazione da parte di professori e studenti competenti nelle diverse
lingue. La banca dati, così disegnata, presenta un’omogeneità in grado di
favorire il lavoro lessicografico: la forte coesione strutturale tra corpora
permette infatti di operare davvero in parallelo.
Tra gli obiettivi del progetto vi è anche quello di implementare strumenti
informatici di gestione e interrogazione dei corpora, che consentano ai
membri del gruppo di effettuare ricerche ed estrarre dati sull’uso lessicale,
fondamentali per lo svolgimento del lavoro lessicografico. Si è dunque
realizzato un software online, per ora accessibile ai soli membri dell’unità di
ricerca, ma in prospettiva disponibile anche per gli utenti, che consenta la
consultazione dei corpora, sia in chiave monolingue che multilingue. Nella
ricerca di soluzioni per l’implementazione di un’installazione del corpus su
apposito server Internet, si è optato per l’ultima release di NoSketchEngine,
versione open source di Sketch Engine.

JADT’ 18

413

3. Il dizionario LBC: processo di definizione del lemmario
La banca dati, così elaborata, si pone quale risorsa di base per lo sviluppo di
un dizionario elettronico multilingue del lessico dei beni culturali, che possa
risultare strumento utile soprattutto in ambito traduttivo e turistico. In vista
della particolare utenza e applicazione, l’intento è quello di fornire una
risorsa lessicografica che presenti le seguenti caratteristiche:
- trattamento dei lemmi più “problematici” del dominio, con inclusione a
lemma di nomi propri ed espressioni multiparola, categorie lessicali
generalmente assenti dalle risorse, tuttavia di particolare rilevanza in virtù
delle difficoltà traduttive e del forte carico culturale;
- attenzione per l’aspetto più prettamente pratico e referenziale del lessico
della cultura, con apertura a quelle voci di arti e mestieri tradizionalmente
trascurate dalla lessicografia italiana, nonché interesse rivolto alle persone,
alle opere e ai luoghi fisici della storia culturale, più che al carattere teorico e
mentale (Harris, 2003) ed estetico generale (De Mauro, 1971) che ha a lungo
connotato il lessico artistico, in particolare quello della critica d’arte;
- inclusione non solo di nomi, ma anche di verbi, di norma esclusi dalle
risorse terminologiche, qui ritenuti di interesse per rendere conto di tecniche
e pratiche;
- impianto corpus-based, non solo per la selezione, descrizione e traduzione
dei lemmi, con individuazione degli equivalenti a partire dall’analisi di
concordanze bilingui, ma anche per l’offerta all’utente, entro la scheda
lessicografica, di esempi e citazioni testuali reali.
L’approccio corpus-based viene adottato sin dal processo di definizione del
lemmario, sviluppato a partire dal corpus LBC-italiano.
Il metodo proposto prevede la combinazione di tre ordini eterogenei di dati:
dato lessicografico; dato testuale quantitativo; dato testuale qualitativo. Il
dato di origine lessicografica, assunto sullo sfondo a frame di riferimento,
viene dunque incrociato con il dato testuale, tanto di livello quantitativo keyword e liste di frequenza- quanto di livello qualitativo -prodotto di ricerche
mirate su corpus e di osservazione dei contesti.
Per quanto riguarda le risorse adottate, la fonte lessicografica scelta è il
Grande Dizionario Italiano dell’Uso (De Mauro, 2007), la più estesa risorsa
lessicografica esistente per la lingua italiana, mentre alla banca dati LBC
viene affiancato, quale corpus generale di riferimento, il corpus Paisà
(www.corpusitaliano.it), costruito nel 2010 tramite web-crawling e raccolta
mirata di documenti da specifici siti web, per un totale di 250 milioni di token,
inteso come rappresentativo della lingua e cultura comune contemporanea
(Lyding et al., 2014). Indirettamente, viene assunto come corpus di
riferimento anche itTenten16, il corpus per la lingua italiana implementato in
Sketch Engine, interamente raccolto tramite web-crawling nel 2016

414

JADT’ 18

(5.864.495.700 token). Riguardo agli strumenti impiegati, l’adozione di un
software di corpus management e query all’avanguardia come Sketch Engine
(www.sketchengine.co.uk) risulta infatti cruciale per il processo di lavoro,
descritto di seguito nel dettaglio.
3.1 Fasi di lavoro
La prima operazione è consistita nell’estrazione dal corpus LBC di una lista
di parole chiave (2000), applicando la funzione keywords di Sketch Engine: le
keyword vengono ordinate in base al keyness score, dato dal rapporto tra la
frequenza normalizzata della parola nel focus corpus (LBC) e la sua frequenza
normalizzata in un corpus generale (itTenten16), previa applicazione di una
costante, denominata simple math parameter1 (Kilgariff et al., 2014).
Alla lista delle keyword è stata affiancata la lista di matrice lessicografica,
estratta dal Gradit selezionando l’insieme dei lemmi etichettati con marca
[TS] (tecnico-specifico) per arte, pittura, scultura e architettura, per un totale
di 2515 lemmi, di cui molti (370) multiparola. In maniera inattesa, dal
confronto tra le due liste emergono solo 24 coincidenze.
Risultando poco pulita, la lista delle keyword è stata sottoposta a uno spoglio
manuale, che ha ridotto i 2000 lemmi candidati a 219, primo vero lemmario
di base (comprendente nomi propri come Mantegna, arcaismi come fregiatura,
tecnicismi come nicchia).
Si è proceduto a questo punto a una serie di confronti, a partire dalla lista di
frequenza lemmatizzata del corpus LBC, come sintetizzato in Tabella 1.
L’incrocio con la lista del Gradit ha restituito 272 lemmi comuni, di cui 235
sono stati accolti previo controllo. Il lavoro di confronto con il corpus generale
Paisà ha seguito invece due linee di sviluppo: lo studio dei lemmi
caratterizzati da più alta differenza di frequenza relativa con peso maggiore
in LBC (i primi 600), da cui sono emersi 77 lemmi di interesse (figura, Firenze,
Raffaello) e lo spoglio dei lemmi presenti in LBC ma non in Paisà, che ha
permesso di individuarne 62 (tecnicismi come scalea e imbasamento, numerosi
arcaismi e varianti arcaiche come scarpellino, Florenzia, Buonarruoto).
L’insieme delle voci della lista Gradit assenti in LBC (ben 2243) è stato inoltre
sottoposto a un esame puntuale, che ha portato ad aggiungere al lemmario
1629 lemmi2. Il corpus LBC è in effetti in fase di sviluppo, per cui molte aree

A seconda dei bisogni dell’utente e della natura dei corpora, la costante può
essere modificata per restituire una lista con candidati a frequenza maggiore o
minore, con 100 come valore consigliato per ottenere parole del vocabolario core e
rumore minimo, qui applicato.
2 Non si sono accolti: lemmi astratti, propri della critica d’arte (asemanticità);
lemmi riferiti a movimenti e tendenze generali (astrattismo); aggettivi o avverbi. Si
1

JADT’ 18

415

di interesse (per esempio il dominio dell’arte contemporanea) non risultano
ancora adeguatamente rappresentate: la lista del Gradit può offrire in questa
direzione materiali utili, in attesa dell’ampliamento del corpus.
Dalla convergenza dei lemmi accolti è stato così possibile arrivare alla
definizione di un primo lemmario, per un totale di 2147 lemmi.
Tabella 1
Risorse
Lista
LBC
Lista
Gradit
Lista
LBC
Lista
Paisà

Lemmi

Lemmi di interesse

Lemmi
estratti

Lemmi
accolti

Lemmi comuni

272

235

600

77

1139

62

2000

219

2243

1629

TOT.

2222
(-75 lemmi
ripetuti)
= 2147

8388
2515
8388
1032178

Lemmi con differenza di
frequenza relativa significativa
Lemmi presenti in LBC assenti in
Paisà

Lista
keywords
<LBC

2000

Tutti

Lista
Gradit

2515

Lemmi assenti in LBC

Il confronto con Paisà, in particolare, ha permesso di individuare una serie di
lemmi di interesse, non rappresentati nella lista lessicografica: nomi propri
(Raffaello, Firenze), tecnicismi (travata), molti lemmi comuni, spesso
semanticamente polivalenti e dotati di accezione specifica (figura, opera), ma
anche arcaismi (reliquiere) e varianti grafiche arcaiche (trivertino), ritenuti utili
in vista della lettura e traduzione di testi.
3.2 Estrazione di lemmi multiparola

sono invece accettati: lemmi che siano forme derivate di lemmi rappresentati in LBC
(aggrottescare, con grottesca presente in LBC); lemmi relativi all’arte contemporanea
(acrilico); unità multiparola ritenute di interesse (arco acuto); verbi ritenuti di interesse
(festonare).

416

JADT’ 18

Il lavoro di analisi testuale si è finora limitato a soli lemmi semplici; l’intento
è tuttavia quello di includere nel dizionario, come entrate autonome, anche
lemmi multiparola. Si tratta di entità problematiche, per le quali sono state
proposte innumerevoli denominazioni, classificazioni e criteri di
identificazione. Relativamente alla lingua italiana, si è parlato di lessemi
complessi (Voghera, 1994), polirematiche (Voghera, 2004, De Mauro, 2005),
espressioni multiparola (Masini, 2009). Alla tradizione anglosassone appartiene
il termine multiword expression (MWE), iperonimo utilizzabile per descrivere
una serie di entità generalmente distinte sul piano teorico, ma accomunate da
proprietà quali restrizione combinatoria, alto grado di lessicalizzazione e
convenzionalità, non-composizionalità e opacità semantica. Tale approccio
generalizzante risulta presente anche entro la tradizione lessicografica, poco
interessata del resto a distinzioni di ordine teorico (multiword lexical units in
Zgusta, 1971). La decisione è quella di adottare una terminologia che sia in
linea con tale tradizione, scegliendo la denominazione lemmi multiparola, e
abbracciando una definizione generale come quella proposta da Calzolari et
al. (2002:1934): «word combinations characterised by different degrees of
fixedness and idiomaticity that act as a single unit at some level of linguistic
analysis, such as idioms, collocations, preferred combinations». La
caratteristica essenziale dei lemmi multiparola è quella di comportarsi come
parole semplici: in ambito lessicografico, dunque, la scelta non può essere che
quella di trattarli al pari di lemmi semplici, introducendoli come entrate
autonome. Tale scelta risulta operativa anche nel Gradit: lo spoglio della lista
lessicografica ha permesso già dunque di accogliere un buon numero di
lemmi multiparola (303). Per quanto riguarda il dato testuale, un primo
tentativo è consistito nell’applicazione della già citata funzione keywords
fornita da Sketch Engine. La funzione permette in effetti di estrarre anche
una lista di term, intesi come noun-phrase dotati di un buon indice di keyness,
ma la scarsa qualità dei risultati ottenuti (solo 2 candidati su 2000 sono
risultati accettabili) e il fatto che la lista includa solo sintagmi nominali e non
verbali ha spinto ad adottare soluzioni alternative.
Si è deciso dunque di partire da un nucleo di lemmi semplici, per estrarne i
word sketch. Si tratta della funzione più caratteristica del software, vera e
propria sintesi corpus-based del comportamento collocazionale di una parola: i
collocati, selezionati e ordinati in base a un indice di associazione (logDice),
vengono mostrati entro categorie, i gramrel, corrispondenti a specifici pattern
definiti da una sketch grammar soggiacente, scritta in CQL. Il nucleo dei
lemmi di partenza (643) è risultato dalla convergenza di tre sottoinsiemi,
scelti in quanto fortemente rappresentativi del dominio ma anche
potenzialmente produttivi in termini combinatori: la lista keyword (in una
versione ripulita, ma ancora comprendente aggettivi e tutti i nomi propri); i

JADT’ 18

417

lemmi comuni tra lista LBC e lista Gradit; i lemmi base delle unità
multiparola della lista Gradit accolte nel lemmario. Per la generazione degli
sketch si è applicata inoltre una versione personalizzata della sketch grammar,
tarata sul dominio di studio: la modifica e l’aggiunta di regole ha permesso
di focalizzarsi su pattern di specifico interesse, non previsti dalla grammatica
del software al momento del lavoro, tra cui in particolare combinazioni
includenti nomi propri, sia come basi (Accademia fiorentina) che come collocati
(fabrica di San Piero). Dall’analisi degli sketch sono emersi 57 lemmi
multiparola, di cui addirittura 46 assenti nella lista del Gradit (80.7%). Le
ragioni dell’assenza possono essere diverse: a volte i lemmi risultano
collocati all’interno delle entrate, altre volte è in gioco una differente
valutazione in merito alla segmentazione delle unità multiparola (per cui, per
esempio, a olio viene collocato a lemma, mentre colore a olio no). È possibile
tuttavia individuare numerosi lemmi che si pongono come decisamente
nuovi rispetto alla lista “ufficiale” proveniente dal Gradit: tra questi, ancora
una volta, nomi propri (Colonna di Traiano) e varianti grafiche (volta a botta),
ma anche nomi comuni con collocati nomi propri (marmo di Carrara) e
tecnicismi del lessico artistico utili a colmare lacune negli elenchi di oggetti o
tecniche risultanti incompleti nella lista del Gradit (nicchia bislunga, nicchia
piana, nicchia quadra, da aggiungere alla serie nicchia a edicola, nicchia a
tabernacolo, nicchia angolare, nicchia finta). I lemmi multiparola estratti dal
corpus sono confluiti nel lemmario definitivo, giunto così a un totale di 2204
lemmi, di cui 360 multiparola. Il lavoro sui lemmi multiparola è comunque in
corso d’opera e l’intento è quello di portare avanti la ricerca su corpus in tale
direzione.
4. Conclusioni e lavoro futuro
Il metodo descritto ha portato allo sviluppo di un lemmario per un dizionario
del lessico dei beni culturali comprendente 2204 lemmi. L’approccio corpusbased ha permesso di individuare una serie di lemmi assenti nella fonte
lessicografica adottata, per un totale di 287 nuovi lemmi.
In futuro, il metodo illustrato si vedrà sottoposto a validazione, con
l’applicazione ad altre lingue della banca dati LBC. Parallelamente, il
processo di lavoro per il lemmario italiano verrà comunque portato avanti, in
due direzioni: per l’aggiunta di ulteriori lemmi multiparola -per esempio,
tramite estrazione di liste di frequenza di n-grammi- e per la sua crescita in
concomitanza con quella del corpus stesso. Anche i lemmi estratti dai corpora
delle altre lingue e non presenti nel lemmario italiano potranno inoltre
contribuire al suo arricchimento.

418

JADT’ 18

References
Billero R., Nicolás Martínez M. C. (2017). Nuove risorse per la ricerca del
lessico del patrimonio culturale: corpora multilingue LBC. In CHIMERA
Romance Corpora and Linguistic Studies, Madrid, UAM, 4.2, pp. 203-16.
Calzolari N., Fillmore C. et al. (2002). Towards best practice for multiword
expressions in computational lexicons. Proceedings of the 3rd International
Conference on Language Resources and Evaluation (LREC 2002). Las Palmas,
Canary Islands, pp. 1934-40.
Cresti, E., Panunzi, A. 2013. Introduzione ai corpora dell’italiano. Bologna, Il
mulino.
De Mauro T. (1971). Senso e significato: studi di semantica teorica e storica.
Adriatica
De Mauro T. (2005). La fabbrica delle parole: il lessico e problemi di lessicologia.
Torino, UTET
De Mauro T. (2007). Grande Dizionario Italiano dell'uso. Torino, UTET.
Harris R. (2003). The necessity of Artspeak. The language of the arts in the Western
tradition. London, Continuum.
Kilgariff A., Jakubíček M. et al. (2014). Finding terms in Corpora for many
languages with the Sketch Engine. In Proceedings of the Demonstrations at
the 14th Conference of European Chapter of the Association for Computational
Linguistics, Sweden, April 2014, pp. 53-56
Lyding, V., Stemle, E. et al. (2014). The PAISÀ Corpus of Italian Web Texts.
In Proceedings of the 9th Web as Corpus Workshop (WaC-9), Association for
Computational Linguistics, Gothenburg, Sweden, April 2014. pp. 36-43.
Masini F. (2009). Combinazioni di parole e parole sintagmatiche. In E.
Lombardi Vallauri e L.Mereu, Spazi linguistici, Roma, Bulzoni, pp. 191-209
Voghera M. (1994). Lessemi complessi: percorsi di lessicalizzazione a
confronto. Lingua e stile 28, pp. 185-214.
Voghera M. (2004). Le polirematiche. La formazione delle parole. In M.
Grossmann e F. Rainer. Tübingen, Max Niemeyer Verlag, pp. 56-68.
Zgusta L. (1971). Manual of Lexicography. Prague-The Hague-Paris,
Academia/Mouton.

JADT’ 18

419

“The grief that doesn’t speak”: Text Mining and
Brain Structure
Daniela Laricchiuta1, Francesca Greco2, Fabrizio Piras3, Barbara Cordella4,
Debora Cutuli5, Eleonora Picerni6, Francesca Assogna7,
Carlo Lai8, Gianfranco Spalletta9, Laura Petrosini10
Sapienza University, IRCCS Fondazione Santa Lucia – daniela.laricchiuta@uniroma1.it
2Sapienza University of Rome, Prisma S.r.l. – francesca.greco@uniroma1.it
3IRCCS Fondazione Santa Lucia - f.piras@hsantalucia.it
4 Sapienza University of Rome – barbara.cordella@uniroma1.it
5Sapienza University of Rome, IRCCS Fondazione Santa Lucia – debora_cutuli@yahoo.it
6Sapienza University, IRCCS Fondazione Santa Lucia – eleonora.picerni@uniroma1.it
7IRCCS Fondazione Santa Lucia - f.assogna@hsantalucia.it
8Sapienza University of Rome – carlo.lai@uniroma1.it
9IRCCS Fondazione Santa Lucia – g.spalletta@hsantalucia.it
10Sapienza University of Rome, IRCCS Fondazione S. Lucia – laura.petrosini@uniroma1.it
1

Abstract
Contemporary neurosciences have shown that emotions, thought and
language involve the functioning of connected brain areas, which allow the
recognition and expression of one's own feelings. The scope of this pilot
study is to investigate the link among the verbal expression of emotional
experiences (assessed with the Toronto Structured Interview for Alexithymia
- TSIA -), the linguistic structure and the brain structure. To this aim, 9
healthy adult subjects of both sexes were interviewed by means of the TSIA
and the cortical and subcortical structural measures were detected. The TSIA
transcripts were analysed by using a cluster analysis and, subsequently, a
correspondence analysis, and the values of factors were correlated with
cortical and subcortical structural measures as well as TSIA scores,
evidencing significant associations. The study highlighted that in healthy
subjects it is possible to identify a link between the manner in which people
express their experiences, recognize and use their emotions and the brain
structural correlates.
Abstract
Le neuroscienze contemporanee hanno evidenziato come le emozioni, il
pensiero e il linguaggio coinvolgono il funzionamento di aree cerebrali
differenti connesse tra loro, le quali consentono il riconoscimento e
l’espressione dei propri sentimenti. Scopo di questo studio pilota è di
indagare il nesso tra l’espressione verbale delle proprie esperienze emotive
(valutata con la Toronto Structured Interview for Alexithymia – TSIA -), la

420

JADT’ 18

struttura del linguaggio utilizzato e la struttura cerebrale. A questo scopo 9
soggetti sani di entrambi i sessi sono stati intervistati con la TSIA e sono state
rilevate le misure strutturali corticali e sottocorticali. Le interviste sono state
sottoposte ad analisi dei cluster e successivamente ad analisi delle
corrispondenze e i valori dei fattori sono stati correlati con le misure
strutturali corticali e sottocorticali e con i punteggi della TSIA. I risultati
evidenziano associazioni significative, che mettono in luce come in soggetti
sani sia possibile individuare un nesso tra il modo in cui le persone
raccontano le proprie esperienze, la loro capacità di riconoscere e utilizzare le
loro emozioni e la loro struttura cerebrale.
Keywords: Text mining, brain imaging, alexithymia, TSIA
1. Introduction
According to the multiple code theory, the emotional information is
represented in verbal, non-verbal symbolic and non-verbal sub-symbolic
multiple systems (Bucci, 1997). The verbal system is a communication and
reflection code through which the emotional, private and subjective
experience can be shared with others. It refers to the capacity of language to
direct and regulate ourselves, activate imagination and emotions, stimulate
actions and control them. The multiple channels of the non-verbal systems
include representations and proceedings related to implicit elaboration
associated with visceral, somatic, sensory and motor modalities. While in the
non-verbal symbolic system the information is processed in images, in the
sub-symbolic one, rapid and complex computations are carried out in an
implicit continuous path. These computations contribute to recognize slight
facial expressions modifications, identify body movement or vocal quality
changes, and distinguish visceral states. The multiple code theory is in line
with the contemporary neurosciences (LeDoux, 2012; Damasio et Carvalho,
2013), suggesting that in the presence of the affective experience it is possible
to discriminate between emotions and feelings. Emotions occur at a
physiological and motor-expressive level, involving bodily systems and
subcortical and cortical somato-sensory brain areas. Feelings are based on
complex symbolization and cognitive processes related to the functioning of
prefrontal and associative cortices. The integration of emotions and feelings,
as well as of the verbal and non-verbal systems, depends on the so-called
referential processes, which transform the non-verbal symbolic and subsymbolic materials into words, and vice versa.
The referential processes are the core factors contributing to the
development, maintenance and promotion of health, since a deficit in these
processes generates dysfunctional conditions and pathologies, characterized

JADT’ 18

421

by multifactorial and bio-psycho-social etiology as well as marked
somatization. Among these dysfunctional conditions, alexithymia is a
psychological construct represented by impairment in cognitive-emotional
and affective processing (Bagby et al., 1994). It describes people with
deficiencies in identifying or describing subjective emotions or feelings,
difficulty in distinguishing between bodily sensations of emotional arousal
and feelings, and limited affect-related fantasy and imagery. People with
alexithymic traits have a tendency to focus on facts without affective
involvement rather than inner experiences, exhibiting a “concrete and realitybased cognitive style”. They often avoid social situations, seem cold, show a
lack of intimacy and warmth and are insecurely attached to others. Although
alexithymia is not a psychological disorder per se, it is associated with a low
quality of life and enhanced risk of psychological impairment and it is
present in a broad spectrum of psychosomatic disorders (Taylor et Bagby,
2004). Neuroimaging studies have indicated that people with high
alexithymic traits show less activation in brain areas associated with
emotional awareness and volumetric variations in brain areas associated
with emotional and somato-sensory and sensory-motor processing
(Laricchiuta et al., 2015a, and see for a literature review Laricchiuta et al.,
2015b). Therefore, the aim of this pilot study is to investigate a complex biopsycho-social pattern of verbal expression of the emotional experiences,
alexithymia levels and brain structure.
2. Data collection and analysis
A sample of 9 (males=5) healthy adult subjects of both sexes was recruited for
the pilot study at the IRCCS Fondazione Santa Lucia, Rome. Participants
were selected according to the following inclusion criteria: age between 18
and 70 years and suitability for structural Magnetic Resonance Imaging
(MRI) scanning. Exclusion criteria included the suspicion of cognitive
impairment or dementia; the subjective complaint of memory difficulties or
of any other cognitive deficit, regardless of interference with daily activities;
major medical illnesses; current or reported psychiatric or neurological
disorders; known or suspected history of alcoholism or drug dependence and
abuse; and MRI evidence of focal parenchymal abnormalities or cerebrovascular diseases.
To assess the cortical and subcortical structural measures, participants
underwent an imaging protocol that included standard clinical sequences
(FLAIR, DP-T2-weighted) and a volumetric whole-brain 3D high-resolution
T1-weighted sequence, performed with a 3 T Allegra MR imager, with a
standard quadrature head coil. Volumetric whole-brain T1-weighted images
were obtained in the sagittal plane using a Modified Driven Equilibrium

422

JADT’ 18

Fourier Transform (MDEFT) sequence (Echo Time/Repetition Time -TE/TR- =
2.4/7.92 ms, flip angle 15°, voxel size 1 x 1 x 1 mm3). All planar sequence
acquisitions were obtained in the plane of the anterior-posterior commissure
line. For the volumetric measures, T1-weighted images were processed and
examined using the SPM8 software, specifically the VBM8 toolbox running in
Matlab 2007b. For the cortical thickness, FreeSurfer imaging analysis suite
(v5.1.0) was used for cortical reconstruction of the whole brain. The
segmented, normalized, modulated and smoothed images were used for
analyses. Then, participants were interviewed by using the Toronto
Structured Interview for Alexithymia (TSIA, Bagby et al., 2006; Italian
version Caretti et al., 2011), composed of 24 items referred to four factors of
the alexithymia construct: the Difficulty in Identifying Feelings (DIF); the
Difficulty in Describing Feelings (DDF); the Externally Oriented Thinking
(EOT); and the Imaginal Processes (IP). Each item is assessed by a specific
open-ended question and its response is 3-point scored (coded ‘0’, ‘1’, or ‘2’).
The sum of scores results in a total score that ranges from 0 (low alexithymia
levels) to 48 (high alexithymia levels). The transcripts of the TSIA responses
were used to evaluate the linguistic structure by means of a multivariate
analysis. Namely, the nine TSIA transcripts resulted in a medium size corpus
of 62.792 tokens. In order to check whether it was possible to statistically
process data, two lexical indicators were calculated: the type-token ratio and
the hapax percentage (TTR= 0,10; Hapax= 51,0%). According to the large size
of the corpus both lexical indicators highlighted its richness and indicated the
possibility to proceed with the analysis. First, data were cleaned and preprocessed with the software T-Lab (Lancia, 2017) and keywords selected. In
particular, we used lemmas as keywords instead of type, filtering out the
lemma of the high rank of frequency and those of the low rank of frequency
lower to 9 occurrences (for keyword election see Cordella et al., 2014; Greco,
2016). Then, on the context units per keywords matrix, we performed a
cluster analysis with a bisecting k-means algorithm (Savaresi et Boley, 2004)
limited to ten partitions, excluding all the context units that do not have at
least two keywords co-occurrence. To finalize the text mining a
correspondence analysis (Lebart et Salem, 1994) on the keywords per clusters
matrix was made in order to explore the relationship between clusters and to
identify the latent dimensions setting the interviews.
Then parametric associations between TSIA scores and regional volumes or
cortical thickness, and between lexical scores (resulted by the correspondence
analysis) and TSIA scores or brain structural measures were calculated by
means of Pearson’s correlations in order to identify the possible direction and
extent of the linear relationship between the variables.

JADT’ 18

423

3. Main results
The results of the cluster analysis show that the 369 keywords selected allow
for the classification of 96.8% of the corpus. According to the theoretical
framework (Cordella et al., 2014) we choose the solution with four cluster.
The correspondence analysis detected three latent dimensions. In table 1, we
can appreciate how the clusters are placed in the factorial space produced by
three factors. The first factor represents the experience that could be personal
(negative pole) or social (positive pole); the second factor reflects the thought
that could be made on feelings (negative pole) or on a rational reasoning
(positive pole); and the third factor represents the aim of the thinking process
that could lead to make a speculation (negative pole) or a choice (positive
pole).
Table 1  Cluster coordinates on factors (the percentage of explained inertia
is reported between brackets under each factor).
Cluster Label
1

Think

CU classified

Factor 1

Factor 2

Factor 3

Experience

Thought

Aim

34,24%

Reasoning

Speculate

0,03

0,46

-0,64
Speculate

2

Feel

18,09%

Personal

Feeling

-0,13

-0,65

3

Relationship

23,15%

Social

4

Remember

24,51%

-0,18
Choice

0,52

0,08

0,24

Personal

Reasoning

Choice

-0,73

0,23

0,30

CU = context units classified.

The four clusters are of different sizes and reflect the general approach to the
emotional experience solicited by the TSIA. The first cluster reflects the
reasoning on the life event and hypothesis resulting in a rationale thinking
process; the second cluster highlights the capacity to reflect on the experience
of personal feelings; the third cluster represents the relationships
characterizing social life; and the fourth cluster gets back to memories,
reasoning on personal choices that were made (table 2).

424

JADT’ 18
Table 2  Cluster (the percentage of context units classified in the cluster
is reported between brackets).
Cluster 1

Cluster 2

Think

Cluster 3

Feel

Cluster 4

Relationship

Remember

keyword

CU keyword

CU keyword

CU keyword

pensare

104 sentire

100 persona

163 immaginare

CU
84

vedere

71 proprio

77 parlare

163 vedere

71

scrivere

45 riuscire

77 sentimento

149 prendere

64

amico

38 momento

67 provare

116 mettere

61

chiedere

32 capire

66 persone

114 positivo

39

chiamare

26 situazione

51 capire

100 casa

31

trovare

26 vivere

50 cercare

80 bello

28

tempo

21 piacere

50 situazione

71 problemi

27

ragazzo

20 succedere

33 amico

70 tornare

27

diverso

20 rabbia

32 trovare

45 portare

24

CU = context units classified in the cluster.

The correlation coefficient between the lexical scores of the three factors
(resulted from the correspondence analysis) and the TSIA scores, as well as
the brain volumes and the cortical thickness are reported in table 3. Namely,
DIF, DDF, EOT and total alexithymia scores were positively associated with
the second factor. At neurobiological level, the first factor was negatively
associated with volumes of right caudate and thickness of right medial
orbitofrontal cortex, as well as positively associated with the thickness of
right lateral occipital cortex. Finally, the third factor was negatively
associated with volumes of middle anterior, central and middle posterior
cerebral cortices, as well as with thickness of right postcentral cortex and left
posterior cingulate cortex. Conversely, the third factor was positively
associated with thickness of the right posterior cingulate cortex. Finally, the
IP scores (TSIA) were positively correlated (r = 0.72; p = 0.03) with the left
entorhinal cortical thickness values.
4. Discussion
Although this is a pilot study and it is not possible to generalise the findings,
the present data suggest that the methodology proposed (in order to identify
the connections among verbal expression, alexithymia levels and brain
structure) seems to be promising for a deeper understanding of the biopsycho-linguistic connections. In fact, results indicate that high alexithymia
scores were associated with a thought modality characterized by a rational
(and not emotional) reasoning. Furthermore, the tendency to be engaged in

JADT’ 18

425

personal (not social) experience was associated with large volumes of right
caudate and thickness of right medial orbitofrontal cortex.
Table 3  Correlation coefficients between lexical factors and TSIA scores
as well as cerebral structure values.

Variables

Factor 1 Factor 2 Factor 3

Difficulty Identifying Feelings (TSIA Factor 1)
Difficulty Describing Feelings (TSIA Factor 2)
Externally Oriented Style of Thinking, (TSIA Factor 3)
TSIA Total score
Right-Caudate
Right Hemisphere Medialorbitofrontal Thickness
Right Hemisphere Lateraloccipital Thickness
Left Hemisphere Posteriorcingulate Thickness
Mid Posterior Cortical Cortex
Mid Anterior Cortical Cortex
Central Cortical Cortex
Right Hemisphere Postcentral Thickness
Right Hemisphere Posteriorcingulate Thickness

.83
.71
.68
.77
-.71
-.78
.70

In the table are reported only the correlation coefficients with a p<0,05.

Conversely, the tendency to be engaged in social (not personal) experiences
was associated to great thickness of right lateral occipital cortex. The
speculative thinking processes (negative pole of the third factor) was
associated with large volumes of middle anterior, central and middle
posterior cerebral cortex, as well as with great thickness of right postcentral
cortex and left posterior cingulate cortex. Finally, thinking processes related
to a choice was associated with great thickness of the right posterior
cingulate cortex.
Overall the study indicates that the organizational factors of thought and
language, conveying also the emotional meaning of the text, are related to the
structure of cerebral areas involved in somato-sensory associative processes
(postcentral and lateral occipital cortices), in emotional awareness (entorhinal
and posterior cingulate cortices), and in emotional control and feelings
(orbitofrontal cortex). Just such functions are compromised in the presence of
high levels of alexithymia, because an altered referential process can lead to
somato-sensorially perceive but not-verbally express the emotions.
Furthermore, in the present study most of associations were found between
first and third factor (resulted from the correspondence analysis) and the
macro-structural measures in the right brain hemisphere, totally fitting the

-.71
-.75
-.75
-.78
-.78
.68

426

JADT’ 18

proposal of Bucci (1997) that suggests the right hemisphere as the
neurophysiological substratum underlying the processing of emotional
information and referential process. On this vein, alexithymia may be
considered an embodiment process related to altered perception of
physiological correlates (viscero- and somato-motor responses) of the
emotional activation resulting in a deficit in the emotional awareness. In fact,
a dysfunctional referential process can lead to a lack of words for the
emotions, up to being without symbols for the somatic states.
References
Bagby R.M., Parker J.D. and Taylor G.J. (1994). The twenty-item Toronto
Alexithymia Scale--I. Item selection and cross-validation of the factor
structure. J Psychosom Res. 38(1):23-32.
Bagby R.M., Taylor G.J., Parker J.D. and Dickens S.E. (2006). The
development of the Toronto Structured Interview for Alexithymia: item
selection, factor structure, reliability and concurrent validity. Psychother
Psychosom. 75(1):25-39.
Bucci W. (1997). Psychoanalysis and cognitive science: A multiple code
theory. Guilford Press.
Caretti V., Porcelli P., Solano L., Schimmenti A., Taylor G.J. and Bagby R.M.
(2011). Reliability and validity of the Toronto Structured Interview for
Alexithymia in a mixed clinical and nonclinical sample from Italy.
Psychiatry Research, 187:432-436.
Cordella B., Greco F. and Raso A. (2014). Lavorare con Corpus di Piccole
Dimensioni in Psicologia Clinica: Una Proposta per la Preparazione e
l’Analisi dei Dati. In Nee E., Daube M., Valette M. and Fleury S., editors,
Actes JADT 2014 (12es Journées internationales d’Analyse Statistque des
Données Textuelles, Paris, France, Juin 3-6, 2014), pp. 173-184.
Damasio A. and Carvalho G.B. (2013). The nature of feelings: evolutionary
and neurobiological origins. Nat Rev Neurosci., 14(2):143-152.
Greco F. (2016). Integrare la disabilità. Una metodologia interdisciplinare per
leggere il cambiamento culturale. Franco Angeli.
Lancia F. (2017). User’s Manual: Tools for text analysis. T-Lab version Plus 2017.
Laricchiuta D., Petrosini L., Picerni E., Cutuli D., Iorio M., Chiapponi C.,
Caltagirone C., Piras F. and Spalletta G. (2015a). The embodied emotion in
cerebellum: a neuroimaging study of alexithymia. Brain Struct Funct.,
220(4):2275-2287.
Laricchiuta D., Lai C. and Petrosini L. (2015b). Alexithymia: From
Neurobiological Basis to Clinical Implications. In Bryant M.L., editor,
Handbook on Emotion Regulation: Processes, Cognitive Effects and Social
Consequences, Nova Science Publishers.

JADT’ 18

427

Lebart L. and Salem A. (1994). Statistique Textuelle. Dunod
LeDoux J. (2012). Rethinking the emotional brain. Neuron, 73(4): 653-676.
Savaresi S. M. and Boley D. L. (2004). A comparative analysis on the bisecting
K-means and the PDDP clustering algorithms. Intelligent Data Analysis,
8(4): 345-362.
Taylor G.J. and Bagby R.M. (2004). New trends in alexithymia research.
Psychother Psychosom., 73(2):68-77.

428

JADT’ 18

Icone gay: tra processi di normalizzazione e di
resistenza. Ricostruire la semantica degli hashtag
Gevisa La Rocca1, Cirus Rinaldi2
1

Kore University of Enna – gevisa.larocca@unikore.it
2University of Palermo – cirus.rinaldi@unipa.it

Abstract 1
The mediatization of emotions emerges as an affordance of social media, the
study of which involves paying attention to digital practices and the
formation of the sense of public affection, of connected audiences expressing
their participation through expressions of sentiment. This happens both for
the great events and for the daily demonstrations of support or of its
negation. Here we choose to analyze the tweets in which the fans express
their opinions on the participation in the reality shows of their “icons”:
Vladimir Luxuria and Cristiano Malgioglio. To reconstruct the hashtag
semantics we use: the NodeXL software for network analysis and Iramuteq
for the extraction of lexical worlds.
Abstract 2
La mediatizzazione delle emozioni emerge come una affordance dei social
media, il cui studio implica il porre attenzione alle pratiche digitali e alla
formazione del senso dell’affetto pubblico, dei pubblici connessi che
esprimono la loro partecipazione attraverso le espressioni del sentimento.
Questo accade tanto per i grandi avvenimenti che per le quotidiane
manifestazioni di sostegno o della sua negazione. Qui si sceglie di analizzare
i tweets in cui i fans si esprimono in merito alla partecipazione ai reality
show delle loro «icône» : Vladimir Luxuria e Cristiano Malgioglio. Per
ricostruire la semantica degli hashtag si usa il software NodeXL per la
newtork analysis e Iramuteq per l’estrazione dei mondi lessicali.
Keywords: LGTB, rappresentazione sociale, semantica degli hashtag,
network analysis.
1. Introduzione
All’interno della più generale questione della rappresentazione dei tipi gay
(Dyer, 1992), la riflessione teorica ha rilevato la complessità che si cela dietro
una facciata che è solo apparentemente semplice. In particolare, la
costruzione di tipi e icone LGBT offre codificazioni simboliche ampie e
sovrasignificazioni persino contradditorie, dal momento che la produzione

JADT’ 18

429

culturale subculturale e comunitaria gay non soltanto ha da sempre articolato
temi e significati già esistenti (Hall, 1996), ma è stata a sua volta incorporata
in
rappresentazioni
eteronormative.
Guardando
alle
principali
trasformazioni sociali in atto relative ai processi di acquisizione della
cittadinanza nei contesti neo-liberisti e alle istanze di «normalizzazione»
tuttora in discussione all’interno della sfera pubblica e del movimentismo
LGBTQI1 (matrimonio egualitario, coppie non-bianche, dis-abilità e
coscrizione militare), bisogna prestare attenzione ai processi di costruzione
simbolica di (nuove e inedite) maschilità complici assimilate all’interno della
produzione del consumo neo-liberista2. Il neo-liberismo pone la propria
enfasi sulla libertà individuale e sui diritti, e sulla regolazione degli interventi
centrali dello Stato: una delle principali modalità di governance neo-liberista
consiste nelle diverse tecniche di normalizzazione, attraverso le quali
vengono individuate e riprodotte norme di comportamento specifiche
indicate dai governi liberisti che i “cittadini” dovranno interiorizzare a fini
auto-regolativi3. Di particolare interesse per la presente trattazione, i modelli
di cittadinanza sostenuti sono finalizzati a rafforzare le dimensioni normative
di genere e sessualità (matrimonio, coscrizione militare, etc.); la
rivendicazione di «eguaglianza» si è fortemente intrecciata con la negazione e
il distanziamento da parte degli omosessuali dalle tradizionali
rappresentazioni che li associavano a individui «amorali», «inferiori»,
«subordinati», e «peccaminosi». La tentazione di essere normali sta producendo
effetti paradossali sulla visibilità dei gay, perché se da un lato assistiamo a
quanto viene definito de-omosessualizzazione o eterosessualizzazione delle
omosessualità4, dall’altro invece, osserviamo un rafforzamento delle
dicotomie e delle gerarchie di genere che circoscrivono fortemente
l’omosessualità attraverso l’interiorizzazione di un’economia eteronormativa

Per un approfondimento si rinvia a Warner, M. (1999). The trouble with the
normal. Sex, politics and the ethics of queer life, New York, The Free Press.
2 Cfr. Duggan, L. (2003). The Twilight of Equality?: Neoliberalism, Cultural Politics,
and the Attack on Democracy, Boston, Beacon Press e Chasin, A. (2000). Selling Out: The
Gay and Lesbian Movement Goes to Market, New York, St. Martin’s Press.
3 Richardson, D. (2004). Locating sexualities: from here to normality, «Sexualities», 7,
4, 391- 411.
4 Bersani, L. (1998). Homos (1995), tr. It. Homos, Milano, Pratiche Editrice, 38, 131.
Sui processi di eterosessualizzazione, si consideri la critica fornita in Ingraham, C.
(2012), Atti innaturali: disciplinare l’eterosessualità, in C. Rinaldi (ed.). Alterazioni.
Introduzione alle sociologie delle omosessualità, Milano-Udine, Mimesis, 97-119.
1

430

JADT’ 18

del desiderio5. Il presente lavoro intendere problematizzare le forme di
tipizzazione LGBT a partire dall’analisi delle differenti rappresentazioni che
hanno avuto le partecipazioni ai reality show di personaggi pubblici,
dichiaratamente parte della comunità LGBT, come Vladimir Luxuria e
Cristiano Malgioglio.
2. Processi di normalizzazione e di resistenza
I processi di assimilazione eteronormativa da parte delle comunità gay, e la
conseguente costruzione normalizzante di alcune soggettività LGBT, hanno
portato a nuove gerarchie di desiderabilità all'interno delle medesime
comunità LGBT. Uno dei principali effetti è la depoliticizzazione e la
privatizzazione della cultura movimentistica LGBT, ancorata nella
riproduzione e nella protezione del domestico (la famiglia/le famiglie e la
nazione o il Nuovo Ordine Globale), nella riproduzione iperbolica della
dicotomia di genere normativa, nel consumo di tali rappresentazioni e nella
partecipazione attiva al nuovo ordine globale. A corroborare questa
prospettiva, i processi politici globali in atto (sia di tipo securitario che di
valorizzazione di alcune soggettività) – insieme alla globalizzazione delle
maschilità nelle forme della loro mercificazione e delle immagini veicolate
dai mass media transnazionali – sembrano fondarsi sulla definizione anche
di standard omonormativi e omonazionalisti che non soltanto si
contrappongono al nemico da combattere (la figura ambigua dell’indecoroso,
del terrorista o dell’immigrato) ma che si discostano, nel contempo, da tutte
quelle corporeità terroristiche che lo standard nazionalista individua come
eccessive/eccedenti (a meno che anche esse non scelgano di diventare docili ).
Uno degli effetti principali delle nuove forme di normalizzazione, è un
diretto rafforzamento degli assetti di genere normativi, un rafforzamento
della maschilità e dei suoi indicatori in termini generali e, soprattutto, una
diretta complicità nella marginalizzazione degli omosessuali effeminati non
soltanto nella società più vasta a ma anche all’interno delle comunità gay.
La costruzione delle nuove maschilità si fonda principalmente
sull’epurazione del femminile e dell’effeminatezza, attraverso una serie di
contesti e pratiche – dalle rappresentazioni nei mass media, al marketing
della «pink economy» sino agli annunci personali nei social network, nelle
app per incontri – che attestano quanto le maschilità omosessuali egemoni e
normative considerino accettabili e giustificabili le condotte e gli

5 Martino, W. (2006). Straight-acting masculinities: normalization and gender
hierarchies in gay men’s lives, in Kendall, C., Martino, W. (eds.), Gendered outcasts and
sexual outlaws. Sexual oppression and gender hierarchies in queer men’s lives, The Haworth
Press, New York, 35-60.

JADT’ 18

431

atteggiamenti effeminofobi. La storica associazione dell’omosessualità con
l’effeminatezza e con il femminile di fatto non stimola un ripensamento della
maschilità omosessuale (e delle forme di strutturazione della maschilità in
generale), ma la fa arroccare su posizioni difensive e distanzianti e su forme
contro-reattive e contro-culturali che, di fatto, riproducono la norma. La
paradossalità della maschilità omosessuale consiste nella ricerca
dell’affrancamento culturale attraverso la proposta di una maschilità vera,
senza metterne in discussione però misoginia, patriarcato e checcofobia. Una
maggiore visibilità delle omosessualità implica un costo da pagare, nello
specifico una differenza che si dissolve in eguaglianza e nella costituzione di
un blocco egemonico in grado di rendere meno percettivamente visibili le
divisioni di genere del patriarcato.
Se pensiamo in termini relazionali le maschilità (eterosessuali e omosessuali),
una loro ibridazione giustificata su base consumistica, se permette alle prime
di apparire in modo meno tradizionale e rigido e alle seconde di acquisire in
risorse simboliche della tradizione virilista, tuttavia riproduce –attraverso
entrambe – l’illusione della scomparsa del (dividendo del) patriarcato. Dal
momento che entrambe le forme di soggettivazione maschile partecipano in
modo selettivo e differenziale all’accaparramento e all’appannaggio di
risorse materiali e simboliche egemoniche, esse non saranno definibili
esclusivamente in termini oppositivi, piuttosto si alimenteranno e
rafforzeranno in maniera reciproca. Questa polarizzazione virilista non
soltanto perpetra forme di annichilimento nei confronti della manifestazione
della maschilità frocia ma impedisce, di fatto, anche la eventuale costituzione
pubblica di maschilità eterosessuali al di fuori delle rappresentazioni
eteronormative. Tuttavia, il costo da pagare perché la maschilità gay sia
accettata socialmente è la negazione del suo carattere sessuale, della sua
diversità in termini di desideri erotici, di pratiche, della materialità della sua
espressione sessuale. L’assimilazione prevede, di conseguenza, la
riproduzione dei sistemi di divisione di genere, di classe e di età presenti
nella società più vasta, attraverso la riproduzione di modelli di consumo
esistenti all’interno del più vasto ordine globale.
3. La mediatizzazione e gli hashtag
Con l’introduzione del termine e del concetto di mediatizzazione ci si
riferisce a quel processo in base al quale le istituzioni sociali e culturali e le
modalità di interazione sono cambiate, cambiano e cambieranno come
conseguenza della crescita dell’influenza dei media, tenendo conto, però,
delle circostanze, ovvero di come mutano la cultura e la società. Si tratta di
quel costante contatto comunicativo con gli altri, la cui esplorazione avviene
in modi del tutto inediti (Cardoso, 2008; Boccia Artieri, 2012; Colombo, 2013),

432

JADT’ 18

che trasforma la condizione del vissuto in un nuovo orizzonte di senso
sociale (Boccia Artieri et al., 2017) producendo una metamorfosi delle
relazioni sociali. Si tratta di quell’insieme di pratiche o di habitus
caratterizzate da una regolarità dell’agire in relazione a specifici bisogni, che
porta con sé un intero mondo di capacità, vincoli e potere (Couldry, 2012).
Non è difficile accettare che l’uso di un mezzo di comunicazione per un
periodo di tempo continuato determini il carattere stesso della conoscenza da
comunicarsi, e che a causa della sua stessa pervasività porti all’emergere di
una nuova civiltà, ovvero renda possibile lo strutturarsi di una forma
particolare con cui si manifesta la vita materiale, sociale e spirituale di un
popolo (Innis, 1951; Ong, 1982; La Rocca, 2017). L’aspetto sociale dei social
media è, ormai, evidente da sé; essi sono parte di una società in cui svolgono
una pluralità di funzioni di intermediazione (Colombo, 2003; 2013), essi,
infatti, sembrano essere pensati per rendere possibili le collaborazioni
partecipative, cioè dal basso, e allo stesso tempo, alimentano una
socievolezza di simmeliana memoria, che è riassumibile nelle caratteristiche
oggi individuate da Peter Dahlgren (2009) nella talkative society.
Si tratta di considerare emoticons, emoji, hashtag, commenti, rimandi, foto, link,
video, tutti quegli strumenti che consentono di ricollocare il testo nelle
intenzioni enunciative di chi lo ha posto in essere o condiviso. Sono fenomeni
di cui si occupa con più interesse la discourse analysis, ma che non possono
essere ignorati se l’obiettivo è un’analisi del contenuto digitale dei new e social
media. Questo per una ovvia considerazione, com’è possibile limitare
l’osservazione solo allo scritto e non estenderlo ai suoi elementi accessori, se
l’obiettivo è conoscere il senso di quanto viene detto a proposito di un dato
argomento o fenomeno in rete? Solo in questo modo l’analisi del contenuto si
apre alla possibilità di includere il linguaggio utilizzato come una meta
risorsa tecnologizzata. È chiaro che intesa in questo modo l’analisi del
contenuto si avvicina più alla discourse-ethnografic (Androutsopoulos, 2010;
2011) piuttosto che a un’analisi delle frequenze; perché innanzitutto
ricostruire i percorsi e le emozioni di un topic online non è semplice e questo
è dovuto alla struttura della grammatica e della sintassi della costruzione dei
messaggi, alla commistione linguistica che quindi richiede nel momento
dell’analisi un software capace di elaborare testi in più lingue, ma anche di
ricodificare le emoticons e di valutare in relazione a esse le intenzioni del
testo. Si tratta di sviluppare un approccio di analisi del contenuto multimodale,
indicando con questo termine come anche in questo settore sia necessario
attuare ciò che è avvenuto nello studio del discorso (Jewitt, 2014; Kress, van
Leeuwen, 2001), dove si presta attenzione al modo in cui il linguaggio
interagisce con altri sistemi semiotici; sostituendo al “linguaggio”, la
costruzione del contenuto, che inevitabilmente interagisce con gli altri sistemi

JADT’ 18

433

semiotici. In questo caso è chiaro che un approccio in cui è il ricercatore a
svolgere manualmente tutte queste operazioni o a ricodificare le espressioni
riconducendole a categorie di senso condivise diventa la soluzione più
opportuna. Ma appare altresì veritiero che un lavoro di questo tipo, a meno
di ricevere un grosso finanziamento a supporto della ricerca, rimane sempre
legato a un numero limitato di osservazioni.
3.1. Per una ricostruzione della semantica degli hashtag
Si sceglie quindi di lavorare in una duplice ottica: da un lato si scaricano i
tweets relativi alla partecipazione ai reality show quali: L’isola dei famosi e
Grande Fratello VIP per i due soggetti individuati, e si ricostruisce il network
dei topic, dall’altro si esplorano mediante l’analisi lessicale i contenuti dei
tweets individuati come perni connettori.
Il ventaglio dei sentimenti che gli individui agganciano a questi messaggi è
ampio e variegato e attiene alla natura umana, è possibile interpretarlo
seguendo l’impostazione data da Korina Giaxoglou e Katrin Döveling (2018)
nella loro special issue dedicata alla mediatizzazione delle emozioni sui social
media. La mediatizzazione delle emozioni emerge come una effordance dei
social media il cui studio implica il porre attenzione alle pratiche digitali e
alla formazione del senso dell’affetto pubblico, dei pubblici connessi (Boyd,
2010) che esprimono la loro partecipazione attraverso le espressioni del
sentimento (Papacharissi, 2016). Finora, gli studi hanno esaminato la
formazione e la veicolazione dei sentimenti in rete legandoli a grandi
avvenimenti sociali, momenti storici che prefigurano cambiamenti epocali. Si
tratta, senza dubbio, di storie di connessione ed espressione, dove gli hashtag
servono come significanti vuoti che invitano ad una identificazione
ideologica a vasto orientamento polisemico (Colleoni, 2013; Papacharissi,
2016). I post promossi dai singoli, individuati e volti a sostenere o denigrare,
quelle che qui sono state individuate come icône gay, si sostanziano o forse
meglio si foraggiano sicuramente di un senso emotivo che è personale.
Come per Zizi Papacharissi (2016) anche le nostre interpretazioni, in questo
contesto, sono guidate dalla comprensione dell’affetto come una forma di
intensità pre-emotiva soggettivamente sperimentata e connessa a processi di
premeditazione o anticipazione di eventi prima del loro verificarsi. Ci sono
radici emotive legate alla percezione e sperimentazione dell’affetto che
provengono da contesti socioculturali cui gli individui appartengono, in
questo senso le emozioni mediatizzate sono delle forme espressive di culture
più profonde. Per realizzare l’analisi dei network degli hashtag si utilizza NodeXL,
per l’analisi lessicale Iramuteq.

434

JADT’ 18

References
Androutsopoulos J. (2010). Localising the global on the participatory web:
Vernacular spectacles as local responses to global media flows. In
Coupland, N. (ed.), Handbook of Language and Globalization (203-231).
Oxford: Wiley-Blackwell.
Androutsopoulos J. (2011). From variation to heteroglossia in the study of
computer-mediated discourse. In Thurlow, C. e Mroczek, K. (eds.), Digital
Discourse: Language in the New Media (277–298). London: Oxford
University Press.
Boccia Artieri G. (2012). Stati di connessione. Milano: Franco Angeli.
Boccia Artieri G., Gemini L., Pasquali F., Carlo S., Farci M., Pedroni M. (2017).
Fenomenologia dei social network. Presenza, relazioni e consumi mediali degli
italiani online. Milano: Guerini.
Boyd D. (2010). Social network sites as networked publics: Affordances,
dynamics, and implications. In Papacharissi, Z. (ed.), A Networked Self:
Identity, Community, and Culture on
Social Network Sites (39-58). New York: Routledge.
Cardoso G. (2008). Preference for online social interaction: A theory of
problematic Internet use and psychosocial wellbeing. Communication
Research, 30: 625-648.
Chasin A. (2000). Selling Out: The Gay and Lesbian Movement Goes to Market.
New York: St. Martin’s Press.
Colleoni E. (2013). Beyond the differences: The use of empty signifiers as
organizing device in the #occupy movement. In Proc. Of Material
Participation: Technology, the Environment and Everyday Publics, Università
di Milano, Maggio.
Colombo F. (2003). Introduzione allo studio dei media. Roma: Carocci.
Colombo F. (2013). Il potere socievole: storia e critica dei social media. Milano:
Bruno Mondadori
Couldry N. (2012). Media, Society, World: Social Theory and Digital Media
Practice. Cambridge: Polity.
Dahalgren P. (2009), Media and Political Engagement. Citizens, Comunication,
and Democracy. New York: Cambridge University Press.
Duggan L. (2003). The Twilight of Equality?: Neoliberalism, Cultural Politics, and
the Attack on Democracy. Boston: Beacon Press.
Giaxoglou K., Döveling K. (2018). Mediatization of emotion on social media:
forms and norms in digital mourning practices. Social Media + Society, pp.
1-4.
Ingraham C. (2012). Atti innaturali: disciplinare l’eterosessualità. In Rinaldi,
C. (Ed.), Alterazioni. Introduzione alle sociologie delle omosessualità (97-119).
Milano-Udine: Mimesis.

JADT’ 18

435

Innis H. A. (1951). Empire and communication. Oxford: University of Oxford
Press.
Jewitt C. (2014). The Routledge Handbook of Multimodal Analysis. London:
Routledge.
Kress G., Van Leeuwen T. J. (2001). Multimodal Discourse: The Modes and Media
of Contemporary Communication. London: Arnold.
La Rocca G. (2017). Cantami o diva. La bazzecola del dicunt in rete. Sociologia
della comunicazione, 53: 56-74.
Martino W. (2006). Straight-acting masculinities: normalization and gender
hierarchies in gay men’s lives. In Kendall, C., Martino, W. (Eds.), Gendered
outcasts and sexual outlaws. Sexual oppression and gender hierarchies in queer
men’s lives (35-60). New York: The Haworth Press.
Ong W. J. (1982). Orality and Literacy. The Tecnologizing of the Ward. London:
Routledge.
Papacharissi Z. (2016). Affective publics and structures of storytelling:
Sentiment, events and mediality. Information, Communication & Society, 19:
307–324.
Puar J. (2007). Terrorist Assemblages: Homonationalism in Queer Times. Durham,
N.C.: Duke University Press.
Richardson D. (2004). Locating sexualities: from here to normality. Sexualities,
7(4): 391- 411.
Rinaldi C. (2012). Globalizzazione della maschilità e maschilizzazione dei
processi globali. In G.E.M. Scichilone (Ed.), L’era globale: linguaggi,
paradigmi, culture politiche (173-189). Milano: Franco Angeli.
Rinaldi C. (2007). De-gener(azioni): riflessioni per una sociologia del
transgenderismo. In Antosa, S. (a cura di), Spazio e identità queer (127-148).
Omosapiens vol. 2. Roma: Carocci.

436

JADT’ 18

Looking for topics: a brief review
Ludovic Lebart1
1

Telecom-ParisTech – ludovic@lebart.fr

Abstract
This paper presents a brief review of several endeavors to identify latent
variables (axes or clusters). When dealing with textual data, these latent
variables (clusters or axes) are sometimes designated ex ante by the term
“topic”. The first attempts to identify interpretable latent variables dates back
to factor analysis at the beginning of last century. Recent years have
witnessed a series of algorithmic attempts such as non-negative matrix
factorization (NMF) or Latent Dirichlet Allocation (LDA). In the meantime,
latent variables are also identified through several hybridizations and
synergies of principal axes methods and clustering techniques. A single
medium-sized classical corpus (Shakespeare’s 154 Sonnets) will serve as a
benchmark to sketch and compare in a compact way some characteristic
features of several methods.
Keywords: Topic Modelling, NMF, LDA, Correspondence Analysis, Factor
Analysis, Clustering.
1. Introduction
There is a profusion of new disciplines around the industrial applications
involving texts, with subsequent proliferations of tools and disparities of
terminologies. There are also disparities in the attitude towards the texts,
sometimes influenced by the availability and the user-friendliness of
software. The problems entailed by huge sets of newsgroup or tweets are
quite different from those encountered when dealing with literature, political
discourses, psychological surveys. Because they are well known, translated
in almost every language, deeply studied and commented, we will use
Shakespeare’s Sonnets as a benchmark to briefly compare the ability of
several techniques to recognize topics in a corpus.
2. An outline of the contents of Shakespeare’s sonnets
The 154 sonnets of William Shakespeare deal with themes such as love,
friendship, effects of time, beauty, treason, lust, death. Note that the
definition of topics in Text Mining is a pragmatic one, and may also recover
the concepts of theme and motif.
2.1. Theme, Topic, Motif

JADT’ 18

437

Usually, a topic is an objective explanation of the subject matter, whereas a
theme represents the deeper underlying message. A motif is simply a
recurring idea or pattern used to reinforce the main theme. Schematically,
topics answer the questions:, "What's the story about? Who? What? How?"
and themes answer: "Why was the story written?". Topics in literature are
easier to identify than themes.
Three main contiguous series of sonnets are generally recognized as three
dominant themes:
Sonnets 1 to 17: (Procreation). These sonnets celebrate the beauty of a young
man who is urged by the poet to marry so as to perpetuate that beauty.
Sonnets 18 to 126: (Young man). This longest sequence concerns the same
young man (not definitively identified), the destructive effect of time, the
force of love, friendship and poetry.
Sonnets 127 to 154: (Dark Lady).These sonnets are mostly addressed to a dark
haired woman, not without some irony and cynicism (the two last sonnets
153 and 154 are specific epigrams in an ancient style; they should deserve in
fact a specific category).
2.2. Eight themes derived from expert commentaries
The themes Young man and Dark lady could contain five sub-themes. While
the first theme (Procreation) remains untouched, the new Young man and Dark
lady themes will comprise only those sonnets which were not assigned to the
five new categories below (Absence, Storm, Rivalry, Death, Eternal poetry).
Table 1. List of eight a priori themes/topics with the corresponding sonnets numbers

Procreation
YoungMan
DarkLady
Absence
Storm
Rivalry
Death
Etern_poetry

1 - 17
20-25, 33-38, 40-42, 46, 47, 49, 53-55, 59-60,62-70, 75-77, 88106, 108-112, 115-125,
127-136, 139, 140, 143-146, 153,154
26-32, 39, 43-45, 48, 50-52, 56-58, 61, 113-114
141,142,147-152
78-87
71-74
18,19,81

The partition of sonnets given in Table 1 is inspired by the works of Alden
(1913) and Paterson (2010) but not explicitly mentioned by these authors.
Figure 1 shows however that, after a blind correspondence analysis ignoring
these themes, most of their locations are statistically significant on the
principal plane of visualization.

438

JADT’ 18

Figure 1. Locations of 7 themes/topics in the principal plane of the correspondence analysis of
the lexical table (154 sonnets x 173 words, [min. frequency = 10]) as supplementary
categorical variables. Conservative bootstrap confidence ellipses [drawing with replacement of
sonnets] show the significant distances between several pairs of a priori themes. Note that the
theme “Eternal Poetry”, too much overlapping with others, is missing in this graphical
display.

Evidently, the following attempts to find topics into the corpus of sonnets
will ignore that a priori partition into themes. We do not expect either to
retrieve automatically these themes.
However, the knowledge of these themes issued from literary criticism can
provide us with a template for reading and interpreting the results more
easily. Note that statistical tools mainly based on frequencies detect almost
indifferently topics, themes or motifs.
3. Six selected methods for topic research
Among the six procedures selected in the present paper, three (RFA, FCA,
LOA, LSA) make use of the Singular Values Decomposition (SVD). The
remaining two methods (NMF and LDA), less geometrical, involve a specific
model and more complex algorithms.
RFA (Rotated Factor Analysis) is historically the first attempt to identify

JADT’ 18

439

unobserved “latent factors” (Thurstone, 1947, after the pioneering papers of
Spearman,1904, and of Garnet, 1919). RFA involves SVD in one of the most
popular algorithms known as Principal Factor Analysis. In this case, the
topics are the words characterizing each kept factors. Initially conceived for
numerical values, it could be adapted to sparse frequency tables. [R packages
‘psych’ and ‘GPArotation’].
FCA (Fragmented Correspondence Analysis), in the vein of ALCESTE
methodology (Reinert, 1986), is based on the CA of fragments of texts [in our
case 7 consecutive lines, i.e.: half a sonnet), cf. Lebart (2012). The principal
axes of CA serve to cluster these fragments (hybrid clustering using
Hierarchical Classification –Ward criterion – and k-means). At the end of the
process, the topics are defined by the series of words that characterize each
cluster (software ‘DtmVic’).
LOA (LOgarithmic Analysis) (Kazmierczak, 1985) is similar to Spectral
Mapping (Lewi, 1976) thanks to a difference of weighting. Both methods, like
CA, comply with the principle of distributional equivalence (stability of the
results vis-à-vis fusions of similar columns or rows). Applied to contingency
or frequency tables, LOA often produces results similar to those of CA, with
less sensitivity towards outliers as a consequence of the logarithmic
shrinkage. A clustering of sonnets (similar to that of FCA) is then performed.
The topics are then the words characterizing each cluster (software:
‘DtmVic’).
LSA (Latent Semantic Analysis (or Indexing), Deerwester et al., 1990) which is
basically a SVD of the matrix of Tf.Idf coefficients (Term frequency x Inverse
of document frequency). A clustering of sonnets (similar to that of FCA and
LOA) is then performed. The topics are then the words characterizing each
cluster. (R package ‘lsa’ [Fridolin Wild] and ‘DtmVic’)
In the domain of text analysis, the two following methods belong more
specifically to the field of “Topic Modelling”.
NMF (non-negative matrix factorization) starts with an equation that
reminds Singular Values Decomposition (SVD): Decomposition of a data
matrix A as the product of two matrices of lower rank, B and C: A = B C. The
marked difference lies in a constraint of positivity of the coefficients of B and
C (those of A being already supposed > 0) (Lee and Seung, 1999; Berry et al,
2007; after Paatero &Tapper, 1994. See also Gaujoux, 2010). In the topic
modeling context, the main output of NMF is a set of topics characterized by
list of words (software ‘scikit-learn’ [Python] by Grisel O., Buitinck L., Yau
C.K; In: Pedregosa et al., 2011).
LDA (Latent Dirichlet Allocation) (Blei et al., 2003; Griffiths et al., 2007) is a
generative statistical model (involving unobserved topics, words, and
document) devised to uncover the underlying semantic structure of a

440

JADT’ 18

collection of texts (documents, supposed to be a mixture of a small number of
topics). The method is based on a hierarchical Bayesian analysis of the texts.
(package R: ‘topicmodels’, and software ‘scikit-learn’ [Python]).
At this stage, we have limited our investigation to six techniques out of a
great number of approaches likely to identify topics. Among these
approaches let us mention the direct use of CA without fragmentation of the
texts, the techniques of clustering (used in FCA and LOA) which contain
many more methods and variants, the already mentioned Alceste
methodology (Reinert, 1986). The present piece of research evidently needs to
be extended. In fact, each method involves also a series of parameters
(threshold of frequency for the words; preprocessing options such as
lemmatization/stop words; size of fragments or context units, number of
iterations). The following experiment limited to six methods will be tersely
summarized. A thorough investigation would need many more pages.
4. Excerpts from the list of 49 topics (limited to two topics per method)
The number of topics detected by each of the six selected methods varies
between six and ten. Only two topics are printed below for each method.
4.1 Rotated Factor Analysis (Rotation Oblimin). (2 topics out of 6)
RFA1 eyes see bright lies best form say days
RFA2 beauty false old face black now truth seem
4.2 FCA (Fragmented Correspondence Analysis) (2 topics out of 7)
FCA1 beauty truth muse age youth praise old eyes glass long seen lies false
time days
FCA2 night day bright see look sight
4.3 Logarithmic Analysis (Spectral mapping) (2 topics out of 8)
LOA1 summer away youth sweet state hand seen age rich beauty time hold
nature death
LOA2 pen decay men live earth verse muse once life hours make give gentle
death
4.4 Latent Semantic Analysis (2 topics out of 8)
LSA1 time heart beauty more one eyes eye now myself art still sweet world
LSA2 end grace leave words lie spirit change shame self could ever decay
write
4.5 NMF topics (2 topics out of 10)
NMF0: love true new hate sweet dear say prove lest things best like ill let
know fair soul
NMF1: beauty fair praise art eyes old days truth sweet false summer nature
brow black live
4.6 Latent Dirichlet Allocation LDA (2 topics out of 10)
LDA0 summer worse praise nature making time like increase flower let copy

JADT’ 18

441

rich year die LDA1 sing sweets summer hear love music eyes bear single
confounds prove shade eternal.
5. A synthesis of produced topics
How to compare the complete lists of topics, since neither the order of topics,
nor the order of words within a topic are meaningful? We deal here with real
‘bags of words’ exemplified by the excerpts of lines in section 4. We will add
the eight a priori themes defined in table 1. Each a priori theme corresponds to
a subset of sonnets. That subset will be described by its characteristics words.
We can then perform a clustering of these 57 topics/themes (49 + 8). The
technique of additive trees (Sattath and Tversky, 1977; Huson and Bryant,
2006) seems to be the most powerful tool for synthesizing in compact form
these 57 topics/themes (figure 2). Let us recall one important property of
additive trees: the real distance between two points can be read directly on
the tree as the shortest path between the two points.
Ideally, we expect to find a tree with as many branches as there are real
topics in the corpus, each branch of the additive tree being characterized by
seven labels: six labels corresponding to the six methods briefly described
above, plus one label corresponding to one a priori theme. Such situation
occurs when each method has uncovered the same real topics. The observed
configuration is not that good, but we can distinguish between six and nine
main branches, which is probably the order of magnitude of the number of
different topics. We note also that several different methods often participate
in the same branch, which suggest that that branch correspond to a real topic
discovered by almost all the six methods. Let us mention that a similar
additive tree performed on the 49 topics (not involving the eight a priori
themes) produces approximately the same branches. Thus, the eight a priori
themes can be considered here as illustrative elements, serving only as
potential identifiers of the branches.
It is remarkable that the eight a priori themes (boxed labels) are well
distributed over the whole of Figure 2. If we except the branch of the tree
located in the upper right part of the display, on the right of the label “Young
man”, all the main branches have as a counterpart one of the a priori themes.
As an example of interpretation of figure 2, the branch in the lower center
part of figure 2: [NMF7, LOA4, RFA3, LDA7, LSA5] is clearly closely linked
to the a priori topic named Rivalry (see section 2.2) (concurrence of five
methods out of six). Most of the branches of the additive tree could be
interpreted likewise. The upper right branch identified by none of the a priori
themes may represent an unforeseen topic. More research and an expertise in
Elizabethan poetry are required to confirm that we are dealing here with an
undetected new theme. To conclude, we can only observe that each of the

442

JADT’ 18

involved method, be it ancient or modern, may contribute to detect topics…
and that exploratory tools are essential to visualize the complexity of the
process and assess the obtained results.

Figure 2. Additive Tree describing the links between the 49 topics provided by the 6 selected
methods and the 8 a priori themes. The identifiers are those of section 4 for the 6 selected
methods. The 3 first letters indicate the method, followed by the index of the produced topic.
The distance between two topics is the chi-square distance between their lexical profiles.
Threshold of frequencies for words: 2. The boxed identifiers of the a priori themes are those
(possibly shortened) of table 1.

References
Alden, R. M. (1913). Sonnets and a Lover's Complaint. New York: Macmillan.
Berry M.W., Browne M., Langville Amy N., Pauca V.P., and Plemmons R.J.
(2007). "Algorithms and applications for approximate nonnegative matrix
factorization". In: Computational Statistics & Data Analysis 52.1: 155-173.
Blei, D., Ng, A., and Jordan, M. (2003). Latent Dirichlet allocation. Journal of
Machine Learning Research, 3: 993—1022.

JADT’ 18

443

Deerwester S., Dumais S.T., Furnas G.W., Landauer T.K.and Harshman R.
(1990). Indexing by latent semantic analysis, J. of the Amer. Soc. for
Information Science, 41 (6): 391-407.
Garnett J.-C. (1919). General ability, cleverness and purpose. British J. of
Psych, 9, 345-366.
Griffiths T.,L., Steyvers M., and Tenenbaum J.,B. (2007). Topics in Semantic
Representation. Psychological Review, 114, 2, 211-244.
Huson D. H., Bryant D. (2006) Application of Phylogenetic Networks in
Evolutionary Studies. Molecular Biology and Evolution, 23 (2): 254 - 267.
Software available from www.splitstree.org.
Kazmierczak J.-B. (1985). Analyse logarithmique : deux exemples
d'application. Revue de Statist. Appl., 33, (1), 13-24.
Lee D.D. and Seung H. S. (1999). Learning the parts of objects by nonnegative matrix factorization. Nature, 401: 788-791.
Lebart L. (2012). Articulation entre exploration et inférence. In : JADT_2012.
Dister A., Longree D., Purnelle G., Editors. Presse Universitaire de Liège.
Lewi P.J. (1976). Spectral mapping, a technique for classifying biological
activity profiles of chemical compounds. Arzneim. Forsch. in: Drug
Res. 26, 1295-1300.
Paterson D. (2010). Reading Shakespeare Sonnets. Faber & Faber Ltd. London.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J.,
Passos, A., Cournapeau, D., Brucher, M., Perrot M. and Duchesnay, E.
(2011). Scikit-learn: Machine Learning in Python, Journal of Machine
Learning Research , 12, 2825-2830.
Reinert, M. (1986). Un logiciel d’analyse lexicale: [ALCESTE]. Cahiers de
l’Analyse des Données, 4, 471–484.
Sattath S. and Tversky A. (1977). Additive similarity trees. Psychometrika, vol.
(42), 3: 319-345.
Shakespeare, W. (1901). Poems and sonnets: Booklover's Edition. Ed. The
University Society and Israel Gollancz. New York: University Society
Press. Shakespeare Online. Dec. 2017.
Spearman C. (1904). General intelligence, objectively determined and
measured. Amer. Journal of Psychology, 15, 201-293.
Gaujoux R.et al. (2010). A flexible R package for nonnegative matrix
factorization. In: BMC Bioinformatics 11.1 (2010): 367.
Thurstone L. L. (1947). Multiple Factor Analysis. The Univ. of Chicago Press,
Chicago.

444

JADT’ 18

Analyse Diachronique de Corpus : le cas du poker
Gaël Lejeune1, Lichao Zhu2
1

STIH, Sorbonne Université – gael.lejeune@sorbonne-universite.fr
2 LLSHS, Université Paris XIII – lichao.zhu@univ-paris13.fr

Abstract
In this paper we will investigate a diachronic corpus. We want to highlight
how people’s mentalities evolve regarding the gambling especially the poker
game and how the evolution is correlated with the way that the game is
considered in press articles. We study plain or metaphorical meanings of the
terms in question by using clustering and statistical methods in order to
detect changes of meanings in a relatively large period of time.
Résumé
Dans cet article nous nous intéressons à l'étude diachronique de corpus de
presse dans le but d'illustrer des évolutions dans la vision de la société sur les
jeux d'argent et de hasard ainsi que sur les joueurs. Nous utilisons des
méthodes de statistique textuelles et de clustering pour détecter les grandes
tendances visibles sur noter échelle de temps en nous focalisant sur le poker .
Nous montrons que si le regain de popularité du jeu de poker se traduit par
un traitement médiatique plus important, les métaphores exploitant la notion
de poker restent très fréquentes.
Keywords: analyse diachronique, corpus, jeux d'argent et de hasard
1. Introduction
L'analyse diachronique de corpus opère sur un champ assez large. Nous
pouvons en juger par exemple en observant les nombreux travaux sur
l'évolution des langues, travaux qui passionnent aussi bien la communauté
scientifique (Dediu & de Boer 2016) que les médias si l'on se fie par exemple à
l’intérêt renouvelé porté par ceux-ci sur l’évolution des dictionnaires. Dans le
champ purement scientifique, les intérêts dans le domaine embrassent tous
les niveaux de l'analyse linguistique même si la morphologie (Macaulay
2017) et le lexique (la néologie par exemple chez Gérard et al. 2014). La
sémantique est un autre aspect des études diachroniques notamment pour
étudier les représentations mentales des locuteurs (Hamilton et al. 2016). Le
travail présenté ici s'intéresse à une autre catégorie de représentations
mentales qui est l'image que certaines activités ludiques peuvent prendre au
cours du temps. Nous nous intéressons ici à un jeu d'argent et de hasard qui

JADT’ 18

445

a connu une sorte de nouvelle jeunesse ces dernières années : le jeu de poker.
Dans ce travail, nous nous inspirons de l’analyse de l’usage du lexique dans
(Hamilton et al. 2016), nous souhaitons examiner l’évolution de l’usage d’un
mot, d’un terme particulier au cours du temps. Ce travail, même si notre
ambition est moins large, peut se rattacher aux études sur la néologie
sémantique (Sablayrolles 2002) ou néosémie (Rastier et Valette 2009). Pour
illustrer l’intérêt que représente le poker en tant que phénomène de société,
nous pouvons considérer le retentissement autour du Moneymaker Effect1 ou
encore cette citation du journal Le Monde daté du 22 janvier 2007 qui illustre
le changement d’image de ce jeu: « Considéré il y a encore peu de temps comme
un jeu sulfureux se jouant dans les arrière-salles de bars louches ou dans des
appartements huppés à l'abri des regards indiscrets, le poker fait une entrée en force à
la télévision ». En particulier, dans sa variante à la mode Texas Hold'Em, le
poker est redevenu un jeu dont on parle et dont on parle plutôt positivement.
Notre objectif est d’une part de mesurer à quel point ce regain d’attention a
pu se traduire par une amélioration de l’image du jeu de poker en général.
D’autre part, il s’agit de voir dans quelle mesure les usages métaphoriques
du terme poker, plutôt connotés “négativement” (poker menteur, coup de
poker2…) ont pu évoluer conjointement à cette plus grande popularité du jeu
lui même. Dans la section 2 nous présenterons le corpus que nous avons
constitué pour cette étude. Puis, nous proposerons dans les deux sections
suivantes une analyse statistiques des prédicats puis une analyse sous forme
de clustering. Enfin, nous présenterons nos conclusions et perspectives.
2. Présentation de notre corpus d’étude
De manière à pouvoir s’affranchir des variations de choix éditoriaux entre
journaux, nous nous avons souhaité nous concentrer sur une seule
publication. Nous avons choisi le Monde ce qui nous permettais d’exploiter
des articles dont la publication s’étalent sur 30 ans : 1988-2017. Pour la partie
1988-2005 nous avons utilisé le corpus du monde distribué par ELRA3, nous
avons restreint aux textes contenant le terme poker. Pour les années 2006 à
2017 nous avons extrait d’Europresse4 les articles qui comportait le terme
poker. Dans les deux cas nous avons considéré toutes les variantes possibles
dans la casse. Nous avons ainsi obtenu 3528 textes dont la répartition dans le

Par exemple : http://www.slate.com/articles/news_and_politics/explainer/
2011/06/the_moneymaker_effect.html
2
Dans le sport par exemple, on remarque des contextes de « tentative
désespérée », « dernière chance » ...
3
http://catalog.elra.info/product_info.php?products_id=438&language=fr
4
http://www.europresse.com/fr/
1

446

JADT’ 18

temps est présentée Figure 1. Nous pouvons observer que le nombre
d’articles a connu une chute entre 2005 et 2006. Ceci semble être dû au fait
que nous passions à ce moment précis d’une étude du corpus complet du
monde tel qu’existant auprès d’ELRA à une étude fondée sur la base
Europresse. De fait, sur nos critères de recherche, la base Europresse ne
totalise que 47 articles pour 2003 (contre 129 dans le corpus ELRA), 62 pour
2004 (contre 117) et 67 articles pour 2005 (contre 117). Les contraintes
respectives d‘utilisation de ces deux sources de données nous ont interdit de
pouvoir disposer d’un corpus dont la constitution soit constante. Nous nous
sommes efforcés de s’affranchir de ce biais en adaptant notre méthodologie
(notamment le clustering).

Figure 1 : Répartition du nombre d'articles par année

Nous avons 4353 occurrences du terme recherché, leur répartition est
instructive (Figure 2) : la très grande majorité des articles (2834/3528 soit
80,33%) ne comporte qu’une seule occurrence. Nous pensons que ceci est le
reflet de deux tendances. D’une part le sujet de l’article est rarement le poker
pour lui même, il est question d’un personnage qui par ailleurs joue au poker
par exemple. D’autre part, cette rareté de la répétition révèle un usage
massivement métaphorique, en effet comme l’a montré (Lejeune 2013) une
métaphore perd de sa force en étant répétée. Si un terme est répété, il est très
probable qu’il soit employé dans son sens plein. Si cette observation était
faite sur des noms de maladies infectieuses, il nous semble que ceci est avant
tout lié au genre de texte et que cela s’applique également ici. Si nous allons
un peu plus loin, nous pouvons faire l’hypothèse que la métaphore peut être
filée, mais qu’elle est rare dans les articles expositifs. D’autre part, dans le cas
peu probable d’une métaphore filée, les conventions stylistiques impliquent
de changer le terme employé, le journaliste utilisera plutôt des termes du
même champ lexical.

JADT’ 18

447

Figure 2 : Répartition des d'articles selon le nombre d’occurrences du terme « poker »

La répartition des articles entre ceux qui comportent une et une seule
occurrence et ceux qui en comportent plusieurs montre des variations
importantes dans le temps (Figure 3). Si l’on observe des périodes de 5 ans,
on peut se rendre compte que le nombre d’articles comprenant plusieurs
occurrences de “poker” représente 15% des articles sélectionnés sur la
période 1988-1992, se pourcentage descend à 10% jusqu’en en 2003 puis
remonte progressivement pour finalement rester au-dessus de 20% à partir
de 2004-2008 avec une pointe à 30% pour les périodes 2007-2011 à 2009-2013.

Figure 3 : Répartition par année des articles selon le nombre d’occurrences

3. Prédicats et séquences figées
Dans la théorie linguistique lexique-grammaire de M. Gross (1975) et de G.
Gross (2012), les prédicats sont considérés comme les noyaux d’une phrase
capables de disposer d’arguments, grâce à leurs propriétés
transformationnelles et distributionnelles. Parmi les apports de cette théorie
figurent le « schéma d’arguments » et les « prédicats appropriés ». Nous
relevons dans notre corpus les contextes gauches et droits des séquences

448

JADT’ 18

figées « partie de poker » et « coup de poker » afin de distinguer leurs
emplois métaphoriques et non métaphoriques. Ce travail est fait en étudiant
le premier verbe précédant ou suivant l’expression (sans remonter au-delà
d’une phrase). Nous montrons dans les tableaux 1 et 2 les 20 verbes les plus
fréquents pour chaque contexte se trouvent le plus fréquemment dans ces
contextes (20 dans les contextes gauches, 20 dans les contextes droits).
Tableau 1 : Effectif des verbes dans le contexte gauche de “[partie|coup] de poker”
être (76)

jouer (62)

faire (15)

tenter (14)

gagner (11)

avoir (11)

ressembler (10)

prendre (9)

tenir (8)

lancer (8)

perdre (7)

voir (6)

partir (6)

engager (6)

agir (5)

réussir (4)

livrer (4)

remporter (3)

organiser (3)

mener(2)

Tableau 2 : Effectif des verbes dans le contexte droit de “[partie|coup] de poker”
être (98)

avoir (75)

jouer (16)

pouvoir (13)

devoir (8)

gagner (7)

engager (7)

venir (6)

livrer (6)

faire (5)

vouloir (4)

voir (4)

tenter (4)

tenir (4)

réussir (4)

prendre (4)

monter (4)

bluffer (4)

aller (4)

retrouver (3)

Hormis les verbes « être » et « avoir » qui sont susceptibles d’être des verbes
auxiliaires ou semi-auxiliaires, pour les autres verbes on peut se trouver dans
trois cas de figure :
h) Verbe support
i) Prédicat approprié : le sens littéral de l’expression peut être
activé
j) Prédicat non approprié : le sens métaphorique de
l’expression est activé
Le cas des verbes support n’est pas pertinent pour notre étude. Pour le
second cas, nous observons que le verbe jouer, prédicat approprié pour les
deux séquences décrites, est très souvent lié à un usage métaphorique. Dans
le troisième cas, de loin le plus fréquent. Les verbes « tenter », « s’engager »,
« réussir », « mener », « lancer » voire « remporter » ne sont pas tout à fait
congruents avec le sens premier de la séquence, c’est-à-dire qu’ils ne sont pas
des prédicats appropriés au sens propre du jeu de poker. Des occurrences de
ces verbes dans le corpus confirment cette intuition :
Il leur fallait lancer la partie de poker que Bonn et Paris s'apprêtent à jouer sur le
GATT (1993)
les enjeux de la partie de poker qui s'engagera mercredi à la mi-journée lorsque
l'ambassadeur[...] (2017)
[ils] avaient pu croire un moment que leur coup de poker allait réussir. (1989)

JADT’ 18

449

[Celui qui] est davantage connu pour ses coups de poker financiers continue à mener
sa stratégie (2015)
Elle venait de remporter la partie de poker menteur qui constitue l'essentiel des
premiers hectomètres. (1995)
4. Étude des champs lexicaux par clustering
Si les séquences « partie de poker » et « coup de poker » sont ambiguës dans
le sens où elles figurent dans des champs lexicaux différents, on peut se
demander ce qu’il en est des champs lexicaux du terme « poker » en général.
Pour étudier cette question, nous avons réalisé un clustering de notre corpus.
Nous avons utilisé l’implantation des k-moyennes (K-means) de la
bibliothèque Python scikit-learn. Nous avons fixé le nombre de clusters K à
105 et le nombre maximal d’itérations à 400, la mesure des poids est le tf-idf.
Nous avons extrait tous les n-grammes de mots avec n allant de 1 à 3 puis
seulement nous avons utilisé une stop-list. De sorte que, par exemple, « de »
n’était pas gardé en tant que tel mais que nous le retrouvions dans « coup de
poker » ou « loi de Robien ». Nous avons tout d’abord travaillé sur le corpus
lemmatisé puis nous avons observé que les résultats étaient semblables sans
lemmatisation, nous avons donc supprimé ce pré-traitement. Nous allons
maintenant décrire chaque cluster en donnant la proportion du corpus qu’il
couvre ainsi que les 10 termes les plus significatifs.
Cluster 0, « sport et poker 1 » : 3,1 % (club, football, équipe, Ligue, France,
championnat, saison, joueurs, OM, Marseille) Ce cluster comporte deux
volets : l’un sur les « coups de poker » dans les championnats de football et
l’autre où il est question des championnats de poker eux mêmes.
Cluster 1, « politique » : 18,79 % (ministre, président, politique,
gouvernement, pays, ,État, premier ministre, premier, États, faire). Un cluster
autour de l’action politique, notamment au niveau européen. Un exemple
intéressant de métaphore (filée) ici : « M. Erdogan remet tout en jeu, comme
un joueur de poker fait tapis »
Cluster 2, « fourre-tout » : 38,01 % (être, bien, film, vie, entre, Jean, monde,
France, temps, homme) Le seul de nos clusters qui n’ait pas d’unité ni de
tendance thématique, ici les expressions contenant poker sont pour moitié
métaphoriques.
Cluster 3, « culture_1 » : 5,13 % (film, Booker Prize, roman, prix, livres, livre,
littéraire, base, prix littéraire, attribué). Ce cluster rassemble les livres ayant
trait au poker, les expressions liées sont prises dans leur sens littéral

5

9 et 12.

Selon la méthode du coude (elbow method), la valeur optimale se situait entre

450

JADT’ 18

(l’expression « coup de poker » y est quasi absente).
Cluster 4 « finance »: 4,2 % (Vivendi, marché, groupe, Bourse, marches,
actionnaires, titres, taux, millions, fonds, terme, milliards, prix) Il se
caractérise uniquement par des thématiques associées au domaine de finance
et notamment aux coups de poker boursiers.
Cluster 5 « sport et poker 2»: 5,04 % (Coupe, match, équipe, joueurs, France,
club, football, finale, francs, PSG). Nous avons ici un cluster sur le sport où
environ la moitié des articles concernent toutefois le poker lui même.
Cluster 6 « industrie du poker »: 12,96 % (jeux, paris, ligne, marché,
milliards, euros, millions, Internet, dollars, Bourse) Ici nous avons tout ce qui
est lié à l’industrie du poker et notamment à l’essor des jeux d’argent sur
Internet (dont le poker a été un fer de lance).
Cluster 7 « sport »: 3,26 % (Tour, numéros, France, coureur, étape, peloton,
course, équipe, Tour de France, maillot) Nous avons ici des usages,
massivement métaphoriques, dans le domaine du sport (principalement le
cyclisme). Un exemple avec le terme spécialisé flop : « [P.A.Bosse] avait
trouvé cette image [...] : Si on compare le 1500m au poker, il a un flop
d'avance. »
Cluster 8 « culture_2 » : 7,14 % (blues, musique, CD, rock, John Lee Hooker,
jazz, album, guitare, musiciens, scène) Un usage métaphorique dans le
domaine de la musique avec des expressions telles que « poker face »,
« poker perdant »...
Cluster 9 « culture_3 » : 2,38 % (Dracula, Bram Stoker, vampire, roman, film,
fantastique, Christie, Coppola, comte, Frankenstein) Le cluster 3 était centré
sur le domaine littéraire, ici il est question de cinéma et particulièrement des
personnalités liées au poker. L’usage y est surtout littéral.
Pour ce qui est de la répartition temporelle, il est très intéressant de noter que
le cluster 6 (l’industrie du poker) devient le second plus important derrière le
cluster 2 ( à partir de 2005 (popularisation des jeux d’argent sur Internet) et
plus encore à partir de 2010 (légalisation des paris en ligne). Le cluster 0
(sport et poker) devient plus important à partir de 2004 d’autant qu’en son
sein la thématique poker y est alors largement majoritaire.
5. Conclusion
Nous avons proposé dans cet article une étude diachronique d’articles de
presse contenant le mot « poker ». Notre hypothèse initiale était que ce terme
était souvent employé dans des expressions métaphoriques et que le regain
de popularité de ce jeu depuis quelques années avait du amener une plus
grande proportion d’usage littéral. Nous avons observé que dans plus de
80 % des cas, le terme poker n’apparaissait qu’une fois dans les textes. Nous

JADT’ 18

451

avons montré que ceci était dû à un usage principalement métaphorique, on
ne répète pas une métaphore, mais aussi au fait que le poker est rarement le
sujet central de l’article. Cette tendance change quelque peu à partir de 2005,
le poker devenant lié à des championnats et des retransmissions télévisuelles
plutôt qu’à des tripots et des casinos. Enfin, nous avons montré que les
usages métaphoriques relevaient très majoritairement de 3 domaines : la
finance, la politique et le sport.
References
Dediu D. and de Boer B. (2016)., Language evolution needs its own journal ,
Journal of Language Evolution, Volume 1, Issue 1, 1 January 2016, Pages
1–6
Gérard C., Falk I., and Bernhard D. (2014). Traitement automatisé de la
néologie : pourquoi et comment intégrer l’analyse thématique ? Actes du
4e Congrès mondial de linguistique française (CMLF 2014), Berlin, Pages
2627-2646
Gross, M. (1975). Méthodes en syntaxe: régime des constructions
complétives, Hermann.
Gross, G. (2012). Manuel d'analyse linguistique: Approche sémanticosyntaxique du lexique, Presses Universitaires du Septentrion.
Hamilton W.L., Leskovec J., and Jurafsky D. (2016). Diachronic Word
Embeddings Reveal Statistical Laws of Semantic Change. In Proc. Of the
Association for Computational Linguistics Conference (ACL) 2016
Lejeune G. (2013) Veille épidémiologique multilingue : une approche
parcimonieuse au grain caractère fondée sur le genre textuel, Thèse de
doctorat en Informatique de l'Université de Caen
Macaulay, M. & Salmons. (2017). Synchrony and diachrony in Menominee
derivational morphology, J. Morphology 27: 179
Rastier, F., Valette, M. (2009) « De la polysémie à la néosémie », Le français
moderne, S. Mejri, éd.,
La problématique du mot, 77, 97-116.
Sablayrolles, F. (2002). « Fondements théoriques des difficultés pratiques du
traitement des
néologismes », Revue française de linguistique appliquée, VII-1, pp. 97-111.

452

JADT’ 18

Approche textométrique des variations du sens
Julien Longhi1, André Salem2
Université de Cergy-Pontoise, France – julien.longhi@u-cergy.fr
Université de la Sorbonne nouvelle, France – salem@msh-paris.fr

1
2

Abstract
The use of textometric methods relies on the hypotheses, firtly, that stable
units exist (forms, lemmas or their graphical approximations) and, secondly,
that occurrences of these forms can be retrieved from different parts of a
corpus. Once automatic counting performed, more sophisticated textometric
methods can be employed to focus on textual variations (repeated segments,
collocations, etc.) that occur around the same unit but in different contexts
found within the corpus. This approach leads to the identification of
semantic variations with relation to the context of each occurrence as
highlighted through automatic segmentation. We will illustrate this by using
examples of repeated segments within the corpus that contain the N-gram
/enemy / taken from a widely-studied chronological text series.
Résumé
Pour pouvoir mettre en œuvre les méthodes de la textométrie, il est
indispensable de postuler, dans un premier temps, l'existence d'unités stables
(formes, lemmes ou leurs approximations graphiques), dont on recensera
ensuite les occurrences dans les différentes parties du corpus étudié. Une fois
les dépouillements automatiques réalisés, il est cependant possible d'utiliser
des méthodes textométriques plus élaborées pour accéder aux variations
textuelles (segments, répétés, cooccurrences, etc.) qui peuvent se réaliser
autour d'une même forme dans chacun des contextes particuliers du corpus.
Cette démarche permet d'accéder au repérage de variations sémantiques qui
se rapportent à chacune des occurrences des formes produites par la
segmentation automatique. Nous illustrons notre démarche à l'aide
d'exemples prélevés dans les parties d'une série textuelle chronologique
largement étudiée, des segments répétés du corpus qui contiennent le Ngram /ennemi/.
Keywords:.unité textométrique, sémantique, variation du sens
1. Introduction
Notre étude s’inscrit dans une perspective de prise en compte des
dynamiques du sens à l’œuvre dans les discours, qui tiendrait compte de la

JADT’ 18

453

variation, de l’hétérogénéité, ou encore de l’articulation entre topologie
textuelle et discursive, sens et profilage. Le sens se construit dans différents
champs où il est susceptible de paraître, et s’analyse « par le contexte, sous
forme d’indices de position liés aux modalités de sa mise en place dans le
champ » (Cadiot et Visetti, 2011), la caractérisation sémantique se faisant
alors sur la base de la composition et décomposition des profils disponibles.
L'automatisation du dépouillement de vastes corpus de textes, à des fins
textométriques, nécessite au contraire que le repérage des unités de
décompte puisse être confié à des machines. Pour pouvoir mettre en œuvre
les méthodes de la textométrie, il est indispensable de postuler, dans un
premier temps, l'existence d'unités stables (lexèmes, lemmes ou leurs
approximations graphiques), dont on recensera ensuite les occurrences dans
différentes parties du texte. Cette manière de faire permet d'étudier la
répartition de chacune des unités dans un corpus ou encore de rapprocher les
différents contextes qui contiennent chaque unité textométrique. Ces
simplifications, incontournables dans le premier temps de l'analyse, nous
éloignent de l'étude du sens de chacune des occurrences que l'on peut
élaborer dans chaque contexte particulier. Cependant, une fois les premiers
dépouillements automatiques réalisés, il est possible d'utiliser des méthodes
textométriques plus élaborées pour accéder aux variations textuelles qui
peuvent se réaliser autour d'une même forme dans le corpus (segments
répétés, cooccurrences, etc.). C’est ce croisement de perspectives et ce va-etvient entre approche empirique et théorisation sémantique, que nous
souhaitons mettre à l’épreuve dans la présente étude.
2. Application au corpus Duchesne
Pour illustrer notre démarche, nous appliquons ces méthodes à l'étude de la
ventilation, dans les différentes parties d'une série textuelle chronologique
largement étudiée, des segments répétés du corpus qui contiennent le Ngram /ennemi/.
2.1. Rappels sur l'analyse de la série chronologique Duchesne
La série chronologique Père Duchesne a déjà fait l'objet de nombreuses
analyses textométriques1. Nous avons montré, en particulier, que les

Le corpus Père Duchesne est constitué par la réunion d'un ensemble de
livraisons du journal Le Père Duchesne de Jacques-René Hébert, parues entre 1793 et
1794. Pour une description plus avancée de ce corpus, on consultera, par exemple
(Salem, 1988).
Les analyses dont nous rendons compte ci-dessous, ont été effectuées à l'aide du
logiciel Lexico5. Cedric Lamalle, William Martinez, Serge Fleury ont largement
1

454

JADT’ 18

typologies réalisées à partir d'une partition de ce corpus en huit périodes,
correspondant chacune à un mois de parution, mettaient en évidence un
renouvellement lexical fortement lié à l'évolution dans le temps. On peut
vérifier, sur la figure 1, que les parties correspondant aux périodes
successives de parution sont proches sur les facteurs issus de l'analyse du
tableau (8 parties x 1420 formes dont la fréquence dépasse dix occurrences)2.
La méthode des segments répétés permet de repérer toutes les occurrences de
suite de formes graphiques qui apparaissent plusieurs fois dans un corpus de
textes (Lafon et Salem, 1983 ; Salem, 1986). Pour la présente étude, nous
avons constitué un ensemble d'unités textuelles qui contient outre les formes
graphiques ennemi et ennemis, tous les segments répétés qui contiennent l'une
ou l'autre de ces formes. On a projeté sur la figure 1, en qualité d'éléments
supplémentaires, cet ensemble de segments. La position sur ce graphique des
différents segments montre que ces unités ne sont pas employées de manière
uniforme tout au long des périodes.

Figure 1 : Duchesne. Les segments contenant la séquence ennemi sur
le plan des deux premiers facteurs issus de l'analyse de tableau 8 parties x 1420 formes
(F>=10)

Guide de lecture pour la figure 1 : La figure fournit la représentation des huit
parties du corpus Duchesne, sur les deux premiers axes issus d'une Analyse

contribué au développement des fonctionnalités de ce logiciel. Les auteurs tiennent à
les en remercier.
2 Ce phénomène connu sous le nom d'effet Guttman, a été largement décrit par
Guttman (1941, 1946, 1950), Benzecri (1973) et Van Rijckevorsel (1987).

JADT’ 18

455

des correspondances, réalisée sur l'ensemble des formes dont la fréquence
dépasse 10 occurrences. Les segments répétés du corpus contenant la
séquence de caractères /ennemi / ont été projetés sur ce même plan, en tant
qu'éléments supplémentaires. La figure a été allégée des segments
redondants (ex : segments contenus dans des segments plus longs). Certains
des éléments superposés par l'analyse ont été très légèrement déplacés à fin
de rendre la figure plus lisible.
Ainsi par exemple, le segment plus cruels ennemis trouve toutes ses
occurrences au début du corpus alors que celles du segment ennemis de la
liberté sont plutôt concentrées vers la fin.
L'analyse des projections des différents segments qui contiennent le n-gram
/ennemi/ va nous permettre de dégager des contextes dont la distribution
diffère fortement entre le début et la fin de la période temporelle couverte par
le corpus.
2.2. L'évolution du contexte de la forme ennemi(s)
On peut estimer que le contenu sémantique de la forme ennemi(s) conserve
une valeur relativement stable tout au long des périodes couvertes par le
corpus que nous étudions. Le chercheur confronté à l'analyse de ces textes
retrouvera sans peine, lors de l'examen de chacune des occurrences du terme,
les principaux traits sémantiques décrits dans un dictionnaire de langue à
propos de ce lexème (opposé, hostile, etc.). Cependant, l'analyse de ces
mêmes contextes montre qu'il en va tout autrement pour ce qui concerne les
référents auxquels la forme renvoie, dans chaque période particulière. Aux
plus cruels ennemis, plus mortels ennemis, ennemis du dehors (les puissances
étrangères, les expatriés), des périodes du début, succèdent bientôt les
ennemis du dedans et du dehors, expressions qui peuvent s'analyser comme une
dénonciation du fait que les ennemis du dehors ne constituent pas le seul
danger et qui opère donc une modification manifeste du référent de départ.
Par la suite la mention des ennemis de l'intérieur complètera la notion
d'ennemis du dedans. Il faut noter que les ennemis de l'intérieur sont de plus en
plus souvent précédés de l'article défini les qui les désigne comme une réalité
dont l'existence est présupposée (elle n’est plus à démontrer).
Progressivement, nos ennemis, deviennent vos ennemis, puis les ennemis. Dans
la dernière période les ennemis, désormais désignés, de manière
préférentielle, au pluriel, ne sont plus qualifiés par leur localisation ou par
leur rapport aux destinataires du message (nos/vos ennemis) mais par des
valeurs supposée communes auxquelles ils sont censés s'opposer : ennemis du
peuple, ennemis de la république, ennemis de la révolution, ennemis de la liberté,
ennemis de l'égalité.

456

JADT’ 18

3. La sémantique de ennemi(s)
Les variations constatées montrent que la forme ennemi(s) prend différents
sens selon les contextes dans lesquels elle s'inscrit, en ce qu’ils sont associés à
des référents distincts. Plutôt que de représenter le sens comme la somme des
cooccurrences constatées, nous souhaitons analyser ces valeurs comme un
sous-ensemble prélevé sur un ensemble de valeurs acquises. Les espaces
sémantiques déterminés et caractérisés par l’analyse statistique jouent un rôle
fondamental qui, au-delà des synonymies, ou des polysémies, se
renouvellent « en étant confronté aux textes – ce qui impliquerait de prêter
attention à d’autres corrélations » (Visetti 2004 : 11). La description
sémantique que nous proposons s’inscrit dans le champ de la sémantique
lexicale3, du côté des approches qui envisagent la construction des référents
comme extrinsèque. Cependant, alors que ces approches mobilisent en
général des analyses phrastiques, et travaillent sur des exemples forgés, nous
introduisons une perspective statistique qui précède la représentation du
sens. La description de l’objet ennemi(s) n’est pas séparée des rapports que
l’on entretient avec lui, et sa description suppose une prise en compte
différenciée de ses propriétés extrinsèques (relatives à ces rapports), et de ses
propriétés intrinsèques (supposées stables et indépendantes).

Figure 2 : Niveaux et unités d'analyse

3

Cadiot et Némo (1997 : 127-128)

JADT’ 18

457

L’intérêt de cette démonstration textométrique est pour nous de fournir des
résultats concrets et matériels pour l’analyse des sens d’une unité lexicale.
Ceci a plusieurs conséquences pour la mise en œuvre d’une sémantique
soucieuse de l'exploitation des constats empiriques :
1) la représentation des variations du sens en contexte nous a permis
d’identifier la manière donc les propriétés sont introduites et
attribuées dans le corpus. Le référent change au fil du temps,
puisque les ennemis, initialement définis comme du dehors, et
introduits par nos, deviennent vos ennemis, et se présentent
finalement sous la forme ennemi(s) de + N. Le besoin d’être déterminé
par un complément du nom, ou son équivalent, qui indique avec
quoi le terme « relatif » se trouve mis en relation », cette
complémentation explicitant « ainsi la référence identitaire »
(Steuckardt, 2008).
2) L’évolution dans le corpus au fil du temps permet de rendre compte
de la dynamique sémantique à l’œuvre, laquelle rend compte
diachroniquement des évolutions de sens. La textométrie permet
ainsi de saisir les processus, et donc de donner du sens à la
dimension potentiellement « hétéroclite » des propriétés des
référents.
Ainsi, au plan linguistique, le passage du référent 1 ou référent 2 se fait par
l’intermédiaire d’une transformation des propriétés de ennemi(s) : défini de
manière situationnelle (du dehors) et relative (nos, nos plus cruels), il acquiert
des propriétés plus polémiques (vos, du dedans et du dehors), pour s’intégrer
ensuite dans un processus discursif qui construit le référent (ennemi de + N :
ennemi de la liberté ; ennemi du peuple), par l’introduction de termes à fort
charge axiologique. Le référent introduit alors un point de vue, qui n’est pas
strictement géographique ou institutionnel, mais aussi politique et
idéologique. L'approche statistique dévoile, en outre, que c’est le pluriel qui
est prioritairement mobilisé.
3. Conclusion
De manière désormais classique, les méthodes de la textométrie permettent
de mettre en évidence les variations du vocabulaire qui surviennent au cours
des périodes successives d'une même série textuelle chronologique. Dans la
présente étude, nous avons appliqué les méthodes d'analyse statistique
multidimensionnelle (AFC) à l'étude d'un ensemble particulier, celui des
segments répétés réunis sur la base du fait qu'ils contenaient tous une même
unité graphique (en l'occurrence, le n-gram /ennemi/).
La confrontation des segments ainsi sélectionnés nous permet d'observer des
variations autour des formes graphiques ennemi et ennemis. L'analyse de ces

458

JADT’ 18

variations dans le temps nous conduit à distinguer des référents qui varient
en fonction des périodes réunies dans le corpus.
Au-delà des séries textuelles chronologiques, la méthode que nous avons
présentée est susceptible de recevoir des applications dans l'étude de
nombreux types de corpus. L'extraction semi-automatique des unités dont les
contextes varient fortement en fonction des parties d'un corpus textuelle peut
également être envisagée.
References
Benzécri J-P. and coll. (1981). Pratique de l'analyse des données, Linguistique et
lexicologie. Dunod.
Cadiot P. and Nemo F. (1997). Propriétés extrinsèques en sémantique
lexicale. Journal of French Language Studies, 7(2): 127-146.
Cadiot P. and Visetti Y.-M. (2001). Pour une théorie des formes sémantiques.PUF.
Guttman L. (1941). The quantification of a class of attributes: a theory and
method of a scale construction. In P. Horst, The prediction of personal
adjustment, SSCR New York.
Lafon P. and Salem A. (1983). L’Inventaire des segments répétés d'un texte.
Mots. Les langages du politique, 6 : 161-177.
Lamalle C, Martinez W, Fleury S, and Salem A. (2002). Les dix premiers pas
avec
Lexico3.
Outils
lexicométriques.
http://www.cavi.univparis3.fr/Ilpga/ilpga/tal/lexicoWWW
Lebart L. and Salem A. (1994). Statistique textuelle. Dunod.
Longhi J. (2008). Objets discursifs et doxa. Essai de sémantique discursive.
L’Harmattan, coll. « Sémantiques ».
Rastier F. (2011). La mesure et le grain. Sémantique de corpus. Honoré
Champion, coll. « Lettres numériques ».
Salem A. (1987). Pratique des segments répétés. Klincksieck.
Salem A. (1988). Approches du temps lexical. Mots. Les langages du politique,
17 : 105-143.
Steuckardt A. (2008). Les ennemis selon L’Ami du peuple, ou la catégorisation
identitaire par contraste. Mots. Les langages du politique [En ligne],
69 | 2002. http://journals.openedition.org/mots/10023
Van Rijckevorsel J. (1987). The application of fuzzy coding and horseshoes in
multiple correspondances analysis. DSWO Press.
Visetti Y.-M. (2004). Le Continu en sémantique : une question de formes.
Texto ! juin 2004. http://www.revuetexto.net/Inedits/Visetti/Visetti_Continu.html

JADT’ 18

459

ADT et deep learning, regards croisés. Phrases-clefs,
motifs et nouveaux observables
Laurent Vanni1, Damon Mayaffre, Dominique Longree2
1

UMR 7320 : Bases, Corpus, Langage - prenom.nom@unice.fr
2L.A.S.L.A. - prenom.nom@uliege.be

Abstract 1
This contribution confronts ADT and Machine learning. The extraction of
statistical key-passages is undertaken following several calculations
implemented using the Hyperbase software. An evaluation of these
calculations according to the filters applied (taking into account only positive
specificities, only substantives, etc.) is given.
The extraction of key passages obtained by deep learning - passages that
have the best recognition rate at the time of a prediction - is then proposed.
The hypothesis is that deep learning is of course sensitive to the linguistic
units on which the computation of the key statistical sentences are based, but
also sensitive to phenomena other than frequency and other complex
linguistic observables that the ADT has more difficulty taking into account as would be the case with underlying patterns (Mellet et Longrée, 2009). If
this hypothesis is confirmed, it would on the one hand permit better
understanding of the black box of deep learning algorithms and on the other
hand to offer the ADT community a new point of view.
Abstract 2
Cette contribution confronte ADT et Deep learning. L’extraction de passagesclefs statistiques est d’abord proposée selon plusieurs calculs implémentés
dans le logiciel Hyperbase. Une évaluation de ces calculs en fonction des
filtres appliqués (prise en compte des spécificités positives seulement, prise
en compte de substantifs seulement, etc) est donnée. L’extraction de
passages-clefs obtenus par deep learning - c’est-à-dire des passages qui ont le
meilleur taux de reconnaissance au moment d’une prédiction - est ensuite
proposée. L’hypothèse est que le deep learning est bien sûr sensible aux
unités linguistes sur lesquelles le calcul des phrases-clefs statistiques se
fondent, mais sensible également à d’autres phénomènes que fréquentiels et
d’autres observables linguistiques complexes que l’ADT a plus de mal à
prendre en compte - comme le seraient des motifs sous-jacents (Mellet et
Longrée, 2009). Si cette hypothèse se confirmait, elle permettrait d’une part
de mieux appréhender la boîte noire des algorithmes de deep learning et
d’autre part d’offrir à la communauté ADT de nouveaux points de vue.

460

JADT’ 18

Keywords: ADT, deep learning, phrase-clef, motif, spécificités, nouveaux
observables
1. Introduction
Pour des raisons techniques avant tout, l’ADT s’est constituée à partir des
années 1960 autour du token, c’est-à-dire du mot graphico-informatique.
Depuis lors, la discipline n’a cessé de varier et d’élargir ses observables,
convaincue que le token seul rendait difficilement compte du texte dans sa
complexité linguistique. Ainsi la tokenisation en particules graphiques
élémentaires reste l’acte informatique premier des traitements
textométriques, et le calcul des spécificités lexicales reste l’entrée statistique
privilégiée de nos parcours interprétatifs. Cependant, la recherche d’unités
phraséologiques élargies et complexes, caractérisantes et structurantes des
textes, est devenue le programme d’une discipline désormais adulte.
Historiquement, dès 1987, le calcul des segments répétés (Salem, 1987) ou les ngrams a représenté une avancée puisque les segments significatifs du texte,
de taille indéterminée, étaient automatiquement repérés ; et aujourd’hui la
détection automatique, non supervisée, de motifs (Mellet et Longrée, 2009;
Quiniou et al., 2012; Mellet et Longrée, 2012; Longrée et Mellet, 2013) - objets
linguistiques complexes à empans variables et discontinus - apparait un
enjeu décisif. C’est dans cette perspective que cette contribution travaille et
met à l’épreuve l’idée de passages-clefs du texte, tels qu’ils sont implémentés
dans les deux versions d’Hyperbase (locale développée par Etienne Brunet et
web développée par Laurent Vanni) que l’UMR Bases, Corpus, Langage
produit en collaboration avec le LASLA. La démonstration se fait en deux
temps. D’abord, nous proposons une extraction statistique de
\textit{passages-clefs}, avec évaluation de leur pertinence interprétative sur
un corpus français et un corpus latin. Ensuite une confrontation
méthodologique avec le deep learning est mise en œuvre puisque le
traitement deep learning attribue, après apprentissage, les passages de texte à
leur auteur avec un taux de réussite éprouvé : par déconvolution nous
repérons alors au sein de ces passages les zones d’activation, en soupçonnant
qu’il s’agit, d’un point de vue linguistique, de motifs remarquables.
2. Les passages-clefs en ADT
2.1. Terminologie
Si nous préférons le terme de passage-clef à celui de phrase-clef c’est que les
traitements ici présentés n’ont pas de modèle syntaxique, et que la
ponctuation forte qui délimite habituellement la phrase est un jalon utile
mais non-nécessaire à nos traitements. La notion de passage a été fortement

JADT’ 18

461

théorisée par (Rastier, 2007) dans un article éponyme et désigne une
« grandeur » du texte dont la valeur textuelle c’est-à-dire interprétative est
patente. Un passage est donc un morceau de texte jugé suffisamment parlant,
notamment par sa taille qui gagne à dépasser le mot, le segment voire la
phrase, pour prétendre rendre compte d’un texte. Le passage-clef, quant à
lui, s’appuie sur la définition rastirienne mais est une unité de surcroit
textométrique ; c’est-à-dire une unité dont la pertinence est calculable et
l’extraction automatique.
2.2. Implémentations
Les logociels ADT comme Hyperbase, Dtm-Vic, Iramuteq implémentent des
calculs et l’extraction de passages-clefs. Dans tous les cas, les calculs proposés
reposent sur l’examen des mots spécifiques (Lafon, 1984) : grosso modo, plus
un passage concentre de spécificités, plus ce passage est jugé remarquable.
Nous présentons ici deux types d’approche sur des passages arbitrairement
constitués de 50 mots : un calcul naïf et sans filtre dans lequel tous les mots
du passage sont considérés et un calcul filtré par nos connaissances
linguistiques (sélection a priori des mots à considérer). Une évaluation de ces
deux types d’approche est ensuite donnée.
2.3. Calcul sans filtre
Dans le cadre des études contrastives habituelles en ADT, l’indice de
spécificité de chaque mot (Lafon, 1984) est sommé, qu’il soit positif ou négatif
en postulant que si les mots positifs (les mots sur-utilisés par un auteur par
exemple) doivent promouvoir le passage, il est légitime que les mots négatifs
(les mots sous-utilisés par un auteur) doivent l’handicaper. Chaque passage
du corpus se trouve ainsi doté d’un super-indice de spécificité et Hyperbase
fait remonter en bon ordre les passages les plus caractéristiques des textes
comparés.
Ainsi pour le français, sur le corpus de la présidentielle française 2017, le
passage-clef le plus fortement d’E. Macron (versus les autres candidats) est le
suivant :
[...] nous croyons dans l'innovation, dans la transformation écologique et
environnementale, parce que nous voulons réconcilier cette perspective et l'ambition
de nos agriculteurs, parce que nous croyons dans la transformation digitale, parce
que nous sommes pour une société de l'innovation, parce que nous voulons […]
Quoique naïf, le calcul apparait performant puisque l’interprétabilité sociolinguistique de ce passage est évidente : de fait Macron s’est fait élire sur un
discours dynamique (voulons , innovation (deux fois), transformation (deux
fois), digitale) et un discours rassembleur susceptible de transcender le clivage
gauche/droite (nous (5 fois), réconcilier).

462

JADT’ 18

2.4. Calcul filtré
Par connaissances linguistiques et statistiques, le calcul peut être raffiné. Par
exemple, seules les spécificités positives – et parmi elles, les spécificités les
plus fortes – peuvent être considérées au motif qu’un objet s’identifie mieux
par ses qualités que par ses défauts. Ensuite, les mots outils (conjonctions,
déterminants) peuvent être écartés : ils présentent le double inconvénient
d’avoir de très hautes fréquences (potentiellement déterminante pour le
calcul des spécificités) et d’être peu parlants d’un point de vue sémanticothématique. Et encore, la catégorie grammaticale peut être choisie : par
exemple seuls les noms propres et communs, parfois plus chargés de sens,
sont pris en compte. Ainsi pour le latin un passage-clef de Jules César,
contrasté à de nombreux auteurs contenus dans la base du LASLA, est le
suivant :
[...] partes Galliae uenire audere quas Caesar possideret neque exercitum sine magno
commeatu atque molimento in unum locum contrahereposse sibi autem mirum uideri
quid in sua Gallia quam bello uicisset aut Caesari aut omnino populo Romano
negotii esset his responsis ad Caesarem relatis iterum ad eum Caesar […]
De fait, ce passage de la Guerre des Gaules peut être effectivement considéré
comme très représentatif de l’œuvre de César. On relève des noms propres
connus (Galliae, Caesar, Gallia) ou des noms communs correspondant à la
réalité militaire du moment (bello, commeatu). Toutefois la méthode ne permet
pas de repérer des structures caractéristiques de la langue et du style de
César, comme par exemple une proposition participiale marquant la
transition entre épisodes dans une négociation : His responsis ad Caesarem
relatis, « Ces réponses ayant été rapportées à César ».
2.5. Evaluation
Calcul naïf ou calcul élaboré : nous récapitulons quelques performances.
Dans un corpus contrastif, nous calculons le score de super-spécificité de
chaque passage en fonction des différents auteurs comparés (Tableau 1). Par
exemple pour le français, sans aucun filtre 58% des passages du corpus de la
présidentielle sont attribués justement à leur auteur ; et en ne considérant que
les spécificités positives, le score descend à 52%. A l’opposé, en imposant le
double filtre de la catégorique grammaticale (seulement les substantifs) et de
l’indice de spécificité (seulement les spécificités positives) nous élevons le
taux de bonne attribution à 89% pour le français et 82% pour le corpus latin
du LASLA.

JADT’ 18

463
Tableau 1. Taux d’attribution ADT et taux de prédiction deep learning

3. Deep learning : à la recherche de nouveaux marqueurs linguistiques
3.1. Convolution et déconvolution, les principes
Le découpage du texte en segments de taille fixe est une méthode qui peut
aussi être utilisée pour entraîner un réseau de neurones. Chaque segment
devient alors une image d'un texte que le réseau va utiliser pour apprendre
(Ducoffe et al., 2016) et faire ensuite de la prédiction. Sur nos deux corpus de
référence (français et latin), les taux de précision convergent rapidement et
atteignent le même niveau que ceux obtenus avec l'ADT (Figure 1). Si nous
connaissons les paramètres à faire varier pour optimiser la détection des
passages-clefs ADT, ceux issus du deep learning sont complètement non
supervisés et découverts automatiquement par le réseau. L'idée des réseaux à
convolution est de proposer un modèle capable de faire automatiquement
une abstraction performante des données.1 La convolution utilise pour cela
un mécanisme de filtres qui va lire le texte avec une fenêtre coulissante pour
extraire à chaque fois une partie de la matière linguistique présente dans la
fenêtre (Figure 2). Avec des centaines de filtres de tailles différentes, le texte
est lu en utilisant tous les empans linguistiques possibles et le mécanisme de
back-propagation2 finit par accorder un certain poids à certains éléments du
texte qui le pousse à prendre la bonne décision. Le deep learning est souvent
considéré comme une boîte noire faute de pouvoir mettre en évidence
précisément ces éléments. Nous avons donc ici concentré nos efforts sur la
déconvolution. Ce mécanisme utilisé notamment en analyse d'images permet
de démêler le réseau et de lui redonner une forme interprétable par l'humain.
Notre modèle est composé d'une couche de pré-apprentissage (Mikolov et
al., 2013) pour la représentation des mots en vecteurs, d'une couche de
convolution (Kim, 2014), un maxpooling pour compresser l'information et
enfin un réseau classique de perceptron à une couche cachée pour la
classification (Figure 2). La déconvoltution est en fait une simple copie
partielle de ce réseau (jusqu'à la convolution) à laquelle on ajoute à la fin une
transposée de la convolution. On copie bien sûr le poids de chaque neurone

1 L'abstraction des données peut être considérée comme les saillances lexicales
d'un texte qui lui donnent une identité propre
2 \Correction de l'erreur à chaque phase d'apprentissage.

464

JADT’ 18

après l’entraînement dans cette copie de réseau et on obtient un nouveau
réseau dont la couche de sortie correspond au résultat de chaque filtre de la
convolution. Une simple somme de ces filtres pour chaque mot nous donne
un indice d'activation du mot dans son contexte. Au final nous observons ici
des zones de texte s'activer plus ou moins suivant l'importance que leurs a
accordée le réseau.

Figure 2. Convolution et déconvolution d’un passage du discours d’E. Macron

3.1. Résultats et perspectives
A la lecture des résultats, nous voyons que le modèle identifie, sans surprise,
des mots que le traitement statistique avait calculés comme spécifiques. Mais
pas seulement. Certaines zones éclairées par le réseau semblent relever d’une
nouvelle forme de lecture du texte. Nous pouvons illustrer ce constat avec un
extrait des vœux d’E. Macron le 31 décembre 2017:
[...] une transformation en profondeur de notre pays advienne à l'école pour nos
enfants , au travail pour l' ensemble de nos concitoyens , pour le climat , pour le
quotidien de chacune et chacun d' entre vous . Ces transformations profondes ont
commencé et se poursuivront avec la [...]
Dans ce passage, les mots transformation et notre, fortement spécifique de
Macron, sont activés : ici il n’y a pas de plus-value heuristique par rapport à
l’ADT. De même, le segment répété chacune et chacun, très spécifique, est
repéré par le réseau. Mais il y a aussi les mots pays et advienne qui ne sont pas
statistiquement spécifique de Macron et qui ont pourtant fortement contribué
à la reconnaissance du passage. Si l’on regarde maintenant les activations
autour de ces mots ciblés, on voit que c’est une expression formée de
plusieurs mots, pas forcément contigus, qui est repérée par le réseau. Il
semble donc que le deep learning ait identifié des structures phraséologiques
ou motifs linguistiques sensibles aux occurrences et à leur organisation
syntagmatique. Plus loin, la visualisation du passage dans son ensemble met
au jour une topologie textuelle ou un rythme auxquels le deep a été sensible
(Figure 3).

JADT’ 18

465

Figure 3. Déconvolution : observation de la topologie d’un passage

3. Conclusion
L’ADT et le deep learning ne sont peut-être pas des continents étrangers l’un
à l’autre (Lebart, 1997). Cette contribution en croisant approche statistique et
réseau de neurones nous a permis d’identifier des passage-clefs et peut-être
des motifs susceptibles de nourrir nos traitements textuels. Si les observables
qui ont présidé à la détection de passages-clefs par l’ADT (les spécificités
lexicales) sont connus et éprouvés, les zones d’activation du deep learning
semblent relever de nouveaux observables linguistiques. Rappelons que la
matière linguistique et la topologie des passages ne sauraient renvoyer au
hasard : les zones d’activations permettent d’obtenir des taux de
reconnaissance de plus de 90% sur le discours politique français et de 85%
sur le corpus du LASLA ; soit des taux équivalents ou supérieurs aux taux
obtenus par le calcul statistique des passages-clefs. Reste désormais à
améliorer le modèle et à en comprendre tous les aboutissants mathématiques
comme linguistiques. La première amélioration que l’on se propose
désormais d’implémenter est l’injection d’informations morphosyntaxiques
dans le réseau afin de mettre à l’épreuve des motifs linguistiques toujours
plus complexes.
References
Ducoffe, M., Precioso, F., Arthur, A., Mayaffre, D., Lavigne, F., et Vanni, L.
(2016). Machine learning under the light of phraseology expertise : use
case of presidential speeches, de Gaulle - Hollande (1958-2016). Actes de
JADT 2016, pages 155–168.
Kim, Y. (2014). Convolutional neural networks for sentence classification.
EMNLP, pages 1746–1751.
Lafon, P. (1984). Dépouillements et statistiques en lexicométrie. Genève-Paris,
Slatkine-Champion.
Lebart, L. (1997). Réseaux de neurones et analyse des correspondances.
Modulad, (INRIA Paris), 18, pages 21–37.

466

JADT’ 18

Longrée, D. et Mellet, S. (2013). Le motif : une unité phraséologique
englobante ? Etendre le champ de la phrase ́ologie de la langue au
discours. Langages 189, pages 65–79.
Mellet, S. et Longre ́e, D. (2009). Syntactical motifs and textual structures.
Belgian Journal of Linguistics 23, pages 161–173.
Mellet, S. et Longrée, D. (2012). Légitimité d’une unité textométrique : le
motif. Actes de JADT 2012, pages 715–728.
Mikolov, T., Chen, K., Corrado, G., et Dean, J. (2013). Efficient estimation of
word representations in
vector space. ArXiv : 1301.3781.
Quiniou, S., Cellier, P., Charnois, T., et Legallois, D. (2012). Fouille de
données pour la stylistique : cas des motifs séquentiels émergents. Actes de
JADT 2012.
Rastier, F. (2007). Passages. Corpus 6, pages 25–54.
Salem, A. (1987). Pratique des segments répétés. essai de statistique textuelle. Paris
: Klincksieck.

JADT’ 18

467

Déconstruction et reconstruction de corpus... À la
recherche de la pertinence et du contexte
Lucie Loubère
Lerass Université de Toulouse – lucie.loubere@iut-tlse3.fr

Abstract
Faced with corpora of large sets of texts, we propose a method of selection,
based on the identification of segments of texts relevant to a topic by
successive classification, then recomposition of the corpus with all the texts
having at least one relevant segment . This approach makes it possible to
preserve the contextualizations and narrative discourses surrounding a
theme while excluding off-topic texts.
Résumé
Face aux corpus constitués de grands ensembles de textes, nous proposons
une méthode de sélection, basée sur l’identification de segments de textes
pertinents à une thématique par classification successive, puis recomposition
du corpus avec l’intégralité des textes ayant au moins un segment pertinent.
Cette démarche permet ainsi de conserver les contextualisations et discours
narratifs entourant une thématique tout en excluant les textes hors-sujet.
Keywords: Big corpus, Reinert classification, Iramuteq
1. Introduction
La multiplication d’outils d’extraction de contenus numériques ou
l’abonnement des universités aux bases de données de presse, sont autant de
raisons favorisant la création de corpus de grande taille. À ces facilités
grandissantes s’opposent de nouvelles difficultés. L’hétérogénéité des
contenus mis à disposition par une communauté, les algorithmes de
recherche de bases de données, ou simplement les limites d’ambiguïté de
requêtes génèrent de nombreux bruits à nos corpus. Nous proposerons ici
une méthode s’appuyant sur une identification de contenu par classification
successive (Ratinaud et Marchand, 2015), puis une régénération du corpus
par concaténation de l’intégralité des articles contenant au moins un segment
de texte (ST) dan le matériel identifié comme pertinent.
2. Problématique
La sélection de corpus par classifications successives, en utilisant comme

468

JADT’ 18

unité le segment de texte, permet d’obtenir un sous corpus pertinent avec
une thématique (Loubère, 2014; Ratinaud et Marchand, 2015). Cependant,
lorsque le corpus de départ est constitué de textes au contenu narratif
structuré et délimité (article de presse, blog, argumentaires dans une
concertation…) ce processus peut supprimer les éléments périphériques au
thème étudié. Ces contenus restent portant pertinents pour la compréhension
de l’objet d’étude, mais peuvent être classés avec le bruit des textes hors sujet
dès les premières étapes de sélection. L’objectif de cettte méthode est donc
d’exclure le bruit de textes hors-sujets tout en conservant le contexte
d’évocation de la thématique principale.
3. Méthodologie
Le processus proposé ici se décompose en trois étapes :
k) Numérotation des textes par un identifiant en méthadonnée
l) Extractions des segments de textes propres à notre thématique par
classifications successives. Cette étape repose sur la classification
hiérarchique descendante (CHD) de type Reinert (Reinert, 1983)
proposée par le logiciel Iramuteq (Ratinaud, 2009). En permettant de
faire émerger les mondes lexicaux, ce traitement nous permet de
sélectionner les segments concernant notre thématique, puis de les
re-soumettre à une CHD afin de préciser le corpus. Cette étape est
reconduite jusqu’à avoir une classification dont toutes les classes
concernent la thématique étudiée.
m) Re-composition du corpus par concaténation des articles
apparaissant au moins une fois dans l’extraction finale de l’étape 2
4. Exemple empirique
Dans les parties qui suivront, nous présenterons une mise en application de
cette méthode sur un corpus utilisé lors de notre thèse (Loubère, 2018). Il est
constitué d’une extraction d’article de presses quotidiennes nationales
(libération, l’humanité, le monde, la croix, le figaro) portant sur la thématique
du numérique éducatif du 01/01/2000 au 31/12/2014. Afin de couvrir le plus
d’informations possible la requête exécutée sur la base de donnée
d’Europresse retournait tous les articles contenant au moins un terme
éducatif dans la liste : collège, lycée, école, éducation et au moins un terme
numérique dans la liste : numérique, informatique, multimédia, TICE.
4.1. Les classifications successives
Cette extraction retourna 18 804 articles, auxquels nous avons retiré 875
doublons. LE corpus exploité ici est donc constitué de 17 929 articles
représentant 450 815 segments de textes, sur lesquels nous avons apposé en

JADT’ 18

469

méthadonnée le numéro de l’article source. Nous allons présenter ici les
classifications successives
Nous avons effectué une CHD de 20 classes en phase 1 et un minimum de
1000 ST par classe, nous obtenons 16 classes représentant 99,72 % du corpus.
Le résultat obtenu est présenté sur le dendrogramme en illustration 1

Illustration 1 : dendrogramme de la première CHD

Ce premier découpage montre une séparation en 3 blocs. Le premier est
composé des personnalités publiques, le second est composé par des
thématiques extérieures à notre sujet. En effet, de nombreux articles
contiennent les termes de notre requête sans être pour autant dans le
domaine éducatif (ou numérique). Ainsi, les classes 9 et 8 regroupent les
actualités ou dossiers portant dans le domaine de la culture. Nous citerons
comme exemple non exhaustif d’article de ce domaine un article du journal
Le monde commentant les sorties cinématographiques dans lequel nous
relèverons « les enfants privés d’école jouant dans les rues », et pour un autre
film « les décors numériques ». Nous retrouvons sur le même principe les
classes 6, 5 et 13 traitant des conflits armés détruisant les lycées et relatant
une infériorité numérique.Enfin, le troisième bloc présente une classe centrée
sur le numérique (classe 12), deux classes centrées sur l’éducatif (11 et 10) et
deux classes sur l’aspect législatif et économique (classes 1 et 2). Afin de
pouvoir affiner ces thématiques et les possibles interactions, nous avons
choisi de conserver le bloc entier, soit les segments composant les classes 1, 2,
10, 11, 12 et 14. L’export précédent nous a permis d’obtenir 194 966 segments
de texte sur lesquels nous avons effectué une deuxième CHD de 15 classes en
phase 1 et seuil minimal de 100 ST. Nous obtenons 14 classes portant sur
99,97 % des segments. Le résultat est présenté en illustration 2.
Ce deuxième découpage reprend une structure en trois groupes. Ici, nous
relevons le contexte économique du marché du numérique (classe 14, 5 et 6).

470

JADT’ 18

Illustration 2 : dendrogramme de la deuxième CHD

Le second bloc (classe 4, 3, 7, 8, 10) est constitué des différents discours
témoins de la numérisation de la société. Le troisième groupe séparé du reste
du corpus par le premier facteur est centré sur-le-champ éducatif. Les trois
premières classes à se détacher partagent un discours sur l’après-formation et
le recrutement (classes 9, 2 et 1). La classe 11 constituant 10,3 % du corpus est
entrée sur l’éducation primaire et secondaire, alors que la classe 12 porte sur
l’enseignement supérieur et la recherche. Notre étude portant sur le système
scolaire secondaire, nous ne conserverons que la classe 11 pour l’étape
suivante.
L’export de cette dernière constitue un corpus de 20 167 segments de texte
sur lesquels nous avons effectué une CHD de 15 classes en phase 1 et un
minimum de 100 ST par classe. Nous obtenons 8 classes rapportant 99,22 %
des segments.. Ce dendrogramme, structuré en deux blocs, nous montre une
séparation entre un discours centré sur l’aspect structurel de l’éducation
(classes 8, 6, 4, 3) et celui traitant de l’enseignement (classes 2, 1, 5, 7).

Illustration 3: dendrogramme de la troisième CHD

JADT’ 18

471

Dans la partie structurale nous retrouverons les segments de texte traitant
des réformes sous un angle gouvernemental (classe 8), suivie de tout le
discours se regroupant des aspects temporels, comme le temps de travail
mais également les rythmes scolaires (classe 6). La classe 3 constitue un
discours sociologique sur l’éducation, nous y retrouvons de nombreuses
statistiques étudiant les répartitions sociales dans les différents cursus. Enfin,
la classe 4 traite des établissements scolaires dans leurs diversités.
Les autres classes portent toutes sur le domaine pédagogique : la classe 7
concerne les contenus d’enseignement. La classe 5 traite de la mise en place
d’outil numérique parascolaire (jeux éducatifs, fiche de révision) alors que la
classe 2 est centrée sur la mise en place de formations à distance. Enfin, la
classe 1 est le discours portant sur le numérique dans l’éducation, les mots
clés employés dans notre requête y sont tous surreprésentés. Nous ne
conserverons dons que les segments composant cette classe.
L’extraction de cette dernière classe nous permet d’obtenir 2072 segments sur
lesquels nous avons effectué une CHD de 20 classes en phase 1 avec un seuil
de 100 ST par classe. Cette classification nous a montré une réelle stabilité de
la thématique. En effet, les 8 classes exposée portent chacune sur un aspect
du numérique éducatif.

Illustration 4 : dendrogramme de la quatrième CHD

4.2. Classification du corpus recomposé
Le corpus recomposé des 2902 articles contenant au moins un segment de
texte dans la classe 1 de la troisième CHD est constitué de 72460 segments.
Une CHD de 20 classes en phase 1 et un minimum de 800 ST par classe nous
donne le dendrograme suivant :

472

JADT’ 18

Illustration 5 : dendrogramme de la CHD sur le corpus recomposé

Nous y retrouvons donc au-delà de discours sur l’utilisation du numérique
dans les établissements, un discours sur l’économie reflétant le marché du
numérique éducatif et les frais engendrés par les dotations des
établissements. Un discours à la frontière de la culture et de l’éducation, avec
les formations de ces domaines empreinte de numérique. Mais également un
discours sur l’actualité géopolitique mondiale contextualisant des initiatives
où le numérique apporte des solutions éducatives lors de ségrégation
ethniques, ou éloignements géographiques. Tous ces mondes lexicaux
constituent des éléments du discours social sur notre sujet, qu’une étude
réduite aux segments ciblés lors des CHD successives ne permettrait pas
d’explorer.
5. Conclusion
Le principe des CHD successives, s’il nous permet d’accéder finement aux
segments contenant le discours sur le numérique éducatif, nous éloigne
d’une compréhension globale du sujet. En effet, interroger les bases de
données de presse sur une longue période et une sélection de presse
généraliste apporte une quantité importante de documents hors contexte. Ces
données portent des éléments contextuels communs avec les articles traitant
de notre sujet (personnalités politiques, discours économique…), la proximité
lexicale des segments de ces champs structure les classes de discours
communes aux articles portant sur notre sujet ou non. Cette hétérogénéité
associée à l’insécurité d’un grand ensemble (Geffroy et Lafon, 1982) nous
empêchant une connaissance du corpus antérieure à l’analyse lexicométrique
conduit « à tracer un peu trop vite une autoroute » (Geffroy et Lafon, 1982, p.
140) jusqu’à notre classe 1 finale. Ce phénomène questionne la constitution
d’un corpus sur une dimension architextuelle, alors même que l’outil de
classification utilisé ici joue sur un niveau intertextuel et cotextuel (Rastier,
2015), rapprochant des passages de textes en fonction de leur structure
lexicale. La présence de textes aux sujets hétéroclites fait ressortir de façon

JADT’ 18

473

précoce des thématiques indépendamment de leur hypothétique poids dans
le corpus qu’aurait constitué une sélection de textes centrés sur notre sujet.
Ainsi, les segments traitant de sujets de politique générale ou exposant le
contexte social d’un pays dans les articles traitant du numérique éducatif
sont classés avec ceux des articles hors sujets. Cette difficulté éloigne le
chercheur de la compréhension d’un discours. La démarche que nous venons
de présenter nous permet de se rapprocher d’un positionnement de
textomètre (Pincemin, 2012), sélectionnant les segments pertinent par une
démarche inductive, mais en conservant l’unité sématique du texte dans la
construction du corpus final.
Bibliography
Geffroy, A., & Lafon, P. (1982). L’insécurité dans les grands ensembles.
Aperçu critique sur le vocabulaire français de 1789 à nos jours d’Etienne
Brunet. Mots, 5(1), 129-141.
Loubère, L. (2014). Le traitement des TICE dans les discours politiques et
dans la presse. In Présenté à 12èmes Journées internationales d’Analyse
statistique des Données Textuelles.
Pincemin, B. (2012). Sémantique interprétative et textométrie. Texto! Textes et
Cultures, 17(3), 1-21.
Rastier, F. (2015). Arts et sciences du texte. Paris: Presses universitaires de
France.
Ratinaud, P. (2009). IRAMUTEQ : Interface de R pour les Analyses
Multidimensionnelles de TExtes et de Questionnaires. Consulté à
l’adresse http://www.iramuteq.org
Ratinaud, P., & Marchand, P. (2015). Des mondes lexicaux aux
représentations sociales. Une première approche des thématiques dans
les débats à l’Assemblée nationale (1998-2014). Mots. Les langages du
politique, (2), 57-77.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique:
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, 8(2), 187-198.

474

JADT’ 18

L’apport du corpus-maquette à la mise en évidence
des niveaux descriptifs de la chronologie du sens.
Essai sur une Série Textuelle Chronologique du
Monde diplomatique (1990-2008).
Heba Metwally
Université d’Alexandrie, Égypte – heba.metwally77@gmail.com

Abstract
Chronological corpora and particularly time series (Lebart et Salem 1994)
organize the textual data in corpora according to their natural sequence in
time. Today, scholars are interfacing increasingly with chronological corpora
following the democratization of access to big data. The lexicometry develops
into stylometry, textometry and logometry. And statistical data analysis
integrates the observation of co-occurrential systems and lexical networks in
their complexity. This improves the analysis of semantic contents according
to their localisation in the semantic strata. This contribution aims to enhance
the description of the chronology of meaning. The study is based on a corpus
of more than 5000 articles (ca 11 millions of tokens) published in the Monde
diplomatique between January 1990 and December 2008.
To analyze big chronological corpora we propose a scale model of the
chronological corpus by compressing the initial corpus to its most frequents
nouns. The compression procedure is duplicated in the four sub-corpuses of
relevant semantic stability. We obtain two descriptive levels of chronology:
the synthetic level of dominant contents and the analytical level of the four
chronological phases of meaning. The two levels are intended to respond to
different investigations on time and meaning. Working on sets of scale models
that are either connected horizontally (chronological sequence) or vertically
(the synthetic perspective clarified by an analytic perspective) enlarges our
field of observation and deepens our understanding of chronological data in
particular and the unfolding of text in general.
Keywords: chronological corpus – logometry – logogenesis – clustering –
method Reinert – corpus semantics – media analysis

JADT’ 18

475

Résumé
Les corpus chronologiques et a fortiori les Séries Textuelles Chronologiques
(Lebart et Salem, 1994) organisent les données textuelles dans le corpus selon
leur enchaînement naturel dans le temps. La banalisation des corpus textuels
et l’accès facilité et accéléré au big data multiplient les corpus
chronologiques, puisque finalement toute production textuelle s’étale dans le
temps. La lexicométrie – au sens classique – doublée de la stylométrie, de la
textométrie voire de la logométrie, et la statistique occurrentielle enrichie par
un outillage cooccurrentiel (Viprey, 1997), (Mayaffre, 2014), la voie est
ouverte aujourd’hui à une observation améliorée des contenus sémantiques
qui gagnent en visibilité grâce aux tentatives parfois incontrôlées de leur
objectivation. Cette contribution a pour objectif de contribuer à la description
de la chronologie des contenus sémantiques. On s’appuie sur un corpus
d’articles du MD (1990-2008). On compte plus de 5000 articles et plus de 11
millions d’occurrences. On propose pour cela le recours à un corpus-maquette,
une compression du corpus chronologique intégral à partir des noms les plus
fréquents. Cette démarche de compression est reproductible dans les souscorpus des périodes de stabilité sémantique. On obtient deux niveaux
descriptifs de la chronologie, à savoir le niveau global, synthétique des
contenus dominants et le niveau subordonné, analytique des sens particuliers
des phases transitoires du discours. Les deux niveaux infèrent sur un
questionnement différent sur le temps en multipliant les pistes
d’interrogation et en articulant le niveau synthétique et son niveau
analytique.
Mots-clés: corpus chronologique – logométrie – logogénétique –
classification – méthode Reinert – sémantique de corpus – Analyse de
discours médiatique
1. Introduction
Dans la tradition lexicométrique, les STC (Séries Textuelles
Chronologiques) problématisent les investigations sur le temps1. Ce type de
corpus est né, dans les études à caractère historique, du questionnement sur
le changement dans le discours au fil du temps. Et les travaux d’André

1 « Nous appelons séries textuelles chronologiques ces corpus homogènes
constitués par des textes produits dans des situations d'énonciation similaires, si
possible par un même locuteur, individuel ou collectif, et présentant des
caractéristiques lexicométriques comparables. » (Lebart et Salem, 1994 : 217)

476

JADT’ 18

Salem2 témoignent de l’intérêt porté à la description des corpus textuels
chronologiques. Pour ce faire, André Salem généralise les STC, décrit la
particularité des sorties machines des analyses statistiques qu’elles
produisent (AFC ; calcul de spécificités), introduit la notion de «temps
lexical », et conçoit une gamme de calculs visant, dans un premier temps, la
« mise en évidence et la mesure du stock lexical au cours du temps » (Salem,
1988 : 118) et, dans un second temps, la caractérisation des périodes dans une
STC. Plus généralement, la particularité des STC est de concilier la linéarité
du texte, du temps et la sérialité du corpus. Si tous les corpus sont
partitionnés en séries pour permettre la comparaison, ces séries ont
l’avantage de conserver l’ordre naturel des textes qui s’échelonnent – sans
conflit – dans le corpus et dans le temps.
Aujourd’hui, le champ des observables est constamment élargi grâce à
l’évolution des outils informatiques et au progrès de la tokenisation pour
embrasser progressivement des niveaux descriptifs textuels que le chercheur
filtre ou articule à sa guise. La lexicométrie est enrichie et mise à jour par la
textométrie et la logométrie dont le projet est de dépasser la lexie vers les
textes, le discours et le sens. Le sens est objectivable grâce à la formalisation
de la cooccurrence, et à son baptême comme unité minimale de
contextualisation, i.e de sens (Mayaffre, 2008). Dès lors, la statistique
occurrentielle se double de la statistique cooccurrentielle. La cooccurrence
devient unité de décompte généralisée à laquelle s’applique les calculs
statistiques traditionnels (Brunet, 2012). Des applications d’ADT de tradition
benzécriste se développent pour appréhender les réseaux lexicaux dans leur
complexité. La cooccurrence généralisée (Viprey, 1997, 2005, 2006) se donne une
visée exploratoire et la méthode Alceste (Reinert, 1983, 1993) procède à la
démarche classificatoire des réseaux lexicaux structurants des textes. C’est
dans ce cadre des progrès de la méthodologie et de la technologie qu’une
sémantique de corpus (Rastier, 2011) est envisageable.
Ce champ d’investigation intéresse naturellement les études chronologiques
qui peuvent désormais observer le mouvement des contenus sémantiques
dans le temps pour comprendre l’impact du temps dans la thématisation
d’une Série Textuelle Chronologique3. Pour l’objectivation des fonds

Cf. (Salem, 1988, 1991, 1993, 1994)
Ce point précisément constitue la problématique de notre thèse de doctorat
intitulée « Les thèmes et le temps dans Le Monde diplomatique (1990-2008) », soutenue
le 11 décembre 2017 à l’Université Côté d’Azur (UCA) à Nice.
2
3

JADT’ 18

477

sémantiques4 du discours, on sollicite la méthode Alceste implémentée dans le
logiciel libre Iramuteq (Ratinaud et Marchand, 2012) qui s’articule à
Hyperbase. Pour une visualisation améliorée des topics du discours, on
propose de recourir à une maquette du corpus et de ses sous-corpus. Au sens
propre, la maquette est une représentation en trois dimensions, à échelle
réduite qui reste fidèle dans ses proportions. Ici, dans le cas des corpus
textuels, la maquette est une compression du corpus intégral qui se réduit à
ses noms les plus fréquents. A partir d’une STC du Monde diplomatique (19902008), cette contribution se donne deux objectifs. Dans un premier temps, elle
vise à mettre en exergue les deux niveaux descriptifs complémentaires de la
chronologie du sens : chronologie des contenus dominants (3.) et la
logogénétique (4.) tout en relevant l’intérêt de étude conjointe de ces deux
niveaux. Dans un second temps, il s’agit également de mettre à l’épreuve
notre proposition de la maquette. On recherche une visualisation améliorée
des contenus sémantiques structurants grâce au recours à une maquette,
reproduction grossière et fidèle des textes dont l’usage spécifique sera illustré
dans les lignes suivantes.
2. Du corpus intégral à la maquette du sens et du temps
Le choix du Monde diplomatique pour l’étude de l’évolution du sens s’appuie
sur la richesse et la stabilité de son contenu. La période couverte par cette
étude marque un moment historique important, à savoir le monde après la
chute du Mur de Berlin. En plus, cette période se caractérise par une
continuité éditoriale5. Bref, nous avons affaire à un discours stable, sans
complexe qui à l’examen multidimensionnel épouse un schéma évolutif
classique sans ruptures6. On estime que la stabilité du discours est un facteur
indispensable à l’étude de l’évolution, celle-ci reposant principalement sur la
continuité.
La finalité de ce travail, à savoir l’étude de la chronologie du sens d’un gros
corpus textuel, préside à la conception de la maquette. La taille du corpus

Les fonds sémantiques sont les isotopies ou les macrostructures sémantiques sur
lesquelles se détachent les formes sémantiques que sont les thèmes. Cf. (Rastier, 2011 :
24)
5 Il s’agit du mandat d’Ignacio Ramonet qui est directeur de la publication de
janvier 1990 à mars 2008.
6 Par examen multidimensionnel on entend l’AFC de la distance entre les textes
qui dans le cas des données sérielles reproduit une forme parabolique baptisée
parabole Guttman et qui est symbolique du mouvement linéaire des données
ordonnées dans le temps. Cf. (Salem, 1991).
4

478

JADT’ 18

intégral excédant 11 millions d’occurrences (voir ci-dessous Tableau 1) pose
immédiatement le problème de son interprétation comme il nous confronte à
la difficulté de l’appréhension des fonds sémantiques structurants du corpus.
En ADT, les chercheurs procèdent assez souvent pour des raisons pratiques à
des sélections au sein de la population statistique étudiée. A notre tour, on
propose un mode de réduction qui se fonde sur la finalité herméneutique et
perpétue la pratique d’une sémantique interne. On pose ici – sans généraliser
– que le discours médiatique par sa vocation informative et sa référence au
monde structure son contenu d’une manière privilégiée autour des noms. La
classe nominale (noms communs et noms propres) est la classe grammaticale
la plus importante dans le corpus ; elle couvre 28,9% de la surface du corpus.
Elle connaît également une stabilité distributionnelle au fil de la STC.
L’importance numérique absolue et la distribution équilibrée attestent le
critère de la représentativité statistique7. Aussi une comparaison avec
d’autres corpus8 entre les listes des lemmes les plus fréquents triés par
catégorie grammaticale confirme le pouvoir caractérisant de la classe
nominale en général et des noms propres en particulier. On s’appuie donc
sur la classe nominale et l’argument fréquentiel pour réduire le corpus
intégral à ses 380 noms les plus fréquents. La démarche laisse intacts les
partitions du corpus et l’enchaînement des textes pour respecter la structure
séquentielle des textes et la conception chronologique du corpus. L’une et
l’autre garantissent au corpus textuel son authenticité ; seul leur maintien
autorise l’examen de l’hypothèse de travail présidant à la conception du
corpus textuel. Pour expliquer un peu ce travail philologique simple dans
son principe, la démarche consiste à mettre un cache sur tout le texte à
l’exclusion des 380 noms les plus fréquents. Cette procédure est à reprendre
dans les sous-corpus de stabilité sémantique. Celle-ci se laisse mesurer d’une
manière endogène à l’aide du calcul de la distance entre les textes à partir de
la forme minimale de signification thématique, la cooccurrence. La distance
intertextuelle calculée sur les cooccurrences au sein des noms de la maquette
donne à voir quatre périodes qui fondent les quatre sous-corpus, ceux-ci
réduits à leur tour à des maquettes. Cette périodisation endogène fonde le

Dans notre travail doctoral (Metwally, 2017), nous avons étudié les contenus
des classes de fréquences du corpus intégral pour une compréhension de la hiérarchie
numérique du lexique. Aussi avons-nous analysé la structure grammaticale des
données et leur distribution dans la STC.
8 (Labbé et Monière, 2003); (Mayaffre, 2004).
7

JADT’ 18

479

temps sémantique9 selon lequel on remodèle le corpus intégral et sa maquette.
Le tableau 1 (ci-dessous) synthétise la structure lexicale du corpus, des souscorpus et de leurs maquettes. Celles-ci couvrent chacune approximativement
9,8% de la surface leurs corpus originaux respectifs. Cette stabilité de
représentativité numérique autorise la comparaison entre les données.
Tableau 1: Tableau synthétique de la structure lexicale du corpus,
des sous-corpus et de leurs maquettes
corpus et souscorpus
taille
(N=occurrences)
vocabulaire
(V=mots)
maquette et sousmaquettes
(V=noms)
maquette et sousmaquettes (taille)

1990-1993

1994-1997

1998-2001

2002-2008

1990-2008

2697013

2402434

2552998

3765908

11418356

67989

67571

70954

86032

140690

307

282

290

375

380

266439

218643

229119

382298

1115311

On obtient donc finalement un dispositif complexe à deux niveaux : le niveau
global des contenus sémantiques de l’ensemble de l’empan chronologique
étudié dont on peut étudier la dynamique (3.); et le niveau analytique,
d’ordonnancement chronologique, des phases sémantiques stables et qui
permet et l’observation du mouvement des contenus sémantiques et la
confrontation avec le niveau global synthétique (4.). L’étude des fonds
sémantiques est concevable en mobilisant la statistique cooccurrentielle qui
met en évidence les structures sémantiques pertinentes. A l’issue de la CHD
appliqué à la maquette et ses sous-maquettes, sont observables les mondes
lexicaux stabilisés (Reinert, 1993, 2008) du sens global et de ses phases
transitoires (voir les dendrogrammes Fig. 1, 3, 4).
3. La dynamique des contenus dominants
La démarche habituelle dans les études chronologiques repose d’abord sur
une étude statique première du sens global pour procéder ensuite à une vue
dynamisée. Les vues statiques relèvent d’un artifice méthodologique
provisoire destiné à mettre en évidence les contenus sémantiques stabilisés
9 On s’est permis de parler de temps sémantique à la suite du temps lexical d’André
Salem (1988). Le temps sémantique est le rythme selon lequel s’organisent dans le
temps les contenus sémantiques et que mesure ici la distance intertextuelle calculée
sur la cooccurrence.

480

JADT’ 18

au bout d’un mouvement dynamique. La saisie du sens global répond au
questionnement sur les contenus dominants, consensuels d’une période à
l’autre, qui survivent au cours de 19 ans de production d’articles. Pour
l’analyse de la structure sémantique de la maquette, on donne à Iramuteq la
maquette globale, où les 380 noms les plus fréquents s’organisent sur l’axe
syntagmatique selon l’ordre de leur apparition, et dont les partitions assurent
au corpus une structure chronologique adaptée au temps sémantique du
corpus. Une fois Iramuteq mobilisé, il se met à découper le texte en segments
de textes paramétrables. Le choix de l’étendue des segments de textes (ST) est
capital, car ce sont les ST qui constituent les énoncés analysés et classés par la
méthode Alceste. Pour ces unités de contexte on a estimé la succession de 10
noms dans le corpus-maquette comme l’équivalent dans le corpus intégral de
la fenêtre contextuelle de 33 mots10. On vise par là un espace intermédiaire
entre la phrase et le paragraphe. Une fois Alceste activé, il procède à une
CHD qui croise les ST et les noms pour effectuer un classement partant du
caractère lexical prédominant des ST.

Figure 1 : Les mondes lexicaux de la maquette (1990-2008)11

On impose à l’algorithme un paramétrage exigeant qui nous garantit une
grille de lecture assez riche. Avec 15 classes demandées à l’issue de la phase

Cette estimation repose sur le pourcentage de la classe nominale dans
l’ensemble du corpus (28,9%). Voir (Metwally, 2017).
11 Dans ces listes, on peut repérer quelques verbes (partir, produire, revenir,
sentir, passer). Il s’agit d’une erreur due à une lemmatisation effectuée par Iramuteq
malgré les tentatives de dissuasion. Il s’agit plutôt de substantifs (parti, produit,
revenu, sens, passé).
10

JADT’ 18

481

1, 8 se trouvent stabilisées (Figure 1). Les sorties machines de la CHD sont
multiples. La représentation en dendrogramme correspond au classement
stricto sensu ; et elle est enrichie d’informations supplémentaires qui mettent
en valeur la CHD. On commence par l’identification rapide de la structure
sémantique du discours et de la hiérarchie de l’information.
Le dendrogramme, par sa logique binaire de représentation, oppose les
contenus économiques, les plus importants avec 41,5% des ST classés, aux
contenus non-économiques. Ceux-ci distinguent les thématiques politiques
(35,2% des ST classés) et les thèmes de l’Homme (23,3% des ST classés),
thématiques socio-culturelles qui traitent de sujets historiques et culturels et
de questions sociétales. Suivant la logique hiérarchique descendante de la
classification, des classes spécialisées se stabilisent pour mieux caractériser
les trois domaines sémantiques identifiés. Au sein des classes économiques se
spécialise une classe socio-économique dédiée aux questions de l’emploi et
du travail (classe 8 ; « emploi », « travail », « chômage », « salaire »,
« syndicat ») ; celle-ci se distingue des deux classes de la macro-économie qui
traitent de l’économie domestique (classe 2), de la machine économique des
pays (« développement », « industrie », « concurrence », « secteur »), et
l’économie mondiale (classe 7) qui couvre les questions des finances et de la
performance économique des pays sur le marché mondial (« dollar »,
« banque », « dette », « prix », « croissance »). Attachés à la même branche
des thèmes politiques, les mondes lexicaux de l’Homme connaissent une
variation qui différencie les questions philosophiques et/ou idéologiques sur
l’histoire et la culture (classe 1 ; « histoire », « siècle », « monde », « culture »,
« sens », « conscience », « passé ») du quotidien des êtres humains dans ce
monde (classe 6 ; « femme », « enfant », « victime », « quartier », « violence »,
« police », « vie », « école »). Si l’analyse du sens passe nécessairement par la
suspension provisoire de la structure sérielle du corpus, l’interrogation des
partitions de la maquette sur leur part aux classes lexicales restitue la
temporalité définitoire du corpus. Une projection des classes dans les
périodes de stabilité sémantique met en évidence la dynamique des classes,
la thématisation de chaque période pour permettre finalement d’inférer sur
l’évolution du sens.
Les classes lexicales poursuivent différentes tendances au cours du temps.
Les thèmes du pouvoir (classes 4 et 5) est un axe informatif important qui ne
subit guère de variations quantitatives. La classe des politiques
internationales (classe 3) connait un pic positif exceptionnel dans la dernière
période.

482

JADT’ 18

Figure 2 : Périodes et classes de la maquette (écarts en Chi2)

Ce sont les contenus économiques et socio-historiques qui sont traversés par
deux logiques évolutives opposées. L’ordonnancement des bâtons positifs
met en relief les pics positifs importants et exclusifs de deux classes
économiques dans les deux premières périodes. Cette importance s’évanouit
progressivement. Dans la dernière période les déficits les plus importants
sont ceux des classes économiques. Face à la régression des contenus
économiques, la progression est réservée aux contenus socio-historiques
(classes 1 et 6). Il s’ensuit une couleur thématique changeante d’une période à
l’autre. Les contenus économiques qui marquent les 19 ans qui ont suivi la
chute du Mur de Berlin proviennent majoritairement des deux premières
périodes, tandis que les deux périodes suivantes connaissent des centres
d’intérêt socio-historiques qui se mêlent dans la troisième période à des
thèmes économiques et dans la dernière période aux événements globaux de
politiques internationales. A l’œil nu, l’histogramme de la dynamique du
sens global se laisse diviser en deux moments évolutifs distincts et
asymétriques. Sur le plan quantitatif, le sur-emploi de la première moitié de
la série n’est jamais égalé par un sur-emploi pareil dans la deuxième moitié.
Sur le plan qualitatif, les contenus majoritaires de la première partie sont des
contenus techniques et relèvent de l’axe informatif le plus important, un axe
technique qui relève des visions macro. Par contre, les contenus dominants
de la deuxième moitié de la série sont plus variés et traduisent un intérêt
croissant aux sujets philosophiques et humanistes. Un mouvement général
semble déplacer le focus de l’ordre mondial vers les hommes et le sens de
leur vie dans le monde.

JADT’ 18

483

La description de la chronologie du sens touche à ses limites. Car les
contenus dominants qu’on observe ici sont précisément les contenus
consensuels, ceux qui trouvent toujours leur expression d’une période à
l’autre selon un dosage qui leur garantit finalement la supériorité
quantitative. Le mouvement dynamique de ces contenus revient donc à une
interrogation sur leurs périodes spécifiques. Ceci dit, on pose que la
dynamique des contenus dominants repose nécessairement sur les sens
particuliers de ces périodes. L’étude du niveau subordonné de la génétique
du discours (tout de suite ci-dessous) est certes instructive pour une analyse
plus détaillée de la spécificité sémantique de chaque période. L’étude de la
formation du sens nous renseigne également sur le rapport entre le sens
particulier, temporaire et le sens général, dominant. Elle est indispensable
pour compléter et éclairer nos observations sur l’évolution.
4. La logogénétique ou la génétique du discours
Le mot logogénétique reprend le mot anglais logogenesis dont Halliday (1994)
explicite la signification et l’intérêt en termes suivants :
“It is helpful to have a term for this general phenomenon
– i.e. the creation of meaning in the course of the unfolding of text. We shall call
it logogenesis, with ‘logos’ in its original sense of ‘discourse’ (see Halliday &
Matthiessen, 1999: 18; Matthiessen, 2002b). Since logogenesis is the creation of
meaning in the course of the unfolding of a text, it is concerned with patterns
that appear gradually in the course of this unfolding; and the gradual
appearance of patterns is, of course, not limited to single texts but is rather a
property of texts in general instantiating the system of language.” (Halliday,
1994 : 601)

La logogénétique ou la génétique du discours permet de renouer avec les
modèles linguistiques qui traversent le texte et contribuent à sa formation.
Concrètement ici, on voit dans l’observation et la confrontation ordonnée
dans le temps des CHD des quatre sous-maquettes un grand intérêt pour
rétablir les modèles sémantiques propres des périodes de stabilité
sémantique et qui fondent le mouvement général du sens et sa stabilisation
au niveau global au cours du temps. On reprend les mêmes paramètres
utilisés pour la CHD de la maquette globale dans les quatre sous-maquettes
pour obtenir les dendrogrammes ci-dessous (Fig. 3, 4). Un examen attentif de
la structure interne des sous-maquettes du sens est susceptible d’offrir des
grilles de lectures analytiques des contenus dominants, de leur dynamique et
de leur formation. On ne saura pas épuiser la valeur heuristique de ces

484

JADT’ 18

dendrogrammes. Et on se contente de souligner l’apport principal de cette
démarche à la description du sens sans prétendre effectuer une analyse
fouillée du sens. Celle-ci devrait reposer sur une étude systématique des
réseaux lexicaux ce qui dépasse l’objectif de cette contribution.

Figure 3 : Les mondes lexicaux des deux premières périodes

JADT’ 18

485

Figure 4 : Les mondes lexicaux des deux dernières périodes

La première remarque à souligner est la permanence des fondamentaux du
discours et le nombre fixe de mondes lexicaux qui se stabilisent d’une
période à l’autre. Cette stabilité de la structure sémantique ratifie la
pertinence de l’étude de l’évolution. Celle-ci s’effectue nécessairement au
sein d’un environnement stable. Observons l’évolution de la hiérarchie de

486

JADT’ 18

l’information d’une période à l’autre. Le graphique ci-dessous (Figure 5) rend
compte de l’importance de chaque domaine sémantique au sein des ST
classés. La comparaison est instructive d’une période à l’autre, et entre le
niveau des sous-maquettes et le niveau supérieur de la maquette globale.

Figure 5 : L’évolution de l’importance des fondamentaux du discours au cours du temps (en
pourcentages)

Quelle que soit la période, les contenus politiques restent les plus dominants.
A l’examen de la répartition interne des classes politiques on note
l’importance des classes de politiques internationales qui sont constamment
au nombre de deux (Fig. 3, 4) par opposition au niveau global qui ne connaît
qu’une seule classe (Fig. 1, classe 3). C’est l’ampleur des classes de politiques
internationales dans les sous-maquettes qui fait la supériorité des thématiques
politiques. Et pourtant, ce n’est pas le cas au niveau global. Ceci est dû
principalement à la nature conjoncturelle des événements internationaux : les
guerres américaines de la première et dernière période, les questions
sécuritaires d’actualité en Europe après la chute du mur de Berlin, la guerre
de Kosovo dans la troisième période, le conflit israélo-palestinien avec ses
variantes et ses flux et reflux au cours du temps (voir le contenu des classes
lexicales, Fig. 3, 4). Tant d’événements spécifiques de certaines périodes et
qui ne parviennent pas tous à se stabiliser au niveau global pour caractériser
les 19 ans. D’où la prédominance des contenus politiques dans les sousmaquettes et leur recul au niveau global.
Par contre, les contenus économiques connaissent une tendance inverse. Au
niveau global, ils occupent le sommet de la pyramide hiérarchique avec trois

JADT’ 18

487

classes. Au niveau subordonné des sous-maquettes, ils viennent en deuxième
rang pour passer dans la dernière période au troisième rang. Le nombre de
leurs classes fluctue entre trois et un. Ce qui est curieux est que la variété
maximale du nombre des classes économiques finit par se stabiliser au
niveau global. À la différence des thématiques de politiques internationales,
les thématiques économiques connaissent des prolongements plus pérennes.
Il suffit d’observer les dendrogrammes des sous-maquettes pour localiser dans
le temps les sources des trois classes économiques de la maquette globale.
Comme le montre bien l’évolution de la hiérarchie de l’information (Fig. 5),
les thèmes socio-historiques continuent à s’amplifier pour dépasser les
thématiques économiques dans la dernière période. Ce constat est bien
compatible avec la dynamique du sens global (Fig. 2) où on a observé les
déficits record des thèmes économiques et le sur-emploi significatif des
classes socio-historiques. Notons également que ces dernières croissent
quantitativement et qualitativement. C’est exclusivement dans la dernière
période qu’on a affaire à deux classes socio-historiques. Dans cette dernière
période la classe 6 caractérisée par « enfant » et « femme » ressemble à la
classe 6 de la maquette globale (Fig. 1), tandis que la classe voisine (classe 2)
lexicalisée par « science », « recherche », « individu », « pratique » n’a pas
d’équivalent lexical au niveau global. Il s’agit de contenus émergents qui ne
trouvent pas de précédents dans la STC. Le vocabulaire de la classe 2 se situe
à mi-chemin entre le sociétal et le social. Le ST le plus caractéristique de la
classe nous éclaire sur sa particularité rhétorique. A l’occasion du Sommet G8
2007 dont le thème est ‘croissance et responsabilité’, le MD lance un tract
appelant à une révolution culturelle généralisée. On élargit la fenêtre de
l’observation au-delà des limites du ST12 pour améliorer l’identification du
contenu sémantique:13
« A quand, là encore, la lancée d’initiatives mondiales de la part de
quelques pays courageux – on attend la France – pour prendre à contrepied la vieille tentation d’inféoder la recherche aux désignations

Tandis que le ST se limite à la succession de 10 noms parmi les 380 noms les
plus fréquents du corpus, la lecture ne s’arrête pas aux frontières des ST mais elle en
part. Selon Rastier (2007), le passage - îlot de pertinence – « n’a pas de bornes fixes et
son empan dépend évidemment du point de vue qui a déterminé sa sélection » (p. 31).
Notre paramétrage cible le paragraphe, i.e la période, qui relève du niveau
mésotextuel, lieu de l’observation et de l’objectivation des thèmes. Et la lecture
poursuit sur l’axe syntagmatique le développement d’un thème d’un ST à l’autre.
13 Sont mis en rouge uniquement les noms spécifiques de la classe 2.
12

488

JADT’ 18

d’objectifs par quelques manipulateurs, et pour lancer les chercheurs, au
contraire, à l’assaut des nouvelles questions vitales : telles, en sciences
humaines, les formes de légitimité anthropologique, politique et
démocratique qui conviendraient à une société-monde en formation ;
telle, en sciences technologiques, la rupture nécessaire avec les grands
systèmes énergivores, laquelle permettrait demain aux sociétés – locales,
urbaines, régionales – d’assurer leur autonomie alimentaire et
énergétique sans se désengager de la conversation mondiale autorisée
par la circulation instantanée des données ? Bref, le pire des réflexes de
solidarité défensive ne parvient plus à occulter les questions désormais
immédiatement planétaires : celle qu’on ne tergiversera plus à nommer
simplement la nature, ce support de la vie terrestre devenu poste de
résistance principal pour le mirage de la valeur argent ; celle de la
culture, aussi bien identitaire et artistique que scientifique, et qui
constitue – au moins à l’égal de la production matérielle désormais
technologisée – un vaste univers d’activités essentielles, dont la logique
ouverte ne peut être inféodée au rendement de type industriel ou
financier sans péril pour l’humanité civilisée, et pour sa pluralité
démocratique ; et enfin la question cruciale des sociétés plus autonomes
par rapport au tourbillon techno-chrématistique, et qui seront dans
l’avenir autant de sources d’emplois plus stables, d’activités moins
gaspilleuses d’énergie et moins polluantes, et aussi de conversations
politique plus proches des citoyens. » (Août 2007)
Le ST le plus spécifique fait partie d’un passage qui fait appel à une
révolution culturelle généralisée. Celle-ci se charge de poser les questions
sociétales et civilisationnelles les plus urgentes et de promouvoir les
alternatives-solutions. La révolution est celle de la culture scientifique. Est
urgente une refonte de la pensée dominante et unique dans tous les
domaines. Tout est à réinventer : des théories de référence pour une sociétémonde autre que la mondialisation, des théories économiques au service des
sociétés et des hommes, d’autres technologies bioéthiques qui respectent la
nature, ceci pour rester fidèle à la culture démocratique. Ce passage donne
une idée sur la couleur sémantique de cette classe exclusive de la dernière
période et qui échappe au sens global. D’une manière générale, les contenus
socio-historiques connaissent un tournant qualitatif au cours du temps. Sur
les dendrogrammes (Fig. 3, 4) on identifie leur emplacement libre entre les
thèmes politiques et les thèmes économiques d’une période à l’autre. Dans
les deux premières périodes, les questionnements sur l’histoire et la condition

JADT’ 18

489

de l’Homme sont mobilisés par la situation politique, tandis que les contenus
économiques régressifs des deux dernières périodes attirent les thèmes sociohistoriques.
5. Conclusion
Rapporter la structure sémantique des sous-maquettes à la dynamique des
contenus dominants nous éclaire sur la formation du sens global et sur sa
logique. Autrement dit, la dynamisation du sens global par la projection des
classes lexicales sur la chronologie constitue un niveau intermédiaire entre le
niveau des sous-maquettes, celui des phases sémantiques stables et de leurs
sens particuliers d’un côté et le niveau synthétique du sens qui finalement se
stabilise au niveau global après l’accumulation des sens particuliers. Ce
qu’on voulait illustrer ici c’est ponctuellement l’intérêt du recours à une
maquette, réduction raisonnée du corpus à ses noms les plus fréquents,
modèle à échelle réduite repris dans les sous-corpus de stabilité sémantique.
Cet usage couplé à une statistique cooccurrentielle ciblant les réseaux
lexicaux structurants permet un accès rapide aux fonds sémantiques, condition
première pour pratiquer une sémantique de corpus. La maquette balise une
sémantique de corpus qui va du global au local (Rastier 2001). Plus
concrètement, si la cooccurrence est l’interprétant minimal saisi au sein du
passage (Rastier 2007), on lui a assigné la mission de mesurer le temps
sémantique pour déterminer les phases de stabilité sémantique où l’on peut
observer les mondes lexicaux stabilisés (Reinert 1993, 2008). Ceux-ci sont les
interprétants maximaux objectivables au niveau de la maquette et des sousmaquettes. La maquette telle qu’on la conçoit ne renvoie pas à un modèle
généralisable mais à un usage généralisable. Un usage qui pour chaque
corpus contribue à la reconstitution de son modèle sémantique quelle que
soit sa spécificité et à réaliser la vocation de sa conception. Ici, dans le cas des
corpus chronologiques, la maquette réconcilie l’étude du sens et l’étude du
temps. Tandis que la première passe par délinéarisation du texte et la capture
de la structure non-séquentielle du texte, la seconde poursuit l’organisation
séquentielle des textes. La maquette en tant que dispositif destiné à un usage
prédéfini intègre l’étude du non-séquentiel dans le séquentiel et efface le faux
contraste entre eux.
Références
Brunet E. (2008). Les séquences (suite). JADT 2008.
Brunet E. (2012). Nouveau traitement des cooccurrences dans Hyperbase.
Corpus (11).

490

JADT’ 18

Halliday M. A. (1994). Introduction to Functional Grammar. London : Edward
Arnold.
Lebart L. et Salem A. (1994). Statistique textuelle. Paris : Dunod.
Mayaffre D. (2008a). Quand ‘travail’, ‘famille’, ‘patrie’ co-occurrent dans le
discours de Nicolas Sarkozy. Etude de cas et réflexion théorique sur la
cooccurrence. JADT 2008.
Mayaffre D. (2008b). De l’occurrence à l’isotopie. Les co-occurrences en
lexicométrie. Sémantique & synatxe (9).
Mayaffre D. (2014). Plaidoyer en faveur de l’Analyse des Données
co(n)textuelles. Parcours coocurrentiels dans le discours présidentiel
français (1958-2014). JADT 2014.
Metwally H. (2017), Les thèmes et le temps dans Le Monde diplomatique (19902008). Thèse de doctorat, Université Côté d’Azur.
Rastier F. (2001). Arts et sciences du texte. PUF.
Rastier F. (2007). Passages. Corpus (6), pp. 25-54.
Rastier F. (2011). La mesure et le grain. Sémantique de corpus. Paris : Champion.
Ratinaud P. et Marchand P. (2012). Application de la méthode ALCESTE aux
« gros » corpus et stabilité des « mondes lexicaux » : analyse du «
CableGate » avec IRAMUTEQ. JADT 2012.
Reinert M. (1983). Une méthode de classification descendante hiérarchique :
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données. 8(2), pp. 187-198.
Reinert M. (1993). Les « mondes lexicaux » et leur « logique » à travers
l’analyse statistique d’un corpus de récits de cauchemars. Langage et société
(66), pp. 5-39.
Salem A. (1988). Approches du temps lexical. Statistique textuelle et séries
chronologiques. Mots (17). pp. 105-143.
Salem A. (1991). Les séries textuelles chronologiques. Histoire & Mesure, VI
(1/2). pp. 149-175.
Salem A. (1993). De travailleurs à salariés. Repères pour une évolution du
vocabulaire syndical (1970-1993). Mots(63). pp. 74-83.
Salem A. (1994). La lexicométrie chronologique. Dans Actes du colloque de
lexicologie politique ‘Langages de la Révolution’. Paris : Klincksieck.
Viprey J.-M. (2005). Corpus et sémantique discursive : éléments de méthode
pour la lecture des corpus. Dans A. Condamines, Sémantique et corpus.
Paris : Lavoisier.
Viprey J.-M. (2006). Structure non-séquentielle des textes. Langages (183).

JADT’ 18

491

Séries textuelles homogènes
Jun Miao 1, André Salem 2
Université Lumière de Lyon 2, France – miaojun@miaojun.net
2 Université de la Sorbonne nouvelle - Paris 3, France – salem@msh-paris.fr
1

Abstract
Textometric methods, widely used for the study of large corpora, are applied
here to a set of small texts, which, however, present homogeneous
characteristics. Our study focuses on a chronological textual series consisting
of reports of successive congresses of the CCP (Chinese Communist Party)
during the period 1982-2017. The textometrical methods are firstly used to
highlight the changes occurred during the 2017 congress.
Secondly, we apply these same methods to the subcorpora consisting of a
collection of fragments, automatically extracted from each congress and
related to the same topic. This subcorpora thereby constituted make it
possible to observe, with greater efficiency, the contextual variations that
occur over time around the same type. The method can be extended to any
corpora consisting of fragment systems that present a certain level of
homogeneity among them.
Keywords: Textual series, Chinese political speeches, homogeneous
subcorpora
Résumé
Nous appliquons ici des méthodes textométriques, largement utilisées pour
l'étude de vastes corpus, à des ensembles de textes dont la taille est réduite
mais qui présentent de fortes caractéristiques d'homogénéité. Notre étude
porte sur une série textuelle chronologique constituée par les rapports
successifs des congrès du PCC (Parti Communiste Chinois) durant les années
1982-2017.
Les méthodes de la veille textuelle textométrique sont d'abord mises en œuvre
pour mettre en évidence les changements survenus lors du congrès de 2017.
Dans un deuxième temps, nous appliquons ces mêmes méthodes à des souscorpus, constitués par la réunion de fragments extraits de chacun des congrès
et relatifs à un même thème. Les sous-corpus ainsi constitués permettent
d'observer avec une efficacité accrue des variations contextuelles qui
surviennent au fil du temps autour d'une même forme-pôle. La méthode
peut être appliquée à tout corpus constitué de systèmes de fragments
présentant une certaine homogénéité entre eux.

492

JADT’ 18

Mots-clés: Séries
homogène.

textuelles,

discours

politique

chinois,

sous-corpus

1. Introduction1
Le développement des capacités textométriques permet désormais d'explorer
avec profit des ensembles de textes extrêmement vastes et souvent variés.
Nous avons, cependant, insisté, avec d'autres, sur l'intérêt qu'il y a à
appliquer ces mêmes méthodes à des corpus constitués par la réunion de
productions textuelles présentant de fortes caractéristiques d'homogénéité et
forcément plus réduites de ce fait (Salem 1991). Au delà des séries
chronologiques, auxquelles nous empruntons nos exemples, la démarche que
nous présentons peut être appliquée à différents types de corpus.
Depuis quelques décennies, le Congrès national du Parti communiste chinois
(PCC) a lieu une fois tous les cinq ans. Il constitue la plus haute instance de
ce Parti, dans laquelle sont annoncées les décisions importantes2. Dans la
dernière décennie, les commentaires et les analyses quantitatives, portant sur
les textes de congrès du PCC, plus ou moins appuyés sur des méthodes
d'analyse statistiques, se sont multipliés dans la presse et sur différents sites
de l'Internet.
Le corpus que nous étudions est constitué d'un ensemble des textes produits
lors des congrès du PCC, entre 1982 et 2017. Pour des raisons que nous
analysons, les textes produits durant cette dernière période présentent une
grande homogénéité, tant du point de vue de leur taille que de celui des
thèmes qu'ils abordent et du style qu'ils emploient. Nous commençons par
étudier de manière classique la série chronologique PCC1982-2017 divisée en
congrès afin de mettre en évidence des variations dans l'emploi du
vocabulaire. Nous proposerons ensuite une méthode qui permet, selon nous,
d'étudier au plus près les variations du contexte immédiat d'un terme donné.
2. Analyse chronologique de la série PCC1982-2017
Le corpus ainsi constitué compte au total 115 1338 occurrences pour 7365

Les analyses dont nous rendons compte ci-dessous, ont été effectuées à l'aide
du logiciel Lexico5. Cedric Lamalle, William Martinez, Serge Fleury ont largement
contribué au développement des fonctionnalités de ce logiciel. Les auteurs tiennent à
les en remercier.
2 L’article de Salem et Wu (2008) constitue une étude chronologique portant sur
l'intégralité des congrès du PCC survenus depuis sa fondation 1921 jusqu'à l'année
2012. Au-delà des évolutions chronologiques qu'elle avait permis de mettre à jour,
cette étude montre le caractère hétérogène de la forme congrès considérée sur une
échelle aussi large.
1

JADT’ 18

493

formes différentes3. La division en congrès amène une partition du corpus en
huit parties. Les longueurs des parties, pour chaque congrès, s’échelonnent
entre 2 400 et 2 900 occurrences. La forme de fréquence maximale est toujours
la forme的 (de, DE1), dont on peut vérifier la forte diminution au fil des
congrès4.
2.1 Le congrès 2017
Lorsque survient un nouveau congrès qui complète une série chronologique
pré-existante, la méthode des spécificités permet de répondre à la question :
Quelles sont les principales évolutions lexicales survenues lors du dernier congrès de
la série ? C'est une opération de veille lexicale. Le calcul des spécificités
appliqué au congrès de 2017 signale des spécificités positives, dont le
contenu revêt un caractère nettement lexical : 时代 (shídài, ère, S +24), 治理
(zhìlǐ, gérer, S +21), 生态 (shēngtài, écologie, S +15), 梦 (mèng, rêve, S +14)5. A
l'inverse, les formes de spécificités négatives, pour cette même période, sont
plutôt des formes grammaticales, telles que 的 (de, DE1, S -38) , 这 (zhe, ce, S 22), 地 (de, DE2, S -14).
Le même calcul appliqué aux segments répétés du corpus permet de préciser
les modifications survenues lors ce même congrès. La mise en vedette du
terme 新 时代 (xīn shídài, nouvelle ère), employé 36 fois lors du congrès de
2017, a été largement commentée par les analystes qui se sont penchés sur ce
texte6. Le recensement systématique des segments fortement spécifiques pour
cette même période permet de mettre en évidence des séquences répétées
dont certaines ont pu échapper aux commentateurs et qui constituent
également des néologismes par rapport aux congrès précédents : 新 时代

La séquence textuelle continue des textes chinois, composés de caractères
juxtaposés (scriptio continua, dans laquelle les mots ne sont pas séparés par des
espaces), a été soumise à un segmenteur automatique NLPIR (Zhang, 2016), très
largement utilisé dans le monde sinophone, afin d'être segmentée en mots graphiques.
4 Nous expliquons dans une étude parallèle comment cette diminution
progressive peut être mise en rapport avec l'évolution du style d’écriture.
5 Dans nos exemples, la forme native chinoise est suivie de sa transcription en
pinyin, puis d'un équivalent français (lequel ne peut prétendre au statut de traduction
satisfaisante pour chacune des occurrences du terme). Un coefficient de spécificité,
positive ou négative, de forme S +/- xx indique enfin le degré de spécificité de la forme
dans la partie du texte considérée.
6 De nombreux articles publiés à cette occasion ont explicitement mentionné la
3

fréquence (36 occurrences) de la formule新 时代 (xīn shídài, nouvelle ère) ex :
Vandnepitte (2017). D'autres sites ont proposé aux internautes de classer congrès par
fréquence d'apparition de plusieurs termes répétés dans chaque congrès (Qian, 2017).

494

JADT’ 18

中国 特色 社会主义, (le socialisme à la chinoise dans la nouvelle ère, 13 occ., S +12)
, 治理 体系. (le système de gouvernance, 13 occ., S +12). Plus remarquable à nos
yeux, certaines expressions extrêmement courantes dans les périodes
précédentes ont complètement disparu du texte du dernier congrès. Tel est le
cas, par exemple pour des segments comme :
有 中国 特色,
(posséder des caractéristiques chinoises, 0
occ., S -7),
有 中国 特色 社会主义
(avoir un socialisme à la chinoise, 0 occ., S 5).
L'analyse des spécificités permet également de localiser des parties du texte
dans lesquelles le renouvellement lexical se révèle particulièrement
important. Sur la figure 1, une carte des sections a été établie pour chacun des
congrès, divisé en chapitres. Les sections apparaissent d'autant plus sombres
qu'elles renferment de nombreuses occurrences de termes spécifiques pour le
dernier congrès. La représentation permet de vérifier que le renouvellement
ne se fait pas de manière uniforme, dans le dernier congrès. Une partie du
vocabulaire spécifique du congrès de 2017, était déjà largement présente dans
les deux congrès précédents. La carte permet en outre de localiser
précisément les chapitres du dernier congrès qui font le plus fortement l'objet
d'un renouvellement lexical.
La figure 2 ci-dessous permet d'apprécier l'évolution du vocabulaire
survenue dans la dernière période en combinant une représentation
factorielle sur l'ensemble des congrès et les spécificités calculées pour le
dernier congrès. Une analyse réalisée sur les huit congrès met en évidence la
progressivité des changements lexicaux. On a projeté en qualité d'éléments
supplémentaires les formes spécifiques positives de la dernière partie. Ce type
de représentation peut être articulé avec les cartes de section, présentées cidessus pour illustrer les changements lexicaux.
3. Utiliser la structure des documents
Dans chacun des textes de l'édition originale des congrès, des repères
éditoriaux (intertitres, numérotation de sous-parties, etc.) permettent
d'effectuer un découpage en unités plus petites que nous appellerons
chapitres. Chaque chapitre correspond à l'évocation d'un thème particulier
(développement économique, perspectives internationales, état des forces
armées, etc.). Lors de chacun des congrès, ces thèmes sont abordés tour à
tour, souvent dans un ordre similaire qui peut conduire à proposer une
description globale de l'ordonnancement de ces textes de congrès.

JADT’ 18

495

Guide de lecture:
A gauche, on trouve
une carte des sections
réalisée à partir d'un
découpage
en
chapitres.
Chaque
ligne regroupe les
chapitres relatifs à
un même congrès.
Les carrés les plus
foncés
correspondent aux
chapitres les plus
chargés en formes
spécifiques dans le
dernier
congrès
(S+>10).
En bas : le texte du
deuxième chapitre
du dernier congrès,
qui
figure
au
dessous de la carte
est signalé comme
particulièrement
chargé
en
formes
spécifiques.

$<chap1=19-00># <para=19-05>同志 们 ： ¶ # <para=19-06>现在 , 我 代表
第十八 届 中央 委员会 向 大会 作 报告 . ¶ # <para=19-07>中国共产党 第十九
次 全国 代表大会 , 是 在 全面 建成 小康 社会 决胜 阶段 * 中国 特色 社会主义

进入 新 时代 的 关键 时期 召开 的 一 次 十分 重要 的 大会 . ¶ # <para=1908>大会 的 主题 是 ： 不 忘 初心 , 牢记 使命 , 高举 中国 特色 社会主义 伟大
旗帜 , 决胜 全面 建成 小康 社会 , 夺取 新 时代 中国 特色 社会主义 伟大 胜利 ,

为 实现 中华民族 伟大 复兴 的 中国 梦 不懈 奋斗 /... /. ¶.
Figure 1 : Repérage des portions caractéristiques pour le dernier congrès (2017)

496

JADT’ 18

Figure 2 : Spécificités positives du congrès 2017 mises en évidence dans l’AFC

Guide de lecture: Sur la figure 2, les différents congrès s'échelonnent dans le
temps selon une parabole. Cet échelonnement résulte d'un renouvellement
important du vocabulaire au fil des congrès. Les formes les plus spécifiques
pour le dernier congrès ont été projetées en qualité d'éléments
supplémentaires.

JADT’ 18

497

3.1 Analyse en chapitres
Lorsqu'on soumet à des analyses typologiques, le même corpus divisé, cette
fois, en chapitres, on constate que les chapitres correspondant aux mêmes
thèmes, mais appartenant à différents congrès, ont une forte tendance à se
regrouper, du fait qu'ils emploient des vocabulaires proches. La structure
chronologique, mise en évidence par l'analyse en congrès s'efface, dans ce
cas, devant une typologie d'ordre thématique. La figure 3 montre les résultats
d'une Analyse factorielle des correspondances effectuée à partir du corpus
PCC1982-2017 divisé cette fois en 89 chapitres. Sur cette figure, les
identificateurs des chapitres sont constituées de deux parties. Le premier
nombre indique le numéro du congrès dont le chapitre est extrait. Le second,
l'ordre du chapitre à l'intérieur du congrès. Comme on peut le vérifier sur
cette figure les chapitres correspondant à un même thème ont tendance à se
regrouper fortement.

498

JADT’ 18

Figure 3 : Analyse factorielle des correspondances sur le corpus divisé en chapitres

A titre d'exemple, nous avons agrandi les portions du graphique qui
correspondent à deux groupes thématiques :
a) le groupe un pays deux systèmes qui correspond à une orientation
politique constante du PCC, réaffirmée à chaque congrès ;
b) un groupe de chapitres correspondant à l'analyse des relations
internationales, qui constitue également un moment incontournable
pour chaque congrès, à partir du 14ème.
3.2 Le sous-corpus thématique « un pays deux systèmes »
L'étape suivante consiste à réitérer ces mêmes analyses à partir de souscorpus réduits, rassemblant les seules chapitres relatifs à une même
thématique. Les analyses textométriques effectuées sur ces sous-corpus
homogènes débouchent sur des résultats particulièrement lisibles. Lors de
l'analyse de ce type de corpus, la dimension chronologique revient au
premier plan. Le sous-corpus qui rassemble les passages relatifs au thème un
pays, deux systèmes ne compte que deux milles occurrences, sur l'ensemble des
congrès. L'analyse des formes qui apparaissent spécifiquement dans les
contextes de ce terme, montre cependant une nette évolution du contexte
immédiat de ce terme. Le Congrès de 1987, présente la formule comme un
principe à mettre en œuvre. Dans les congrès suivants, on voit apparaître les
verbes maintenir et continuer (2002) puis mettre en œuvre sans faille (2007). En
2017, il s'agit d'appliquer intégralement et avec précision le principe un pays, deux
systèmes. La figure 4 montre une projection des différents segments qui
contiennent l'expression sur l'analyse réalisée à partir du sous-corpus7.
7 Le graphique a été légèrement modifié pour permettre une plus grande
lisibilité. Les segments redondants ont été écartés; les points superposés ont été
légèrement déplacés.

JADT’ 18

499

Figure 4 : Variations lexicales autour de l'expression : un pays deux systèmes

4. Conclusion
Nos expériences nous amènent à conclure que l'analyse textométrique opérée
à partir de regroupements de fragments homogènes, prélevés autour d'un
même thème durant les années couvertes par une série chronologique conduit
à des résultats dont l'interprétation se révèle particulièrement aisée. La
grande homogénéité lexicale des fragments rapprochés permet alors
d'observer des variations très fines. Elle compense largement la taille réduite
du corpus, peu favorable, a priori, dans le cas d'études textométriques.
Au delà des applications aux seules séries textuelles chronologiques, la méthode
pourra être utilisée pour toute sorte de corpus, dans une large variété de
langues, à la condition qu'il soit possible de distinguer des sous ensembles
thématiques homogènes
Références
Miao J. (2012). Approches textométriques de la notion de style du traducteur Analyses d'un corpus parallèle français-chinois : Jean-Christophe de Romain
Rolland et ses trois traductions chinoises. Thèse doctorale dirigée sous la
direction de M. André Salem, Paris 3.
Qian G. (2017). 中共历届党代大会报告语象分析 (Analyses lexicales des
rapports de tous les congrès du Parti communiste chinois).Lianhe Zaobao
du 19 novembre 2017.
Salem A. (1991). Les séries textuelles chronologiques. Histoire & Mesure
Année, Vol. (6) : 149-175.

500

JADT’ 18

Salem A., Wu Li-Chi. (2008). Essai de textométrie politique chinoise. In
André Salem et Serge Fleury, éditeurs, Lexicometrica – Explorations
textométriques,
Vol.
(1).
URL :
http://lexicometrica.univparis3.fr/numspeciaux/special8.htm (consulté le 5 février 2017).
Vandepitte M. (2017). Quatre choses à savoir sur la Chine – dans le cadre du
XIXème congrès du Parti. Traduit par Anne Meert en français du
néerlandais. Investig’Action du 15 novembre 2017. URL : goo.gl/8fgSkq
(consulté le 25 novembre 2017).
Logiciels utilisés :
Zhang H.P. (2017). Segmenteur automatique chinois NLPIR. URL :
http://www.nlpir.org/
Salem A. (2017). L’outil d’analyse textométrique Lexico 5. URL :
http://www.lexi-co.com/index.html

JADT’ 18

501

TaLTaC in ENEAGRID Infrastructure
Silvio Migliori1, Andrea Quintiliani1, Daniela Alderuccio1,
Fiorenzo Ambrosino1, Antonio Colavincenzo1, Marialuisa Mongelli1,
Samuele Pierattini1, Giovanni Ponti1 Sergio Bolasco2,
Francesco Baiocchi3, Giovanni De Gasperis4,
1

ENEA DTE-ICT – silvio.migliori@enea.it, 2 Sapienza Università di Roma,
3 Staff TaLTac - info@taltac.it, 4 Dip. DISIM Università dell‘Aquila

Abstract
The aim of this joint ENEA-TaLTaC project is to enable the TaLTaC User
Community and the Digital Humanists to have remote access to the TaLTaC
software through the ENEAGRID Infrastructure. ENEA's research activities
on the integration of Language Technologies (Multilingual Text Mining
Software and Lexical Resources) in the ENEA distributed digital
infrastructure provide a "community Cloud" approach in a digital
collaborative environment and on an integrated platform of tools and digital
resources, for the sharing of knowledge and analysis of textual corpora in
Economic and Social Sciences and e-Humanities. Access to the TaLTac
software in Windows and Linux version will exploit the high computational
capacity (800 Teraflops) of the e-infrastructure, to which users access as a
single virtual supercomputer.
Riassunto
Obiettivo del progetto congiunto ENEA-TaLTaC è consentire alla comunità
degli utenti TaLTaC e ai ricercatori nelle Digital Humanities l’accesso remoto
al software TaLTaC attraverso l'infrastruttura digitale ENEAGRID. Le
attività di ricerca dell'ENEA sull'integrazione delle tecnologie linguistiche
(software di Text Mining per testi multilingue e risorse lessicali) in
ENEAGRID forniscono un approccio "community Cloud" in un ambiente
collaborativo digitale e su una piattaforma integrata di strumenti e risorse
digitali, per la condivisione delle conoscenze e l'analisi di corpora testuali in
Scienze Economiche e Sociali ed e-Humanities. L’accesso al software TaLTac
in versione Windows e Linux sfrutterà l’elevata capacità computazionale (800
Teraflops) dell’infrastruttura di calcolo, a cui gli utenti accedono come ad un
unico supercomputer virtuale.
Keywords: Text Mining Software, Cloud Computing, Digital-Humanities,
Socio-Economic Sciences, Big Data.

502

JADT’ 18

1. Introduction
“TaLTaC in CLOUD” is a joint ENEA-TaLTaC project for the set-up of an ICT
portal on the ENEA distributed e-Infrastructure1 (Ponti et al., 2014), hosting
TaLTaC Software (Bolasco et al., 2016, 2017). Users will access TaLTaC
software (Windows and Linux versions) in a remote and ubiquitous way,
and the computational power (800 Teraflops) of ICT ENEA distributed
resources, as a single supercomputer. The aim of this joint ENEA-TaLTaC
project is to enable the TaLTaC User Community and Digital Humanists to
have remote access to TaLTaC software through ENEAGRID Infrastructure,
integrating ICT inside Digital Cultural Research.
ENEAGRID offers a digital collaborative environment and an integrated
platform of tools and resources assisting research collaborations, for sharing
knowledge and digital resources and for storing textual data. In this virtual
environment, TaLTaC software evolves from a stand-alone uniprocessor
software toward a multiprocessor design, integrated in an ICT research einfrastructure. Furthermore, it evolves towards implementing ancient
language lexical and semantic knowledge and e-resources, facing research
needs and implementing solutions also for Digital Humanities communities.
2. TaLTaC Software
The TaLTaC software package, conceived at the beginning of the 2000s, has
been progressively developed to date in three major releases: T1 (2001), T2
(2005) and T3 (2016); it is widespread among the text analysis community in
Italy and abroad with over 1000 licenses, including two hundred entities
between university departments, research institutions and other
organizations.
The 2018 release of the software, T3, implemented the following priority
objectives: i) the processing of big data (around of a billion words), achieving
the independence from the dimensions of the text corpora, limited only by
hardware resources; ii) the automatic extraction on multiple layers of results
from text parsing (tokenization): layer zero (text in the original version), layer
1 (recognition of words with automatic corrections of the accents), layer 2
(pre-recognition of most common Named Entities), layer 3 (reconstruction of
pre-defined multiwords); iii) computing speed, taking advantage of the
power of the multi-core processing readily available on current computers

The ENEAGRID infrastructure is based on several software components which
interact with each other to offer an integrated distributed system. The ENEAGRID
infrastructure allows access to all these resources as a single virtual system, with an
integrated computational availability of about 16000 cores, provided by several
multiplatform systems.
1

JADT’ 18

503

(personal or cloud).
Table 1 shows the processing times of three parsing, up to layer 2, for larger
corpora on PC (1-core and 8-cores) and on ENEAGRID. Preliminary results
on ENEAGRID (1core-CRESCO) show that with increasing corpus size there
is an even greater saving of time.
TALTAC was installed in ENEAGRID infrastructure, but the computational
capabilities of the HPC system are not yet exploited because the current
version of the software does not support multi-core. Therefore, the present
ENEAGRID capabilities allow only multi-users access and computation;
future versions of the software will be tested for multi-core capabilities to
exploit the real power of ENEA ICT High Performance Computing.
Table 1. Preliminary results of processing times of three parsing on PC and on ENEAGRID.
ENEAGRID
1 core
(CRESCO)
in minutes

millions
74

0,41

3,4

1,1

0,33

3,5

284

1,55

13,0

3,8

0,29

13,2

tokens
1 "La Repubblica " (100 th Artic.)
2 "La Repubblica " (400 th Artic.)
3 Italian and French Press
4 Various Press Collection

MAC i7 (7th generation)
8core
1 core 8 cores
/1core
in minutes
in %

size
of file
GB

535

2,89

37,4

8,8

0,24

41,3

1.138

6,18

88,2

14,0

0,16

54,7

For the characteristics of the technological architecture of the TaLTaC3
platform, see previous works (Bolasco et al. 2016, 2017), that can be
summarized here as: a1) HTML 5 for the GUI and jQuery with its derived
Javascript frameworks to encapsulate the GUI user interaction functions for
the MAC and Cloud solution; a2) Windows native DotNET desktop
application; b) JSON (JavaScript Object Notation): as an inter-module language
standard, with a structured and agile format for data exchange in
client/server applications; c) Python / PyPy: advanced script/compiled
programming language, mostly used for textual data analysis and natural
language processing at the CORE back end; d) No-SQL: high performance
key/value
data
structure
storage
server
Redis
adopted
for
vocabularies/linguistic resources persistence; e) RESTful: interface standard
for data exchange over the HTTP web protocol; f) MULTI-PROCESSING:
exploiting in the best possible way multi-core hardware, distributing
processing power among different CPU cores.
The choice of the Python language allowed to develop a cross-platform
computational core running on Windows, Linux, macOS. In particular, the
overall system of software processes runs smoothly over a linux-based cloud
computing facility, like the ENEAGRID. Furthermore, the Python code

504

JADT’ 18

compiled through the 64bit PyPy just-in-time-compiler allows very efficient
macro operations over a large set of data, stored as hash dictionaries, so that
the upper limits of performance and capacity is only due to the physical limit
of the host machine, in terms of RAM and number of cores and OS kernel
scheduler. In our test each node in the ENEAGRID infrastructure hosted a
single Redis instance and a number of 24 logic cores, with 16GB of RAM.
3. ENEAGRID Infrastructure
ENEA activities are supported by its ICT infrastructure, providing advanced
services as High Performance Computing (HPC), Cloud and Big Data
services, communication and collaboration tools. Advanced ICT services are
based on ENEA research and development activities in the domains of HPC,
of high performance networking and data management, including the
integration of large experimental facilities, with a special attention to public
services and industrial applications. As far as High Performance Computing
is concerned, ENEA manages and develops ENEAGRID, a computing
infrastructure distributed over 6 ENEA research centers for a total of about
16000 cores and a peak computing power of 800 Tflops.
HPC clusters are mostly based on conventional Intel Xeon cpu with the
addition of some accelerated systems as Intel Xeon/PHI and Nvidia GPU.
Storage resources includes RAID systems for a total of 1.8 PB, in
SAN/Switched and SRP/Infiniband configuration. Data are made available by
distributed and high performances files systems (AFS and GPFS).
ENEA Portici Center has become one of the most important italian HPC
center in 2008 with the project CRESCO - Computational RESearch Center for
COmplex Systems. CRESCO HPC clusters are used in many of the main
ENEA research and developments activities, such as energy, atmosphere and
sea modeling, bioinformatics, material science, critical infrastructures
analysis, fission and fusion nuclear science and technology, complex systems
simulation. CRESCO clusters have provided in 2015 and 2016 more than 40
million core hours each year to ENEA researchers and technologists and to
their external partners (external users account for about 30% of the total
machine time).
CRESCO6, the new HPC cluster recently installed in Portici in the framework
of the 2015 ENEA-CINECA agreement, provides a peak computing power of
700 Tflops and is based on the new 24 cores Intel SkyLake cpu. Its nodes will
be connected by the new Intel OmniPath high performance network,
providing a 100 Gbps bandwidth.
ENEA ICT department provides also general purpose communication,
elaboration and collaboration tools and services as Network management, EMail, Video Conferencing and Voip services, Cloud Computing and Storage.

JADT’ 18

505

A friendly user access to scientific and technical applications (as Ansys,
Comsol, Nastran, Fluent) is provided by dedicated web portals (Virtual
laboratories) relying on optimized remote data access tools as NX
technology.
4. TaLTaC in ENEAGRID Infrastructure
4.1 Software Installation and Access on ENEA e-Infrastructure
The software TaLTaC is available on Windows and Linux through
ENEAGRID via AFS in a geographically distributed file system, which
allows remote access to each computing node of the HPC CRESCO systems
and Cloud infrastructure from anywhere in the world.
This provides three capabilities: i) data mining, sharing and storage; ii) ICT
services necessary for the efficient use of HPC resources, collaborative work,
visualization and data analysis; iii) the implementation of software and its
settings for future data processing and analysis. Moreover, the availability of
the software on the ENEA ICT infrastructure can benefit of the advantages of
AFS such as scalability, redundance, backup and so on.
Through the ACL rules it can be possible to manage the accessibility of the
software to the community of users in compliance of the license policies that
will be put in place. The following two options are provided for TaLTaC
running: the first one is to use the applications installed in the windows
system and the second one is to use FARO2 – Fast Access to Remote Objects
(the general purpose interface for hardware and software capabilities by web
access) to directly access the applications installed in the Linux environment
and that refer to the data in AFS.
4.1.1. TaLTaC2 (Windows) on Remote Desktop Access
The software TaLTaC2 is available on “Windows Server 2012 R2” by remote
desktop access to a virtual machine that can be reached by the ThinLinc
general-purpose and intuitive interface. All the users involved in the project
activities can access the server but only the person in charge of developing
and installing the application can obtain administrator privileges. For this
reason, AFS authentication is always required. Every TaLTaC2 user with AFS
credentials can access ENEAGRID to run the software and to manage data on
AFS own areas via web and from any remote location. In the AFS
environment, an assigned disk area with a large memory capacity is defined.
This area is mainly used for storage and sharing of large amounts of data
(less than 200 MB) (analysis, reports and documents) that come from running
the software on a single processor, in serial mode, or for future parallel data
mining applications.

506

JADT’ 18

4.1.2. TaLTaC3 (Linux) on CRESCO System
On the CRESCO systems, that is accessible from ENEAGRID infrastructure,
TaLTaC3 is available on CentOS Linux nodes and then it is possible to
leverage the overall computing power dedicated to the activities of TaLTaC
and Digital Humanists communities. Every user can start own work session
allocating a node with a reserved Redis instance and as many computing core
as needed.
Performance improvements are obtainable through the parallelization so that
a single user can use the full capacity of the assigned node, in terms of
number of computing cores. The TaLTaC3 package is automatically started
as the user login to the node by a shell script. The opensource Mozilla Firefox
web browser makes the user interface in the current beta version. The access
to the TaLTaC3 portal use the ThinLinc remote desktop visualization
technology that allows an almost transparent remote session on the HPC
system, including the graphical user interface, thanks to the built-in features
such as load-balancing, accelerated graphics and platform-specific
optimisations. Input and output data can be accessed through the
ENEAGRID filesystems and therefore easily uploaded and downloaded.
4.2 Case Studies
ENEA distributed infrastructure (and cloud services) enables the
management of research process in Economic-Social Sciences and Digital
Humanities, providing technology solutions and tools to academic
departments and research institutes: building and analyzing collections to
generate new intellectual products or cultural patterns, data or research
processes, building teaching resources, enabling collaborative working and
interdisciplinary knowledge transfer.
4.2.1. TaLTaC User Community
The current (2018) community of TaLTaC over the years aggregated users
from the computer laboratories of automatic analysis of texts and text
mining, also carried out within the institutional courses of bachelor and
magistral degrees, plus Ph.D. students from doctoral degree courses at the
universities of Rome "La Sapienza" and "Tor Vergata", of Padua, Modena,
Pisa, Naples and Calabria (a total estimate of over 1300 students over the last
eight years); furthermore, there is another set of users that subscribed to
specific tutorial courses dedicated to TaLTaC (more than 60 courses for a
total number of 750 tutorial participants).
A call about the opportunity of using "remotely" the software via the ENEA
distributed computing facilities, received the manifestation of interest by 40
departments and other research institutes.

JADT’ 18

507

4.2.2. Digital Humanities Community as TaLTaC user
In collaboration with academic experts, ENEA focused on Digital Humanities
projects in Text Mining & Analysis in Ancient Writings Systems of the Near
East and used TaLTaC2 to perform quantitative linguistic analysis in
cuneiform corpora (transliterated into latin alphabet) (Ponti et al., 2017).
Cuneiform was used by a number of cultures in the ancient Near East to
write 15 languages over 3,000 years. The cuneiform corpus was estimated to
be larger than the corpus of Latin texts but only about 1/10 of the extant
cuneiform texts have been read even once in modern times. This huge
cuneiform corpus and the restricted number of experts lead to the use of Text
Mining and Analysis, clustering algorithms, social network analysis in the
TIGRIS Virtual Lab for Digital Assiriology2, a virtual research environment
implemented in ENEA research e-infrastructure. In TIGRIS V-Lab
researchers perform basic tasks to extract knowledge from cuneiform
corpora. (i.e. dictionaries extraction with word list of toponyms, chrononyms,
theonyms, personal names, grammatical and semantic tagging,
concordances, corpora annotation, lexicon building, grammar writing, etc.).
5. Conclusions
Researchers and their collaborators will use computational resources in
ENEAGRID to perform their work regardless of the location of the specific
machine or of the employed hardware/ software platform.
ENEAGRID offers computation and storage resources and services in a
ubiquitous and remote way. It integrates a cloud computing environment
and exports: a) remote software (i.e. TaLTaC); b) Virtual Labs: thematic areas
accessible via web, where researchers can find set of software (and
documentation regarding specific research areas); c) remote storage facilities
(with OpenAFS file system). In this virtual environment, TaLTaC software
evolves from a uniprocessor software toward a multiprocessor design,
integrated in an ICT research e-infrastructure.
This project leads to the TaLTaC evolution from a stand-alone software
(allowing Text Mining & Analysis to search for linguistic constructions in
textual corpora, showing results in a table or concordance list) to a software
“always and anywhere on”, that also can be accessed, providing an interface
where you can visualize results, create interpretative models, collaborate
with others, combine different textual representations and storing data, codeveloping research practices. Furthermore, this project reflects the shift

2 TIGRIS - Toward Integration of e-tools in GRId Infrastructure for e-aSsyriology
http://www.afs.enea.it/project/tigris/indexOpen.php
http://www.laboratorivirtuali.enea.it/it/prime-pagine/ctigris

508

JADT’ 18

from the individual-researcher-approach to a collaborative research
community-approach, leading to a community-driven software design,
tailor-made on specific research community needs and to Community Cloud
Computing.
This
interdisciplinary
knowledge
transfer
enables
creating/activating new knowledge from big (cultural and socio-economic)
data, both in modern and ancient languages.
References
Bolasco, S., Baiocchi, F., Canzonetti, A., De Gasperis, G. (2016). “TaLTaC3.0, un
software multi-lessicale e uni-testuale ad architettura web”, in D. Mayaffre, C.
Poudat, L. Vanni, V. Magri, P. Follette (eds.), Proceedings of JADT 2016, CNRS
University Nice Sophia Antipolis, Volume I, pp. 225-235.
Bolasco S., De Gasperis G. (2017). “TaLTaC 3.0 A Web Multilevel Platform for
Textual Big Data in the Social Sciences” in C. Lauro, E. Amaturo, M.G.
Grassia, B. Aragona, M. Marino. (eds.) Data Science and Social Research Epistemology, Methods, Technology and Applications (series: Studies in
Classification, Data Analysis, and Knowledge Organization) Springer Publ.,
pp. 97-103.
Ponti G., Palombi F., Abate D., Ambrosino F., Aprea G., Bastianelli T., Beone F.,
Bertini R., Bracco G., Caporicci M., Calosso B., Chinnici M., Colavincenzo A.,
Cucurullo A., Dangelo P., De Rosa M., De Michele P., Funel A., Furini G.,
Giammattei D., Giusepponi S., Guadagni R., Guarnieri G., Italiano A.,
Magagnino S., Mariano A., Mencuccini G., Mercuri C., Migliori S., Ornelli P.,
Pecoraro S., Perozziello A., Pierattini S., Podda S., Poggi F., Quintiliani A.,
Rocchi A., Sciò C., Simoni F., Vita A. (2014) “The Role of Medium Size
Facilities in the HPC Ecosystem: The Case of the New CRESCO4 Cluster
Integrated in the ENEAGRID Infrastructure”. In: Proceedings of the
International Conference on High Performance Computing and Simulation, HPCS
(2014), ISBN: 978-1-4799-5160-4.
Ponti G., Alderuccio, D., Mencuccini, G., Rocchi, A., Migliori, S., Bracco, G., Negri
Scafa, P. (2017) “Data Mining Tools and GRID Infrastructure for Text
Analysis” in “Private and State in the Ancient Near East” Proceedings of the
58th Rencontre Assyriologique Internationale, Leiden 16-20 July 2012, edited by
R. De Boer and J.G. Dercksen, Eisensbrauns Inc. - LCCN 2017032823 (print) |
LCCN 2017034599 (ebook) | ISBN 9781575067858 (ePDF) | ISBN
9781575067841.
ENEAGRID http://www.ict.enea.it/it/hpc Laboratori Virtuali http://www.ict.enea.it/it/laboratori-virtualixxx/virtual-labs
TIGRIS Virtual Lab http://www.afs.enea.it/project/tigris/indexOpen.php
TaLTaC: www.taltac.it

JADT’ 18

509

The dimensions of Gender in the International
Review of Sociology. A lexicometric approach to the
analysis of the publications in the last twenty years
Isabella Mingo, Mariella Nocenzi
Sapienza University of Rome – isabella.mingo@uniroma1.it; mariella.nocenzi@uniroma1.it

Abstract 1 (in English)
The Social Sciences and, specifically, the sociological research has
progressively assumed the gender factor as one of the strategic keys to
understand contemporary phenomena. In fact, as a variable for sociostatistical analysis or as a characterizing trait of individual identity, it is a
decisive factor in the interpretation of the deep social transformations and it
inspires the self-reflection of the sociologists about the analytical tools of their
discipline. The contribution proposes, through a lexicometric approach, an
analysis of the articles published in the last two decades by the oldest Journal
of Sociology, published by Routledge. The main aim is to highlight the
different ways in which gender issues are declined in the international
sociological researches presented in the repertoire of the International
Review of Sociology and to outline, both on the lexical level and on the topic
level, the changes occurred over time.
Abstract 2 (in French, Italian or Spanish)
Le scienze sociali e, nello specifico, la ricerca sociologica hanno
progressivamente assunto il fattore del genere come una delle più strategiche
chiavi di lettura dei fenomeni contemporanei. Si tratta, infatti, di un fattore
che, quale variabile per l’analisi socio-statistica o come tratto caratterizzante
dell’identità individuale, si rivela dirimente nell’interpretazione delle
profonde trasformazioni sociali in atto e spunto per un’autoriflessione degli
stessi sociologi sugli strumenti di analisi della loro disciplina. Il contributo
propone, mediante un approccio lessico-metrico, un’analisi degli articoli
pubblicati nelle ultime due decadi dalla più antica rivista di sociologia, edita
da Routledge, con l’obiettivo di evidenziare i diversi modi con cui il concetto
di genere viene declinato nelle ricerche sociologiche internazionali presentate
nel repertorio dell’International Review of Sociology e di delineare, sia sul
piano lessicale che su quello delle tematiche, i cambiamenti intervenuti nel
corso del tempo.
Keywords: Gender, International Review of Sociology, Lexicometric
Analysis, Textual Analysis, Social Change, Sociological Analysis

510

JADT’ 18

1. Introduction and the hypothesis of the paper
From 1955, when in a relevant paper the American scholar John Money (et
al., 1955) coined the term of gender for the definition of “those things that a
person says or does to disclose himself or herself as having the status of boy
or man, girl or woman”, the social sciences have developed entire subfields
and a wide range of topics to analyse it with a variety of research methods.
Sociologists, in particular, had outlined specific theoretical approaches and
had led many detailed studies to understand firstly what gender is and the
difference with sex. They had shared that if the meaning of sex is the biological
classification based on body parts, gender, on the other hand, is the social
classification based on one’s identity, presentation of self, behavior, and
interaction with others. Sociologists, hence, view gender as a learned behavior
and a culturally produced identity, and, for these reasons, they define it as a
“social” category. It has always been a very relevant category for the critical
analysis of the social construction because one of the most important social
structures is the status and one of the most strategic statuses is just gender.
In the last decades, the sociological theories and researches based on gender
are become more and more widespread, articulated, integrated with other
subfields of sociology and of the other social sciences. One of the most
representative indicator of this research development and specialization is
not only the common recognition and, then, institution of the sociology of the
gender as a subfield of the sociology, but the most frequent use of gender as
reference concept for all the other sociological theoretical approaches to the
analysis of the social system. The same sociology of gender has studied many
topics, with multiple research methods, including identity, social interaction,
power and oppression, and the interaction with race, class, culture, religion,
and sexuality, among others.
This paper aims to observe and, if possible, to interpret this progressive
diffusion and specialization in the use of gender as a theoretical and research
category through the publications of the International Review of Sociology, a
sociological journal, edited by Routledge with a worldwide online and paper
diffusion, during the last two decades. This journal, the oldest review in the
field of sociology in Europe, founded by René Worms in 1893 in Paris, still
maintains – as the “Aims and scope of the Review” state – «the traditional
orientation of the journal as well as of the world’s first international academic
organization of sociology which started as an association of contributors to
International Review of Sociology: it assumes that sociology is not conceived
apart from economics, history, demography, anthropology and social
psychology. Rather, sociology is a science which aims to discover the links
between the various areas of social activity and not just a set of empty
formulas. Thus, International Review of Sociology provides a medium through

JADT’ 18

511

which up-to-date results of interdisciplinary research can be spread across
disciplines as well as across continents and cultures»1.
The Authors proposes to highlight the different ways in which gender issues
are declined in the international sociological researches, through an analysis
of the articles published in the last two decades (1997-2017) in International
Review of Sociology. We consider the last two decades of publication not only
because of the best accessibility to the International Review of Sociology
catalogue. For the sociology, indeed, the recent gender studies and researches
have registered a deeper specialization in terms of connection with other
disciplines, unusual application of the gender approach to some social
phenomena, exploration of new research frontiers (multiple gender
identities, gender sensitive data arrangement, the non-alignment statuses of
sex and gender et similia).
2. Data and Methods
The analysis of the International Review of Sociology papers was carried out
mainly through a lexicometric approach, integrated with hermeneutic
analysis useful both in the first and in the last phase of the study. The first
phase has regarded the collection of the corpus, while the last one has
concerned the interpretation of the results obtained from quantitative and
automatic procedures. The lexicometric analyses, supported by the software
IRaMuTeQ2, were carried out to extract the most relevant forms/lemma and
to apply some exploratory techniques for identifying the main lexical-textual
dimensions, the relationships between some keywords, the recurring topics,
and possible differences over the time analysed.
2.1. The Corpus: Selection Criteria and Preliminary Analysis
The texts analyzed in this study have been collected from the archive of the
International Review of Sociology, considering the papers published from 1997
to 2017.
In the first stage, they have been extracted all the papers which propose the
term gender in title, abstract, body text and/or key words).They were 235,
distributed over the past 20 years, as shown in Table 1.
Then, they have been selected only those papers which present a relevant

See at the International Review of Sociology web site, page “Aims and scope”,
https://www.tandfonline.com/action/journalInformation?show=aimsScope&journalC
ode=cirs20.
2 IRaMuTeQ is a open software, distributed under license GNU GPL, based on R
statistical software and on Python language. It has now reached version 0.7 alpha 2
and it is still under development (Ratinaud, 2009).
1

512

JADT’ 18

reference to gender as theoretical or empirical category – and not only as a
composing part of a title of some sources, a statistical variable, or synonym –
in order to outline meaningful remarks for the aims of each article. This
selection has been supported by a hermeneutic analysis, based on careful
reading of the papers to evaluate the centrality of the gender issues in their
hypotheses and theses, as in the implementation of the theoretical and/or
empirical methodologies. They resulted 67, distributed over the past 20 years,
as shown in Table 1.
Table 1 - Extracted and Selected Papers

1997-1999
2000-2002
2003-2005
2006-2008
2009-2011
2012-2014
2015-2017
Total

Extracted Papers (EP)

Selected Papers (SP) SP/EP%

19
18
22
21
45
55
55
235

2
3
3
3
20
15
21
67

10,53
16,67
13,64
14,29
44,44
27,27
38,18
28,51

The incidence of the selected papers on the extracted ones (SP/EP%)
highlights the increased relevance of the term gender over time: it is used
more and more often as analytic category in sociological research, rather than
as a synonym or to indicate only a demographic characteristic of individuals.
The corpus, submitted to the subsequent analyzes, includes therefore 67
selected papers, and has the following lexicometric measurements:
dimension N=495470, word types V=21680; Type/token ratio TTR= 4,38%;
Hapax/V= 41,56%; Hapax/N=1,82%.
These characteristics show that the corpus can be considered sufficiently
large for a quantitative approach analysis (Bolasco, 1999, p.203).
2.2. Strategy of Analysis
The analyzes on the corpus, carried out with IRaMuTeQ, will be the
following:
1- Lexicon Analysis: exploration of the lexicon used in the corpus and
identification of theme-words/lemma;
2- Analysis of the specific lexicon: individuation of specific words/lemma
by time and by author/authors gender;
3- Correspondence Analysis: extraction of lexical dimensions starting from
the Aggregated Lessical Table (ALT) Lemma/Texts (Lebart, Salem 1994),

JADT’ 18

513

in which the texts were identified according to the different years of
publication (Y = 1997 ..., 2017;) and the gender of the author/authors (G =
1-Female; 2-Male; 3-Male and Female)
4- Cluster Analysis: identification of main topics through descending
hierarchical analysis (Reinart 1983) applied to the Binary Lexical Table
(BLT), Text segments / Lemma.
5- Similarity Analysis: description of the clusters obtained in point 4),
through graphic representation starting from the proximity matrix
between forms or lemmas.

References
Bolasco S. (1999). Analisi multidimensionale dei dati. Metodi, strategie e criteri
d'interpretazione, Roma, Carocci
Lerbart L., Salem S. (1994). Statistique textuelle, Paris, Dunod.
Money, John; Hampson, Joan G; Hampson, John (1955). “An Examination of
Some Basic Sexual Concepts: The Evidence of Human Hermaphroditism”.
Bull. Johns Hopkins Hosp. Johns Hopkins University. 97 (4), pp. 301–19.
Ratinaud, P. (2009). IRAMUTEQ: Interface de R pour les Analyses
Multidimensionnelles
de
Textes
et
de
Questionnaires.
http://www.iramuteq.org.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique :
application à l’analyse lexicale par contexte. Les Cahiers de l’Analyse Des
Données, 8, 187–198.

514

JADT’ 18

The Rhythm of Epic Verse in Portuguese
From the 16th to the 21st Century
Adiel Mittmann, Alckmar Luiz dos Santos
Universidade Federal de Santa Catarina (Florianópolis, Brazil)
adiel@mittmann.net.br, alckmar@gmail.com

Abstract
The verses of most epic poems in Portuguese have been written following the
example of the Italian endecasillabo: a verse whose last stressed syllable is the
tenth, which usually means, in both Italian and Portuguese, that most verses
have a total of eleven syllables. In addition to the tenth, other syllables may
be stressed within the verse as well, and the specific distributions of stressed
and unstressed syllables make up different rhythmic patterns. In this article,
we investigate how such patterns were used in six epic poems written in
Portuguese, ranging from the 16th to the 21st century, for a total of 52,412
verses. In order to analyze such a large amount of verses, we used Aoidos, an
automatic scansion tool for Portuguese. By using supervised and
unsupervised machine learning, we show that, though the influence of earlier
poets (especially Camões) is ever present, poets favor different rhythmic
patterns, which can be regarded as their rhythmic signature.
Keywords: Epic poetry, Portuguese, Scansion.
Résumé
Les vers de la plupart des épopées en portugais ont été écrits à l’instar de
l’endecasillabo italien : un vers dont la dernière syllabe accentuée est la
dixième, ce qui signifie généralement, en italien et en portugais, que la
plupart des vers ont onze syllabes au total. En plus de la dixième, des autres
syllabes peuvent aussi être accentuées dans ce vers, chaque combinaison de
syllabes accentuées et non accentuées représentant un standard rythmique.
Dans cet article, nous examinons comment ces standards ont été utilisés dans
six épopées écrites en portugais, du XVIème au XXIème siècles, dans un total de
52.412 vers. Pour analyser une telle quantité de vers, nous avons employé
Aoidos, un outil automatique de scansion pour le portugais. En utilisant des
apprentissages supervisés et non-supervisés, nous concluons que, encore que
l’influence de poètes précédents (surtout celle de Camões) se fasse toujours
remarquer, chaque poète préfère de différents standards rythmiques, qui
peuvent être considérés comme sa signature rythmique.
Mots-clés: Epopée, Portugais, Scansion.

JADT’ 18

515

1. Introduction
Poets are frequently compared to one another, but over the centuries rarely
have such comparisons been made objectively, especially with respect to
verse structures. When critics state that a poet has followed the steps of
another too closely and has therefore produced unoriginal and derivative
work, they can seldom rely on objective facts. Works such as that of Chociay
(1994), who manually analyzed and tabulated more than 1,500 verses, are not
the rule, but the exception. It is indeed a tedious and tiresome task for any
human to carry out; but looking at a great amount of text from afar and
extracting relevant information from it constitutes a core element of distant
reading (Moretti, 2013).
Table 1: Poems included in the corpus. The code is derived from the poem’s title.

Code
L
M
C
A
B
F

Author
Luís de Camões
Francisco de Sá de
Santa Rita Durão
Fagundes Varela
Carlos Alberto Nunes
José Carlos de Souza

Born in
Portugal
Portugal
Brazil
Brazil
Brazil
Brazil

Poem
Os Lusíadas
Malaca
Caramuru
Anchieta
Os Brasileidas
Famagusta

Year
1572
1634
1781
1875
1938
2016

Verses
8,816
10,656
6,672
8,484
8,504
9,280
52,412

In this article, we turn our attention to the verse most commonly used in epic
poetry in Portuguese, the decassílabo, which was borrowed from Italian1. It is
the verse used by Dante in his Divina Commedia and by Petrarch in his
Canzoniere. Stressed syllables are distributed in the verse according to certain
rules; in particular, the 10th syllable (which defines the length of the verse)
must always be stressed. Other syllables may also be stressed, producing
many possible rhythmic patterns—which are, both in Portuguese and Italian,
required to have their 6th or, less commonly, their 4th syllable stressed
(Versace, 2014). We identify such patterns by indicating the syllabic positions
that are stressed within a given verse, so that a pattern like 3-6-10 means that
the 3rd, 6th and 10th syllables are stressed.
We are interested in tracking which rhythmic patterns poets have favored
In both Italian and Portuguese, this kind of verse always has its 10th syllable
stressed and typically has a total of eleven syllables, since most words in both
languages have a stress on the penult. However, in Italian this verse is called
endecasillabo because of the total number of syllables, whereas the Portuguese term
decassílabo emphasizes the fact that the 10th is the last stressed syllable in the verse.
1

516

JADT’ 18

over the centuries and whether such patterns are characteristic to each poet.
For this purpose, we have assembled a corpus consisting of six poems, whose
publication dates range from the 16th to the 21st century, for a total of 52,412
verses (about 300,000 words). In order to analyze such an amount of verses,
we have used our automatic scansion tool, Aoidos (Mittmann et al., 2016),
which is capable of scanning thousands of verses in a few seconds and
producing rhythmic information. The next section describes the corpus we
used in our experiments; Section 3 reports the results obtained with our
analyses; finally, Section 4 presents our conclusions and discusses future
work.
2. Corpus
The poems chosen to compose the corpus for this article are summarized in
Table 1. We adopted two criteria in order to select these poems. Firstly, we
searched for an important—and thus well known—or exemplary epic poem
in each century, from the 16th up to the present. Secondly, we required
trustful and reliable digital editions; in one case (17th century), we produced
a digital edition especially for this article, since no suitable candidate was
found.
Camões’ poem Os Lusíadas is by far the most important epic poem ever
written in Portuguese. Its influence can be felt, for instance, even in 20thcentury lyrical poets such as Jorge de Lima. Meneses’ Malaca Conquistada and
Durão’s Caramuru follow very closely the Camonean model: they use
identical rhyme schemes, they have a similar argument and they celebrate a
protagonist in like manner. Nevertheless, we would like to investigate
whether the two authors innovated with respect to rhythm, even though they
kept the overall model of the Camonean epic. These three poems in our
corpus were written by Portuguese citizens (Durão was born in colonial
Brazil and died before the country’s independence), while the remaining
three poems were written by Brazilian poets.

14610
610
1610
2-

1
Ce-

2
ssem

3
do

4
sá-

5
bio

6
Gre-

7
go
e

8
do

9
Troi-

10
a-

11
no

As

na-

ve-

ga-

ções

gran-

des

que

fi-

ze-

ram;

Ca-

le-

se

de
A-

le-

xan-

dro
e

de

Tra-

ja-

no

A

fa-

ma

das

vi-

tó-

rias

que

ti-

ve-

ram;

JADT’ 18
610
24610
24610
136810
14610

517

Que
eu

can-

to o

pei-

to i-

lus-

tre

Lu-

si-

ta-

no,

A

quem

Nep-

tu-

no e

Mar-

te
o-

be-

de-

ce-

ram:

Ce-

sse

tu-

do
o

que
a

Mu-

sa
an-

ti-

ga

can-

ta,

Que
ou-

tro

va-

lor

mai-

s al-

to

se
a-

le-

van-

ta.

Figure 1: Scansion produced by Aoidos.

Fagundes Varela’s Anchieta, a romantic piece of the 19th century, would not
be, at a first glance, an epic poem, since its subject is the telling of New
Testament stories to Brazilian Indians by priest José de Anchieta. However,
as historian Maria Aparecida Ribeiro and others remark, Anchieta is a kind of
“religious epopee” (Ribeiro, 2003), which drives our attention to the
Romantic effort to renew the ancient models inherited from Classical or
Neoclassical literature (although it clearly returns to the Greek epic model, as
it does not adopt regular sized stanzas). Despite some important differences
in the narrative logic, the verses reproduce the most important invariants of
the genre: the honoring of a protagonist (Anchieta) and the use of the
decassílabo (blank ones, in this case). As for Carlos Alberto Nunes’ Os
Brasileidas, this poem also presents some invariants that characterize the
traditional epic poem: blank decassílabo verses; several cantos, beginning with
the proposition; the intention of celebrating an individual hero, in this case
Antônio Raposo Tavares, a 17th-century Brazilian trailblazer. In addition to
the absence of rhymes, in order to emphasize the differences in relation to the
Camonean epic style, there is no regular stanza division in each one of the
nine cantos (ten, if we consider the epilogue), as in Anchieta, although they
may vary significantly, from seven up to sixty five or more verses. Finally,
regarding Famagusta, by José Carlos de Souza Teixeira, one quickly notices
that it is a curious combination of traditional epic elements from different
ages. In addition to the epic intention of celebrating an historical event and

518

JADT’ 18

some sort of heroic action, its formal elements are, to say the least, very
heterogeneous. For instance, it takes the Camonean eight verse stanza but
adopts a different rhyme scheme, resulting no more in the well-known ottava
rima (ABABABCC), but in the medieval Sicilian stanza called strambotto
romagnuolo (ABABCCDD), scarcely used in Brazilian literature2.
3. Analysis
In order to analyze the corpus, we used Aoidos, an automatic scansion tool
for Portuguese (Mittmann et al., 2016), much like Métromètre (Beaudouin
and Yvon, 2004) and Anamètre (Delente and Renault, 2015) for French.
Starting from the written word, Aoidos produces a phonetic transcription for
each verse and then applies many rules (such as elision or syncope) to
produce a series of alternative scansion. By examining the poem as a whole,
the system then selects the most appropriate alternative and, by applying a
set of heuristics, proposes a rhythmic pattern for each verse. The scansions
generated by Aoidos have been manually verified to be correct in 99.0% of
cases (Mittmann, 2016). Figure 1 shows the output produced by the system
for the 3rd stanza of Camões’ Os Lusíadas.

2-4-8-10

1-3-6-8-10

1-4-6-8-10

4-6-8-10

1-4-8-10

4-8-10

1-6-10

1-6-8-10

7.7
7.1
7.6
9.4
9.6
10.5

4-6-10

2-6-8-10

15.2
12.2
11.2
7.3
7.7
11.7

1-4-6-10

2-4-6-10

9.0
9.6
7.1
11.3
13.2
16.2

1-3-6-10

2-6-10

10.3
11.9
10.3
14.4
14.0
13.8

2-4-6-8-10

3-6-10

L
M
C
A
B
F

7.6 11.0 6.2
8.2 9.5 5.2
8.5 9.1 7.7
9.0 4.0 6.1
7.3 5.1 6.8
9.0 8.0 7.0

7.9
5.7
6.1
7.2
5.1
4.0

6.2
5.5
3.6
5.4
5.2
4.5

1.1
6.5
8.1
6.5
4.6
0.2

4.5
3.8
5.8
4.2
3.5
4.3

5.0
3.9
4.6
3.3
2.7
3.1

4.1
3.6
2.9
2.5
3.5
2.7

0.5
2.2
3.6
4.6
3.4
0.1

0.4
1.8
2.2
1.2
3.1
0.1

1.3
1.1
0.4
1.6
2.1
1.8

1.0
0.8
0.8
1.3
1.4
1.2

3-6-8-10

Poem

Table 2: Rhythmic pattern usage (%) for each poem.

A total of 42 different rhythmic patterns were found among all 6 poems.
Table 2 shows how frequently patterns with an average usage of at least 1%
were employed in each poem. In each row, the bold number indicates the
pattern most favored by that row's poem. Although some patterns, such as

The Brazilian-born baroque poet Manoel Botelho de Oliveira did use this
stanza in some madrigals written in Spanish, such as this one: Si Cupido me inflama, /
Si desdeñas mi empleo; / En amorosa llama, / En nieve desdeñosa el Etna veo, / Con
amor, y tibieza / Tenemos su firmeza, / Y en disonancia breve / Suspiro fuego yo, tu
brotas nieve.
2

JADT’ 18

519

Figure 2: Dendrogram built from all cantos of all poems.

3-6-8-10 and 1-3-6-10 remain more or less constant, many others display a
wide range of relative usage: pattern 2-6-10 ranges from 7.1% to 16.2%, and
pattern 1-4-8-10 from 0.1% to 3.1%. Whereas Camões (L) does seem to set the
tone for the following poems, there are clear differences when one considers
patterns such as 2-4-6-10 and 2-4-8-10. In fact, pairs such as Malaca (M) and
Caramuru (C) or Anchieta (A) and Os Brasileidas are more similar between
themselves than Camões’ Os Lusíadas (L) is to any other poem. By looking at
numbers from one century to the next, twice a change of more than 5% can
be seen: from Caramuru (C) to Anchieta (A) there was a decrease of 5.1% for
the pattern 2-4-6-8-10, and from Os Lusíadas (L) to Malaca (M) the pattern 2-48-10 increased in usage by 5.4%.
An interesting question arises at this point: do smaller parts of the poems
reflect the overall distribution shown in Table 2? In other words, given a
smaller part of a poem, could we tell from which work it was taken simply
by looking at its rhythmic signature? To answer this question, we divided
each poem into its cantos, for a total of 72 divisions, with an average of 727.9
verses per canto. We then extracted the usage frequency of the rhythmic
patterns, thus producing a feature vector for each canto. By iteratively
clustering such vectors, we obtained the dendrogram shown in Figure 2;
complete linkage was used. Each canto in the figure is indicated by a letter
(the poem code) and a number (the canto number within the poem). Cantos
from the same poem are also displayed with the same color. The closer to the
center that two branches link together, the more different the cantos they
contain are. We can immediately see that, in general, cantos that belong to
the same poem are located next to each other. All cantos of Camões’ Os

520

JADT’ 18

Lusíadas (L), in particular, are tightly grouped in their own branch. It is also
interesting to note that, except for Famagusta (F), whenever a smaller group of
cantos from the same poem were placed far from the larger group of cantos,
there is a certain order: it was the first three cantos of Caramuru (C) were
separated; the last four of Anchieta (A); and the first two of Os Brasileidas (B).
Two cantos from Famagusta (F1 and F16) are only linked with other nodes at
a great distance; this stems from the fact that these two cantos are the shortest
ones in all of the corpus: the first canto has only 24 verses, the sixteenth 112.
Such small amounts of verses produce poor feature vectors.
In order to further investigate how well the cantos reflect the poems, we
employed a nearest centroid classifier. In this case, each of the 72 feature
vectors (the rhythmic signatures of the cantos) was labeled with the poem
they belong to. We then used stratified k-fold cross validation, with k = 4 and
100 repetitions to assess the classifier’s performance. The mean precision
obtained was 96.5%, mean recall 95.9% and mean F1 score 95.5%; the mean
accuracy was 95.6%. This means that, given a sample of 54 cantos (because k
= 4), the classifier guesses the right poem for the other 18 cantos in about 96%
of the cases.
4. Conclusion
The frequency with which poets employ certain patterns of stressed and
unstressed syllables in their verses can be regarded as a rhythmic signature—
at least in epic poems, the subject of this article. In this work, we have
subjected 72 individual cantos to a hierarchical clustering technique (Figure
2), which shows that rhythmic patterns do reflect an author’s preferences
(unconscious as they might be). Furthermore, a nearest centroid classifier
obtained a mean accuracy of 95.6%, which is also evidence for the existence
of a rhythmic signature. This kind of analysis is possible thanks to automatic
scansion systems, such as Aoidos, which allow a large amount of verses
(more than 50,000 in this case) to be scanned and analyzed.
Although Camões, whose poem Os Lusíadas is the oldest in our corpus, has
influenced newer generations of poets, this article shows that, at least
rhythmically, each poet in our corpus took their own path. In fact, Camões’
verses are the ones most easily distinguished from the others (see Figure 2).
Lesser-known poems, such as Malaca or Os Brasileidas, have not failed to
produce rhythmic signatures that, in most cases, set them apart from other
works. In addition to the rhythmic signature, we would like to investigate, in
the future, additional features that could be extracted from verses and used
in stylometric analyses. In particular, the decassílabo usually falls into one of
two categories: either the 6th syllable has the dominant stress or—less
commonly—the 4th; in the former case, the verse is heroic; in the latter,

JADT’ 18

521

Sapphic. A verse whose rhythmic pattern includes the 6th syllable, but not the
4th, is heroic; but one that includes both the 6th and the 4th could be either
heroic or Sapphic. It would be interesting to resolve this ambiguity and
evaluate how well these categories characterize a poet’s style.
Although this article has only considered epic poems, there is no reason to
believe that rhythmic signatures are limited to this genre. In the future, we
would like to explore how well the approach shown here fares when applied
to other verses and other genres.
Acknowledgments
For the nearest centroid classifier we employed Scikit-learn (Pedregosa et al.,
2011). For the dendrogram, we used Dendextend (Galili, 2015) and Circlize
(Gu et al., 2014).
References
Beaudouin, Valérie and Yvon, François (2004). “Contribution de la métrique
à la stylométrie”. 7èmes Journées internationales d’Analyse statistique des
Données Textuelles. (2004), pp. 107–118.
Chociay, Rogério (1994). A Identidade Formal do Decassílabo em “O
Uraguai”. Revista de Letras 34, 229–243.
Delente, Éliane and Renault, Richard (2015). Projet Anamètre : Le calcul du
mètre des vers complexes. Langages 3.199, 125–148.
Galili, Tal (2015). dendextend: an R package for visualizing, adjusting and
comparing trees of hierarchical clustering. Bioinformatics 31 (22), 3718–
3720.
Gu, Zuguang et al. (2014). circlize implements and enhances circular
visualization in R. Bioinformatics 30 (19), 2811–2812.
Mittmann, Adiel (2016). “Escansão Automático de Versos em Português”.
PhD thesis. Universidade Federal de Santa Catarina.
Mittmann, Adiel, Wangenheim, Aldo von, and Luiz dos Santos, Alckmar
(2016). “Aoidos: A System for the Automatic Scansion of Poetry Written
in Portuguese”. 17th International Conference on Intelligent Text
Processing and Computational Linguistics. (2016).
Moretti, Franco (2013). Distant reading. London: Verso.
Pedregosa, F. et al. (2011). Scikit-learn: Machine Learning in Python. Journal of
Machine Learning Research 12, 2825–2830.
Ribeiro, Maria Aparecida (2003). Anchieta no Brasil: Que Memória? História
Revista 8, 21–56.
Versace, Stefano (2014). A Bracketed Grid account of the Italian endecasillabo
meter. Lingua 143, 1–19.

522

JADT’ 18

Le vocabulaire des campagnes électorales
Denis Monière1, Dominique Labbé2
Université de Montréal (denis.moniere@umontreal.ca)
2 PACTE CNRS - Université de Grenoble (dominique.labbe@umrpacte.fr)
1

Abstract
After having done a first presidential term, V. Giscard d’Estaing, F.
Mitterrand, J. Chirac and N. Sarkozy were candidates for a second term. In
this study, their electoral speeches are compared with their presidential ones
drawing attention to the specific nature of the vocabulary used. It would
appear that this calculation is mainly biased by grammatical categories and
word frequency. We present modifications of the classical formulae which
make it possible to neutralize the influence of grammatical categories and, at
least partially, that of word frequency. Electoral discourse privileges the verb
over the name, as such speech is more personalized than governmental
discourse, it focuses on the country and its inhabitants, the rest of the world
being pushed into the background. Finally, in recent years, the polemical
dimension is becoming predominant.
Résumé
Après un premier mandat présidentiel, V. Giscard d’Estaing, F. Mitterrand, J.
Chirac et N. Sarkozy ont été candidats à un deuxième mandat. On compare
leurs discours électoraux avec leurs discours présidentiels à l’aide des
spécificités du vocabulaire. Il apparaît que ces spécificités dépendent surtout
des catégories grammaticales et des effectifs des mots. On présente des
modifications du calcul classique qui permettent de neutraliser l’influence
des catégories grammaticales et, au moins partiellement, celle des fréquences.
Le discours électoral privilégie le verbe au détriment du nom, il est plus
personnalisé que le discours au pouvoir, il se centre sur le pays et ses
habitants, le reste du monde passant au second plan. Enfin, ces dernières
années, la dimension polémique devient prédominante.
Keywords: lexicometry ; political discourse ; French presidential campaigns ;
specific vocabulary ; spécificités du vocabulaire.
1. Introduction
Le discours électoral diffère-t-il du discours de gouvernement et en quoi ? La
réponse est difficile car il faut neutraliser l’effet des personnalités et des
conjonctures pour isoler l’effet sur le discours des choix stratégiques du

JADT’ 18

523

locuteur. L’idéal serait de pouvoir étudier les mêmes hommes à peu près
simultanément dans les deux positions de gouvernant puis de candidat. Le
corpus des discours des présidents français depuis 1958 remplit ces deux
conditions (présentation du corpus dans Arnold et al 2016). En effet, pour 5
présidents (C. de Gaulle, V. Giscard d’Estaing, F. Mitterrand, J. Chirac et N.
Sarkozy), ce corpus contient leurs interventions lorsqu’ils étaient présidents
et leurs discours de campagne pour leur réélection. Certes, en 1965, de Gaulle
n’a pratiquement pas fait campagne (Labbé 2005), mais ses successeurs ne
l’ont pas imité en 1981, 1988, 2002 et 2012 (corpus en annexe).
Pour comparer ces corpus, le calcul des "spécificités" semble l’outil le plus
adapté (Lafon 1980 et 1984). Il rapporte le vocabulaire d’un sous-ensemble de
textes (sous-corpus) à un corpus de référence. Mais il se heurte à une double
difficulté : la spécificité éventuelle d’un vocable est liée à sa catégorie
grammaticale et à sa fréquence d’emploi (Labbé, Labbé 1994 ; Monière et al.
2005), comme nous allons le vérifier d’abord avec le cas de Sarkozy en 2012
(Sur cette campagne : Labbé, Monière 2013). Dès lors, la mesure des
spécificités doit neutraliser, autant que possible, ces deux inconvénients.
2. Les catégories grammaticales du discours électoral
Le discours présidentiel de Sarkozy s’étend de son investiture (16 mai 2007)
au 12 février 2012 (annonce de sa candidature). La campagne s’étend
jusqu’au soir du second tour (6 mai 2012). Le corpus complet (P) compte 1074
interventions, soit au total 3 221 259 mots avec 21 602 vocables différents. A
partir de sa déclaration de candidature, Sarkozy est intervenu 110 fois (souscorpus E), soit 369 808 mots et un vocabulaire de 8 511 vocables différents.
Ces interventions sont d’abord marquées par un net changement de style
(tableau1).
Tableau 1. Densités des catégories grammaticales dans les interventions de Sarkozy lors de la
campagne de 2012 comparées à ses interventions comme président 2007-2012 (en ‰)
Catégories
Verbes
Futurs
Conditionnels
Présents
Imparfaits
Passés simple
Participes passés
Participes présents
Infinitifs
Noms propres
Substantifs
Adjectifs

P-E (CorpusSous corpus)
159.2
7.0
3.2
82.9
6.4
0.6
20.8
2.1
36.3
27.9
178.4
54.0

E Sous corpus
169.4
7.2
2.8
89.3
6.4
0.3
23.8
2.1
37.6
23.0
176.0
46.6

(P-E)/P
+6.4
+1.6
-11.2
+7.7
-0.2
-55.2
+14.6
+2.9
+3.6
-17.3
-1.3
-13.7

Indice
+
+
+
≈
+
≈
+
-

524
Adj. participe passé
Pronoms
Pronoms personnels
Déterminants
Articles
Nombres
Possessifs
Démonstratifs
Indéfinis
Adverbes
Prépositions
Coordinations
Subordination

JADT’ 18
5.2
124.3
65.4
181.6
131.9
18.7
14.5
7.6
8.9
67.1
150.1
29.1
25.9

4.5
132.6
69.6
182.5
128.1
20.9
17.0
7.8
8.7
68.9
145.6
25.4
27.9

-13.1
+6.7
+6.5
+0.5
-2.9
+11.9
+17.3
+2.7
-2.4
+2.7
-3.0
-12.7
+8.0

+
+
+
+
+
+
+
+

Dans le discours présidentiel, on rencontre 159 verbes en moyenne pour 1
000 mots ; dans les discours électoraux, cette proportion passe à 169‰, soit
une augmentation de +6,4%, ce qui est un écart significatif avec moins de une
chance sur 10 000 de se tromper (signe + en dernière colonne). Les lignes
suivantes donnent le détail des temps et des modes. Le recul le plus
significatif concerne le conditionnel (le discours électoral ne doit pas
connaître le doute). En revanche, le participe passé connait l’augmentation la
plus forte (le président sortant peut difficilement éviter de défendre sa
gestion).
Les pronoms, les adverbes et les conjonctions de subordination évoluent
dans le même sens que les verbes. Ils sont réunis dans le "groupe du verbe".
A l’inverse, les substantifs, adjectifs, articles et prépositions suivent la
tendance inverse : groupe du nom. Le tableau 2 donne les densités des deux
groupes chez les 4 présidents.
Tableau 2. Densités des groupes du verbe et du nom (en ‰) dans les discours électoraux (E)
comparés aux discours présidentiels (P-E).
Catégories
Sarkozy (2007-2012)
Groupe du verbe
Groupe du nom
Giscard d’Estaing (1974-1981)
Groupe du verbe
Groupe du nom
Mitterrand (1981-1988)
Groupe du verbe
Groupe du nom
Chirac (1995-2002)
Groupe du verbe
Groupe du nom

P-E (CorpusSous corpus)

E Sous
corpus

(P-E)/E

Indice

376.6
621.1

398.9
599.2

+5.9
-3.5

+
-

351.5
646.1

392.5
604.5

+11.7
-6.4-

+
-

386.4
611.0

427.1
569.8

+10.5
-6.7

+
-

329.5
668.8

333.2
665.1

+1.1
-0,6

+
-

JADT’ 18

525

Chez tous les présidents en campagne, il se produit une augmentation du
groupe du verbe et un recul de celui des noms. Statistiquement, ces
mouvements sont significatifs (avec = 1%). L’écart le plus fort est observé
chez Giscard d’Estaing puis chez Mitterrand. Cependant, Chirac tranche sur
les autres avec une densité du verbe beaucoup plus faible et une campagne
présidentielle presqu’aussi distanciée que ses interventions lors de son
premier mandat, marqué par une cohabitation de 5 ans (1997-2002) avec un
Premier ministre socialiste (Jospin). Dans son discours électoral, la densité
des verbes augmente nettement (+3,6%) mais se trouve en partie compensée
par un recul des pronoms, ce qui accentue le caractère dépersonnalisé des
propos de Chirac à l’opposé des trois autres.
En conséquence, pour les 4 présidents, les principaux verbes apparaissent en
spécificités positives du discours électoral et il ne s’en trouve que quelquesuns en spécificités négatives. Il en est de même pour les pronoms et les
adverbes. La situation inverse se constate pour les adjectifs, les substantifs,
etc. Autrement dit, si un mot appartient à une catégorie sous-employée dans
le sous-corpus (par rapport à sa densité d’utilisation dans le corpus entier), ce
vocable a toute chance d’apparaître dans les spécificités négatives (et
positives dans le cas inverse). Il est possible de neutraliser ce biais.
3. Neutralisation de la catégorie grammaticale
Le calcul standard est le suivant. Soit :
- le corpus de référence (P) long de Np mots ;
- le sous-corpus E long Ne mots dont on recherche les spécificités par rapport
àP;
- un vocable i avec Fip occurrences dans P et Fie dans E. Si sa répartition est
uniforme, ce vocable apparaîtra Eie(u) fois dans le sous-corpus E :
E ie(u)  Fip * U avec U =

Ne
369 808

 0.113
N p 3 223 570

(1)

La probabilité pour que le vocable i soit observé Fie fois dans E suit une loi
hypergéométrique de paramètres Fip, Fie, Ne, Np :
Fip   N p  Fip 
 

Fie   N e  Fie 

P( X  Fie ) =
N p 
 
N e 

(2)

L’indice de spécificité (S) est la somme des probabilités – calculées avec (2) –

526

JADT’ 18

de survenue des J valeurs entières de X variant de 0 à Fie {X=0 ; X= Fie} :
j = Fie

S = P( X  Fie ) =

 P(X  j)

(3)

j= 0

Si au seuil , Fie excède Eie(u), le vocable est « spécifique plus » (S+) ; S- dans le
cas contraire. Avec ce calcul, la plus grande partie des verbes usuels de
Sarkozy apparaissent donc en S+ de sa campagne électorale et la majorité des
substantifs en S-, parce que, dans ses discours électoraux, la première
catégorie est privilégiée par rapport au discours de gouvernement où elle est
moins utilisée (à l’inverse des substantifs). Pour corriger ce biais, le calcul
prend en compte les catégories grammaticales (g). La modification est
présentée dans : Monière, Labbé, Labbé 2005 ; Mayaffre 2006 et Monière,
Labbé 2012.
Soit : Nge et Ngp le nombre de mots appartenant à la catégorie grammaticale G
respectivement dans le sous-corpus E et le corpus entier P. Les formules (1) et
(2) deviennent :
E ie(u)  Fip * U avec U =

N ge
N gp

Fip   N gp  Fip 

 
Fie   N ge  Fie 
P( X  Fie ) =
 N gp 


 N ge 

(4)

(5)

Les formules (4) et (5) appliquées aux 4 corpus aboutissent à un équilibre
relatif, au sein de chaque catégorie, entre les S+ et les S- (tableau 4). Ces
formules neutralisent donc la liaison entre spécificités et densité des
catégories grammaticales. Comme indiqué dans Monière & Labbé 2012, cette
modification change drastiquement la liste des "mots spécifiques" mais elle
laisse subsister la liaison entre spécificité et fréquence.
4. Questions de seuils
Le calcul porte sur une minorité du vocabulaire et il est asymétrique. En effet,
avec = 1% :
- l’effectif minimal pour être S+ est de 5 occurrences ("seuil de spécificité
positive"), toutes dans les discours électoraux (E) et à condition que Eie(u) < .5,
ce qui signifie que Nge < 0.10Ngp. Par construction le calcul élimine donc tous
les vocables d’effectifs inférieurs à 5. Dans le corpus Sarkozy, cela représente

JADT’ 18

527

plus de la moitié du vocabulaire (54 % des vocables). Autrement dit,
seulement 46% du vocabulaire peut être S+ ;
- le "seuil de spécificité négative" correspond à la situation suivante : un
vocable i absent de E (Fie = 0) alors qu’on en attend au moins 5 (Eie(u) ≥ 5). En
pratique, cela signifie que son effectif dans P est égal ou supérieur à 5*1/U,
soit ici 40. Autrement dit, pour le discours électoral de Sarkozy, 83% du
vocabulaire de P ne peut apparaître en S-.
Dès lors, les vocables dont les effectifs dans P sont compris entre 5 et 39
peuvent être S+ mais pas S- dans E. On s’attend donc à ce qu’il y ait plus de
vocables S+ que S-.
5. Liaison entre spécificité et fréquence
9 876 vocables apparaissent 5 fois ou plus dans P. Si ce corpus était
homogène (hypothèse nulle H0), une distribution normale des vocables
laisserait attendre - avec = 1% - environ 100 vocables spécifiques. Le
tableau 3 compare les résultats observés et attendus (avec H0).
Tableau 3. Effectifs des vocables classés par catégories grammaticales et par spécificités
Verbes
Mots à majuscule
Substantifs
Adjectifs
Pronoms
Adverbes
Déterminants
Prépositions
conjonc.
Total

&

Effectifs (Fip ≥ 5)
1 540
1 501
4 175
2 065
52
411
72
60

H0
15
15
42
21
1
4
1
1

S+
176
112
455
140
18
20
21
21

S143
142
468
115
13
57
12
9

Total S
319
254
923
255
31
77
33
30

9 876

100

963

959

1 922

Il y a donc vingt fois plus de vocables spécifiques que n’en laisse attendre H0
(répartition homogène des mots entre corpus et sous-corpus). A priori, cela
signifie simplement que discours électoral et discours de gouvernement sont
fortement contrastés. En fait, ce décalage provient essentiellement des
vocables les plus fréquents (tableau 4 et Figure 1).
Tableau 4. Proportion des vocables spécifiques de E dans l’ensemble du vocabulaire (P) classé
par fréquence absolues.
Classe de
fréquence (P)
5-9
10-14
15-19
20-29

Vocables spécifiques de
E dans la classe
64
68
55
89

Total vocables de P
dans la classe
2 759
1 237
757
987

Proportion des vocables
de P spécifiques de E
2,3
5,5
7,3
9,0

528

JADT’ 18
30-49
50-99
100-199
200-499
500+
Total

143
317
332
398
473
1 939

997
1 054
799
686
640
9 916

14,3
30,1
41,6
58,0
73,9
19,6

Figure 1. Liaison entre la spécificité et la fréquence

Au-dessus du seuil de spécificité positive (ici 40), la proportion de vocables
spécifiques est directement corrélée avec la fréquence : la courbe suit la
diagonale du tableau et le coefficient de détermination de Y par X est égal à
0,997, ce qui indique une liaison rigide et linéaire. Il en est toujours ainsi :
plus un vocable donné est fréquent dans un corpus, plus il a de chances
d’être "spécifique" à l’une quelconque des parties de ce corpus. Cette
dépendance peut être interprétée de deux manières. D’une part, l’essentiel
des choix thématiques seraient véhiculés par les vocables les plus fréquents
et la variation dans leurs fréquences d’emploi seraient la principale
manifestation de ces choix. Cependant, dès que le corpus atteint une certaine
longueur, l’observateur se trouve noyé dans des listes qui contiennent la plus
grande part du vocabulaire usuel, ce qui en rend l’interprétation difficile.
D’autre part et à l’inverse, on peut penser que le raisonnement probabiliste qui sous-tend ce calcul - doit être adapté à cette liaison manifeste entre
spécificité et fréquence.
6. Neutralisation de la liaison entre fréquence et spécificité
Les limites des classes de fréquence du tableau 5 et de la figure 1 ont été
fixées selon une échelle proche d’une progression géométrique, ce qui assure

JADT’ 18

529

aux classes des effectifs sinon égaux du moins suffisamment proches et
importants. Ceci correspond à une particularité dite "loi de Zipf" - ou "ZipfMandelbrot" - selon laquelle le nombre d’occurrences d’un mot dans un texte
est lié à son rang dans la distribution des fréquences (Zipf 1935 ; Mandelbrot
1957).
Dès que le corpus atteint une longueur suffisante (au moins un demi-million
de mots) et que le sous-corpus est égal à au moins d’un dixième du corpus,
on peut découper le vocabulaire en quelques classes de fréquence. Pour un
corpus de la dimension de celui de Sarkozy (et des trois autres présidents),
trois classes suffisent : vocables "rares" (inférieurs à 100 occurrences) ;
"fréquents" (de 100 à moins de 500) ; "très fréquents" (500 et plus). Dans ces
trois classes, les vocables sont classés par catégorie grammaticale puis en
fonction de leur indice de spécificité et, dans chacune des classes, seuls les
plus caractéristiques sont retenus. Le tableau 5 donne les 5% les plus
caractéristiques du discours électoral de Sarkozy comparé à son discours
présidentiel, pour trois catégories grammaticales.
Tableau 5. Spécificités les plus remarquables du discours électoral de Sarkozy par rapport à son
discours présidentiel (par catégories grammaticales en trois classes de fréquence)
<100
Vocables significativement sur-employés :
Verbes :
voler, cotiser, détester,
casser, éduquer,
suspendre, démolir
Mots à majuscule

Mélenchon, Le Pen,

Substantifs

honte, rassemblement,
héritier, socialiste,
colère, délit, amalgame

100 – 499
adresser, bénéficier,
apprendre, souffrir,
supprimer, régulariser
François, Polynésie,
Hollande, Schengen,
TVA
jeunesse, souffrance,
gauche, destin, erreur,
étranger, salaire,
outremer

Vocables significativement sous-employés :
Verbes
admirer, illustrer,
progresser, témoigner,
expérimenter, inaugurer, évoquer, marquer,
associer
Mots à majuscule

Bush, Poutine,
Roumanie, Quatar

Russie, Inde, Iran,
Barroso

Substantifs

refondation, coalition,
scientifique, lycéen

processus, visite,
équipe, conférence,
planète, gouvernance,
alliance,

500+
dire, vouloir, parler,
vivre, proposer,
changer, respecter,
défendre
France, Français, Corse

travail, entreprise, droit,
république, vie, emploi,
ami, enfant, territoire,
peuple,
être, devoir, savoir,
comprendre, trouver,
attendre, remercier,
essayer
Afrique, G20,
Méditerranée, Merkel,
Paris, Chine
pays, monsieur,
président, état, ministre,
politique,
gouvernement, question

530

JADT’ 18

Chez Sarkozy, le discours électoral est affaire de volonté, il se centre sur le
pays, ses habitants mais aussi l’adversaire – la gauche, Hollande - dont il
dénonce les amalgames et les erreurs. Les spécificités négatives indiquent
que le discours électoral n’est pas affaire de devoir ou de connaissance ; il
"oublie" le reste du monde et ses dirigeants, les institutions du pays comme le
gouvernement et les ministres, etc.
7. Conclusions
Lorsqu’un président entre en campagne, il doit descendre dans l’arène et
adopter un discours de combat qui se caractérise avant tout par une
augmentation de la densité des verbes, une forte personnalisation et un recul
de la place accordée aux substantifs et aux adjectifs. Ces caractéristiques se
retrouvent dans les discours électoraux des Premiers ministres canadiens
(Monière, Labbé 2010). Cependant, en campagne ces derniers insistent sur le
"nous" car, dans un système parlementaire, il s’agit de faire élire une majorité
de députés, alors que les présidents français privilégient le "je"… Enfin, ces
dernières années en Amérique du nord comme en France, la forte présence
de la construction négative et la désignation des adversaires (noms propres)
soulignent le caractère polémique du discours électoral.
Le calcul des spécificités – tel qu’il est utilisé en analyse des données
textuelles – enregistre la catégorie grammaticale du vocable analysé et sa
fréquence d’emploi et non pas les choix thématiques du locuteur. La
neutralisation de la catégorie grammaticale est aisée si les mots ont été
étiquetés. En revanche, l’effet de la fréquence est susceptible de plusieurs
interprétations. Toutefois, si l’on souhaite ne pas être enseveli sous les listes
produites par le calcul classique, la solution réside dans le classement des
vocables en classes de fréquence –selon une échelle géométrique - et, au sein
de chacune de ces classes, dans la sélection des vocables les plus singuliers. A
ce prix, les singularités d’un sous-corpus peuvent être identifiées sans avoir à
effectuer des tris discutables dans des listes trop longues.
References
Arnold E., Labbé C. & Monière D. (2016). Parler pour gouverner : Trois études
sur le discours présidentiel français. Grenoble : Laboratoire d'Informatique
de Grenoble, 2016.
Labbé C., Labbé D. (1994). Que mesure la spécificité du vocabulaire ? Grenoble :
CERAT, décembre 1994. Reproduit dans Lexicometrica, 3, 2001.
Labbé D., Monière D. (2010). Quelle est la spécificité des discours électoraux?
Le cas de Stephen Harper. Canadian Journal of Political Science, 43:1, p. 69–
86.
Labbé D., Monière D. (2013). La campagne présidentielle de 2012. Votez pour

JADT’ 18

531

moi ! Paris : l’Harmattan.
Lafon P. (1980). Sur la variabilité de la fréquence des formes dans un corpus.
Mots, 1, p. 127-165.
Lafon P. (1984). Dépouillements et statistiques en lexicométrie. Genève-Paris :
Slatkine-Champion.
Mandelbrot B. (1957). Étude de la loi d'Estoup et de Zipf Fréquences des mots
dans le discours. Apostel L et al. Logique, langage et théorie de l'information.
Paris, PUF, p. 22-53.
Mayaffre D. (2006). Faut-il pondérer les spécificités lexicales par la
composition grammaticale des textes ? Tests logométriques appliqués au
discours présidentiel sous la Vème République. Condé C., Viprey J.-M.
Actes des 8e Journées internationales d'Analyse des données textuelles.
Besançon : Presses universitaires de Franche Comté, II, p. 677-685.
Monière D., Labbé C., Labbé D. (2005). Les particularités d'un discours
politique : les gouvernements minoritaires de Pierre Trudeau et de Paul
Martin au Canada. Corpus, 4, p.79-104.
Monière D., Labbé D. (2012). Le vocabulaire caractéristique du Premier
ministre du Québec J. Charest comparé à ses prédécesseurs. Dister A. et
al. (éds). Proceedings of the 11th International Conference on Textual Data
Statistical Analysis. Liège : LASLA - SESLA, p.737-751.
Zipf G. K. (1935). La psychobiologie du langage. Paris : CEPL, 1974.

532

JADT’ 18

Faire émerger les traces d’une pratique imitative dans
la presse de tranchées à l’aide des outils
textométriques
Cyrielle Montrichard
ELLIADD, UBFC – cyrielle.montrichard@edu.univ-fcomte.fr

Abstract
The main goal of this paper is to show how textometric tools can help to
reveal the imitative usage of genres. During the Great War, soldiers must not
criticize the hierarchy or the governement. Trench press is written by and for
French soldiers in which we can find a great number of media and literary
genres. Plus, we assume that writers use a number of discursive schemes to
implicitly tell their point of view on the war, the governement and the
« sacred union » discours which has become the mainstream speech in the
public space in the early begining of the war. Therefore a corpus of this press
seems to be the perfect place to search the notion of imitative usage of genres.
To put into perspective the results given by the textometric tools we use a
sample corpus from the national french press.
Résumé
L’objectif de cette contribution est d’interroger la pratique imitative des
genres médiatiques et littéraires. Pour ce faire, nous mobilisons un corpus de
presse de tranchées dans lequel se déploient de nombreux genres et sousgenres. Portant notre attention tout particulièrement sur les genres des
dépêches et du roman-feuilleton nous montrons, en comparant ce corpus à
un corpus échantillon de textes parus dans la presse quotidienne nationale en
quoi la presse de tranchées copie les genres instaurés dans la presse civile. La
seconde partie interroge le corpus au niveau syntagmatique pour tenter de
faire émerger les registres ludiques et satiriques ayant court dans cette
presse.
Keywords : presse écrite, genre, pratique imitative, première guerre
mondiale, presse de tranchées.
1. Introduction
La presse de tranchées est un type de document né pendant la première
guerre mondiale. Cette presse a la particularité d’être écrite par et pour les
combattants (Audoin-Rouzeau, 1986). La censure ainsi que le discours

JADT’ 18

533

doxique d’union sacrée tenant place dans l’espace publique durant la période
du conflit ne permettent pas aux locuteurs d’exprimer ouvertement leur
opinion (Forcade, 2016). L’objectif de cette communication est de montrer
comment émergent les registres ludiques et satiriques dans la presse de
tranchées à travers l’inscription de discours dans des genres faisant écho à la
matrice générique médiatique et littéraire.
Comment repérer à l’aide des outils textométriques les traces discursives
d’une pratique imitative des genres médiatiques et littéraires dans la presse
de tranchées ?
Cette communication vise à interroger la « pratique imitative » c’est-à-dire les
« différentes formes ou genres qui permettent à un auteur de produire un
texte (T2) attribué, sérieusement ou non, et de manière plus ou moins
explicite, au modèle dont il s’est inspiré (T1) » (Aron, 2013). Pour ce faire,
nous avons réuni en corpus cinq titres de presse de tranchées au format
XML-TEI pour plus de 500 000 occurrences permettant une analyse du
discours outillée.
À l’aide des outils textométriques et de la plateforme TXM (Heiden et al.,
2010) nous proposons de montrer comment les textes s’inscrivent et
reprennent les codes établis des genres médiatiques et littéraires. Ensuite,
nous proposons des pistes d’analyse visant à faire émerger le registre ludique
ou satirique usité par les rédacteurs pour détourner le genre.
2. Contexte de la recherche et présentation du corpus
Notre étude propose d’investir la notion de pratique imitative. Cette dernière
est proche de l’hypertextualité et de l’imitation (Genette, 1982) c’est-à-dire la
reproduction d’un style, d’une manière. En analyse du discours, D.
Maingueneau (1984) a investi la notion de pastiche, confirmant que celui-ci
peut s’opérer sur un genre. Mais le pastiche pour G. Genette (1982) est
associé principalement à une fonction ludique et dans le cadre de notre
étude, la question entre registre satirique et registre ludique reste ouverte,
c’est pourquoi nous nous cantonnerons donc à la notion de « pratique
imitative ». Il n’existe, à notre connaissance, pas de travaux visant à
interroger la pratique imitative en analyse du discours outillée. Xavier
Garnerin (2009), pasticheur, tente de déterminer les méthodes des
pasticheurs qui se situent selon lui « entre analyse et intuition » ce qui dénote
toute la difficulté pour le chercheur à mettre au jour de façon systématique
les liens unissant un texte T2 imitant un texte T1. Nous proposons de mettre
à l’épreuve les outils textométriques pour tenter de percevoir la pratique
imitative des genres.
Notre corpus se compose de cinq titres de presse de tranchées parus entre
1915 et 1918. Nous avons mis en place des variables permettant d’investir les

534

JADT’ 18

genres et les sous-genres (Rastier et Malrieu, 2002). La variable genre scinde
le corpus en deux parties : le genre littéraire (287184 occurrences, 747 articles)
et le genre médiatique (216534 occurrences pour 1005 articles).
Afin d’opérer une étude fine, nous avons aussi catégorisé les textes en sousgenre permettant ainsi de distinguer les romans-feuilletons, les nouvelles, les
poèmes, etc., au sein du genre littéraire et les brèves, filets, dépêches, échos,
faits divers, etc. dans le genre médiatique. L’espace de la contribution ne
nous permet pas d’analyser chacun de ces sous-genres de façon particulière,
c’est pourquoi nous concentrons notre étude sur un sous-genre littéraire, le
roman-feuilleton et un sous-genre médiatique, la dépêche. Afin de mettre en
perspective les résultats obtenus, nous avons constitué un corpus échantillon
donnant à voir 38 dépêches parus entre 1915 et 1918 dans deux quotidiens
nationaux (Le Petit Journal et Le Matin) et trois romans-feuilletons1. Ce corpus
échantillon sera principalement mis à profit pour observer les constructions
syntaxiques et la place des catégories morphosyntaxiques dans les deux sousgenres. Ainsi, la taille des effectifs n’est pas déterminante.
3. L’ancrage dans les moules discursifs médiatiques et littéraires
Dans cette partie, nous montrons comment les textes reprennent les codes
établis dans la presse et dans la littérature à travers l’étude des catégories
morphosyntaxiques et du lexique.
3.1. Les catégories morphosyntaxiques
Le graphique AFC ci-dessous donne à voir la distribution des catégories
morphosyntaxiques (point-ligne en bleu) dans le sous-corpus du genre
littéraire partitionné en sous-genres (point-colonne en rouge). On remarque,
dans cette représentation graphique, que l’axe 1 contribue pour 60,63% à la
structure du graphique. Cet axe semble structuré par le temps des verbes. En
effet, à gauche du graphique on trouve les verbes au présent et au futur alors
qu’à droite, on retrouve les temps du passé (passé simple, imparfait). On
remarque que le roman-feuilleton se situe du côté des verbes au passé,
respectant ainsi les caractéristiques du genre usant des temps du récit.
De plus, si l’on regarde la distribution des verbes en pourcentage dans la
presse de tranchées et la Presse Quotidienne Nationale (PQN), on repère la
proximité dans les temps employés.

1 Entre deux âmes (1912) de Delly paru dans L’Echo de Paris, Le Château noir (1914)
et Confitou (1916) de Gaston Leroux parus dans Le Matin.

JADT’ 18

535

Figure 1. AFC des catégories morphosyntaxique du sous-corpus littéraire partitionné en sousgenre dans le corpus de presse de tranchées.

Figure 2. Graphique représentant pour cent verbes les temps utilisés dans les romans
feuilletons parus dans la presse de tranchées (à gauche) et ceux parus dans la PQN (à droite)

Du côté du genre médiatique, le calcul des spécificités sur les catégories
morphosyntaxiques indique que les dépêches dévoilent un score positif pour
les noms communs (2) alors que les adverbes et les pronoms personnels sont
en sous-emploi (respectivement des scores de -5,4 et -8,7). Ces résultats sont à
mettre en lien direct avec les caractéristiques de la dépêche :
[..] l’auteur de la dépêche se plie à un modèle de représentation qui
doit faire l’économie des ressources stylistiques propres au
littéraire : ni dialogue, ni focalisation interne, ni commentaire sur
l’évènement rapporté. (Kalifa et al., 2011 : 738)
On comprend ainsi le sous-emploi des adverbes et des pronoms personnels,
souvent usités pour introduire un commentaire, alors que l’objectivation de
l’information et l’effacement énonciatif préfèrent les catégories nominales

536

JADT’ 18

aux catégories verbales (Rabatel, 2004). D’ailleurs, on observe sur le
graphique ci-dessous une proximité dans l’emploi des catégories
morphosyntaxiques entre les dépêches de la presse de tranchées et celles de
la PQN.

Figure 3. Graphique qui montre la proportion des grandes catégories morphosyntaxiques
utilisées dans les dépêches parues dans la presse de tranchées (en bas) et celles parues dans la
PQN (en haut)

L’observation de la ventilation des catégories morphosyntaxiques laisse
entrevoir que presse civile et presse de tranchées usent des mêmes catégories
morphosyntaxiques selon les genres.
3.2. Le lexique et les segments répétés
Dans la presse du début XXème, la dépêche débute souvent par une ligne
indiquant le lieu et le jour de l’évènement. Les dépêches de notre corpus de
presse de tranchées suivent cette règle et reprennent cette mise en scène de
l’information. On le voit à travers de nombreux noms de lieux en spécificité
positive comme : « Londres » (4,9), « Paris » (4,2), « Berlin » (2,3), etc. Les
dépêches de la PQN confirment cette tendance avec une moyenne de 4 noms
de lieux par article.
L’escamotage de l’auteur passe d’abord par la mise au point d’un
système d’énonciation à double détente : soit la source de
l’évènement est indiquée – renvoyant toujours à un point de vue
neutre – soit l’évènement est rapporté directement, sans mention
manifeste de la source. (Kalifa et al., 2011 : 738)
Les combattants improvisés journalistes mentionnent souvent une source que
l’on peut percevoir à travers le suremploi des formes graphiques
« communiqué » (score de 16,5) ou « dépêche » (score de 2). De plus, lorsque
l’on s’intéresse aux segments répétés, on remarque que 7 dépêches de
l’Argonnaute débutent par « Communiqué officiel de l’intérieur téléphoné par
[…] ». Du côté de la presse civile on retrouve les formes « dépêche » et
« annonce » justifiant respectivement de 9 et 6 occurrences ainsi
qu’ « Havas » (17 occurrences). Pour le roman-feuilleton dans la presse de

JADT’ 18

537

tranchées, on repère des termes indiquant là aussi le respect de la mise en
scène du roman en « chapitre » (score de 49) et le format feuilleton avec les
termes « suite » (score de 37,4) et « suivre » (22,4).
4. Repérer la pratique imitative
À ce stade de notre étude, nous avons montré la proximité entre presse de
tranchées et PQN mais ni l’étude des catégories grammaticales ni l’étude
lexicale n’a permis de mettre au jour les registres ludiques et/ou satiriques
signes d’une imitation et non d’une inscription dans le genre. Fort de ce
constat, il apparaît nécessaire d’effectuer des recherches qui soient plus
larges que celles du lemme mais plus précise que celles menées jusqu’alors
sur les catégories morphosyntaxiques. Dès lors, une recherche au niveau
syntagmatique semble s’imposer.
4.1. Constructions syntaxiques en suremploi pour les dépêches
Nous avons effectué des recherches pour obtenir les constructions
syntaxiques enchaînant deux catégories morphosyntaxiques sur l’ensemble
du corpus partitionné en sous-genre. Les résultats des premiers syntagmes
en spécificité positive confirment ce que nous avons déjà pu voir : la
catégorie préposition suivie d’un nom propre présente un score de +10,3 et
un retour au texte confirme qu’il s’agit de la présentation du lieu de
l’évènement (« à Londres », « de Paris », etc.). Aussi, on trouve une
construction syntaxique qui induit une construction passive (verbe au
présent suivi d’un verbe au participe passé) indiquant encore l’effacement
énonciatif (Rabatel, 2004). Dans la liste des spécificités positives nous
trouvons la combinaison nom suivi d’adjectif (score de +2,3). La liste éditée
donne à voir 74 syntagmes. Quatorze d’entre eux (soit 19%) ont attiré notre
attention de part, soit l’invraisemblance du dire (« homme volant »,
« provision inépuisable »), soit parce que leur présence ne fait pas sens dans
le genre dans lequel ils se déploient (« bicyclette usagée », « cellules
nerveuses », « chauffage central », « crayon ennemi »). À noter le syntagme
« agence Ivile » jouant de l’homonymie avec « agence civile ». Le retour au
texte permet de mieux comprendre l’usage de ces syntagmes par les
rédacteurs jouant souvent sur le double sens des mots.
Plusieurs saucisses boches (de Francfort) ont été capturées à la
devanture d’un charcutier par un audacieux homme volant.
(Argonnaute, 15 mars 1916)
Le syntagme « saucisses boches » peut renvoyer en 1916 à deux signifiés : le
produit de charcuterie ou le projectile ennemi. C’est sur cette ambiguïté
qu’est basée l’énoncé accentuée par la présence du nom « charcutier » et du

538

JADT’ 18

participe passé « capturées » qui indique chacun une possibilité
d’interprétation différente. Enfin, l’« homme volant » peut être entendu
comme un briguant ayant dérobé de la charcuterie où un homme ayant la
capacité de voler dans les airs et ayant capturé les projectiles ennemis avant
l’impact. Cet exemple dévoile comment les rédacteurs par un registre
ludique créent de la connivence avec le lecteur qui partage les mêmes
références. Un autre exemple permet d’introduire l’idée d’un registre
satirique avec la critique du discours dominant dans l’espace publique.
[…]Paris, 31 avril
[…]Rue du Paon-Blanc (14h.) Paris gronde. Le régime a vécu. Vive
la révolution ! Les bains de la Samaritaine sont en état de siège. Le
syndicat de la Grande Presse n'autorise plus que la parution d'un
bulletin relatant le Communiqué. La censure s'est tranchée la gorge
avec ses ciseaux. L'héroïsme sacré fait battre les cœurs.[…] C'est
l'union sacrée. Concierges, locataires et propriétaires s'embrassent
aux portes des immeubles. (Rigolboche, 10 mai 1917)
L’article remet ici en cause la censure, les festivités parisiennes et fait
également écho aux désaccords entre les propriétaires et les locataires
mobilisés remettant ainsi en cause le discours d’union sacrée tout en
réinvestissant ses dires (Authier-Revuz, 1984). La recherche de syntagmes
nous permet donc d’entrer dans le corpus au niveau du texte et de percevoir
ce qui dans les articles semblent détourner le genre à des fins ludiques et
satiriques.
4.2. Construction syntaxique en suremploi pour le roman-feuilleton
Le roman-feuilleton tient une place importante dans la presse du XIXème
siècle et du XXème (Kalifa et al, 2011). Le conflit ne modifie pas la place de
cette fiction.
La guerre pénètre très rapidement dans le « rez-de-chaussée », et le
roman-feuilleton, sous la forme de récits patriotiques, se mue en
instrument destiné à entretenir et intensifier la mobilisation de la
population en faveur de l’effort de guerre. (Erbs, 2016 : 740)
Voici ce qui est donné à lire aux combattants qui reçoivent et lisent la presse
civile (Gilles, 2013). Nous avons, comme pour les dépêches tenter d’effectuer
une recherche sur les syntagmes de deux occurrences à travers les spécificités
selon les catégories grammaticales. Ces recherches n’ont pas été fructueuses
pour le roman feuilleton. Nous avons donc étendu la recherche à trois

JADT’ 18

539

occurrences. La construction syntagmatique « verbe au passé simple +
déterminant + nom » avec un score de +52 a attiré notre attention. Sur les 130
syntagmes, 24 nous ont interpellés, soit 14% d’entre eux. D’abord, nous
avons repéré des syntagmes qui semblent construits sur des expressions
figées mais où l’un des termes a été modifié comme « fouilla l’horizon » ou
« coupa la pipe ». Nous avons aussi repéré des syntagmes qui ne semblent
pas faire sens comme « revêtit l’ampleur » ou « trancha les jours ».
Alors une colère terrible parut animer l'Armada toute entière.
Proue baissée, les navires foncèrent sur le pirate boche ...
Cependant une première torpille alla frôler par bâbord le vaisseau
amiral ; une deuxième, lancée trop haut, coupa la pipe du
commandant qui flegmatiquement, sortit d'un étui une cigarette
qu'il ajusta au tuyau mutilé de sa pipe. […] (« Krotufex »,
Rigolboche 10/12/1917)
La torpille coupe littéralement la pipe du commandant alors qu’on aurait pu
s’attendre à ce que ce dernier casse sa pipe dans un tel contexte. Cela renvoie
au registre ludique avec le jeu sur l’expression figée mais certainement aussi
au registre satirique offrant ici une critique des romans-feuilletons
patriotiques décrivant des batailles sanglantes sans jamais que le héros ne
succombe. En étudiant les mêmes syntagmes dans le sous-corpus romanfeuilleton dans la PQN, on repère la présence abondante de noms renvoyant
à une partie du corps (« leva les yeux », « prit la main », « secoua la tête »,
« tendit la main ») : sur les dix premiers syntagmes six ont cette
caractéristique. On observe également la présence du corps dans ces
syntagmes dans la presse de tranchées mais ceux-ci semblent une fois encore
surréalistes et usés à des fins ludiques, copiant le genre en le détournant :
« cala les joues », « déchaussa son pied », « frotta la mandibule », « tomba le
torse », etc.
5. Conclusion
Notre contribution avait pour objectif d’investir la pratique imitative avec les
outils textométriques sur un corpus singulier de presse de tranchées mis en
perspective avec un corpus échantillon issu de la PQN. Nous avons pu
montrer dans un premier temps comment les genres sont imités en reprenant
les codes établis dans la presse civile. Pour faire émerger les traces d’une
pratique imitative, il nous a semblé nécessaire d’interroger le corpus, à l’aide
du logiciel textométrique TXM, au niveau syntagmatique. Cette recherche a,
dans le cas de notre étude, permis de faire émerger les registres ludiques et
satiriques ayant court dans la presse de tranchées. Cette presse est un lieu

540

JADT’ 18

énonciatif où l’implicite et la connivence tiennent une place importante au
vue de la censure mais aussi des liens particuliers qui unissent lecteurs et
rédacteurs. Il serait intéressant de voir si, en usant de la même méthodologie,
sur des textes et des genres différents, des résultats similaires peuvent être
observés.
Références
Aron, P. (2013). Le pastiche et la parodie, instruments de mesure des
échanges littéraires internationaux. In Gauvin, L. dir., Littératures
francophones : Parodies, pastiches, réécritures. ENS Éditions.
Audoin-Rouzeau, S. (1986). 14-18, les combattants des tranchées : à travers leurs
journaux. A. Colin.
Authier-Revuz, J. (1984). Hétérogénéité(s) énonciatives. Langages, vol.(73) :
98-111.
Erbs, D. (2016). Le roman-feuilleton français et le serial britannique pendant le
premier conflit mondial, 1912-1920. (thèse de doctorat).
Forcade, O. (2016). La censure en France pendant la Grande guerre. Fayard.
Garnerin, X (2009). Le pastiche, entre intuition et analyse. Modèles
linguistiques, vol.(60): 77-91.
Genette, G. (1982). Palimpsestes. Seuil.
Gilles, B. (2013). Lectures de poilus: livres et journaux dans les tranchées, 19141918. Ed. Autrement.
Heiden, S., Magué, J-P. and Pincemin, B. (2010). TXM : Une plateforme
logicielle open-source pour la textométrie – conception et développement.
In Sergio B. et al. editors, Proc. of JADT 2010 (10th International Conference
on the Statistical Analysis of Textual Data), pp. 1021-1032.
Kalifa, D., Régnier, P., Thérenty, M.-E. et al. (2011). La civilisation du journal :
histoire culturelle et littéraire de la presse française au XIXème siècle. Nouveau
monde éditions.
Maingueneau, D. (1984). Genèses du discours. Madraga.
Malrieu, D & Rastier, F. (2002). Genres et variations morphosyntaxiques.
Traitement automatique des langues vol.(42) : 548-577.
Rabatel, A. (2004). Effacement énonciatif et effets argumentatifs indirects
dans l’incipit du Mort qu’il faut de Semprun. Semen, vol.(17) : 111-148.

JADT’ 18

541

Evolución diacrónica de la terminología y la
fraseología jurídico-administrativa en los Estatutos
de autonomía de Catalunya de 1932, 1979 y 2006
Albert Morales Moreno
Università Ca’ Foscari Venezia / Université de Genève – albert.morales@unige.ch

Abstract
During the first half of 2017, research was carried out at the Institut de
Lingüística Aplicada of the Universitat Pompeu Fabra thanks to a grant from
the Generalitat de Catalunya’s Institut d’Estudis de l’Autogovern in order to
study diachronically the Statutes of Autonomy of Catalonia (EAC acronym,
in Spanish) approved in 1932, 1979 and 2006.
As in other countries and traditions, the negotiation of such an important law
is a challenge in the historical moment in which it occurs, both in legal and
political terms (see Abelló (2007) for the EAC of 1932, Sobrequés (2010) for
the 1979 EAC and Serrano (2008) for the 2006 EAC).
We take lexicometrics as an analytical methodology and the communicative
theory of terminology (Cabré, 1999) as the grounds for our research to study
the use of legal and administrative terminology with respect to the
assignment of competences from a diachronic approach. Specifically, we are
interested in combining the study of repeated segments and the study of
specificities to identify the terms, positions and key institutions of each EAC,
as well as the use of some locutions between 1932 and 2006 in Catalan
statutory discourse.
Resumen
Durante la primera mitad de 2017, se desarrolló una investigación en el
Institut de Lingüística Aplicada de la Universitat Pompeu Fabra para el
Institut d’Estudis de l’Autogovern de la Generalitat de Catalunya (EAC) para
estudiar diacrónicamente los diferentes Estatutos de Autonomía de Cataluña
(EAC), aprobados en 1932, 1979 y 2006.
Al igual que en otros países y tradiciones, la negociación de los proyectos de
regulación de esta escala es un reto en el momento histórico en que ocurre,
tanto en términos legales y políticos (Abelló (2007) para el EAC de 1932,
Sobrequés (2010) para la de 1979 y Serrano (2008) para el de 2006Partimos de
la lexicometría como metodología analítica y de la teoría comunicativa de la
terminología (Cabré, 1999) para estudiar el uso de la terminología jurídica y
administrativa con respecto a la asignación de competencias y materiales a

542

JADT’ 18

partir de un enfoque diacrónico. En concreto, nos interesa combinar el
estudio de segmentos repetidos con el estudio de especificidades para
identificar los términos, cargos e instituciones clave de cada EAC, así como el
uso de algunas locuciones entre 1932 y 2006 en el discurso estatutario catalán.
Keywords: discourse analysis, legal discourse, Catalan statute of autonomy,
repeated segments, terminology, diachronic analysis
1. Introducción
El presente artículo presenta un estudio enmarcado dentro de un proyecto
más amplio de análisis diacrónico de la redacción normativa en catalán. En
dicha investigación, realizada gracias a una financiación posdoctoral del
Institut d’Estudis de l’Autogovern de la Generalitat de Catalunya, se han
estudiado los Estatutos de Autonomía de Catalunya (EAC) de 1932, 1979 y
2006 se han llevado a cabo estudios lexicológicos, estadísticos,
terminológicos, traductológicos y pragmáticos de los distintos EAC.
En esta en concreto, nos hemos centrado a estudiar, desde un punto de vista
terminológico, los segmentos repetidos para evaluar si esta es una estrategia
válida para identificar la evolución de la fraseología especializada relativa a
un ámbito especializado como el Derecho a través del estudio de los
segmentos repetidos específicos de cada EAC. Asimismo, nos proponemos
comparar dichas unidades para ver cuál ha sido la evolución, desde un punto
de vista diacrónico.
Así pues, después de un exhaustivo estudio lexicométrico del corpus, hemos
seleccionado unidades terminológicas especializadas (UTE) relativas al
ámbito jurídico-administrativo que contribuyen a establecer las competencias
de Catalunya en los diferentes EAC, con términos como competència/es,
correspon o atribucioó/ons.
Para dicho análisis, hemos partido de los índices estadísticos que ha arrojado
la exploración lexicométrica desarrollada con Lexico3.6 y como marco teórico
hemos empleado la Teoría Comunicativa de la Terminología (Cabré et al.
1999).
2. Los EAC de 1932, 1979 y 2006
En primer lugar, cabe definir el estatuto de autonomía como una unidad
relativa al ámbito del derecho constitucional que se define como la “norma
institucional básica de las comunidades autónomas” (Diccionario del español
jurídico (DEJ), Real Academia Española).
Numerosos juristas reconocen funcionalmente al EA de las comunidades
como “equivalente a la constitución de un estado miembro de una
federación, porque regula las instituciones autonómicas, establece las

JADT’ 18

543

competencias que deben tener y no puede ser modificado por ninguna otra
ley, ni autonómica ni estatal: sólo puede reformarse por el procedimiento que
el mismo Estatuto prevé, característica propia de las constituciones y no de
las leyes” (Albertí, et al. 2002:111). El Estatuto, pues, “tiene rango de ley
orgánica estatal, forma parte del bloque de la constitucionalidad y está
sometido a unos procedimientos agravados de aprobación y reforma, y sus
previsiones disfrutan de unas garantías reforzadas que no proporciona la
legislación ordinaria” (Pons y Pla 2007:187). En Cataluña, a principios del
siglo XX, con la Mancomunitat, comienza la recuperación del autogobierno.
En el marco de dicha institución, se redacta un primer proyecto de Estatuto
de autonomía, aunque este no se llega a debatir, “porque el 27 de febrero de
1919 se suspendían las sesiones parlamentarias como consecuencia de la
huelga de la Canadenca” (Fontana 2014:327). Debido al desarrollo histórico
convulso de los años posteriores y de la dictadura de Miguel Primo de
Rivera, los proyectos autonomistas se paralizan. Hay que esperar a 1931, la
República, para que se redacte el primer EAC. Aquel texto se debate en las
Cortes en mayo de 1932. Abelló afirma que aquel texto prevé “la inserción de
Cataluña en una república federal” (2007:35) y lo define como “moderado”
(2007:44). A pesar de los recortes que sufre, “se convirtió en una herramienta
útil, que, con la reconquista de las instituciones catalanas de autogobierno,
facultaría una legislación propia, a pesar de que esta fuera limitada” (Abelló
2007:187). La Generalitat de Catalunya asume las competencias durante poco
tiempo, y el 6 de octubre de 1936 el EAC 1932 se suspende parcialmente; con
la llegada de las tropas franquistas a Cataluña, Franco aprueba la ley de
derogación del EAC el 5 de abril de 1938. Con la dictadura de Franco, el
Estado se concibe desde una óptica recentralizadora y, como ya se ha
señalado, se abole la autonomía de las comunidades. Hay que esperar hasta
la muerte del dictador, el 20 de noviembre de 1975, para que, según
Sobrequés (2010: 11), España y Cataluña iniciaran el proceso que tenía que
cambiar su historia: la Transición. Durante esta, se sella el pacto
constitucional de 1978 (la Constitución entra en vigor el 29 de diciembre de
ese año) y se construyen los cimientos jurídicos del Estado autonómico con
un ordenamiento que, a través de los estatutos de autonomía –al menos
desde un enfoque teórico–, se da a los gobiernos autonómicos bastante
autogobierno. El proyecto de redacción comienza el el 8 de septiembre de
1978 y su texto final se aprueba en referéndum el 25 de octubre de 1979.
A principios del siglo XXI, sin embargo, un sector considerable del espectro
social y político catalán percibe el EAC 1979 como un modelo sin recorrido
(la conocida como doctrina Argullol, que supone releer de manera menos
centralista la CE), pero rápidamente se comprueba “hay un número
importante de competencias que, a pesar de ser incluidas en el Estatuto de

544

JADT’ 18

autonomía, no han sido objeto de desarrollo legislativo” (BOPC 2002:89). Por
ese motivo, tras las elecciones autonómicas de 2003, la coalición tripartita
integrada por PSC, ERC e ICV-EUiA inicia en 2004 la tramitación
parlamentaria para la reforma estatutaria. Ello implica una primera
negociación para que se aprobara en el Parlament de Catalunya el 30 de
septiembre de 2005, y una segunda negociación para aprobarlo en las Cortes
Generales (en esa segunda fase, tal y como se expone en Morales (2015), se
producen los cambios más significativos).
El texto final se aprueba en sede parlamentaria el 10 de mayo de 2006, día en
el que el Pleno del Senado aprueba el nuevo estatuto con 128 votos a favor,
125 en contra y 6 abstenciones. El 31 de julio de 2006, Federico TrilloFigueroa y Martínez-Conde (junto con 98 diputados más del PP) presenta el
31 de julio de 2006 un recurso de inconstitucionalidad contra la mayoría de
artículos del nuevo Estatuto (Bosch 2013: 44) porque, entre otras razones,
“aplicaba el término nación en Cataluña, imponía el catalán, establecía una
serie de derechos y deberes que restringían las libertades de los ciudadanos
de Cataluña […] y cuestionaba la unidad de España” (Segura 2013: 217-218).
El 28 de junio de 2010, el Tribunal Constitucional hace pública parte de la
sentencia 31/2010 sobre la constitucionalidad del Estatuto, que declara
inconstitucionales algunas de las partes del EAC 2006. Según numerosos
politólogos e historiadores, esa fecha es clave para la historia política
contemporánea porque “fue el día de la ruptura sentimental con España, el
día en que [muchos catalanes] se convencieron de que Cataluña y los
ciudadanos de Cataluña no tenían cabida en España” (Segura 2013: 32) y para
muchos ciudadanos supuso el salto del autonomismo al independentismo,
sin pasar por el nacionalismo (Segura 2013: 241).
El corpus constituido es, pues, representativo para estudiar diacrónicamente
la evolución del discurso estatutario en lengua catalana a través de los
diferentes Estatutos aprobados a lo largo de la Historia. Para concluir, cabe
añadir que según André Salem (1991:149) este corpus se considera una “serie
textual cronológica”, puesto que son textos lingüística y pragmáticamente
comparables de un arco temporal que permite extraer conclusiones sobre la
evolución del discurso estatutario en lengua catalana de los últimos ochenta
años.
3. Marco teórico y metodológico
Desde la restauración de las instituciones de autogobierno, ha habido
numerosas iniciativas, tanto públicas como privadas, de modernización del
discurso normativo catalán. Cabe destacar el trabajo del Grupo de Estudios
de Técnica Legislativa (GRETEL), de la Dirección General de Política
Lingüística, del TERMCAT, de la Escuela de Administración Pública de

JADT’ 18

545

Cataluña o del Parlament de Catalunya. El modelo que se sigue es el de
Québec, adoptando –y adaptando– las directrices de Spar y Schwab Rédaction
des lois: rendez-vous du droit et de la culture. Según Montolío, se aprovecha para
renovar dicha tradición:
Un caso especial lo constituyen las otras lenguas oficiales del Estado
español (gallego, vasco y catalán). Para estas tres lenguas, la
renovación del lenguaje jurídico ha venido impulsada por una
motivación adicional: la voluntad de recrear una tradición jurídica
truncada tras cuarenta años de prohibición. Entre ellas, cabe destacar
la renovación del lenguaje jurídico catalán.
(Montolío y Albertí 2012:99)
Por ese motivo, los criterios y principios de la que parte la normalización del
lenguaje jurídico catalán son el de economía, el de claridad y el de precisión
en la expresión (DGPL 1999: 7). La falta de estudios lingüísticos exhaustivos
de un componente esencial del discurso normativo catalán como es su
Estatuto de autonomía, ha motivado este trabajo. Este trabajo nace de la
necesidad de analizar combinando la estadística textual y el análisis del
discurso, con una perspectiva diacrónica, los diferentes EAC que ha habido
en vigor hasta la fecha, a partir de una disciplina consolidada: la Lingüística
de Corpus.
De acuerdo con la revisión presentada en Morales (2015:101-175), se han
empleado dichas metodologías para estudiar textos similares. Para garantizar
una selección de las unidades análisis objetiva, pertinente y representativa
basada en criterios estadísticos, nuestro trabajo se desarrolla a partir de la
lexicometría. Dicha escuela ha permitido caracterizar, entre otros, el
vocabulario de personajes sociopolíticos, y de movimientos sociales e
históricos.
Dentro de la lexicometría, nuestra aproximación parte de una aproximación
lexicométrica formalista, puesto que nuestra unidad básica de análisis es la
forma. Posteriormente, hemos normalizado el texto (a partir de metodologías
como las de Arnold (2008:110) y Menuet (2006:157)) para corregir las formas
con errores (gramaticales o de escritura) y evitar que haya conteos
duplicados debido a diferencias mínimas en la ortotipografía. Por último,
hemos insertado en nuestro corpus las marcas estructurales requeridas por
Lexico3.6 para identificar los diferentes EAC.
De las múltiples funcionalidades que incluye el programa, han arrojado
resultados especialmente interesantes el estudio de las concordancias, de los
segmentos repetidos y de especificidades.
Tras la primera exploración lexicométrica, hemos analizado algunos términos

546

JADT’ 18

clave identificados con el análisis de segmentos repetidos para ver si nos
permite caracterizar la fraseología y terminologías propias del ámbito.
4. Análisis
El corpus analizado presenta las principales características lexicométricas
siguientes1:
Identificador
01_1932
02_1979
03_2006

Documento
EAC 1932
EAC 1979
EAC 2006

Ocurrencias
4.242
10.580
40.011
54.833

(7,7 %)
(19,3 %)
(73,0 %)
(100 %)

Formas
1.009
1.766
3.457
4.226

Hápax
606
935
1.546
1.804

Esta parte del análisis se centra en analizar los ya señalados segmentos
repetidos (SR), es decir, las secuencias de formas repetidas con una
frecuencia superior a 5.
La exploración lexicométrica ha arrojado 2.398 segmentos repetidos, pero nos
centraremos en algunos de los más significativos. Su distribución en relación
con su longitud es:
Longitud
2

Secuencias
1282

3

660

4

281

5

98

6

31

7

23

8

10

Ejemplos
de Barcelona
les llibertats
la coordinació
de la Constitució
de seguretat pública
en aquest Estatut
les lleis de Catalunya
a les Corts Generals
el president o presidenta
de conformitat amb les lleis
els poders públics han de
correspon a la generalitat la competència
d ‘ acord amb allò que
sens perjudici d ‘ allò que disposa
el president o presidenta de la generalitat
els poders públics han de vetllar per la
impost sobre la renda de les persones físiques

1 Debido a las diferencias de tamaño obvias, aplicamos, gracias a la profesora
Arjuna Tuzzi, técnicas de análisis estadístico que las tienen en cuenta a la hora de
hacer los cálculos de representatividad y selección esperados, a partir de, entre otros
Tuzzi (2003:128-129) o Van Gijsels, Speelman, y Geeraerts (2005:1).

JADT’ 18

547

Longitud
9

Secuencias
7

10
11

4
11

Ejemplos
en una votació final sobre el conjunt del text
en el diari oficial de la generalitat de Catalunya
correspon a la generalitat la competència exclusiva en matèria de
de l ‘ apartat 1 de l ‘ article 149 de
la carta dels drets i els deures dels ciutadans de Catalunya

De las 20 más frecuentes, por ejemplo, solo cinco tenían interés para nuestro
estudio lingüístico en tanto que unidades con semántica plena, como la
Generalitat, de Catalunya o la competència.
Además de aislar segmentos como de les quals (10), els altres (23), la resta (17),
les quals (18) o en el termini (25) o la seva (57) –que podrían ser interesantes
para investigaciones estilométricas o de atribución de autoría–, a
continuación analizamos algunas de las unidades con una frecuencia
superior.
El sistema ha permitido identificar, por ejemplo, algunos sintagmas relativos
a cargos e instituciones previstos estatutariamente, como les Corts (46) (y les
Corts Generals (33)), Poder Judicial (46), la Comissió Mixta d’Afers Econòmics i
Fiscals Estat-Generalitat (14), l’Agència Tributària de Catalunya (10), el Consell de
Justícia de Catalunya (19), el Govern (50), el President (38), el President o
Presidenta de la Generalitat (26), la Unió Europea (31) i el Parlament de Catalunya
(24). Ha dado buenos resultados, pues, para identificar sintagmas relativos a
unidades muy lexicalizadas como cargos o instituciones.
Uno de los SR más frecuentes es correspon a la Generalitat. Dicho segmento
presenta la distribución siguiente en el corpus:
SR: correspon a la Generalitat
FA
FR (x10000)

EAC 1932
1
4,4

EAC 1979
9
8,5

EAC 2006
144
36,0

Su uso es, como se constata, paradigmático del EAC 2006 (E+11) (presenta
especificidad negativa en los EAC 1932 (E-05) y EAC 1979 (E-07)) y, tal y
como se expone en Morales (2018, en prensa), el ámbito de la atribución
competencial (de la que el segmento repetido es una de las expresiones
lingüísticas más características, al menos en la redacción estatutaria
contemporánea) es de las que más singularidades presenta en el EAC 2006 y
que más cambios ha presentado en el corpus estudiado desde un punto de
vista diacrónico.
Otro de los SR más frecuentes (105 ocurrencias) es la Constitució, que se
reparte de la manera siguiente:

548

SR: la Constitució
FA
FR (x10000)

JADT’ 18

EAC 1932
17
40,1

EAC 1979
42
39,7

EAC 2006
46
11,5

En la mayoría de ocasiones se trata de contextos que hacen referencia a un
artículo concreto de la CE 1978. Son fórmulas que sirven para restringir el
alcance estatutario y establecer una remisión con la Carta Magna española. Es
interesante señalar que el análisis de especificidades denota un uso específico
positivo de dicho SR en los EAC 1932 (E+04) y EAC 1979 (E+07):

JADT’ 18

549

Otras remisiones legislativas que hemos identificado gracias al estudio de los
segmentos repetidos han sido aquest Estatut (96), l’article 149 (de la Constitució)
(26) o el Títol V del mismo EAC (12).
Al tratarse de un corpus legislativo, el análisis también ha permitido
identificar como SR numerosas unidades pertenecientes al lenguaje jurídicoadministrativo que se rigen según el patrón determinante + sustantivo o
sustantivo + adjetivo, como l’article, l’estatut, la legislació, una llei, llei orgànica,
administracions públiques, l’administració, aquest article, comunitat autònoma, de
catalunya, de seguretat, del règim jurídic, disposició addicional, domini públic, dret
civil, el control, el foment, el règim, els àmbits, els articles, els deures, els
mecanismes, els principis, els procediments, els processos, la llei, la llengua, la
majoria, la normativa, la propietat, la salut, les activitats, les actuacions, les
administracions, les administracions públiques, les comunitats, les empreses, les
entitats, les iniciatives, les matèries, les normes, les organitzacions, les polítiques, les
universitats, llei del parlament, polítiques públiques, règim jurídic, serveis públics,
serveis socials, tributs estatals y una llei del parlament.
El aspecto en el que el presente estudio ha proporcionado resultados más
interesantes es, sin lugar a dudas, el relativo a las locuciones más empleadas
en alguno de los EAC, y que en algunos casos se usan de manera
especializada. Algunas de las unidades que hemos estudiado en profundidad
han sido en matèria de/d’, si escau, d’acord amb, en tot cas, en els termes que o sens
perjudici.
El SR en tot cas presenta especificidad en el EAC 2006. Su uso es especifico
positivo del EAC 2006 (E+05) y negativo de los EAC 1932 (E-04) y EAC 1979
(E-03). Sus 95 ocurrencias se distribuyen de la manera siguiente:
SR: en tot cas
FA
FR (x10000)
Esp

EAC 1932
–
–
E-04

EAC 1979
10
9,5
E-03

EAC 2006
85
21,2
E+05

550

JADT’ 18

En la tesis (Morales 2015:398-400), se comprobó que esta es una cláusula
bastante usada en el discurso estatutario catalán contemporáneo y
describimos los usos de dicha cláusula. El libro de estilo del Parlament, una
referencia básica para la redacción estatutaria contemporánea, la define así:
en tot cas
Locució adverbial, equivalent a en qualsevol cas, que es pot emprar
amb valor concessiu o amb el sentit de ‘en tots els casos’. Quan té
aquest sentit, per raons de claredat i precisió, és preferible substituirla per sempre o en tots els casos o, si escau, prescindir-ne.
(SAL 2014:272)
Otra cláusula identificada con el análisis de SR es en els termes, que se
distribuye en el corpus de la manera siguiente:
SR: en els termes
FA
FR (x10000)
Esp

EAC 1932
–
–
E-04

EAC 1979
12
11,3
–

EAC 2006
63
15,7
E+03

El análisis de especificidades indica que su uso es característico positivo en el
EAC 2006, mientras que en los otros no presenta especificidad (EAC 1979) o
bien presenta especificidad negativa (E-04, en el EAC 1932). Al leer
detalladamente las concordancias, se comprueba que aparece sobre todo en
contextos como en els termes que disposin/determini/estableix o similares (en els
termes establerts…):

JADT’ 18

551

Cabe señalar que el EAC 1979 presenta más variedad en relación con el uso
de esta cláusula (las 12 ocurrencias presentan 12 realizaciones diferentes),
mientras que el EAC 2006 se constata menos variación; de los 63 contextos en
los que aparece, las que acumulan más ocurrencias son:
- en els termes que estableix/estableixen/estableixi/estableixin + [les lleis, la
legislació…]: 41
- en els termes que determinin/determinen + [la llei orgànica, la legislació…]: 7
Se comprueba, pues, una fijación más alta. Habría que analizar corpus más
grandes para verificar esta hipótesis, pero esta tendencia a tener un discurso
estatutario más fijado en el EAC 2006 parece confirmarse. Hemos constatado,
sin embargo, que en la mayoría de segmentos repetidos se observa un
comportamiento lingüístico diferente entre los EAC 1932 y 1979 por un lado,
y el EAC 2006 por el otro. Por lo tanto, estos resultados confirman la
hipótesis planteada inicialmente y confirmada con el estudio de distancia
intertextual llevado a cabo por la Dra. Arjuna Tuzzi (Università degli Studi di
Padova).
Otro de los segmentos identificados que equivale a una locución es sens
perjudici, que presenta la distribución siguiente en el corpus:
SR: sens perjudici
FA
FR (x10000)
Esp

EAC 1932
1
2.4
–

Algunas de sus concordancias son:

EAC 1979
28
26.5
E+09

EAC 2006
23
5.7
E-06

552

JADT’ 18

Ya hemos visto en el apartado dedicado al pronombre allò que, en algunos
casos, este SR forma parte de la locución sens perjudici d’allò que. Carles Viver
Pi-Sunyer afirma que (2007:37) el uso de dicha cláusula está relacionado con
la técnica legislativa que se expone a continuación:
L’Estatut d’Andalusia i les propostes de Canàries i de Castella la
Manxa apliquen la mateixa tècnica que l’Estatut de Catalunya,
malgrat que en alguns casos no totes les submatèries que en l’Estatut
de Catalunya es consideren exclusives tenen la mateixa consideració
en els altres tres. Per contra, els estatuts o projectes d’estatut de la
Comunitat Valenciana, d’Aragó, de les Illes Balears i de Castella i
Lleó no identifiquen submatèries exclusives dins d’àmbits materials
en què l’Estat fins ara ha pogut dictar bases, però, en canvi, com hem
vist, en d’altres casos declaren exclusives «sens perjudici»
competències bàsiques estatals, àmbits en els quals es clar que l’Estat
pot establir bases perquè així ho diu expressament la Constitució.
(Viver Pi-Sunyer 2007:37)
Aunque hemos constatado que aparece en el EAC 2006 en 23 ocasiones, la
bibliografía indica que al redactar dicho Estatuto se produjo una innovación
en la técnica legislativa relacionada con el uso de la cláusula en cuestión (sens
perjudici), tal y como afirma Ernest Benach:
Em sembla que [l’EAC 2006] és important, per «la seva nova tècnica
legislativa d’assumpció de competències, que renuncia a la clàusula
del “sens perjudici” i opta per la definició casuística i detallada, dins
de cada àmbit competencial, de submatèries o perfils competencials».
I hi afegeixo jo que a ningú no el podrà sorprendre que, després de
vint-i-cinc anys de patir els perjudicis del «sens perjudici», els
redactors de la Proposta del nou Estatut hagin optat per una tècnica
legislativa moderna que precisa amb claredat l’abast de les
competències de la Generalitat.
(Benach 2006:20)

JADT’ 18

553

Es un cambio, pues, que se comprueba que es fruto de la modernización del
discurso legislativo en redacción estatutaria para obtener en el EAC 2006 un
blindaje competencial más amplio del que se había conseguido con el EAC
1979.
5. Conclusiones
El estudio presentado, como ya se ha señalado, se enmarca dentro de un
proyecto de investigación postdoctoral más amplio realizado durante la
primera mitad del año 2017 en el Institut Universitari de Lingüística Aplicada
de la Universitat Pompeu Fabra gracias a la financiación del Institut
d’Estudis de l’Autogovern de la Generalitat de Catalunya. En dicho estudio
hemos llevado a cabo varios análisis lingüísticos (riqueza léxica, distancia
intertextual, especificidades…) de un corpus de discurso jurídico en lengua
catalana integrado por los Estatutos de autonomía de Catalunya aprobados
en 1932, 1979 y 2006.
Como ya se ha señalado, se han analizado los segmentos repetidos (SR) que
genera el análisis lexicométrico de Lexico3.6. Puesto que los resultados que
generaba eran 2.398 y muchas de las unidades no eran representativas para,
desde el punto de vista del análisis del discurso, estudiar la evolución del
discurso normativo, se ha optado por analizar cualitativamente algunos de
los SR que presentan especificidad en alguno de los subcorpus. Además, el
estudio ha permitido identificar las unidades léxicas y terminológicas más
empleadas en la redacción estatutaria en catalán, así como las instituciones y
cargos que se regulan en dicho EAC.
Hemos identificado que, en el caso de Correspon a la Generalitat es un SR
específico del EAC 2006 que se ha convertido, como ya se ha analizado en
Morales (2018, en prensa) en una de las estructuras formulaicas más
empleadas en la redacción de leyes en catalán. Asimismo, hemos identificado
que, mientras en el EAC 2006 el sintagma la Constitución presenta
especificidad negativa, en los otros dos EAC estudiados sí que se emplea por
encima de las veces esperadas estadísticamente. Habrá que realizar
investigaciones más amplias para entender dicha evolución en la redacción
estatutaria en catalán.
El ámbito en el que la presente investigación ha resultado útil ha sido en la
identificación de locuciones, que en algunos casos se emplean como unidades
de conocimiento especializado (UCE, en terminología de Cabré (1999)). Las
más características, en positivo, del EAC 2006 son en tot cas y en els termes que,
mientras que sens perjudici se tendía a utilizar más en la redacción del EAC
1979. En la bibliografía hemos identificado las motivaciones de dichos
cambios.
Así pues, este estudio ha permitido identificar, cruzando dos análisis

554

JADT’ 18

lexicométricos obtenidos con Lexico3.6 (el de segmentos repetidos y el de
especificidades), algunas unidades lingüísticas (locuciones, términos y
unidades poliléxicas del discurso estatutario y jurídico-administrativo, así
como cargos e instituciones) que han presentado evolución en el discurso
normativo catalán en el periodo 1932-2006. En futuras investigaciones,
ampliaremos el estudio de este tipo de n-grams y ampliarlo a unidades
fraseológicas y estructuras formulaicas, porque parece que podrían aportar
resultados interesantes para describir el discurso estatutario catalán desde
una aproximación cronológica.
Bibliografía
[BOE] Boletín Oficial del Estado. Constitución española. Madrid: Agencia
Estatal Boletín Oficial del Estado, 1978.
[BOPC] Butlletí Oficial del Parlament de Catalunya. "Moció 187/VI del
Parlament de Catalunya, sobre l'exercici de l'autogovern." Butlletí Oficial
del Parlament de Catalunya. 366. Barcelona: Parlament de Catalunya, 2002.
89.
[DGPL] Direcció General de Política Lingüística. Criteris de traducció de textos
normatius del castellà al català. Barcelona: Generalitat de Catalunya.
Departament de Cultura, 1999.
[SAL] Serveis d’Assessorament Lingüístic. Llibre d’estil de les lleis i altres textos
del Parlament de Catalunya. Barcelona: Parlament de Catalunya, 2014.
Abelló Güell, Teresa. El debat estatutari del 1932. Barcelona: Parlament de
Catalunya, 2007.
Albertí, Enoch, et al. Manual de dret públic de Catalunya. Barcelona: Generalitat
de Catalunya. Institut d'Estudis Autonòmics, 2002.
Arnold, Edward. "Le sens des mots chez Tony Blair (people et
Europe)." JADT 2008: actes des 9es Journées internationales d’Analyse
statistique des Données Textuelles, Lyon, 12-14 mars 2008: proceedings of 9th
International Conference on Textual Data statistical Analysis, Lyon, March
12-14, 2008. Eds. Heiden, Serge, Bénédicte Pincemin and Liliane
Vosghanian. Lió: Presses Universitaires de Lyon, 2008. 109-19.
Benach, Ernest. L'Estatut: una aposta democràtica i moderna: Barcelona, 7 de
novembre de 2005. Barcelona: Parlament de Catalunya, 2006.
Bosch, Jaume. De l'Estatut a l'autodeterminació: esquerra nacional, crisi
econòmica, independència i Països Catalans. Barcelona: Base, 2013.
Cabré Castellví, M. Teresa. La terminología. Representación y comunicación.
Elementos para una teoría de base comunicativa y otros artículos. Sèrie
Monografies, 3. Barcelona: Institut Universitari de Lingüística Aplicada,
Universitat Pompeu Fabra, 1999.
Fontana, Josep. La formació d'una identitat. Una història de Catalunya. Vic:

JADT’ 18

555

Eumo Editorial, 2014.
Lamalle, Cédric, et al. Manuel d'utilisation. Lexico3 (Version 3.41 - Février 2003).
París: SYLED–CLA2T. Université de la Sorbonne nouvelle–Paris 3, 2003.
Menuet, Laëtitia. "Le discours sur l’espace judiciaire européen: analyse du
discours et sémantique argumentative." Université de Nantes, 2006.
Montolío, Estrella, and Enoch Albertí. Hacia la modernización del discurso
jurídico: contribuciones a la I Jornada sobre la Modernización del Discurso
Jurídico Español. Barcelona: Publicacions i Edicions de la Universitat de
Barcelona, 2012.
Morales Moreno, Albert. "Estudi lexicomètric del procés de redacció de
l’Estatut d’Autonomia de Catalunya (2006)." Tesi doctoral no publicada.
Universitat Pompeu Fabra, 2015.
Pons, Eva, and Anna M. Pla. "La llengua en el procés de reforma de l'Estatut
d'autonomia de Catalunya." Revista de Llengua i Dret.47 (2007): 183-226.
Real Academia Española. Consejo General del Poder Judicial.
"[DEJ] Diccionario del español jurídico." Madrid.
Salem, André. "Les séries textuelles chronologiques (1)." Histoire et mesure.VI1/2 (1991): 149-75.
Salem, André, M. Teresa Cabré, and Lydia Romeu. Vocabulari de la
lexicometria: català, castellà, francès. Barcelona: Centre de Lexicometria,
Divisió de Ciències Humanes i Socials, 1990.
Segura, Antoni. Crònica del catalanisme: de l'autonomia a la independència.
Barcelona: Angle Editorial, 2013.
Sobrequés, Jaume. L'Estatut de la Transició: l'Estatut de Sau (1978-1979).
Barcelona: Parlament de Catalunya, 2010.
Tuzzi, Arjuna. L’analisi del contenuto. Introduzione ai metodi e alle tecniche di
ricerca. Roma: Carocci, 2003.
van Gijsel, Sofie, Dirk Speelman, and Dirk Geeraerts. "A Variationist, Corpus
Linguistic Analysis of Lexical Richness." Proceedings of the Corpus
Linguistics 2005 Conference, July 14-17, Birmingham, UK 1.1 (2005): 1-16.
Viver Pi-Sunyer, Carles. "Les competències de la Generalitat a l'Estatut de
2006: objectius, tècniques emprades, criteris d'interpretació i comparació
amb els altres estatuts reformats." La distribució de competències en el nou
Estatut. Eds. Viver i Pi-Sunyer, Carles, et al. Barcelona: Institut d’Estudis
Autonòmics, 2007. 13-52.

556

JADT’ 18

Comment penser la recherche d’un signe pour une
plateforme multilingue et multimodale français
écrit / langue des signes française ?
Cédric Moreau
Grhapes EA 7287 - INS HEA - UPL – cedric.moreau@inshea.fr

Abstract 1 (in English)
This article examines the access to the signs in French Sign Language (LSF)
within a corpus taken from the collaborative platform Ocelles, from a
multilingual French bijective/LSF perspective. There is currently no
monolingual dictionary in SL, so deaf users must necessarily master the
written language of the country to access SL contents. Most of the available
tools are based on a hypothetical conceptual relationship of equivalence
between the signs of SL and the words of the dominant vocal languages. This
approach originates in works that ask deaf speakers to translate a lexema
outside the context of the spoken language into the signed language. This
corpus is subsequently used for an inventory of minimal pairs, in which
configurations, locations and movements are widely represented. This
approach is thus the anchorage point for a phonological hypothesis of SL in
which the previous equivalence ‘sign – word’ is dominant and decisive in the
conception of dictionaries. This study lies within a completely different
paradigm, that of the semiotical model which stems from the description of a
typology and the identification of the three main transfer structures (size and
form, situational, and personal). According to Cuxac, the signer can thus
‘make visible’ the experience by relying on the maximal resemblance
sequence of signs/experience, or use the lexical unit without resemblance
with the referent. This model, which is also integrative, therefore takes into
account the diachronic link existing within language under the influence of
pressures between transfer structures and lexical units. The morphemic
approach to the study of lexical units is in this case legitimate since their
compositionality does not rely on strict phonology but, in the first place, on
complex morphology. First of all, we shall present our paradigm and the
origins of the Ocelles multilingual and multimodal platform (written, oral,
and signed languages), out of which our French written/LSF corpus is built.
We will then describe a process likely to enable users to search for an LSF
signifier and to relate this result to that of the corresponding written French
signifier.

JADT’ 18

557

Abstract 2 (in French)
Cet article interroge l’accès aux signes de la langue des signes française (LSF)
d’un corpus dans une perspective multilingue bijective français / LSF à partir
de la plateforme collaborative Ocelles. Actuellement il n’existe pas de
dictionnaire monolingue en LS, les utilisateurs sourds doivent donc
nécessairement maîtriser la langue écrite du pays pour accéder à un contenu
en LS. La plupart des outils à disposition s’appuient sur une hypothétique
relation d’équivalence conceptuelle entre les signes des LS et les mots des
langues vocales dominantes. Cette démarche prend sa source dans des
travaux qui interrogent les locuteurs sourds en leur demandant de traduire
un lexème hors contexte de la langue vocale en langue signée. Ce corpus est
ensuite utilisé dans l’élaboration d’un inventaire de paires minimales, dans
lequel les configurations, leurs emplacements et leurs mouvements sont
largement représentés. Cette approche est ainsi le point d’encrage d’une
hypothèse phonologique des LS dans laquelle l’équivalence « signe – mot »
précédente est dominante et déterminante dans l’élaboration de
dictionnaires. Notre étude s’inscrit dans un tout autre paradigme, celui du
modèle sémiologique qui prend ses origines dans la description d’une
typologie et de la mise en évidence des trois structures de transfert
principales (de taille et de forme, situationnel et personnel). Selon Cuxac, le
signeur peut ainsi « donner à voir » l'expérience en s'appuyant sur la
ressemblance maximale séquence de signes/expérience, ou utiliser l’unité
lexicale sans ressemblance avec le référent. Ce modèle, également intégratif,
prend donc en considération le lien diachronique qui existe au sein de la
langue sous l’influence de pressions entre structures de transferts et unités
lexicales. L’approche morphémique pour l’étude des unités lexicales est dans
ce cas légitime, leur compositionnalité ne relevant pas d’une phonologie au
sens strict mais bien, en premier lieu, d’une morphologie complexe.
Nous exposerons tout d’abord notre paradigme et les origines de la
plateforme multilingue et multimodale (langues écrites, orales et signées)
Ocelles sur laquelle notre corpus français écrit / LSF se constitue. Nous
décrirons ensuite un processus susceptible de permettre aux utilisateurs la
recherche d’un signifiant en LSF et de lier ce résultat à celui du signifiant en
français écrit correspondant.
Keywords: Collaborative platform, Multilingualism, Multi-modality, French
Sign Language, LSF, deaf, Signs research, Semiological model, Ocelles
1. Introduction
Lorsqu’un locuteur de la langue des signes souhaite accéder à une ressource
dans sa langue, notamment pour rechercher une définition dans un

558

JADT’ 18

dictionnaire de langue des signes (LS), il est confronté à deux obstacles. Le
premier repose sur le fait que très peu d’outils présentés comme étant des
dictionnaires numériques de langue des signes ne sont que des lexiques.
Parmi 105 sites répertoriés sur le web, une majorité utilise le qualificatif
« dictionnaire », or seulement 17 d’entre eux présentent des définitions
écrites. Parmi ces 17, uniquement 7 donnent des définitions en LS. La
quantité de dictionnaires en LS est donc extrêmement faible. De plus le
nombre de définitions ne dépasse pas 5 000, nous sommes ainsi très éloignés
des 135 000 proposées par le dictionnaire Larousse en ligne (Moreau, 2012).
Le second obstacle porte sur la difficulté, pour l’utilisateur sourd d’accéder
aux contenus mêmes d’un dictionnaire de ce type. En effet, nous avons
constaté que dans la grande majorité des cas, les entrées proposées sont
étroitement liées à la connaissance de la langue écrite du pays. Un prérequis
nécessaire est donc la maîtrise de cette langue, ce qui constitue un obstacle
majeur pour les personnes sourdes qui ont la LS pour langue première et la
langue écrite, souvent mal maîtrisée, comme langue seconde. Parmi les 7
sites précédemment évoqués, seulement 2 offrent une entrée via les
paramètres linguistiques de la LS (Moreau, 2012).
Cette question prend également un écho particulier lorsque nous
interrogeons le mode de transmission des LS. Il ne s’agit pas d’un mode de
transmission héréditaire, puisqu’environ 95 % des sourds ont des parents
entendants qui, pour la majorité, ne pratiquent pas la LS. L’apprentissage de
la langue a donc lieu dans des contextes variés, à tout âge, souvent sans la
référence fixe d’un adulte proche.
Le pourcentage restant (environ 5 %) est donc constitué de sourds de parents
sourds. Des parents qui, eux-mêmes pour la plupart, font partie de la
catégorie précédente issus de familles entendantes. Seule 0,02 % de la
population sourde signante est en effet composée d’une généalogie comptant
trois générations successives de sourds signeurs. La norme d’apprentissage
des LS ne peut donc pas être comparée à celles des entendants (Cuxac et
Pizzuto, 2010).
En outre, la langue des signes française (LSF), marquée par plus d’un siècle
d’interdiction comme langue d’enseignement, n’est reconnue comme langue
de la République que depuis 2005. C’est dans ce contexte qu’est né le projet
collaboratif multilingue et multimodale Ocelles1, qui ambitionne de définir
tous les concepts, dans tous les champs de la connaissance et dans toutes les
langues (écrites, orales ou signées) (Moreau, 2017).

1 https://ocelles.inshea.fr Projet sous l’égide et avec l’aide de la Délégation
générale à la langue française et aux langues de France (DGLFLF) et du ministère de
l’Éducation nationale.

JADT’ 18

559

2. Affrontement de deux paradigmes
2.1. Une hypothèse phonologique des LS
Susan Goldin-Meadow a mis en évidence, à partir d’une étude basée sur la
communication préscolaire, entre petits enfants sourds et leur entourage
entendant, la création de gestes appelés « home signs » (Goldin-Meadow et
Mylander, 1991) (Goldin-Meadow, 2003). Pour tenter de rentrer en
communication avec leur entourage, ces enfants les réalisent dans l’univers
perceptivo-pratique. Ces productions permettent de faire l’hypothèse de
stabilisations conceptuelles pré linguistiques, à la différence des productions
d’enfants entendants du même âge, pour lesquels le lien entre la langue et ces
savoirs perceptivo-pratiques n’existe pas. Une fois scolarisé, ces enfants
entrent ensuite en contact avec une langue des signes institutionnalisée.
Selon Golwin-Meadow dans la mesure où les formes signifiantes des langues
des signes institutionnalisées ont un statut phonologique, les composants des
« home signs » de l’enfant perdraient alors leur statut de morphèmes pour
devenir des équivalents de phonèmes. Cette hypothèse peut être envisagée
comme point de départ à l’affrontement de deux paradigmes. L’iconicité est
alors comparée à de la gestuelle co-verbale illustrative, reléguée au rang de
pantomime en dehors de tout phénomène linguistique.
C’est dans ce paradigme que s’inscrivent la plupart des « dictionnaires » de
langues des signes actuellement. Leurs entrées sont majoritairement définies
à partir d’une hypothétique équivalence conceptuelle entre les mots des
langues vocales dominantes et celles des unités lexématiques (UL) des
langues signées. (Fusellier-Souza, 2006). L’origine de cette méthodologie
prend racine dans des travaux qui interrogent les locuteurs sourds en leur
demandant de traduire un lexème hors contexte de la langue vocale en
langue signée. Ce corpus est ensuite utilisé dans l’élaboration d’un inventaire
de paires minimales, dans lequel les configurations (formes de la main), leurs
emplacements et leurs mouvements sont largement représentés (Klima et
Bellugi, 1979).
2.1. Une hypothèse morphémique des LS
Notre travail s’inscrit dans un tout autre paradigme dans lequel la
conséquence de la surdité n’est plus un simple effet de changement de canal.
La possibilité de dire et de montrer étant le seul fait du canal visuo-gestuel a
conféré aux langues des signes une architecture différente de celle des
langues vocales.
Selon Cuxac (Cuxac, 2000), deux stratégies discursives d'énonciations
coexistent en LSF. Le signeur via le canal visuo-gestuel, choisit de dire sans
montrer ou bien de dire en montrant. Il peut ainsi « donner à voir »
l'expérience en s'appuyant sur la ressemblance maximale séquence de

560

JADT’ 18

signes/expérience, ou utiliser l’UL sans ressemblance avec le référent. Le
modèle sémiologique (Cuxac et Pizzuto, 2010) prend ses origines dans la
description d’une typologie et dans la mise en évidence des trois structures
de transfert principales :
 les volumes des entités (transferts de taille et de forme (TTF)),
 les déplacements d’actants par rapport à des locatifs stables, à
l’image d’un environnement en quatre dimensions (les trois de
l’espace et le temps) recréé devant le locuteur (transferts situationnels
(TS)),
 l’entité souhaitée par le locuteur, qui devient alors cette entité
(transferts personnels (TP))
(Cuxac, 2000; Sallandre, 2003). Des expériences imaginaires ou réelles sont
ainsi anamorphosées par le locuteur.
Le modèle sémiologique, prend donc en considération le lien diachronique
qui existe au sein de la langue sous l’influence de pressions entre structures
de transferts et UL. Lien qui se retrouve parfois dans l’étymologie de
certaines des UL. L’approche morphémique pour l’étude des UL est dans ce
cas légitime, leur compositionnalité ne relevant pas d’une phonologie au sens
strict mais bien, en premier lieu, d’une morphologie complexe.
Lors de la réalisation d’un signe (transfert ou UL), tout le corps du locuteur
prend une valeur sémantique via une organisation des éléments
morphémiques qui le composent (regard expression faciale, posture,
orientation du visage, configuration, le mouvement, l'emplacement (Stokoe
et al., 1965), l'orientation (Friedman, 1977; Liddell, 1980; Moody, 1980; Yau,
1992).
3. Éléments prégnants dans la recherche d’un signe pour une plateforme
multilingue et multimodale français écrit / LSF
3.1. Contexte d’une recherche d’un signe dans un corpus bilingue langue
écrite/LS
Le projet collaboratif Ocelles permet de relier au fil des contributions, des
définitions de concepts à plusieurs signifiants qu’ils soient sous formes
textuelles, orales ou signées. Les entrées ne sont pas contraintes par la langue
d’origine et l’architecture se déploie au fur à mesure des contributions des
usagers. L’entrée textuelle peut donc prendre la forme, d’un mot ou d’une
expression dans le cas où l’origine du dépôt provient d’une structure de
transfert de la langue des signes. La réflexion actuelle porte donc sur le type
d’indexation possible des signes indispensable au processus de recherche
d’un signe dans le cadre d’un corpus bilingue langue écrite/LS.

JADT’ 18

561

3.2. Automatisation de l’indexation
L’indexation d’un signe se fait via l’entrée textuelle correspondante. Il
n’existe pas aujourd’hui d’indexation automatique de corpus collaboratif
dynamique de signes des LS qui pourrait servir de base pour un moteur de
recherche d’une UL ou d’un transfert directement à partir des paramètres
linguistiques des LS. La nature même du signal vidéo, très complexe à
analyser ne permet pas l’indexation automatique. Outre les pertes
d’informations tridimensionnelles liées aux projections de l’espace 3D à celui
2D de la vidéo, ce travail nécessiterait des outils fins d’analyses et de
reconnaissances, des différents composants corporels, intervenant en
parallèle, à des échelles spatiales et temporelles très différentes, mis au point
pour des langues vocales, linéaires et mono source mais par pour les LS
(Braffort et Dalle, 2012).
3.3. Situation actuelle et limite
Aujourd’hui l’entrée à partir des paramètres linguistiques des signes des LS
se fait majoritairement à partir de la configuration. Sur les 105 sites
répertoriés qui proposent des signes en LS seuls 18 offrent une possibilité
d’accéder directement à un signe à partir des paramètres linguistiques de la
langue des signes, sans recours à une langue écrite. Sur ces 18, 17 proposent
une entrée à partir de la configuration (le nombre de ces entrées manuelles
varie d’ailleurs de 9 à 211 en fonction des sites), 6 proposent une entrée à
partir du mouvement, 10 à partir de l’emplacement et 1 pour la symétrie,
l’image labiale et la mimique faciale (Moreau, 2012). Cette indexation
phonologique des LS, avec un tel écart dans le nombre envisageable de
configurations de 9 à 211 par exemple, interroge la gestion de l’erreur
potentielle du locuteur qui recherche un signe qu’il aurait perçu en discours
(ce qui est le cas dans la majorité des cas, compte tenu du caractère oral des
LS). En outre, sur un choix entre 211 configurations, le locuteur a une chance
sur 211 de choisir la bonne ou 210 risques sur 211 de se tromper…
3.4. Description et critères de recherche
L’indexation ne peut donc reposer uniquement sur une approche strictement
phonologique et doit tenir compte de la gestion des erreurs possibles. Notre
hypothèse repose sur une prégnance pour le locuteur de certaines unités
linguistiques dans une approche morphémique mises en jeux lors de la
formulation d’un signe (Moreau, 2012).
Notre approche est fondée sur le principe d’une indexation collaborative qui
permet de rendre compte des perceptions des locuteurs. Le principe est basé
sur le processus suivant :
 prise en compte du ou des type(s) de transfert(s) utilisé(s)

562

JADT’ 18

(TS / TP / TTF) dans la réalisation d’un signe à moins que l’unité
lexématique puisse éventuellement trouver son origine dans l’un de
ces transferts,
 itération dans le choix d’images clés à partir desquelles repose une
description des unités linguistiques prégnantes (Thom, 1988),
 une description plus fine des unités retenues est ensuite proposée
Si aujourd’hui les structures linguistiques ne peuvent être admises comme
familières à l’ensemble des contributeurs, leur prise en compte ne peut être
ignorée. Deux approches sont envisagées. Une première inhérente à l’objectif
premier de la plateforme, repose sur la proposition d’une définition de ces
concepts afin de familiariser progressivement les locuteurs à leurs usages.
Une succession d’anamorphoses possibles de plus en plus précises est
ensuite proposée. Cette approche est cohérente avec l’utilisation de n’importe
quel outil pour lequel un minimum de prérequis sont nécessaires, à l’image
de l’alphabet pour un dictionnaire. Une seconde approche repose sur la prise
en compte de ces lacunes en inscrivant le processus dans un continuum, qui
permet une possible contribution basée sur la sélection puis la description
d’images représentatives du signe du point de vue de l’usager. C’est donc
l’ensemble des descriptions macro-microscopiques, de chaque contributeur
qui sert de base à la pondération des unités linguistiques prégnantes. Ces
données seront ensuite réutilisées comme critère de recherche d’un signe.

JADT’ 18

563

Conclusion ADT et visualisation, pour une nouvelle
lecture des corpus Les débats de 2ème tour des
Présidentielles (1974-2017)
Jean Moscarola1, Boris Moscarola2
1 Université Savoie Mont Blanc, 2 Le Sphinx-Développement

Abstract 1
The progress of textual data analysis leads from a statistical and lexical
description of corpora to their semantic analysis. The software thus offers the
qualitative researchers the opportunity to feed their interpretations on the
basis of substitutes that summarize them or to code them automatically.
Finally, data visualization offers the reader an experience of the corpus
creating the conditions for a critical control. This approach is illustrated on
the analysis of the 2nd round debate in the presidential election conducted
with DataViv the new Sphinx module.
Abstract 2
Les progrès de l’analyse de données textuelles conduisent d’une description
statistique et lexicale des corpus à leur analyse sémantique. Les logiciels
offrent ainsi au chercheur qualitatif la possibilité de nourrir leurs
interprétations sur la base de substituts qui les résument ou de les coder
automatiquement. Enfin la datavisualisation offre au lecteur une expérience
du corpus créant les conditions d’un contrôle critique. Cette approche est
illustrée sur l’analyse des débats de 2ème tour à l’élection présidentielle
effectué avec DataViv le nouveau module de Sphinx.
Keywords: Analyse de discours, statistique lexicale, analyse sémantique,
data visulaisation, logiciel Sphinx
1. Introduction
L’ADT, née d’une rencontre entre la recherche littéraire et la statistique, passe
de l’étude de grandes œuvres à celle des médias de masse et de la
communication politique. Avec le big data et le web sémantique elle
s’enrichie des nouveaux outils de l’IA en abordant tous types de corpus.
Dans les sciences humaines, l’analyse de contenu s’est développée à
l’articulation de la recherche qualitative pure et des méthodes quantitatives
mais sans rapport explicite avec l’ADT. Ce papier s’adresse aux chercheurs et
chargés d’étude qualitative qui restent réticents à l’usage des outils de l’ADT.
Il s’appuie sur l’étude du corpus des débats de 2ème tour à l’élection

564

JADT’ 18

présidentielle et utilise la nouvelle application Dataviv de Sphinx pour
illustrer une nouvelle expérience de lecture.
2. Les méthodes et les techniques
1.1 Des humanités numériques à l’intelligence artificielle
L’outil informatique a depuis longtemps été utilisé pour informatiser les
grands corpus de la littérature (Frantext). C’est ainsi qu’apparaissent dans les
années 60 les humanités numériques (Burdick) et l’utilisation de la statistique
pour caractériser le style de grands auteurs ou leur attribuer des œuvres
anonymes (Muller). Puis dans les années 70 des statisticiens fondent le
courant français de l’analyse de données textuelle qui trouve un écho avec le
structuralisme et l’analyse de discours (Beaudouin). Dans les années 60 aux
Etats Unis une autre voie était ouverte avec la construction de thésaurus
informatisés (Stone) utilisés pour coder le contenu des media de masse.
Ces approches sont à l’origine des techniques que nous allons exposer. Elles
sont enrichies dans les années 2000 par les progrès de l’ingénierie
linguistique, et du traitement automatique des langues (Veronis).
2.1 Analyse de données textuelle
L’examen statistique des textes a évolué du décompte des mots à l’étude de
leurs associations. Dans la tradition des concordanciers, la voie est ouverte à
la recherche des segments répétés (Lebart), émaillant les discours politiques
(Marchand) ou publicitaires (Floch). L’informatique graphique, les cartes
cognitives (Eden) et les nuages de mots donnent une représentation visuelle
de ces concordances. L’influence des contextes et la recherche des spécificités
lexicales complète des descriptions globales (Brunet, Lebart)
Les méthodes d’analyses factorielles (Benzecri) font la synthèse entre la
rigidité des segments répétés et le désordre des nuages de mot. En dégageant
des d’affinités entre termes fréquemment associés, elles offrent une analyse
structurale des textes popularisée par les cartes factorielles disposant les
univers lexicaux révélateurs des thèmes du texte. A l’analyste d’en faire une
lecture sémiotique.
De manière duale à la mise en évidence des univers lexicaux, Reinert propose
le regroupement des unités de signification (réponses, phrases ou séquence
de mots…) pour créer une partition à partir de plusieurs analyses factorielles
utilisées pour progressivement définir des classes homogènes. Cette
méthode, mise en oeuvre avec le logiciel ALCESTE qui lui a donné son nom,
a été reprise et enrichie par d’autres logiciels (IRAMUTEC, SPHINX).
On retrouve des approches voisines chez les anglo-saxon. ‘L’analyse
sémantique latente’ (Landauer) déplace l’attention de l’observation des
cooccurrences vers la recherche de dimension latentes mesurées par les axes

JADT’ 18

565

factoriels. La théorie du cadrage (Frame Analysis) formulée par Goffman
interprète l’usage de certains mots clés et leurs relations comme des
« conceptualisations diffuses » Ces cadres sont une manière d’interpréter les
univers lexicaux.
2.2 Linguistique
A l’origine les logiciels ne repéraient que les formes graphiques (séquence de
lettre ne comportant aucun séparateur) sans parvenir à différencier singulier
et pluriel ou les différentes flexions d’un même verbe.
La lemmatisation a représenté un grand progrès en remplaçant les différentes
graphies d’un mot par son lemme : L’infinitif pour les verbes, le masculin
singulier pour les noms et adjectifs. Puis l’analyse des propriétés
morphosyntaxiques conduit à distinguer les ‘mots pleins’ selon leur statut
grammatical. Les substantifs, donnent les objets des textes ou des discours,
les adjectifs les appréciations et opinions, les verbes renvoient aux actions. La
recherche des syntagmes permet d’identifier les expressions propres au
domaine, formes les plus expressives des concordances (Mayaffre).
2.3 Sémantique
La sémantique s’intéresse au sens en passant du niveau des signifiants à celui
des signifiés.
Malgré leur intérêt théorique, les travaux de linguistique générale n’ont pu
déboucher sur les applications qui marquent, avec la linguistique de corpus,
le véritable essor de l’analyse sémantique.
L’idée est de modéliser les connaissances de domaines particuliers comme
des signifiés définis par l’ensemble des signifiants qui s’y rattachent
(Saussure).
Dès les année 60, « General Inquirer» développe à Harward des ressources
informatiques permettant de coder automatiquement le contenu des médias.
Ces dictionnaires sont toujours accessibles. WordNet® grande base de
données lexicales de l’anglais développée par l’université de Princeton
généralise cette approche en améliorant l’efficacité des dictionnaires par
l’usage de réseaux sémantiques. WordNet peut être considéré comme un
thésaurus généralisé reflet des corpus sur lesquels il est construit. Ces idées
sont reprises par les moteurs sémantiques.
Dans les années 2000, l’ingénierie linguistique et le traitement automatique
des langues Normier) dépasse l’approche purement lexicale en spécifiant les
thésaurus (Da Silva), par des ontologies(Grubert) et réseaux
sémantiques(Godard). Le thésaurus définit l’arborescence des catégories
conceptuelles : les signifiés. Les ontologies sont constituées de la liste des
mots qui documentent ces catégories : les signifiant. Les réseaux sémantiques

566

JADT’ 18

précisent l’affectation des termes aux catégories du thésaurus en fonctions
des liens constatés à partir de corpus de référence : les référents.
Avec l’essor des réseaux sociaux il devenait enfin primordial enfin
d’appréhender la tonalité de messages susceptibles de faire ou défaire les
réputations. Ainsi dans les années 2010 apparaissent des applications de
traitement automatique des langues pour synthétiser les avis et les opinions
du web. Elles ont acquis leur notoriété sous l’appellation de
‘sentiment analysis’ ou ‘d’opinion mining’ (Thelwall). Ces analyses
complètent la reconnaissance des catégories du thésaurus en évaluant les
textes selon leur orientation positive ou négative mesurée sur une échelle
assimilable à une mesure de l’opinion.
L’Analyse de Données textuelles a ainsi évolué d’une approche descriptive
statistique et lexicale à une approche sémantique fondée sur une
modélisation des connaissances. Rendue très accessible par les logiciels
(Boughzala) , elle présente une ressource pour la recherche qualitative ce que
nous allons illustrer sur un exemple de corpus politique.
3. Contributions de l’ADT à l’analyse de corpus.
3.1 l’exemple des débats de 2ème tour
L’analyse des discours politiques est un classique de l’ADT (Marchand,
Mayaffre). Leurs transcriptions analysées à différents niveaux, (les locuteurs,
les tours de paroles ou les phrases) sont traitées comme des données pour
révéler le style, les structures lexicales, les idées et les opinions qui les
caractérisent. Le corpus des 7 débats de deuxième tour couvre de 1974 à 2017,
43 ans de vie politique. Il est analysé à l’adresse suivante
https://www.sphinxonline.net/debats/1974-2017/analyse.htm, qui présente
de manière détaillée ce dont nous donnons qu’un aperçu dans cet article.
Notre but est d’illustrer les méthodes qui viennent d’être évoquées et de
discuter leur pertinence pour la recherche qualitative. Le lecteur est invité à
en faire lui-même l’expérience plus riche que l’aperçu qui suit :
-Les propos des candidats sont précis : les articles définis sont présents dans
2 phrases sur 3. Les embrayeurs ‘je’ et ‘vous’ sont utilisés de manière plus
fréquente que ‘nous’
-Les expressions « premier ministre », « assemblée nationale » « pouvoir
d’achat », « général de gaulle » « milliard d’euro » dominent sur l’ensemble
de la période.
-La carte des univers lexicaux montre une opposition entre l’évocation de la
vie politique d’une part et les termes de l’économie et de la société d’autre
part.
-Sur les 11 thèmes identifiés par la classification automatique, les thèmes
‘Gouvernement-Majorité’, ‘Pays, Français’, ‘Année Nucléaire’, ‘Entreprise

JADT’ 18

567

Salarié’ arrivent en tête.
-Les principaux concepts reconnus par le thésaurus de l’application utilisée1
sont « Vote » « Civilisation » « Emploi et salaire » « Politique fiscale »
« Citoyenneté »…
-La tonalité des propos est neutre pour la moitié des interventions, pour le
reste les prises de position positives sont un peu plus fréquentes.
La référence aux candidats et aux périodes complète la description globale.
-A chacun son style : Jospin Royal et Mitterrand se distinguent par l’usage de
‘je’ ; Chirac par le ‘nous’ plus collectif et Marine Le Pen interpelle son
débateur (vous) à moins qu’elle ne s’adresse à l’audience. Macron fait preuve
de l’usage le mieux équilibré.
-Les mots clés sur représentés dans chaque période marquent bien le
changement de siècle : ‘politique’, ‘gouvernement’ ‘problème’ au XXème,
‘entreprise’ ‘emploi’ ‘européen’ au XXIème.
-Les catégories thématiques de la classification lexicale sont associées à des
groupes de candidats : Sarkozy, Royal et Hollande développent les thèmes
‘Entreprise, Salarié’, ‘Loi’, ‘Crise Priorité’ ‘Pouvoir Président’. Mitterrand et
Giscard d’Estaing, ‘Socialiste Communiste’, ‘Gouvernement Majorité’,
Macron et Le Pen ‘Chômage, Emploi’, ‘Français, Pays’
-Enfin les concepts de l’analyse sémantique distinguent nettement les
périodes : ‘Vote’ ‘Civilisation’ ‘Degré de libéralisme’ au XXème, et ‘emploi’,
citoyenneté’ ‘politique fiscale’ au XXIème
3.2 Contribution à l’analyse qualitative pure
Ces résultats plus abondamment décrits dans l’application en ligne peuvent
être utilisées dans l’esprit de la recherche qualitative pure dès lors qu’on les
envisage dans une démarche descriptive et exploratoire dont la valeur réside
que dans la capacité du chercheur à les lire et à les d’interpréter (Moscarola).
Les mots clé, nuages, cartes, les classifications et les concepts proposés par les
logiciels sont des substituts du corpus. Ils portent la trace des modèles
mentaux (Johnson-Laird) et des représentations et l'influences sociales dont
parlent la théorie des actes de langage (Austin) et la sociolinguistique. L’ADT
permet d’en faire une sorte de radioscopie et de mieux les comprendre. Elle
offre aussi la possibilité d’une lecture distanciée échappant au risque de
récursivité (Dumez) ou donnant la possibilité de le contrôler. En effet les
substituts lexicaux ou sémantiques sur lesquels le chercheur fonde ses
interprétations peuvent être communiqués pour exposer la lecture qu’il en
fait à la critique d’une discussion basée sur des éléments partagés.

1

Thésaurus Larousse (Péchon 1994) intégré à SphinxIQ2

568

JADT’ 18

3.2 Contribution à l’analyse de contenu
L’ADT peut également être vue comme une modalité de l’analyse de contenu
traditionnelle (Belerson, Bardin). Elle s’en distingue par l’automatisme d’une
‘lecture artificielle’ identifiant des catégories établies statistiquement par
apprentissage ou à partir d’un thésaurus. On retrouve ainsi l’approche
inductive conduisant à interpréter à postériori les structures révélées par les
analyses factorielles ou à reconnaître dans le corpus les concepts du
thésaurus.
Chaque unité de signification peut ainsi être codée dans une variable
‘mesurant’ le sens et utilisable selon les procédures classiques de l’analyse
quantitative. Dans notre exemple on peut ainsi chercher les éléments lexicaux
ou sémantique explicatif ou discriminant les appartenances politique des
candidats…
3.3 Retour au texte et ‘data visualisation’
Le recours à l’ADT lexicale ou sémantique comporte deux risques majeurs
malgré son intérêts pratique et scientifique : le risque d’erreur systématique
auquel expose la lecture par une machine et le risque de réduction abusive
imposé par les choix du chercheur, qu’il s’agisse de sa problématique ou des
résultats qu’il choisit de communiquer.
Le premier risque peut être évité par le retour au texte et une lecture de
vérification. C’est la seule manière pour le chercheur et son lecteur de
contrôler le sens des éléments lexicaux ou la pertinence des concepts et
évaluations identifiés par les moteurs sémantiques ? Cette possibilité
apparaît avec les hypertextes. Elle est d’autant plus nécessaire, qu’avec l’aide
des infographies (nuages de mots, cartes) les représentations deviennent de
plus en plus parlantes.
Les méthodes dites de navigation facilitent ce retour au texte et peuvent être
enrichies par les entrées provenant des codifications lexicales et sémantiques
ou par les éléments des représentations visuelles. La navigation lexicale
généralisée dans l’esprit de la datavisualisation (Faulx Briole) donne ainsi au
lecteur la possibilité d’accéder directement aux verbatims associés aux mots
d’un nuage ou d’une carte, aux catégories d’une classification automatique
ou aux concepts et appréciations d’une analyse sémantique. Par exemples à
quel verbatim correspond l’usage des mots ‘gens’ ou ‘français’, sont-ils plutôt
de gauche ou de droite, à quoi correspond le concept ‘citoyenneté’ et est-il
daté par un époque ou spécifique à certains candidats ? Retour au texte, mais
au contexte aussi.
L’analyse des discours politique a été pionnière dans ce domaine. Le Monde
publie le 15-03-2012 une infographie dynamique donnant accès aux discours

JADT’ 18

569

de campagne des candidats (Véronis). L’observatoire du discours politique
(Mayaffre) en est un autre exemple. Il permet à partir d’un nuage de mots
synthétisant le contenu des discours, d’en détailler les significations par du
verbatim et d’en spécifier l’usage selon les différents candidats.
Avec ce type d’application le chercheur qualitatif peut compléter la
communication de ses résultats et de ses interprétations en donnant accès au
corpus par l’expérience d’une navigation interactive proposée au lecteur. Il
peut ainsi vérifier les interprétations de l’auteur et les prolonger par ses
propres conjectures. C’est ce que nous proposons à l’adresse :
https://www.sphinxonline.net/debats/1974-2017/analyse.htm
Y
sont
présentés les substituts et synthèses qui conduisent à conclure à une
profonde transformation du débat politique amorcée au tournant du siècle.
Ces tendances peuvent être expérimentées par le lecteur pour nourrir une
discussion critique ou susciter de nouvelles explorations et conjecture. Le
logiciel utilisé permet ainsi de produire des résultats et en même temps de
donner la possibilité au lecteur de les discuter. C’est le propre de la démarche
scientifique.
Bibliographie
BARDIN L. (1977) L’Analyse de Contenu PUF
BEAUDOUIN V. (2016) Retour aux origines de la statistique textuelle : Benzécri et
l’école française de l’analyse de données JADT 2016
BENZECRI JP. (1992) Correspondance Analysis Handbook Marcel Decker Inc.
1992
BERELSON, B (1952). Content Analysis in Communication Research. Glencoe:
Free Press..
BOUGHZALA Y., HERVE H., MOSCAROLA J. (2014) Sphinx Quali : un nouvel
outil d’analyses textuelles et sémantiques JADT Université de Paris
BRUNET E. (2016) Apports des technologies modernes à l’histoire littéraire HAL
BURDICK A., DRUCKER J & ali. (2012) Digital humanities MIT Press
DA SILVA L. (2006) Thésaurus et systèmes de traitement automatique de la langue,
Documentation et bibliothèque
DUPUY, P.-O. & MARCHAND, P., (2016) Les débats de l’entre-deux-tours de
l’élection présidentielle française (1974-2012) Mots. Les langages du
politique,
EDEN C. (1988). "Cognitive mapping". European Journal of Operational Research
FAULX-BRIOLE A. (2017) Datavisualisation et tableaux de bord interactifs
Solution Business
FLOCH J.M.(1988), The contribution of structural semiotics to the design of a
hypermarket, International Journal of Research in Marketing, 4, 3, Semiotics
and Marketing

570

JADT’ 18

GOFFMANN E. Frame analysis: An essay on the organization of experience Harper
and Row 1974
GRUBER T. (1992) Toward Principles for the Design of Ontologies Used for
Knowledge Sharing. In: International Journal Human-Computer Studies
JOHNSON-LAIRD, P N. (1983) Mental Models: Toward a Cognitive Science of
Language, Inference and Consciousness. Harvard University Press
LANDAUER, T. K., FOLTZ, P. W., & LAHAM, D. (1998) An introduction to
latent semantic analysis. In Discourse processes, Routledge
LEBART L SALEM A. (1988) Analyse de données textuelles DUNOD
MARCHAND, P. (2016). Les représentations sociales dans le champ des médias. In
G. Lo Monaco, S.
MAYAFFRE D. (2005) Analyse du discours politique et Logométrie : point de vue
pratique et théorique Langage et société N° 114
MAYAFFRE D. (2014) Plaidoyer en faveur de l’analyse de données c(n)textuelle.
Parcours coocurrentiels dans le discours présidentiel français. Actes JADT Nice
MOSCAROLA J. (2018) Faire parler les données. Editions EMS
MULLER C. (1979,). Étude de statistique lexicale. Le vocabulaire du théâtre de Pierre
Corneille, Paris, Slatkine
NORMIER B. (2007). L’apport des technologies linguistiques au traitement et
à la valorisation de l’information textuelle. ADBS.
REINERT A., (1983), Une méthode de classification descendante hiérarchique :
application à l’analyse lexicale par contexte” Les cahiers de l’analyse des
données, Tome 8, N°2, pp. 187198.
STONE D.C. DUNPHY, M.S. SMITH, . M. OGILVIE. (1966) The General
Inquirer: a computer Approach to Content Analysis MIT Press
THELWALL M. (2017) Sentiment Analysis for Smal and Big Data.SAGE
VERONIS J. (2014) Le traitement automatique des corpus oraux. In Traitement
automatique des langues. Hermes

JADT’ 18

571

A conversation analysis
of interactions in personal finance forums
Maurizio Naldi
University of Rome Tor Vergata– maurizio.naldi@uniroma2.it

Abstract 1
Interactions on a personal finance forum are investigated as a conversation,
with post submitters acting as speakers. The presence of dominant positions
is analysed through concentration indices. Patterns in replies are analysed
through the graph of replies and the distribution of reply times.
Keywords: Personal finance; Conversation analysis; Concentration indices.
1. Introduction
Decisions concerning personal finance are often taken by individuals not just
on the basis of factual information (e.g., company’s official financial
statements or information about past performance of funds), but also
considering the opinions of other individuals. Nowadays personal finance
forums on the Internet have often replaced friends and professionals in that
role. In those forums the interaction occurs among people who typically do
not know one another personally and know very few personal information (if
any) about other participants. Anyway, they often create online communities
that can bring value to all participants [1]. Examples of such forums are
SavingAdvice (http://www.savingadvice.com/forums/) or Money Talk
(http://www.money-talk.org/board.html).
The actual influence of such forums on individuals’ decisions has been
investigated in sev- eral papers, considering, e.g., how the level of activity on
forums impacts on stock trading levels [2], how participation in such forums
pushes towards a more risky-seeking behaviour [3], or introducing an
agents-based model to determine how individual competences evolve due to
the interaction [4]. It has been observed that such forums may be employed
by more aggressive participants to manipulate more inexperienced ones [5],
establishing a dominance over the forum. In addition to being undesirable
for ethical reasons, such an influence is often contrary to the very same rules
of the forum. Here we investigate the subject by adopting a different
approach from the semantic anal- ysis of [5]. In particular, we investigate the
presence of imbalances in the online discussion and the dynamics of the
interaction between participants. The rationale is that partici- pants wishing
to manipulate others would try to take control of the discussion by posting

572

JADT’ 18

more frequently and being more reactive.
For that purpose we employ two datasets extracted from the two most
popular personal finance threads on the SavingAdvice website. For the
purpose of the analysis the thread is represented as the sequence of
participants taking turns, with dates and times of each post attached.
We conduct a conversation analysis, wishing to assess if: 1) there are any
dominant par- ticipants (in particular the thread starter); 2) repetitive
patterns appear such as sustained monologues or sparring matches between
two participants; 3) replies occur on a short-time scale.
The paper provides the following contributions:
 through the use of concentration indices we find out that, though no
dominance exist, the top 4 speakers submit over 60% of the posts
(Section 3);
 both recurring reply sequences and monologues appear (Section 4);
 reply times can be modelled by a lognormal distribution, with 50% of
the posts being submitted no longer than 14 or 23 minutes (for the
two datasets respectively) after the last one (Section 4).
2. Datasets
We consider the two most popular threads on the SavingAdvice website. The
topics are the following, where we indicate an identifying short name
between parentheses:
1. Should
struggling
families
tithe?
(Struggling)
African-American
Personal
Finance
Gurus
(Guru)
The main characteristics of those datasets are reported in Table 1. For each
thread we identify the set of speakers S = {s1, s2, . . . , sn}, i.e., the individuals
who submit posts. We identify also the set of posts P = {p1, p2, . . . , pm} and a
function F : P → S, that assigns each post to its submitter. For each speaker
we can therefore compute the number of posts submitted by him/her. If we
use the indicator function 1(·), the number of posts submitted by the generic
speaker si is

(1)
Table 1: Datasets

Thread

Creator

No. of speakers

Struggling
Guru

jpg7n16
25
james.hendrickson 18

No. of posts
155
104

JADT’ 18

573

3. Dominance in a thread
In this section we wish to examine if some dominance emerges in a thread.
We adopt concentration indices borrowed from the field of industrial
economics.We analyse domi- nance by considering the frequency of posts: an
individual (or a group of individuals) is dominant if it submits most of the
posts. We first examine how posts are distributed by looking at the rank-size
plot: after ranking speakers by the number of posts they submit, the
frequency of posts is plotted vs the rank of the speaker. In Figure 1, we see
that a linear relationship appears between log N(i) and the rank i, so that a
power law N(i) = k/i (a.k.a. a generalized Zipf law) may be assumed to apply
roughly, where k is a normalizing constant and α is the Zipf exponent (see,
e.g., [6]), measuring the slope of the log-linear curve, hence the imbalances
between the contributions of all the speakers. By performing a linear
regression, we get a rough estimate of α, reported in Table 2.
Table 2: Concentration measures

Thread
Struggling
Guru

Zipf exponent

HHI

CR4

0.2545
0.2501

0.1220
0.1396

61.94%
67.31%

As more general indices to assess dominance position we borrow two from
Industrial Economics: the Hirschman-Herfindahl Index (HHI) [7, 8, 9], and
the CR4 [10, 11]. For a market where n companies operate, whose market
shares are v1, v2, …, vn the HHI is

(2)
The HHI satisfies the inequality 1/n ≤ HHI ≤ 1, where the lowest value
corresponds to the case of no concentration (perfect equidistribution of the
market) and the highest value represents the case of monopoly. Therefore,
the larger the HHI the larger the concentration. Instead, the CR4 measures the
percentage of the whole market owned by the top four companies: similarly,
the higher the CR4, the heavier the concentration.
In our case, the fraction of posts submitted by a speaker can be considered as
his/her market share, so that the HHI can be redefined as

574

JADT’ 18

(3)
Instead, the CR4 is defined as

(4)
For our datasets we get the results reported in Table 2. According to the
guidelines provided by the U.S. Department of Justice, the point of
demarcation between unconcentrated and moderately concentrated markets
is set as HHI = 0.15 [12]. Since the values in Table 2 are below that threshold,
we cannot conclude that there is a significant concentration phenomenon.
However, the CR4 index shows that the top 4 speakers submit more than 60%
of all the posts. Delving deeper into the top 4, we also see the most frequent
speaker typically contributes around 1/4 of the overall number of posts,
which represents a major influence. In the Struggling dataset, the most
frequent speaker is the thread originator itself (with 22.6% of posts), while
that’s not true in the Guru dataset, where the the most frequent speaker
contributes 26.9% of posts and the originator just 2.88%.

Fig. 1: Rank-size plot

4. Replies
After examining dominance, we turn to interactions. In this section we
analyse the pattern of replies, looking for recurrences in the sequence of

JADT’ 18

575

replies and examining the time elapsed before a post is replied to.
We build a graph representing how speakers reply to each other. We
consider each post as a reply to the previous one. We build the replies graph
by setting a link from a node A to a node B if the speaker represented by
node A has replied at least once in the thread to a post submitted by the
speaker represented by node B. The resulting graphs are shown in Figure 2,
which is ordered from the core to the periphery in order of decreasing degree
of the nodes, laid out on concentric rings. Here the degree of a node
represents the number of speaker to which it replies. In both cases an inner
core of most connected nodes appear, which represent the speakers replying
to most other speakers. Reply patterns emerge as bidirectional links (couples
of speakers who reply to each other). Loops represent monologues instead,
i.e., speakers submitting two or more posts in a row.

Fig. 2: Replies graph

Further, we are interested in how fast the interactions are between
contributors to the thread. We define the reply time as the time elapsing
between a post and the subsequent one. The main statistics of the reply time
are reported in Table 3. In both dataset the mean reply time is around 1 hour,
but 50% of the replies take place within either 14 minutes (Guru dataset) or
23 minutes (Struggling dataset), i.e., with a much smaller turnaround. There
is therefore a significant skewness to the right.
A more complete view of the variety of reply times is obtained if we model
the probability density function. In Figure 3, we report the curves obtained
through a Gaussian kernel estimator, an exponential model, and a lognormal
model (whose parameters have been estimated by the method of moments).
By applying the Anderson-Darling test, we find out that the exponential

576

JADT’ 18

hypothesis is rejected at the 5% significance level, while the lognormal one is
not rejected, with a p-value as high as 0.72 for the Struggling dataset and
0.076 for the Guru dataset.

Fig. 3: Reply time
Table 3: Reply time statistics (in minutes)

Thread

Mean

Median

Standard
deviation

95% percentile

Struggling
Guru

70.5
58.9

23
14

156.2
112.7

254.7
406.7

5. Conclusions
We have analysed two major threads within a personal finance forum as a
conversation between submitters acting as speakers, searching for dominance
and interaction patterns. Though no significant concentration exists, the top
four speakers submit over 60% of the posts. Patterns of interaction emerge as
the presence of several couples of speakers who reply to each other, several
monologues, and short reply times (with 50% being below 14 and 23 minutes
for the two datasets, though a significant distribution tail is present).
References
[1] Arthur Armstrong and John Hagel. The real value of online communities.
Knowledge and communities, 74(3):85–95, 2000.
[2] Robert Tumarkin and Robert F Whitelaw. News or noise? internet
postings and stock prices. Financial Analysts Journal, 57(3):41–51, 2001.
[3] Rui Zhu, Utpal M Dholakia, Xinlei Chen, and Ren ́e Algesheimer. Does
online community participation foster risky financial behavior? Journal of

JADT’ 18

577

Marketing Research, 49(3):394–407, 2012.
[4] Loretta Mastroeni, Pierluigi Vellucci, and Maurizio Naldi. Individual
Competence Evolution under Equality Bias. In 2017 European Modelling
Symposium (EMS), Nov 2017.
[5] John Campbell and Dubravka Cecez-Kecmanovic. Communicative
practices in an on- line financial forum during abnormal stock market
behavior. Information & management, 48(1):37–52, 2011.
[6] Maurizio Naldi and Claudia Salaris. Rank-size distribution of teletraffic
and customers over a wide area network. Transactions on Emerging
Telecommunications Technologies, 17(4):415–421, 2006.
[7] Stephen A Rhoades. The Herfindahl-Hirschman Index. Fed. Res. Bull.,
79:188, 1993.
[8] Maurizio Naldi. Concentration indices and Zipf’s law. Economics Letters,
78(3):329–334,
2003.
[9] Maurizio Naldi and Marta Flamini. Censoring and Distortion in the
Hirschman–Herfindahl
Index Computation. Economic Papers: A journal of applied economics
and policy, 2017.
[10] I Pavic, F Galetic, and Damir Piplica. Similarities and Differences
between the CR and HHI as an Indicator of Market Concentration and
Market Power. British Journal of Economics, Management and Trade,
13(1):1–8, 2016.
[11] Maurizio Naldi and Marta Flamini. Correlation and concordance
between the CR4 index and the Herfindahl-Hirschman index. SSRN
Working paper series, 2014.
[12] The U.S. Department of Justice and the Federal Trade Commission.
Horizontal Merger Guidelines, 19 August 2010.

578

JADT’ 18

Analisi testuale, rumore semantico e peculiarità
morfosintattiche: problemi e strategie di
pretrattamento di corpora speciali.
Stefano Nobile
Sapienza Università di Roma – stefano.nobile@uniroma1.it

Abstract 1
The proliferation of text analysis techniques has made possible the combined
use of different software, directed each time to specific needs for analysis and
research. However, the opportunities offered by the different software do not
mitigate a fundamental problem, inherent in the characteristics of some
peculiar corpora. Perfectly suited for analysis on texts written accurately and
based on a supervised style, however these software can not reduce some
issues. Among these, one of the most common concerns the morphosyntactic
rules of the language with its semantic noise. Problems of "noise", such as
that generated in spontaneous conversations, require many precautions for
the preparation of the corpus. This situation is exaggerated with Twitter,
whose ease of access and messaging download has produced analysis that is
not always adequately supported from the theoretical point of view. Poems
and songs present a similar problem. In these kinds of corpora the problem
derives from the structure of this style of communication, which in using
some rhetorical expedients accentuates the critical mass generated by some
words. What strategies are possible to adequately prepare the corpora to be
analysed in these two particular situations? The contribution proposes some
strategies on how to operate in these particular conditions, highlighting the
advantages on the empirical level but also the effects on the theoretical one.
Abstract 2
La moltiplicazione delle tecniche di analisi testuale ha reso possibile l’uso
combinato di software diversi, piegati di volta in volta a singole esigenze di
analisi e ricerca. Tuttavia, l’ampiezza di opportunità offerte dai diversi
software non attenua un problema di fondo, insito nelle caratteristiche stesse
di alcuni corpora peculiari. Perfettamente adatti ad analisi su testi redatti
accuratamente e improntati a uno stile sorvegliato, questi software non
riescono tuttavia a togliere l’utente dall’impaccio nel quale può trovarsi in
alcune circostanze. Tra queste, una delle più comuni riguarda le regole
morfosintattiche della lingua di riferimento e quindi portatrice di quote
elevate di rumore semantico. Problemi di “rumore”, come quello generato

JADT’ 18

579

nelle conversazioni spontanee, richiedono al ricercatore una serie di
accorgimenti per la preparazione del corpus che tengano conto della
necessità di evitare di ottenere dati fortemente distorti. Questo discorso si
esaspera con Twitter, la cui facilità d’accesso e download dei messaggi è da
qualche tempo foriero di analisi non sempre adeguatamente sostenute dal
punto di vista teorico. A questi casi si aggiunge quello di corpora altrettanto
peculiari come quelli delle poesie e delle canzoni. In corpora di questo tipo il
problema deriva dal costrutto stesso di questo genere comunicativo, che nel
servirsi di alcuni espedienti retorici accentua la massa critica generata da
alcune parole, andando così a incidere, tra l’altro, sul calcolo di alcuni
parametri rilevanti e rendendo meno leggibili i risultati. Quali strategie sono
dunque possibili al ricercatore per preparare adeguatamente i corpora da
analizzare in queste due situazioni particolari? Il contributo che si intende
presentare vuole avanzare alcune proposte su come operare in queste
particolari condizioni, evidenziando i vantaggi sul piano empirico ma anche
le ricadute su quello teorico soggiacente agli obiettivi stessi che analisi su
corpora di questo genere possono porsi.
Keywords: rumore semantico, poesia, canzone, retorica, pretrattamento del
corpus, costruttivismo vs. realismo.
1. Rumore semantico e corpora testuali peculiari
La moltiplicazione delle tecniche di analisi testuale ha reso possibile, ai
ricercatori interessati a lavorare in questo ambito, l’uso – anche combinato –
di diversi software, ciascuno con le proprie peculiarità in risposta alle
differenti esigenze di analisi e ricerca. Tuttavia, l’ampiezza di opportunità
offerte dai tanti software in commercio (T-Lab, Taltac, Spad-T, R, eccetera)
non attenua un problema di fondo, insito nelle caratteristiche stesse di alcuni
corpora peculiari: quello delle distorsioni imputabili al rumore semantico
generato sia da elementi irrilevanti dal punto di vista contenutistico, sia da
ridondanze che alterano i rapporti di forza tra parole.
Perfettamente adatti ad analisi testuali su testi redatti accuratamente e
improntati a uno stile sorvegliato come può essere quello delle testate
giornalistiche o di materiali di tipo istituzionale, questi software non riescono
tuttavia a togliere l’utente dall’impaccio nel quale può trovarsi in alcune
circostanze che, più o meno in concomitanza con la diffusione dei social
network, hanno cominciato ad essere egemoniche quanto a produzioni di
testi sul web. Tra queste circostanze, una delle più comuni riguarda quella
che si potrebbe definire oralità scritta, poco o per nulla accorta alle regole
morfosintattiche della lingua di riferimento e quindi portatrice di quote
elevate di rumore semantico, qui inteso come forma leggibile e trattabile di

580

JADT’ 18

testo. Problemi di “rumore” come quello generato nelle conversazioni
spontanee, rinvenibili – nelle forme più disparate – in rete, richiedono al
ricercatore una serie di accorgimenti per la preparazione del corpus che
tengano conto della necessità di evitare di ottenere dati fortemente distorti.
Vale a dire che le forme linguistiche contratte (cmq, nn, xké), gli elementi
espressivi tesi a restituire i toni del parlato (belloooo, bravaaaa), i segni grafici
del tutto peculiari (Ã, ò, ðŸ , ù, é, 🠳), le ridondanze, i retweet, il testo
non in formato Ascii, gli hashtag, i collegamenti multimediali, il linguaggio
di markup, sono addendi di una somma che dà come risultato una
proliferazione di rumore semantico, ai cui effetti si aggiungono quelli
derivanti dalle distorsioni imputabili agli indici prodotti (ricercatezza ed
estensione lessicale) nonché alle misure del corpus (occorrenze, forme
grafiche, hapax). Questo discorso si esaspera con Twitter, la cui facilità
d’accesso e download dei messaggi è da qualche tempo foriero di analisi non
sempre adeguatamente sostenute dal punto di vista teorico (Ebner, Altmann
e Softic, 2011). Accade infatti sempre più spesso che «l’elevato grado di
automatismo delle procedure e la forte tendenza alla modellizzazione
statistica possono esporre l’analisi testuale a stili di ricerca segnati da
un’ingenua rincorsa dell’oggettività tramite l’estremizzazione ossessiva del
calcolo numerico applicato ai testi, con la conseguente grave perdita del
ruolo del contesto» (Tipaldo, 2014: 191; corsivo aggiunto). La necessità di
contrarre il testo in 120 caratteri (raddoppiati soltanto a partire dal novembre
2017, ma la sostanza non cambia) determina infatti negli utenti l’inclinazione
a trovare soluzioni – a volte convenzionali, altre volte originali – per poter
ridurre il testo entro i limiti prefissati, così come si faceva quando gli sms
avevano set limitati di caratteri ed erano relativamente dispendiosi. Da qui,
la produzione di una quantità considerevole di rumore semantico che rende
difficilmente trattabili i dati testuali “naturali”. Ai casi appena passati in
rassegna – oggi largamente diffusi – si aggiunge quello di corpora altrettanto
peculiari, ma del tutto diversi, come quelli delle poesie e delle canzoni
(Nobile, 2012). In testi di questa natura, il problema deriva dal costrutto
stesso di questi generi della comunicazione. Essi, infatti, nel momento in cui
si servono di alcuni espedienti retorici (l’anadiplosi, l’epanalessi, il poliptoto,
l’anafora, l’epanadiplosi e altri ancora), accentuano la massa critica generata
da alcune parole. Ciò finisce con l’incidere sul calcolo di alcuni parametri
rilevanti (specificità tipiche ed esclusive, estensione lessicale, ricercatezza
lessicale, rango delle singole parole, confronto con i lessici peculiari,
eccetera), rendendo meno leggibili i risultati.
Un caso assai frequente, qui portato al parossismo, è il seguente: nella
canzone, alcune parole, per necessità squisitamente ritmiche oppure per
enfatizzare l’effetto-tormentone, vengono ripetute ossessivamente. È quanto

JADT’ 18

581

accade – per fare un solo esempio, dati i margini ridotti entro i quali deve
rimanere questo contributo – con la canzone Pino (fratello di Paolo), nella quale
la parola Pino compare addirittura 60 volte nel giro di pochi secondi,
andando ineluttabilmente a gonfiare tutte le modalità delle variabili (artista,
decennio di pubblicazione, macro e microgenere musicale, sesso) a cui questa
singola canzone è collegata (Nobile, 2012). Per l’uso delle figure retoriche
vale un discorso analogo. Tra le tante possiamo prendere l’anafora a titolo
esemplificativo. L’anafora è una figura retorica che consiste nella ripetizione
di una o più parole all’inizio di una frase o di un verso. Per quanto essa sia
rintracciabile anche nella prosa, è nella poesia e nella canzone che essa
ottimizza le proprie potenzialità espressive. Tra lo sterminato numero di
esempi che potremmo scegliere, uno è quello di Vai in Africa, Celestino!, un
brano che il cantautore Francesco De Gregori ha pubblicato nel 2005: pezzi di
stella, pezzi di costellazione / pezzi d’amore eterno, pezzi di stagione / pezzi di
ceramica, pezzi di vetro / pezzi di occhi che si guardano indietro / pezzi di carne,
pezzi di carbone / pezzi di sorriso, pezzi di canzone / pezzi di parola, pezzi di
parlamento / pezzi di pioggia, pezzi di fuoco spento. In questo caso, è la parola
pezzi a comparire un considerevole numero di volte grazie, appunto,
all’espediente retorico dell’anafora. Non diverso, ovviamente, è il caso della
letteratura, per il quale – a titolo esemplificativo – possiamo scomodare il
celeberrimo III canto (canto e canzone, appunto…) dell’Inferno dantesco: Per
me si va ne la città dolente / per me si va ne l'etterno dolore / per me si va tra la
perduta gente. La poesia e la canzone, dunque, possono presentare delle
caratteristiche strutturali che vanno a incidere sul text mining operabile dai
diversi software, nella misura in cui forniscono informazioni numeriche
alterate. Quantunque la ridondanza di alcuni termini non implichi
necessariamente lo stravolgimento dell’asse sintagmatico (Bolasco, 2005),
ossia della possibilità di ricostruire il senso del testo in ragione di un criterio
di adiacenza delle parole all’interno dei contesti elementari, essa può
compromettere il senso espresso dai dati relativi alla frequenza delle parole
piene, alle peculiarità (sia quelle endogene, esprimibili in termini di
specificità, sia quelle esogene, traducibili in termini di linguaggio peculiare) e
alla numerosità di forme grafiche. Quali strategie sono dunque possibili al
ricercatore per preparare adeguatamente i corpora da analizzare in queste
due situazioni particolari, ossia profluvio di segni grafici e parole ripetute?
Certamente non è sufficiente ripulire ortograficamente il testo né espungere
da esso tutti quei segni, come le emoticons o la sintassi comunicativa propria
di Twitter, che vanno a interferire su molti parametri d’analisi. Né d’altronde
si può “addomesticare” il corpus fino al punto da stravolgerne l’aspetto
precipuo, ossia la spontaneità del simil parlato del primo caso e la struttura
morfosintattica e retorica del secondo.

582

JADT’ 18

2. Strategie di pre-trattamento del corpus
Le soluzioni ai tipi di problemi testé esposti variano a seconda della natura
del problema, delle competenze informatiche dell’utente e della prospettiva
analitica assunta dal ricercatore e dipenderanno dalla combinazione tra
queste tre dimensioni. Vediamole. La pulizia dei caratteri di testi naturali
dipende in larga misura dalle competenze informatiche dell’utente, al netto
delle potenzialità dei software utilizzati. Ad oggi, un utente privo di abilità
informatiche avanzate non è in grado di fare un lavoro di pulizia impeccabile
su corpora testuali molto “sporchi” come sono quelli che provengono da
Twitter. Se da un lato gli potrà essere d’aiuto una elevata quota di pazienza
per utilizzare un correttore ortografico che ripulisca il testo dagli errori di
battitura tipici di testi “naturali”, e quindi non supervisionati, dall’altro
dovrà necessariamente scontrarsi con la ridda di caratteri speciali che sono
stati richiamati in precedenza. Le soluzioni a disposizione sono tre: il livello
base consiste nella sostituzione manuale e in blocco di tutti i segni grafici da
correggere, facendo attenzione – nell’uso di un normale word processor – alle
maiuscole e alle minuscole. Si tratta di un’operazione tanto più lunga e
faticosa quanto più lungo, complesso e ricco di rimandi ipertestuali è il
corpus da ripulire. In alcuni casi, esistono software come Taltac che
possiedono al loro interno una funzione di rimozione di alcuni caratteri
particolari. Una seconda soluzione è quella di programmare delle macro (o,
alternativamente, di usare programmi esterni) che risolvano lo stesso tipo di
problema. La soluzione è più efficace dal punto di vista del risultato finale,
ma altrettanto impegnativa da quello delle competenze e del tempo richiesti.
La terza soluzione è, sulla carta, quella in grado di ottimizzare meglio il
rapporto costi/benefici. Si tratterebbe, in questo caso, di sfruttare le
potenzialità di programmi di ricerca che si sono dati come obiettivo proprio
quello della pulizia di testi originati nel web e utilizzati per analisi testuali.
Vanno in questa direzione progetti come Readability o CleanEval (Baroni et al.,
2008), che tuttavia presentano a loro volta due ordini di problemi: uno legato
ai costi; l’altro alla effettiva possibilità d’accesso. Entrambi, peraltro,
evidenziano problemi di flessibilità rispetto ai diversi formati di corpora da
elaborare (Claridge, 2007; Petri e Tavosanis, 2009). La questione del
trattamento di corpora che devono la loro peculiarità alla struttura
soggiacente, pur non presentando problemi rilevanti di ordine informatico, è
più complessa e implica scelte decisive da parte del ricercatore. Il ricercatore
dovrà infatti operare delle scelte di carattere gnoseologico e teorico rispetto ai
fini che si pone, ben sapendo che le decisioni che prenderà avranno
inevitabili ricadute sul piano delle risultanze empiriche. In altri termini, il
ricercatore che impatta con materiale testuale che non nasce in forma di

JADT’ 18

583

prosa, ma di verso, si trova sostanzialmente a dover operare una scelta tra
una rappresentazione fedele, “fotografica”, delle caratteristiche del corpus
esaminato e quella che invece tiene conto delle ridondanze e di tutti quegli
elementi che possono contribuire a gonfiare alcuni parametri del corpus, a
partire dal conteggio di forme grafiche e a finire con gli hapax. Nel primo
caso gli esiti dell’analisi subiranno l’impatto non solo di quegli elementi
retorici e morfosintattici che possono caratterizzare la forma-canzone o la
forma-poesia, ma soprattutto del ritornello. Accettare questa prospettiva
significa assumere alcune sezioni di testo – nonché gli elementi di esso che
contribuiscono a ispessire alcuni termini per via delle scelte operate sui versi
dagli autori – come elementi che, proprio perché ripetuti, meritano di
svettare in termini parametrici dall’analisi del corpus stesso. Possiamo dire
che in un caso come questo i risultati siano ingannevoli? Dipende, appunto,
dalla prospettiva che si intende assumere. Una rappresentazione
iperrealistica ci porta a scegliere la prima formula, quella del massimo rigore
filologico, dello zelo assoluto: a un certo ammontare di parole, seppur
ripetute a iosa, deve corrispondere il reale valore di frequenza delle parole
stesse, con tutto ciò che questo implica in termini di relazioni tra parole, di
frequenze e di individuazione di topics all’interno del corpus. All’opposto, il
ricercatore potrebbe avere delle ottime ragioni per propendere per una
prospettiva costruttivista, in virtù della quale il dato viene forgiato in ragione
non già della frequenza effettiva delle parole – con le ridondanze che alcuni
corpora si portano dietro per le ragioni già esposte – bensì del testo spurgato
dagli elementi ridondanti. Un esempio che dovrebbe rendere palmare le
implicazioni e la differenza esistente tra le due opzioni può essere tratto da
un recente lavoro sui testi della canzone italiana che costituisce un
aggiornamento in una direzione più spintamente sociolinguistica di un mio
lavoro precedente (Nobile, 2012). Dal corpus1 che raccoglie i testi degli artisti
che sono riusciti a piazzare uno o più dischi nei primi sessanta posti delle
classifiche di vendita tra gli anni ’60 del Novecento e il 2016 selezioniamo i
due che hanno fatto registrare il maggior numero di ingressi2: Mina (170
canzoni) e Renato Zero (177). Da ciascuno dei due corpora andiamo a
estrarre, previa lemmatizzazione e normalizzazione del testo, le parole piene.
A questo punto possiamo assegnare il rango a ciascuna di esse in base al
numero di occorrenze nella prima e nella seconda situazione: quella nella

Il corpus è costituito dai testi di 5940 canzoni, che hanno sviluppato 1.321.994
occorrenze, 43.855 forme grafiche diverse, 22.160 parole piene e 1.905 hapax.
2 Per il criteri di campionamento, si veda Nobile, 2012: 51-53 o anche Nobile,
L’italiano della canzone dagli anni sessanta a oggi. Una prospettiva sociolinguistica, in corso
di pubblicazione.
1

584

JADT’ 18

quale il testo è riportato pedissequamente così come viene cantato (quindi
con tutti gli elementi di ridondanza di cui si è parlato) e quella in cui esso è
stato invece ripulito da questi elementi che determinano una consistente
ripetizione, imputabile appunto alla struttura della canzone, di alcuni
termini3. Il confronto tra i due ranghi, operato rispetto ai due diversi artisti,
suggerisce l’uso del coefficiente di cograduazione di Spearman (ρ). I valori
ricavati dai due confronti forniscono risultati di indubbio interesse: nel caso
di Mina, il valore del ρ di Spearman è di 0,61; in quello di Renato Zero di
0,68. Questa informazione, da sola, ci fonisce un’indicazione su quanto la
pulizia del testo e il rumore semantico generato dalle ridondanze possa
produrre conseguenze più che tangibili nella strutturazione dei dati da
elaborare: una parola che ha basso rango ha più probabilità di essere
selezionata tra le parole chiave, di comparire come termine specifico di un
certo sottoinsieme, di emergere come parola capace di differenziarsi in
ragione del rango che essa occupa in dizionari di riferimento (De Mauro et
al., 1993) e, quindi, di ergersi a indicatore della peculiarità linguistica di un
determinato locutore o di una certa unità aggregata di analisi. Così, nel
corpus di Mina la parola specchio, una volta sacrificati i ritornelli, arriva a uno
scarto di rango di 165 posizioni e la parola rabbia perde 100 posizioni nei due
diversi trattamenti del corpus. Analogamente, nel corpus di Renato Zero la
parola identikit perde 226 posizioni a seconda che il corpus sia ripulito dalle
ridondanze oppure no: essa si trova in una sola canzone (Io uguale io),
ripetuta un’infinità di volte. Stesso discorso con la parola fame, che perde 183
posizioni: essa, pur essendo – al contrario di identikit – del tutto trasversale
nel canzoniere del cantautore romano, ricorre un consistente numero di volte
come tormentone della canzone C’è fame.
3. Conclusioni
In queste pagine si è visto che alla facilità di accesso a una quantità ciclopica
di materiale testuale rinvenibile sul web non corrisponde una altrettanto
disinvolta possibilità di analisi dello stesso. Da una parte, infatti, questo
materiale incorpora le caratteristiche tipiche del linguaggio cosiddetto
naturale e, in quanto tale, va incontro non soltanto ai comuni problemi di
machine learning e di text mining (i più comuni dei quali sono riscontrabili, per
esempio, nei traduttori automatici o nei programmi di riconoscimento
vocale), ma anche a quelli creati dal sovradosaggio di elementi sempre più

3 La pulizia del testo espunto dai versi duplicati è stata realizzata utilizzando
una funzione del programma Excel (dati, rimuovi duplicati) tenendo fissi i riferimenti
alle singole canzoni e ai diversi autori, in modo da evitare la rimozione di versi
duplicati a prescindere dai due parametri di riferimento testé indicati.

JADT’ 18

585

diffusi come emoticons, caratteri speciali, eccetera. A questi problemi se ne
possono aggiungere altri, annoverabili nell’ambito della poesia e della
canzone, che rendono necessaria una fase particolarmente accurata e
meditata del pre-trattamento dei testi stessi, prima che questi vengano
sottoposti ad analisi. Nell’articolo si è cercato di mostrare come le scelte di
ordine gnoseologico compiute a monte dal ricercatore abbiano, nel caso delle
forme linguistiche peculiari di cui si è parlato, ricadute rilevanti sulle stesse
risultanze empiriche. In più, le operazioni di tipo lessicometrico su materiale
testuale con forte rumore semantico rischiano, se non adeguatamente
supportate da una pulizia – tutt’altro che agile – del corpus spesso, di
produrre risultati in cui la quota di rumore semantico rischia di essere
addirittura superiore a quella del testo vettore di effettivo significato (Nobile,
2016).
Riferimenti bibliografici
Baroni M., Chantree F., Kilgarriff A. and Sharoff S. (2008). Cleaneval: A
competition for cleaning webpages. Proceedings of the 6th Conference on
Language Resources and Evaluation (LREC) (pp. 638-643). Elda.
Bolasco S. (2005). Statistica testuale e text mining: alcuni paradigmi
applicativi. Quaderni di Statistica, 7, pp. 17-53.
Chiari I. (2007). Introduzione alla linguistica computazionale. Laterza.
Claridge C. (2007). Constructing a corpus from the web: message boards. In
M. Hundt, N. Nesselhauf, and C. Biewer, Corpus Linguistics and the Web
(pp. 87-108). Rodopi.
De Mauro T., Mancini F., Vedovelli M. and Voghera M. (1993). Lessico di
frequenza dell'italiano parlato. EtasLibri.
Ebner M., Altmann T. and Softic S. (2011). @twitter analysis of #edmedia10 –
is the #informationstream usable for the #mass. Form@re, 11 (74), pp. 3645.
Lancia F. (2004). Strumenti per l’analisi dei testi. FrancoAngeli.
Nobile S. (2012). Mezzo secolo di canzoni italiane. Una prospettiva sociologica
(1960-2010). Roma: Carocci.
Nobile S. (2016). Consenso e dissenso. Le reazioni degli elettori ai post dei
candidati. In Morcellini M., Faggiano M.P. and Nobile S. (a cura di),
Dinamica Capitale. Traiettorie di ricerca sulle amministrative 2016 (pp. 115138). Maggioli.
Pandolfini V. (2017). Il sociologo e l'algoritmo. l'analisi dei dati testuali al tempo di
Internet, FrancoAngeli.
Petri S. and Tavosanis M. (2009). Building a Corpus of Italian Web Forums:
Standard Encoding Issues and Linguistic Features. JLCL, 24 (1), 115-128.
Tipaldo G. (2014). L'analisi del contenuto e i mass media. Il Mulino.

586

JADT’ 18

L’individu dans le(s) groupe(s) : focus group et
partitionnement du corpus
Daniel Pélissier
Université Toulouse 1 Capitole - daniel2.pelissier@ut-capitole.fr

Abstract
Lexicometric analyzes of the focus groups depend in particular on the choice
of partitioning of the corpus by researcher. After having proposed a typology
of possible partitioning, we present the results of an experiment of one of
these approaches on a corpus of ten focus groups. These analyzes highlight
some contributions and limitations of lexicometry compared to
conversational analysis.
Résumé
Les analyses lexicométriques des focus groups dépendent notamment des
choix de partitionnement du corpus par le chercheur. Après avoir proposé
une typologie des partitionnements possibles, nous présentons les résultats
d’une expérimentation d’une de ces approches sur un corpus de dix focus
groups. Ces analyses mettent en évidence certains apports et limites de la
lexicométrie par rapport à l’analyse conversationnelle.
Keywords: Focus groups, partitioning, individual, group.
Mots clefs : Focus groups, partitionnement, individu, groupe.
1. Introduction
La lexicométrie a étudié d’abord des discours écrits (articles de journaux,
discours politiques, etc.) et des réponses à des questions ouvertes (Lebart et
Salem, 1988) puis s’est intéressée aux conversations orales retranscrites
(Rouré et Reinert, 1993; Bonneau et Dister, 2010). L’analyse de ces dernières
est en effet plus délicate en raison de textes en général plus courts, de
syntaxes particulières. Les focus groups appartiennent à cette famille de
données en posant le problème particulier du nombre important de
participants. Selon certains auteurs, ce type de données est difficile à analyser
avec des logiciels de lexicométrie (Duchesne et Haegel, 2014).
Pourtant, l’analyse lexicométrique a été utilisée dans plusieurs études
(Guerrero et al., 2009; Grésillon et al., 2012; Hulin, 2013; Bengough et al.,
2015; Brangier et al., 2015) et des articles méthodologiques ont analysé
l’efficacité des traitements lexicométriques (Dransfield et al., 2004; Peyrat-

JADT’ 18

587

Guillard et al., 2014).
Ainsi, la possibilité de traiter les focus groups par la lexicométrie est établie.
Cependant, les apports spécifiques d’une approche quantitative sont à
préciser dans un domaine dominé par les approches qualitatives dont
l’analyse conversationnelle. Par exemple, le lien entre focus groups et
représentations sociales est mis en avant (Jovchelovitch, 2004) et la
classification descendante hiérarchique (CDH) de Reinert (1983) forme des
mondes lexicaux (Ratinaud et Marchand, 2015) dont la nature est proche des
représentations sociales. Nous insisterons, dans cet article, sur la place de
l’individu dans le(s) groupe(s), problématique que la lexicométrie permet
d’approcher par un jeu de variables adapté. Mais cette analyse suppose de
préparer le corpus avec des méthodes spécifiques.
Nous présenterons ainsi une typologie des méthodes de préparation d’un
corpus de focus groups en complétant les analyses de Peyrat-Guillard et al.
(2014) et en mettant en exergue celles centrées sur l’individu. Puis, nous
analyserons les résultats de l’expérimentation d’une de ces méthodes en
montrant en quoi elle permet une compréhension des discours de l’individu
dans le(s) groupe(s).
2. Typologie des partitionnements d’un corpus de focus groups
Avant de commencer le traitement lexicométrique de focus groups, le corpus
exige une préparation spécifique. En effet, certaines décisions de
partitionnement détermineront notamment les méthodes lexicométriques
employables et les analyses possibles.
Les textes des modérateurs sont souvent supprimés du focus groups
(Guerrero et al., 2009 ; Peyrat-Guillard et al., 2014) car ses interventions, dans
le cadre d’un focus group servent à fluidifier les échanges sans les orienter.
Cependant, il peut être conseillé de comparer les résultats avec ou sans les
interventions du modérateur (Peyrat-Guillard et al., 2014).
La deuxième question porte sur la partition du corpus issu du focus group.
Plusieurs méthodes existent. Une première possibilité est d’analyser le focus
group comme une entité sans prendre en compte les échanges entre les
individus. Soit chaque focus group constitue un texte sans distinction
d’individu (Dransfield et al., 2004) ; l’argument avancé par les utilisateurs de
cette méthode est de faciliter les analyses statistiques mais cela n’est pas une
évidence, le nombre de segments étant stable. Soit le focus group est
partitionné en thèmes à partir d’une analyse de contenu (Bengough et al.,
2015) ; cette approche permet de comparer par exemple les résultats d’une
analyse thématique avec celle proposée au chercheur par la lexicométrie. La
deuxième famille de partition est celle qui souhaite conserver les échanges du
focus group. Soit la partition peut être centrée sur les individus, dite

588

JADT’ 18

decrowded (Peyrat-Guillard et al., 2014) ; les textes des interventions de
chaque individu sont alors rassemblés (Guerrero et al., 2009). Soit chaque
intervention est considérée comme un texte, approche dite crowded (PeyratGuillard et al., 2014).
Chacune de ces méthodes a des avantages et des inconvénients. Nous ne
pensons pas qu’une partition soit à privilégier mais que la décision dépend
des analyses envisagées par le chercheur selon sa problématique. Dans cet
article, nous nous centrerons sur la deuxième famille qui permet d’étudier
l’individu dans le(s) groupe(s) et pas seulement les thèmes abordés.
3. Résultats de l’expérimentation du partitionnement par locuteur
Nous avons pu expérimenter ces méthodes de partition d’un corpus de focus
groups à partir d’une recherche que nous avons menée auprès de jeunes
diplômés de l’enseignement supérieur (niveaux bac+3 et bac+5). Les
discussions des focus groups concernaient la communication numérique de
recrutement des banques et ces jeunes diplômés échangeaient sur les
dispositifs utilisés par les entreprises pour recruter. Nous avons animé puis
restranscrit10 focus groups de 6 à 7 personnes soit 67 locuteurs au total.
3.1. Préparation du corpus et partitionnement
Une fois les textes préparés (anonymisation, intégration des noms propres
(BNP, Facebook, etc.) au dictionnaire, adaptation du dictionnaire selon les
spécificités du discours, etc.), nous avons décidé de supprimer les
interventions du chercheur car elles restaient neutres par rapport aux
discours des jeunes diplômés que nous souhaitions analyser.
Nous avons alors créé une partition par tours de parole selon ce principe :
(variables entre crochets)
[Groupe1, Ingénieurs , NUM1, 18ans, masc]: il y a des choses marquantes, il y a un
site web où on n'a pas beaucoup d'informations et un autre site où il y a beaucoup
d'informations.
[Groupe1, Ingénieurs , NUM2, 20ans, masc]: je suis d'accord avec toi.
En effet, nous souhaitions repérer des discours individuels dans les focus
groups et pouvoir associer des variables de profil à un locuteur.
Les variables utilisées (tableau 1) ont été déterminées selon nos hypothèses
de recherche et leur accessibilité puis ont été associées par un script
automatique à chaque intervention de locuteur.
Tableau 1. Variables du focus groups associées aux locuteurs.
Num
1

Code variable
num

Valeur
1, 2, 3, etc.

2

formation

3IL :

Source

école

Description
Numéro de chaque
intervenant
Désignation
du

JADT’ 18
Num

Code variable

3

groupe

4
5

sexe
participation

6

initial

589
Valeur
d’ingénieur
LPB :
licence
professionnelle
banque
1, 2, 3, etc.
10 groupes au total
M, F
TA, PA, A
TA : très actif
A : actif
PA : pas actif
STS, IUT

Source

Description
groupe

Numéro du groupe

Statistiques SONAL
selon
le
nombre
d’interventions
Données organisme
de formation

Indicateur
quantitatif de la
participation
de
chaque intervenant
Formation initiale
des intervenants

Le corpus se présentait ainsi de cette façon pour être utilisé dans Iramuteq
(Ratinaud, 2009) :
**** *num_44 *formation_LPB *groupe_1 *sexe_M *participation_ A
*initial_STS moi je veux bien commencer. Quand je suis allé sur le site de la
SG, … Les caractéristiques du corpus obtenu et traité à l’aide du logiciel
Iramuteq sont alors les suivantes : 1876 textes allant d’une seule forme (Oui
par exemple) pour les plus courts à 126 formes ou 280 occurrences pour le
plus long, 40404 occurrences et 2094 formes au total, 21,54 occurrences par
texte en moyenne, les hapax représentent 41,26% des formes. Chaque texte
correspond alors à une intervention d’un locuteur dans un focus group.
3.2. Choix méthodologiques
Si la CDH de Reinert est la plus souvent citée dans la littérature (Duchesne et
al., 2010 ; Gresillon et al., 2012; Hulin, 2013; Peyrat-Guillard et al., 2014;
Brangier et al., 2015; Freitas et Luis, 2015, etc.) d’autres techniques sont
impliquées comme l’analyse factorielle (Dransfield et al., 2004; Guerrero et
al., 2009) ou plus rarement l’analyse de similitude (Bengough et al., 2015).
Notre choix de la classification de Reinert est lié à nos hypothèses de
recherche qui associent les discours de ces jeunes diplômés aux
représentations sociales. Or, la CDH de Reinert (1983) favorise le repérage de
représentations sociales (Ratinaud et Marchand, 2015). Nous avons effectué
plusieurs CDH simples sur segments de texte en faisant varier le nombre de
classes demandées, le nombre minimum de segments par classe. Nous avons
choisi de retenir les formes dont la fréquence est supérieure à 3 (soit 687
formes dans ce cas) pour centrer le traitement sur les formes les plus
présentes. Au terme de ces simulations, nous avons retenu une CDH qui
présente 15 classes avec un taux de segments classés de 83,63%.

590

JADT’ 18

3.3. Exemple d’utilisation de variables, groupes et degré de participation
Chaque intervention ayant été associée à des variables de contexte, la
méthode choisie permet de vérifier le lien existant entre les groupes et
chaque classe repérée. Ainsi, pour ce corpus de focus groups, la classe 1
(Chi²=20,82, recherche d’emploi) et la classe 12 (Chi²=16,76, articles de
journaux) sont associées aux étudiants de 3IL. La classe 7 (Chi²=32,17,
Dupuy) et la classe 13 (Chi²=11,44, avantages et valeurs) sont plutôt liées au
groupe des licences banques (fig. 1).

Figure 1. Chi² par classe pour la variable ‘formation’.

De même, la variable sur la participation (tableau 1 et fig 2.) a permis
d’associer certaines classes avec cette caractéristique. Les résultats de la CDH
permettent ainsi de poser une hypothèse sur le degré de consensus entourant
une représentation sociale.

Figure 2. Association de la classe 8 avec la variable participation.

En effet, la classe 8 sur la taille de l’organisation est associée aux locuteurs
qui ont peu participé globalement (Variable PA (Peu Actif), Chi²=4,19 ; fig. 2)
comme pour la classe 3 (mobilité). Les discussions sur la recherche d’emploi
(classe 1), la banque Dupuy (classe 7) ou les classements des sites internet et
témoignages sont dominées par les locuteurs les plus actifs (Variable TA
(Très actif) : Chi²=5,69 pour la classe 1 et Variable A (Actif) : Chi²=7,51 pour la
classe 7). Elles peuvent être perçues comme plus conflictuelles ou engagées.
Les échanges sur la taille ont ainsi laissé plus de places aux locuteurs peu

JADT’ 18

591

actifs avec des discussions plus consensuelles moins conflictuelles que pour
des représentations moins stabilisées. Cette hypothèse renvoie alors à la
structure possible de cette représentation sociale construite autour d’un
noyau central stable qui exigerait des études complémentaires pour être
confirmée.
3.4. Repérage de discours individuels par l’analyse factorielle de
correspondance (AFC)
Le partitionnement effectué permet aussi de repérer des individus dont les
discours sont différents (fig. 3) grâce à une AFC réalisée à la suite d’une CDH
de Reinert. Dans ce cas, deux individus se détachent principalement : 17 et
37. Le retour au texte permet de confirmer ce repérage. L’autre intérêt est
aussi de souligner des regroupements d’individus différents de leur
rattachement à un focus groups. L’AFC, en mettant en évidence des
ensembles de locuteurs, propose une approche qui dépasse la frontière de
chaque focus groups pour proposer une analyse de l’individu dans les
groupes.

Figure 3. AFC à partir de la CDH présentant les variables (F1/F2, 19,57 % de l’inertie).

4. Conclusion
Les méthodes lexicométriques utilisées pour analyser des focus groups

592

JADT’ 18

dépendent notamment de la partition du corpus effectuée en amont. Dans
notre recherche, l’association de variables à chaque intervention de locuteur a
permis de repérer des sous-groupes d’individus à l’intérieur des focus
groups, des discours d’individus isolés ou des sous-groupes associés à
plusieurs focus groups qui n’apparaissaient pas de façon évidente pendant
les échanges. Cette approche a cependant certaines limites. D’abord, la
procédure automatisée d’association des variables utilisée dans cette
expérimentation ne permet pas de repérer l’évolution des thèmes pendant la
discussion, une variable repérant les tours de paroles aurait alors été
nécessaire. Ensuite, le repérage des individus s’est fait sur une AFC qui
explique une faible part de la variance (19,57 %) et les causes de la singularité
des discours est ainsi difficile à associer à la CDH. Enfin, d’autres méthodes
auraient pu être investies (analyse des antiprofils, spécificités, similitudes,
etc.).
Sans remplacer l’analyse conversationnelle qui apporte des nuances
spécifiques, certaines méthodes lexicométriques peuvent ainsi permettre de
comprendre le corpus différemment et compléter la compréhension de ce
type de données riches et profondes en dépassant notamment la frontière de
chaque focus groups et faciliter une approche transversale du sens.
Remerciements : merci à Pascal Marchand, Pierre Ratinaud et Lucie Loubère
pour leur initiation à la lexicométrie et à Iramuteq.
References
Bengough, T., Bovet E., Bécherraz C., Schlegel S., Burnand B., et Pidoux, V.
(2015). Swiss family physicians’ perceptions and attitudes towards
knowledge translation practices. BMC Family Practice, décembre: 1–12.
Bonneau, J., and Dister, A. (2010). Logométrie et modélisation des
interactions discursives, l’exemple des entretiens semi-directifs. Journées
internationales d’Analyse statistique de Données Textuelles, pp. 253–264.
Brangier, E., Barcenilla, J., Bornet, C., Roussel, B., Vivian, R., and Bost, A.
(2015). Prospective ergonomics in the ideation of hydrogen energy usages.
In Proceedings 19th Triennial Congress of the IEA. Melbourne, pp. 1–2.
Dransfield, E., Morrot, G., Martin, J.-F., and Ngapo, T.-M. (2004). The
application of a text clustering statistical analysis to aid the interpretation
of focus group interviews. Food Quality and Preference, 15(4): 477–488.
Duchesne, S., and Haegel, F. (2014). L’entretien collectif. Armand Colin. Paris.
Duschesne, S., Florence Haegel, Elizabeth FRAZER, Virginie Van Ingelgom,
and Guillaume Garcia, André-Paul Frognier. (2010). Europe between
integration and globalisation social differences and national frames in the
analysis of focus groups conducted in France, francophone Belgium and
the United Kingdom. Politique Européenne, 30(1): 67–105.

JADT’ 18

593

Freitas, E. A. M., and Luis, M. A. V. (2015). Perception of students about
alcohol consumption and illicit drugs. Acta Paul Enferm., 28(5): 408–414.
Gresillon, E., and Marianne Cohen, Julien Lefour, Lydie Goeldner et Laurent
Simon. (2012). Les trames vertes et bleues habitantes : un cheminement
entre pratiques et représentations. L’exemple de la ville de Paris (France).
Développement Durable et Territoires, 3: 2-17.
Guerrero, L., Guàrdia, M., and Xicola, J. (2009). Consumer-driven definition
of traditional food products and innovation in traditional foods. A
qualitative cross-cultural study. Appetite, 52(2): 345–354.
Hulin, T. (2013). Enseigner l’activité « écriture collaborative ». Tic&société,
7(1): 89–116.
Jovchelovitch, S. (2004). Contextualiser les focus groups : comprendre les
groupes et les cultures dans la recherche sur les représentations. Bulletin
de Psychologie, 57(3): 245–261.
Lebart, L., and Salem, A. (1988). Analyse statistique des données textuelles.
Dunod. Paris.
Peyrat-Guillard, D., Lancelot Miltgen, C., et Welcomer, S. (2014). Analysing
conversational data with computer-aided content analysis: The
importance of data partitioning. Journées internationales d’Analyse
statistique des Données Textuelles, pp. 519–530.
Pélissier, D. (2016), Pourquoi et comment utiliser la lexicométrie pour
l’analyse de focus groups ?, Présence numérique des organisations,
11/07/2016.
Ratinaud, P. (2009). Iramuteq. Lerass.
Ratinaud, P., and Marchand, P. (2015). Des mondes lexicaux aux
représentations sociales. Une première approche des thématiques dans les
débats à l’Assemblée nationale (1998-2014). Mots. Les Langages du Politique,
108(2): 57–77.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique :
application à l’analyse lexicale par contexte. Les Cahiers de L’analyse Des
Données, 8(2): 187–198.
Rouré, H., and Reinert, M. (1993). Analyse d’un entretien à l’aide d’une
méthode d’analyse lexicale. Journées internationales d’Analyse statistique de
Données Textuelles. ENST, Paris, pp. 418-42

594

JADT’ 18

Using the First Axis of a Correspondence Analysis
as an Analytical Tool. Application to
Establish and Define an Orality Gradient for
Genres of Medieval French Texts
Bénédicte Pincemin1, Céline Guillot-Barbance2, Alexei Lavrentiev3
Univ. Lyon, CNRS, IHRIM UMR5317 - benedicte dot pincemin at ens-lyon dot fr; celine dot guillot
at ens-lyon dot fr; alexei dot lavrentev at ens-lyon dot fr

Abstract
Our corpus of medieval French texts is divided into 59 discourse units (DUs)
which cross text genres and spoken vs non spoken text chunks (as tagged
with q and sp TEI tags). A correspondence analysis (CA) performed on
selected POS tags indicates orality as the main dimension of variation across
DUs. We then design several methodological paths to investigate this
gradient as computed by the CA first axis. Bootstrap is used to check the
stability of observations; gradient-ordered barplots provide both a synthetic
and analytic view of the correlation of any variable with the gradient; a way
is also found to characterize the gradient poles (here, more-oral or less-oral
poles) not only with the POS used for the CA analysis, but also with words,
in order to get a more precise and lexical description. This methodology
could be transposed to other data with a potential gradient structure.
Keywords: textometry, Old French, represented speech, spoken genres,
methodology, correspondence analysis, 1D model, data visualization, XML
TEI, TXM software, DtmVic software.
1. Linguistic issue and preparation of textual data
We investigate spoken language features of Medieval French in a corpus
composed of 137 texts (4 million tokens), taken from the Base de français
médiéval1. The corpus is annotated with part-of-speech (POS) tags at the
word level; speech quotation chunks and speech turns are marked up using
TEI XML tags at an intermediate level between sentences and paragraphs;
and every text can be situated in a 32-genre typology (Guillot et al., 2017).
Our hypothesis is that the features of orality may be related to text chunks
representing speech, and also to text genres, as for instance some text genres

1

Base de français médieval: http://bfm.ens-lyon.fr

JADT’ 18

595

are intended for oral performance. In order to perform a textometric analysis
(Lebart et al. 1998) on our XML-TEI annotated data, we use the TXM opensource corpus analysis platform (Heiden, 2010; Heiden et al., 2010)2.
We divide our corpus into 59 discourse units (DUs) obtained by splitting
every genre into parts which represent speech on the one hand, and the
remaining parts on the other hand (some text genres have no spoken
passages). Discourse unit labels, like q_rbrefLn for instance, combine four
pieces of information: (i) the first letter is either q for quoted speech chunks,
sp for speech turns, or z for remaining (non oral) chunks; (ii) then we have
the short name of the text genre (here, rbref means “récit bref”, i. e. short
narrative); (iii) the uppercase letter stands for the domain3; (iv) the last
character indicates whether this DU is represented in our corpus by one (1),
two (2) or more (n) texts. We linguistically represent our texts with the POS
tags4 they use5. The reliability of POS tags was measured in a previous study
(Guillot et al., 2015) for a subset of 7 texts in which tags had been manually
checked. For the present analysis, we eliminate low-frequency POS tags (freq.
< 1 500), which include many high error rate tags and do not carry much
weight into the quantitative analysis. For the remaining high error rate tags
(with more than 25% wrong assignments), we measure their influence on the
correspondence analysis (CA) by checking their contribution to the first axis.
Then we remove the proper nouns category (NOMpro) which shows both
high error rate and high contribution to the first axis (14.66 %).
A new correspondence analysis enables two additional improvements from a
linguistic perspective. We remove compound determiners (DETcom,
PRE.DETcom, like ledit) as they emerged at the end of the 13th century, so that
they introduce a singular and substantial diachronic effect (high
contributions on the first axis). Moreover, the second axis describes mainly
the association between psalms (z_psautierRn) and possessive adjectives
(ADJpos): this corresponds to very specific phrases with some distinctive
nouns (la meie aneme, li miens Deus, la tue misericorde), and the adjective is
equivalent to a possessive determiner in other contexts, so we merge the two
categories (DETADJpos). We finally get a contingency table crossing 59 DUs
with 33 POS tags to explore with a CA.

Textometry Project and TXM software: http://textometrie.org
There are 6 domains: literature (L), education (D for “didactique”), religion (R),
history (H), law (J for “juridique”), practical acts (P).
4 We use the Cattex2009 tagset, designed for Old French: http://bfm.enslyon.fr/spip.php?article176.
5 We exclude punctuations, editorial markup and foreign words. CQL query:
[fropos!="PON.*|ETR|OUT|RED"]
2
3

596

JADT’ 18

2. Linguistic and methodological results from correspondence analysis
Our study reveals that the first axis can in fact be interpreted as an orality
gradient. The factorial map (Fig. 1) shows z_ DUs on the left hand side of the
first axis, opposed to q_ and sp_ DUs on the right hand side. Some genres
intended for oral performance go to the right with speech chunks (especially
plays –dramatiqueL, dramatiqueR), whereas genres related to written
processing (especially practical acts (P): charters, etc.) go to the left with outof-speech chunks. As this opposition matches the first axis, orality appears as
the first contrastive dimension for Old French (as regards POS frequencies),
as it is in Biber’s experiences with English (Biber, 1988), with the same kind
of linguistic features (Table 1). Then, as a second result, DUs can be sorted
according to their degree of orality, from “less oral” to “more oral” (see
Appendix6). Peculiar positions (for didactic dialogs or psalms for instance)
can be explained by a formal use of language given by the rules of the genre.
The linguistic analysis of the DU gradient is detailed in (Guillot-Barbance et
al., 2017)7.

Figure 1. CA map of the 59 DUs (TXM). 21 DUs with low representation quality (cosine
squared to 1 × 2 plane < 0.3) and no significant contribution to this plane (ctrb1 < 2 % & ctrb2
< 2 %) have been filtered out (macro CAfilter.groovy), so that the figure is clearer.

Appendix is available online as a related file of this paper in HAL archive:
https://halshs.archives-ouvertes.fr/halshs-01759219
7 Improvements made to the statistical processing in 2018 (management of the
second axis with ADJpos and DETpos merging, confidence ellipses) strengthen the
linguistic interpretation published in 2017, no significant change is observed on
gradient given by the first axis, according to the four zones defined by the analysis,
except for a few points which are not related to this axis (low cosine squared).
6

JADT’ 18

597

Figure 2. CA map of the 17 DUs with the largest confidence ellipses (DtmVic). The two
largest ones (q_proverbesD2, q_lapidaireD2) couldn’t be drawn; the following three largest
ones (q_commentaireD1, q_dialogueD2, q_sermentJ1) show that these DU positions cannot be
interpreted; then other smaller ellipses indicate that the 54 remaining DU positions on axes #1
and 2 are stable.

Table 1. The eight POS with the highest contributions on the first axis, for both sides.
“Less oral” pole
“More oral” pole
personal pronoun
PROper
preposition
PRE
general adverb
ADVgen
common noun
NOMcom
negative adverb
+
definite ADVneg
PRE.DETdef preposition
finite verb
VERcjg
determiner
VERppe
adverbial pronoun (en, y)
PROadv
past participle
DETdef
DETADJpos possessive determiner or
definite determiner
DETcar
adjective
CONsub
cardinal determiner
VERppa
subordinating conjunction
VERinf
present participle
CONcoo
infinitive verb
coordinating conjunction

A bootstrap validation (Dupuis & Lebart, 2008, Lebart & Piron, 2016) is
applied to evaluate the stability of DU positions on the first axis (Figure 2).
Sizes of ellipses in the 1×2 map are correlated to sizes of DUs: the fewer the
words there are in the DU, the less data the statistics process, and the greater
is the confidence ellipse (Table 1). Only five DUs are ascribed a big ellipse
which shows their uncertain position (Figure 2): all of them are DUs from
about ten words to about a hundred words, which are DUs for very singular
linguistic usages, and are neither representative nor relevant for this overall
linguistic analysis. The orality gradient is then confirmed throughout a

598

JADT’ 18

statistic validation on our data.
The 2D factorial map provides a synthetic and efficient visualization. The
second axis display reveals that the “more oral” pole is more compact, more
consistent, than the “less oral” pole, which is more heterogeneous (the cosine
squared values corroborate this). But what we want to stress in this
methodological paper, is that the main linguistic result is uniquely provided
by the interpretation of the first axis. Benzécri has illustrated the same kind of
approach by using a 1D CA to reveal the hierarchy of characters in Racine’s
Phèdre (1981 : 68). This method emphasizes the analytic power of CA, which
separates the data (by the mathematical means of Singular Value
Decomposition) into “deep” components (factors), just as a prism breaks
light up into its constituent spectral colors. Despite its main use as a 2D
illustration of a corpus structure in the textual data analysis field, CA is much
more than a suggestive visualization or a quick sketch.
3. Complementary tools to analyse 1D gradient in textual data
We now test new means to gain insight into the causation of this gradient in
our data.
3.1. Gradient-ordered barplot

Figure 3. Gradient-ordered specificity barplot for Personal Pronoun, as example of a POS
which is correlated to the first axis. For readability reasons, the height of specificity bars is
limited to 20.

The first method we propose is to visualize the evolution of POS frequencies
according to the orality gradient using a specificity bar-plot chart where the
DU order on the x-axis is given by the DU order on the first CA axis: this
display visually reveals how much a POS is correlated with speech or non
speech features, and details its affinity with each DU. For instance, personal
pronouns are typical for the more-oral pole: this is displayed as a rising
profile (Figure 3), and one can easily find out which DU have an outlying use
of this POS. Whereas a POS like adjectives (Figure 4), which is not correlated
to the orality gradient, gets a chart with no overall pattern.

JADT’ 18

599

Figure 4. Gradient-ordered specificity barplot for adjectives, as example of a POS which is not
correlated to the first axis. For readability reasons, the height of specificity bars is limited to 20.

3.2. Back-to-text close reading by getting representative words for each side
of the first axis
The second methodological innovation concerns obtaining lexical
information about orality characteristics in our texts. We select two sets of
DUs based on their cosine squared scores for the first CA axis in order to
represent the more-oral (cos21 > 0.4) and less-oral (cos21 > 0.35) poles (Table
2). The cos2 thresholds are adjusted to get two balanced sets with enough
different DUs to get an adequate representativeness. Then, a specificity
computation, which statistically characterizes the distribution of words into
these two sets, reveals lexical features for more oral and less oral poles,
showing typical words as they can be read in texts. Light is thus shed on the
quantitative result throughqualitative observations.
Table 2.
Representative DUs
Less-oral pole
z_journalJ2
z_plaidsP1
z_commentaireD1
z_diversP1
z_registreP2
z_lettreH1
z_dialogueD2
z_rvoyageL1

Table 3a. Adjectives
typical for the less-oral
subcorpus

Table 3b. Adjectives
typical for the more-oral
subcorpus

More-oral pole
q_romanLn
sp_dramatiqueR1
q_rbrefLn
q_bestiaireD2
sp_dramatiqueLn
q_lyriqueLn
z_lyriqueLn
q_chroniqueHn
sp_lyriqueLn
q_hagiographieRn
q_romanDn
q_mémoiresHn

Our example sheds light on the uses of adjective: whereas adjectives are not
related to the orality gradient as a category (Figure 4), they have strong
associations at a lexical level (Table 3). Represented speech makes much use

600

JADT’ 18

of terms of address introducing speech turns (bel, douz – and their formal
variants: biaus, biax, etc.), and evaluative adjectives (grant, mal, boen). For the
less-oral pole, there are more POS tagging errors; adjectives are more diverse
and often associated with a subset of DUs, for instance present, saint, maistre
are typical of two texts.
4. Conclusion
In this contribution, we have shown several ways to take into account the
limits of real data, especially textual data: managing the POS tags reliability
(§1), validation process to identify where data is lacking (§2), refining
morphosyntatic based analysis with lexical information (§3). But our main
objective is to establish a methodology in order to reveal and study any
gradient-like deep structuration of data. A simple seriation (as illustrated in
Dupuis & Lebart, 2008) could provide the same results for the first step, as it
generates the same ordered view of the data. But CA gives much more
information, qualifying the relation of each variable to the gradient with
indicators like contributions and cosines squared. Interpretation can go
further: CA coordinates are controlled with bootstrap and confidence
ellipses, gradient-ordered barplot visualizations are efficient to analyse in
detail the relationship of any individual variable to the overall gradient, and
the gradient poles can be illustrated by words, which add a concrete and
textual account for the deep structure. Thus, on our corpus of French
medieval texts, we discover that orality is the main contrastive dimension
and that it characterizes represented speech as well as text genres. The
methodology could be applied to other data, and is already entirely
implemented using tools freely available to the scientific community.
This research has benefited from the PaLaFra ANR-DFG project (ANR-14-FRAL0006), for corpus extension and POS evaluation. We are also very grateful to
Ludovic Lebart, for his inspiring comments on a preliminary presentation of this
research, and for DtmVic software, which has evolved in order to take into account
the quantitative particularities of our data.
References
Benzécri J.-P. et al. (1981). Pratique de l’Analyse des données, tome 3. Linguistique
& lexicologie. Dunod, Bordas, Paris.
Biber D. (1988). Variation across speech and writing. Cambridge University
Press.
Dupuis F., Lebart L. (2008). Visualisation, validation et sériation. Application
à un corpus de textes médiévaux. In Heiden S. and Pincemin B., eds, Actes
JADT 2008, Presses univ. de Lyon: 433-444.
Guillot C., Heiden S., Lavrentiev A., Pincemin B. (2015). L’oral représenté

JADT’ 18

601

dans un corpus de français médiéval (9e-15e) : approche contrastive et
outillée de la variation diasystémique. In Kragh K. J. and Lindschouw J.,
eds, Les variations diasystémiques et leurs interdépendances dans les langues
romanes -Actes du Colloque DIA II, Éd. de linguistique et de philologie,
Strasbourg : 15-28.
Guillot-Barbance C., Pincemin B., Lavrentiev A. (2017). Représentation de
l’oral en français médiéval et genres textuels, Langages, 208: 53-68.
Heiden S. (2010). The TXM Platform: Building Open-Source Textual Analysis
Software Compatible with the TEI Encoding Scheme. In Otoguro R. et al.,
eds, PACLIC24, Waseda Univ., Sendai : 389-398.
Heiden S., Magué J.-Ph., Pincemin B. (2010). TXM : Une plateforme logicielle
open-source pour la textométrie – conception et développement. In
Bolasco S. et al., eds, Statistical Analysis of Textual Data -Proceedings of JADT
2010, Edizioni Univ. di Lettere Economia Diritto, Rome : 1021-1031.
Lebart L., Piron M. (2016). Pratique de l’Analyse de Données Numériques et
Textuelles avec Dtm-Vic. L2C, http://www.dtmvic.com.
Lebart L., Salem A., Berry L. (1998). Exploring Textual Data. Kluwer academic
pub., Boston.

602

JADT’ 18

Explorer les désaccords dans les fils de discussion du
Wikipédia francophone
Céline Poudat
Université Côte d’Azur, CNRS, BCL, France – poudat@unice.fr

Abstract
This article concentrates on the exploration of French Wikipedia talk pages,
with a focus on conflicts. We developed a typology of speech acts expressing
disagreement, including direct and explicit forms (je ne suis pas d’accord / je
suis en désaccord) as well as indirect acts, which are besides the most
widespread. Disagreement is indeed a negative reaction that may threaten
the face of the addressee. For this reason, disagreements are rather expressed
indirectly in order to protect faces in interaction. A subset of the Wikiconflits
corpus (Poudat et al., 2016) was annotated according to the typology and we
carried on a primary exploration of the data using statistical methods.
Résumé
Cette étude se concentre sur l’exploration de l'encyclopédie Wikipédia, l'un
des plus gros succès du Web 2.0, et spécifiquement sur l’exploration de ses
discussions éditoriales, avec un intérêt particulier pour les conflits. Nous
nous intéressons aux actes de langage exprimant le désaccord, de son
expression la plus directe et la plus explicite (je ne suis pas d’accord / je suis en
désaccord) à ses formes les plus indirectes, et d’ailleurs les plus usuelles ; le
désaccord est effectivement plutôt exprimé de manière indirecte pour
préserver sa face et celle de l’autre. Nous présentons la typologie que nous
avons développée et nous l’appliquons à un sous-ensemble du corpus
Wikiconflits que nous avons développé (Poudat et al., 2016). Le corpus
annoté est ensuite exploré avec les méthodes de l’ADT et nous restituons
certaines de ses caractéristiques.
Keywords: Wikipedia, CMC corpora, Conflicts, Disagreements, Pragmatics,
Semantic Annotation, Text statistics
1. Introduction
Cette étude se concentre sur l’exploration de l'un des plus gros succès du
Web 2.0 : l’encyclopédie Wikipédia, qui rassemble des milliers de
contributeurs à travers le monde, mais qui demeure paradoxalement peu
observée par les études de linguistique, certainement du fait de la complexité

JADT’ 18

603

de l’objet, qui multiplie les versions, les types de pages et les genres textuels.
Nous nous intéressons spécifiquement aux fils des pages de discussion du
Wikipédia francophone, avec un intérêt particulier pour les conflits. Plutôt
abordés par les sciences sociales (cf. Kittur et Kraut, 2008, 2010; Auray et al.,
2009, Sumi et al., 2011, Borra et al., 2014), les conflits dans Wikipédia ont été
peu décrits d’un point de vue linguistique. Nous proposons de les décrire au
moyen d’une annotation en actes de langage, en distinguant entre marqueurs
du (dés)accord et marqueurs du conflit : si tout désaccord ne tourne pas au
conflit, un conflit nait souvent d’un désaccord. Deux entreprises d’annotation
des interactions conflictuelles de Wikipédia ont été menées ces dernières
années (Bender et al., 2011, Fershke et al., 2012), mais elles ne portaient pas
sur le français, et se positionnaient dans un cadre distinct. La présente
communication se concentre spécifiquement sur l’exploration des marqueurs
du désaccord dans Wikipédia, de son expression la plus directe et la plus
explicite (je ne suis pas d’accord / je suis en désaccord) à ses formes les plus
indirectes, et d’ailleurs les plus usuelles ; le désaccord est effectivement
plutôt exprimé de manière indirecte pour préserver sa face et celle de l’autre.
Après avoir présenté le corpus de travail (2.), nous décrirons la typologie
exploratoire que nous avons développée et les marqueurs que nous avons
annotés manuellement (3.). Nous présenterons enfin certaines des régularités
observées (4.).
2. Wikiconflits : pages et fils conflictuels
Le corpus de travail sur lequel se fonde notre étude comprend un sousensemble du corpus Wikiconflits (Poudat et al., 2016), à savoir l’ensemble des
discussions autour de six articles ayant été identifiés par Wikipédia comme
conflictuels : Igor et Grichka Bogdanoff, Chiropratique, Éolienne, Histoire de la
logique, Psychanalyse et Quotient intellectuel. La conflictualité de chaque fil a
été évaluée et annotée avec une variable à trois modalités : si les fils non
conflictuels sont catégorisés C0, C1 signale la présence d’un désaccord et C2
la présence d’un conflit sur le fil.

page

Tableau 1 : Corpus de travail
tokens
messages
Fils C0

Fils C1

Fils C2

Bogdanoff

73864

493

30

16

20

Chiropratique

29919

226

5

3

12

Éolienne

13454

152

2

7

0

Histoire de la logique

3358

46

4

2

0

Psychanalyse

102338

878

54

39

34

Quotient intellectuel

20059

170

10

20

12

604

JADT’ 18

Désaccords et conflits sont deux formes d’affrontement verbal, à cette
différence que le désaccord est un acte réactif qui exprime une réaction
négative relative à une assertion préalablement exprimée (KerbratOrecchioni, 2016) tandis que le conflit est un acte agressif, qui implique la
présence d’au moins une séquence attaque-réplique caractérisée par l’usage de
marqueurs de violence verbale et d’actes de langage agressifs pour la face de
l’allocutaire (Poudat et Ho-Dac, 2018). Ces définitions doivent être précisées
relativement au genre très particulier qu’incarne la discussion Wikipédia, qui
a pour fonction majeure de permettre aux rédacteurs de l’article de se
coordonner et de clarifier leurs éventuels différends. L’article encyclopédique
est ainsi le premier terrain de coopération entre les contributeurs, la
discussion faisant plutôt office de coulisses de la rédaction – beaucoup
d’utilisateurs réguliers de Wikipédia méconnaissent d’ailleurs l’existence de
ces discussions. En d’autres termes, l’article est le genre premier, la
discussion faisant figure de genre lié ou non autonome. Les désaccords et les
conflits que l’on y observe s’adossent ainsi sur l’article, ce qui nous a amenée
par exemple à observer qu’un désaccord pouvait porter sur un passage de
l’article, considéré dans ce cas comme une assertion contestable. De la même
manière, un conflit peut prendre sa source au cours de la rédaction de
l’article, via une suppression ou un retour en arrière litigieux, qui pourra
donner lieu à l’écriture d’une réplique agressive sur la page de discussion.
Notons que nous écartons de notre étude les conflits non verbaux et autres
guerres d’édition, largement observés par les sciences sociales.
Les fils catégorisés C1 portent la trace verbale d’un désaccord tandis que les
fils étiquetés C2 contiennent au moins une attaque manifeste de la face de
l’un des contributeurs du fil. Cette annotation ne va bien sûr pas de soi et
nous a souvent demandé d’arbitrer entre le contenu du message et son
positionnement dans le fil d’interaction. Un message peut ainsi exprimer un
désaccord ou être agressif sans recevoir de réponse, tandis qu’un
contributeur peut être en désaccord avec un point de vue existant qui n’est
pas pour autant celui de l’un de ses co-énonciateurs. Nous n’avons retenu
que les désaccords ou les attaques orientés vers le(s) co-énonciateur(s) / corédacteurs(s), en ce sens qu’un passage très agressif envers un tiers auteur ou
article par exemple, ne sera pas été considéré comme conflictuel.

JADT’ 18

605

3. Le désaccord comme acte de langage : types et marqueurs
Nous nous sommes ensuite concentrée sur l’annotation manuelle des actes de
langage exprimant le désaccord en développant une typologie adaptée aux
caractéristiques du corpus de travail. Le désaccord étant un acte exprimant
une réaction négative, il est potentiellement menaçant pour la face de
l’allocutaire auquel il s’adresse. C’est pourquoi il est généralement exprimé
de manière indirecte. Les chiffres sont éloquents dans notre corpus : 82% des
actes exprimant le désaccord relevés sont indirects, tandis que près de la
moitié des désaccords exprimés directement sont adoucis ou minimisés.
Les deux grands types d’expression indirecte du désaccord les plus
récurrents que nous avons observés consistent à (i) recourir à la concession
pour mettre en scène un accord partiel et (ii) exprimer son désaccord en se
posant explicitement comme source évaluative (personnellement, je ne pense pas
que… ; j’avoue ne pas comprendre, etc.). Comme nous le signalons dans le
tableau 2, nous avons choisi d’annoter les concessions accompagnées d’un
accord explicite comme « Ok, mais des solutions existent (développement de pales
furtives absorbant les ondes radars) » (discussion Éolienne), ce qui explique
peut-être pourquoi au final nous n’en obtenons qu’un petit nombre (9 occ.).
L’expression du désaccord indirect semble privilégier significativement les
actes secondaires de l’incompréhension (48 occ.) et de l’expression d’une
opinion (29 occ.). À titre de comparaison, nous avons systématiquement
annoté les manifestations d’accord explicites rencontrées. Contrairement au
désaccord, l’accord est dans notre culture un acte positif pour la face de
l’allocutaire. Peu employé de manière indirecte, il est plutôt intensifié
qu’atténué (je suis tout à fait d’accord). On relève 57 actes d’accord explicite
dans le corpus ; à titre de comparaison, on rencontre trois fois plus de formes
exprimant un désaccord, ce qui est probablement dû à la dimension
conflictuelle du corpus. Il nous faut enfin souligner que plus des deux tiers
des 270 fils de discussion considérés ne contenaient aucune des formes
observées, ce qui n’est pas surprenant : un quart des fils ne contiennent qu’un
seul message tandis que nous avons conservé les fils catégorisés harmonieux
à titre de contraste.

Attributs
polarité

Valeurs
accord
désaccord
explicite

type
implicite

Tableau 2 : Typologie du désaccord
Exemples
je suis d’accord
Je suis contre l’avis de X
Accord explicite : je suis d’accord, je suis pour X, favorable à
X, tout à fait de votre avis, je suis de ton avis, OK pour X…
Désaccord explicite : pas d’accord, en désaccord, je ne suis
pas favorable, je suis contre, totalement contre
Voir acte indirect.

606

JADT’ 18
oui / non

atténuation
indirect

non
concession

Concéder

avis

émotion

Se poser
comme
source
évaluative

doute
Incompréhension
assertion
négative
forte

Atténuation d’un accord explicite : je suis assez d’accord
Atténuation d’un désaccord explicite : Nous sommes en
désaccord (mineur) sur un point (mineur)
Seuls les actes d’accord explicite accompagnés d’une
concession ont été retenus.
D'accord pour refuser le paragraphe ajouté à partir d'arkiv ;
en revanche la suppression de la participation d'AR à la
mission ne me semblait pas déraisonnable (discussion
Bogdanoff)
« Personnellement, je pense que non », je ne crois pas, je
ne pense pas…
mots-clés : personnellement, pense, crois, trouve
émotion (rare dans le corpus pour exprimer le désaccord)
j'ai été personnellement choqué par les affirmations
gratuites comme "de gauche/de droite" dès le début de l'article,
que je pense tout à fait intempestives et parfaitement corrélées à
la hauteur du QI du contributeur et aux théories raciales de
Rushton, (discussion QI)
Je doute de la pertinence de ce passage dans cet article.
mots-clés : certain, sûr, doute
Je ne vois pas bien quel rapport ta source a avec ce constat.
(discussion Psychanalyse)
Encore une fois, je ne comprends pas le problème.
Ce n'est pas du tout une question de vocabulaire
secondaire (discussion Bogdanoff)

4. Analyses
Le corpus annoté a ensuite été soumis à différentes méthodes de l’analyse de
données textuelles afin d’explorer ses caractéristiques et de mettre en
évidence les relations entre les types de désaccord et la situation du fil,
harmonieuse, dissonante ou conflictuelle. Comme le montre la Figure 1, les
fils identifiés comme lieux d’un désaccord (C1) sont ceux qui contiennent le
nombre le plus significatif de marqueurs d’accord et de désaccord. Au
contraire, les fils identifiés comme conflictuels contiennent significativement
moins de marques d’accord explicite et de marques de désaccord. Nous voilà
donc rassurée par la cohérence de notre annotation.

JADT’ 18

607

Figure 1 : Ventilation des types d’accord et de désaccord d’un type de fil à l’autre (données
Hyperbase Web)

Afin d’évaluer plus précisément la structure de l’ensemble des annotations
apposées sur les textes, nous avons réalisé une Analyse en Composantes
Principales (ACP) sur la table des décomptes d’annotations en prenant le fil
de discussion comme unité textuelle. Nous avons dû procéder à certains
ajustements, (i) en écartant les fils qui ne contenaient aucune annotation ; (ii)
en isolant certaines variables trop marginales (i.e. 2 occ. de la valeur émotion)
et (iii) en distinguant entre les observations restantes celles qui seront
utilisées comme variables actives ou comme variables supplémentaires.
Ainsi, les variables ayant le trait atténuation ont été intégrées à titre illustratif.
Au total, l’ACP a été réalisé sur un ensemble de taille restreinte, à savoir 98
fils * 8 variables actives (et 13 variables supplémentaires). De manière
intéressante, l’ACP met en évidence la présence d’un facteur taille, c’est-àdire que toutes les observations sont corrélées positivement entre elles et se
regroupent donc du même côté du premier axe factoriel. Certains fils de
discussion ont des valeurs fortes pour toutes les variables, tandis que
d’autres ont des valeurs faibles pour toutes les variables.
Si l’on s’intéresse aux facteurs 2 et 3 (Figure 2) sur lesquels on projette le
degré de conflictualité et les pages du corpus à titre illustratif, on observe une
opposition entre accord et désaccord, et dans une moindre mesure entre
explicite et implicite sur le facteur 2. Accords et actes explicites seraient du
côté de l’harmonie et du désaccord tandis que les désaccords en général et les
désaccords indirects en particulier seraient plus caractéristiques du conflit.
Cette dernière remarque, qui devra être éprouvée et confirmée sur des jeux
de données plus importants, nous semble intéressante : est-ce que les
marqueurs implicites du désaccord vont de pair avec les marqueurs du
conflit ? Y a-t-il une corrélation négative entre expression explicite du

608

JADT’ 18

désaccord et attaques personnelles ?

Figure 2 : Facteurs 2 et 3 de l’ACP – 98 fils * 8 variables actives – Dtm-vic

5. Conclusion et perspectives
Nous avons ainsi proposé une première typologie des actes exprimant le
désaccord en français ; cette typologie a été développée dans le cadre d’un
projet plus général d’exploration des conflits dans Wikipédia. Une seconde
typologie, centrée sur les marqueurs de violence verbale et supposément
caractéristique du conflit, est en cours de développement et viendra faire
système avec la typologie du désaccord pour mettre en évidence les
caractéristiques des interactions conflictuelles dans Wikipédia et dans les
CMC.
En ce qui concerne l’annotation présentée, un guide est actuellement en cours
de rédaction ; chaque marqueur sera validé et évalué au moyen d’un kappa
de Cohen. La typologie est encore en cours d’amélioration ; ainsi une
troisième forme d’expression indirecte du désaccord que nous avions
observée consiste à le neutraliser en déplaçant le focus sur une proposition
ou une suggestion, i.e. un acte de langage positif (ne vaudrait-il pas mieux… ?
Il faudrait peut-être d’abord définir ce qu’on entend par..). Ce type de séquence,
plus complexe à identifier car plus ambigu, est en cours d’intégration.
Enfin, reste à mettre en œuvre des parcours interprétatifs adaptés pour
explorer ce type de données annotées avec nos méthodes ADT ; c’est aussi
l’une des pistes que nous poursuivons ces dernières années, dans nos travaux
(Poudat et Landragin, 2017) et dans le cadre du consortium CORLI.

JADT’ 18

609

Références
Auray, N., Hurault-Plantet, M., Poudat, C., & Jacquemin, B. (2009). La
négociation des points de vue : une cartographie sociale des conflits et des
querelles dans le Wikipédia francophone. In Réseaux 2/2009, n° 154: 15-50.
Bender E.M., Morgan J.T., Oxley M., Zachry M., Hutchinson B., Marin, A.,
Ostendorf, M. (2011). Annotating Social Acts: Authority Claims and
Alignment Moves in Wikipedia Talk Pages. In Proceedings of the Workshop
on Languages in Social Media (pp. 48–57). Stroudsburg, PA, USA:
Association for Computational Linguistics.
Borra E., Weltevrede E., Ciuccarelli P., Kaltenbrunner A., Laniado D., Magni
G., Venturini T. (2014). Contropedia - the Analysis and Visualization of
Controversies in Wikipedia Articles. In Proceedings of The International
Symposium on Open Collaboration (pp. 34:1–34:1). New York, NY, USA.
Ferschke O., Gurevych I., Chebotar Y. (2012). Behind the Article: Recognizing
Dialog Acts in Wikipedia Talk Pages. In Proceedings of the 13th Conference
of the European Chapter of the Association for Computational Linguistics (pp.
777–786). Stroudsburg, PA, USA: Association for Computational
Linguistics.
Kerbrat-Orecchioni, C. (2016). Le désaccord, réaction « non préférée » ? Le cas
des débats présidentiels. Cahiers de praxématique, (67).
Poudat C. et Ho-Dac L.-M. (2018). Désaccords et conflits dans le Wikipédia
francophone. In Travaux linguistiques du Cerlico, Presses Universitaires de
Rennes (sous presse).
Poudat C. et Landragin F. (2017). Explorer un corpus textuel. Méthodes –
Pratiques – Outils. Collection Champs linguistiques, De Boeck, Louvain-laNeuve.
Poudat C., Grabar N., Paloque-Berges C., Chanier T. et Kun J. (2017).
Wikiconflits : un corpus de discussions éditoriales conflictuelles du
Wikipédia francophone. In Wigham, C.R & Ledegen, G., Corpus de
communication médiée par les réseaux : construction, structuration, analyse.
Collection Humanités numériques. Paris : L’Harmattan, pp. 19-36.
Sumi, R., Yasseri, T., Rung, A., Kornai, A., & Kertész, J. (2011). Edit wars in
Wikipedia. In: Proceedings of the ACM WebSci'11, Koblenz, Germany.
pp. 1–3.

610

JADT’ 18

Textometric Exploitation of Coreference-annotated
Corpora with TXM: Methodological Choices and
First Outcomes
Matthieu Quignard1, Serge Heiden2, Frédéric Landragin3, Matthieu Decorde2
ICAR, CNRS, University of Lyon – matthieu.quignard@ens-lyon.fr
IHRIM, ENS Lyon, CNRS, University of Lyon – {slh,matthieu.decorde}@ens-lyon.fr
3Lattice, CNRS, ENS Paris, University Sorbonne Nouvelle, PSL Research University,
USPC – frederic.landragin@ens.fr
1

2

Abstract
In this article we present a set of measures – some of which can lead to
specific visualisations – with the objective to enrich the possibilities of
exploration and exploitation of annotated data, and in particular coreference
chains. We first present a specific use of the well-known concordancer, which
is here adapted to present the elements of a coreference chain. We then
present a histogram generator that allows for example to display the
distribution of the various coreference chains of a text, given a value from the
annotated properties. Finally, we present what we call progress diagrams,
whose purpose is to display the progress of each chain throughout the text.
We conclude on the interest of these (interactive) modes of visualization in
order to make the annotation phase more controlled and more effective.
Résumé
Nous présentons dans cet article un ensemble de mesures – dont certaines
peuvent amener à des visualisations spécifiques – dont l’objectif est
d’enrichir les possibilités d’exploration et d’exploitation des données
annotées, en particulier quand il s’agit de chaînes de coréférences. Nous
présentons tout d’abord une utilisation adaptée de l’outil bien connu qu’est
le concordancier, en n’affichant que les maillons d’une chaîne choisie. Puis
nous montrons un générateur d’histogramme qui permet par exemple
d’afficher la répartition des chaînes de coréférences d’un texte à partir d’une
propriété annotée. Nous montrons enfin ce que nous appelons des
diagrammes de progression, dont le but est d’afficher les avancées au fur et à
mesure du texte des chaînes de coréférences qu’il contient. Nous concluons
sur l’intérêt de ces modes (interactifs) de visualisation pour rendre la phase
d’annotation plus maîtrisée et plus efficace.
Keywords: coreference chain, corpus annotation, annotation tool,
visualisation tool, exploration tool, statistical analysis of textual data.

JADT’ 18

611

1. Introduction
The manual annotation of a textual corpus with referring expressions
(Charolles, 2002) and coreference chains (Schnedecker, 1997, Landragin &
Schnedecker, 2014) requires adapted tools. A coreference chain can cover the
whole text; it is therefore a linguistic object for which the existing means of
visualization and exploration are few and often perfectible. The MMAX2 tool
(Müller & Strube, 2006) allows for visualizing the links between referring
expressions using arrows which link markables. The GLOZZ tool (Mathet &
Wildlöcher, 2009) offers several means of visualization: with arrows like
MMAX2, or with a specific marking in the margin or the middle of the text.
The ANALEC tool (Landragin et al., 2012) and its specific extension for
coreference chains (Landragin, 2016) proposes a graphic metaphor based on
the succession of coloured dots. This allows the analyst to configure visual
parameters, for instance the colour which can be linked to any of the
annotated properties. This type of visualization makes it possible to see at a
glance the structural differences between the different reference chains of a
text. That must be useful to the analyst, in addition to manual explorations
and finer linguistic analyses.
2. Linguistic objects and methodology
In the continuity of previous works (Heiden, 2010; Landragin, 2016), we
present here a set of measures – some of which can lead to specific
visualisations – with the objective to enrich the possibilities of exploration
and exploitation of annotated data. We focus in particular on annotations
which concern discursive phenomena like coreference, i.e., annotations
which are necessarily described within two levels: 1. markable, group of
contiguous words to which is assigned some labels, using for instance a
feature structure; 2. set of markables, or links between markables, as is it the
case for any chain of annotations: anaphoric chains, textual organizers chains,
textual structure elements chains, etc. A feature structure can also be
assigned at level 2, i.e., to the set or to the links.
3. A concordancer adapted to annotations chains
As a first visualization mode, we reuse the very classic concordancer to
display the elements which constitute a coreference chain. The use of such a
visualization tool, which is well established in the community of corpus
exploration (Poudat & Landragin, 2017), seemed natural for visualizing
chains of annotations. The last version of TXM (Heiden, 2010) thus includes a
concordancer which makes it possible to display in a column all the elements
(e.g. referring expressions) of a chain (e.g. coreference chain), with left and
right contexts for each elements. Compared to MMAX2 (Müller & Strube,

612

JADT’ 18

2006) and GLOZZ (Mathet & Wildlöcher, 2009) visualisation choices, i.e.
arrows linking marquables which are displayed directly on the text, this
concordancer has the advantage of regrouping all the relevant information in
a small graphic space.

Fig 1: Concordancer with the elements of a coreference chain, dedicated to a character named
“Caillette”.

Fig. 1 shows the list of all referring expression to the character ‘Caillette’.
Sorted in the textual order, the concordancer shows the alternation of the use
of proper nouns, pronouns, possessives, etc. This concordancer may also be
sorted along a given property of the marquable, e.g. its POS label. This
representation may then be exploited to see whether the POS annotation is
consistent or not.
4. Histograms for visualising distributions of annotations chains
A second mode of visualization, also very traditional, is the histogram (bar
plot). The user can select one or several properties – the determination of the
referring expressions, for instance, or the type of referent – and launch
calculations on their occurrences: cross-counts, correlation computation and
so on. TXM now includes a histogram generator, which allows for example to
display the distribution of coreference chains throughout the text, as well as
the distribution of chains according to the number of referring expressions
they include. These calculations and their associated visualizations provide
TXM with integrated functionalities which required in other state-of-the-art
tools the development of scripts, in order to export the relevant data and
exploit them in an external tool like a spreadsheet.

JADT’ 18

613

Figure 2 compares the distribution of grammatical categories of referring
expressions in three texts. Although all texts are all encyclopedical ones, the
Discourse from Bossuet shows a particular profile, with a high number of
proper nouns (GN.NAM).

Fig 2: Comparative barplots of grammatical categories usage by reference units in three texts:
Bossuet, “Discours sur l’histoire universelle” (1681), Diderot, “Essais sur la peinture” (17591766), Montesquieu, “Esprit des lois” (1755).

5. Progression charts for annotations chains
A third (new) mode of visualization consists to graphically show the
progress of each chain throughout the text. The principle is simple, but the
possibilities of exploration and exploitation of the generated graph are
numerous. In a two-dimensional chart the abscissa of which represents the
linearity of the text, chains are displayed point by point (cf. Fig. 3): each
occurrence of a referring expression increases by one notch the ordinate of
the corresponding point. The resulting broken lines are all ascending but can
considerably vary in their areas of progression and flat areas.
When they are visualized simultaneously, it is possible to detect the parts of

614

JADT’ 18

the text where several referents are competitors, or on the contrary those
where several referents appear alternately. Zooming (in and out) as well as
focussing features allows for visualizing the characteristics of each point,
thus enriching the exploration possibilities of these progression chart and the
underlying coreference chains.

Fig 3: Progression graph of the main coreference chains at the beginning of “Essais sur la
peinture” from Denis Diderot. The dots highlighted with symbols correspond to referring
expressions with low accessibility.

6. Discussion
The common points of these new visualization modes is not only to propose
visual representations which are easy to understand (and possibly
interactive, when it is possible to modify on the fly one of the properties), to
allow the visualization of these representations directly in TXM, with no
need to export annotated data and to use external tools, but also to facilitate
the detection by the analyst of intruders, outliers and deviant examples. For
instance potential annotation errors: it can be the case for a referring
expression which has nothing to do in the currently visualised chain. It may
be a peak or a suspect flat in one of the generated histograms. It may be a
zone with a very high slope (or a very long flat) in a progression diagram. In
all three cases, the analyst can directly access the suspicious annotation, in
order to verify it and of course to modify it. The integration of the
measurements and their visualizations in TXM allows this immediate return
to the corpus annotation phase. This is particularly effective when the corpus
is being annotated manually.

JADT’ 18

615

7. Conclusion and future works
One can say that it is by annotating that we can see the mistakes we make,
but we still need appropriate tools to detect these errors. With the new
possibilities of interaction that we propose here, we hope that we are taking a
significant step in this direction. The first tests which we have carried out
demonstrated the relevance of our approach.
References
Charolles M. (2002). La référence et les expressions référentielles en français.
Ophrys, Paris, France.
Heiden S. (2010). The TXM Platform: Building Open-Source Textual Analysis
Software Compatible with the TEI Encoding Scheme. Proceedings of the
24th Pacific Asia Conference on Language, Information and Computation, Nov.
2010. Sendai, Japan, Institute for Digital Enhancement of Cognitive
Development, Waseda University, pp. 389-398, available at
halshs.archives-ouvertes.fr/halshs-00549764.
Landragin F. (2016). Conception d’un outil de visualisation et d’exploration
de chaînes de coréférences. Statistical Analysis of Textual Data – Proceedings
of 13th International Conference Journées d’Analyse statistique des Données
Textuelles (JADT 2016), Nice, France, pp. 109-120.
Landragin F., Poibeau T. and Victorri B. (2012). ANALEC: a New Tool for the
Dynamic Annotation of Textual Data. Proceedings of LREC 2012, Istanbul,
Turkey, pp. 357-362.
Landragin F. and Schnedecker C., editors (2014). Les chaînes de référence.
Volume 195 of the Langages journal, Armand Colin, Paris, France.
Müller C. and Strube M. (2006). Multi-level annotation of linguistic data with
MMAX2. In Braun S., Kohn K. and Mukherjee J., editors, Corpus
Technology and Language Pedagogy: New Resources, New Tools, New Methods,
Peter Lang, Frankfurt, Germany.
Poudat, C. and Landragin, F. (2017). Explorer un corpus textuel : méthodes,
pratiques, outils. Champs Linguistiques. De Boeck Supérieur : Louvain-laNeuve.
Schnedecker C. (1997). Nom propre et chaîne de référence. Klincksieck, Paris,
France.
Widlöcher A. and Mathet Y. (2012). The Glozz platform: a corpus annotation
and mining tool. In Concolato C. and Schmitz P, editors, Proceedings of the
ACM Symposium on Document Engineering (DocEng’12), Paris, France, pp.
171-180.

616

JADT’ 18

Amélioration de la précision et de la vitesse de
l’algorithme de classification de la méthode Reinert
dans IRaMuTeQ
Pierre Ratinaud
LERASS, Université de Toulouse – ratinaud@univ-tlse2.fr

Abstract
This work presents a proposal to improve the accuracy and the speed of
execution of the divisive hierarchical clustering (DHC) algorithm used by the
Reinert method implemented in the IRaMuTeQ free software. The DHC of
the Reinert method is a serie of bi-partitions on a presence / absence matrix
that intersects text segments and words. In the original version of this
algorithm, after each partition, the largest of the remaining classes is selected
to be split. We propose to replace the selection mode of the classes to be
partitioned by a criteria of homogeneity. The complete rewriting of this part
of the IRaMuTeQ code has also been an opportunity to improve its speed by
implementing part of the code in C ++ and paralleling the procedure. An
experiment carried out on 6 corpora shows that the new algorithm based on
these principles is indeed more precise and faster.
Résumé
Ce travail présente une proposition d’amélioration de la précision et de la
vitesse d’exécution de l’algorithme de classification hiérarchique descendante
(CHD) utilisé par la méthode Reinert implémentée dans le logiciel libre
IRaMuTeQ. La CHD de la méthode Reinert est une série de bi-partitions de
matrices de présence / absence qui croise des segments de texte et des formes.
Dans la version originale de cet algorithme, après chaque partition, la plus
grande des classes restantes est sélectionnée pour être à son tour coupée en
deux. Nous proposons de remplacer le mode de sélection des classes à
partitionner par un critère d’homogénéité. La ré-écriture complète de cette
partie du code d’IRaMuTeQ a également été l’occasion d’une amélioration de
sa célérité par l’implémentation d’une partie du code en C++ et la
parallélisation de la procédure. Une expérimentation menée sur 6 corpus
permet de constater que le nouvel algorithme reposant sur ces principes est
effectivement plus précis et plus rapide.
Keywords: méthode Reinert, classification hiérarchique descendante,
IraMuTeQ, précision

JADT’ 18

617

1. Introduction
La méthode Reinert a pour objectif de faire émerger les différentes
thématiques qui traversent un corpus textuel. Sa plus grande originalité est
sûrement l’algorithme de classification hiérarchique descendante (CHD)
proposé par Reinert (1983). Après avoir rappelé les différentes étapes de ce
type d’analyse, nous proposerons une modification de cet algorithme de
classification dans l’objectif d’améliorer la précision de l’ensemble de la
procédure. Le changement proposé concerne le critère de sélection des sousmatrices après chacune des partitions. La description de cette nouvelle
procédure est complétée par une expérimentation sur 6 corpus en français et
en anglais permettant de comparer la nouvelle version de l’algorithme avec
l’ancienne. Les résultats que nous présentons attestent effectivement d’une
augmentation de la précision de l’algorithme, dont la ré-écriture à également
permis une augmentation de la vitesse d’exécution. Avant d’entamer cette
présentation, il nous semble toutefois nécessaire de rappeler que la CHD
n’est pas la seule particularité de la méthode Reinert.
2. Des corpus aux matrices
Une autre originalité de cette procédure est l’unité utilisée dans la
classification. Dans la plupart des situations, la classification ne porte pas sur
les textes dans leur ensemble, mais sur une granularité inférieure. Les unités
classées sont des segments de texte. Dans le logiciel IRaMuTeQ (Ratinaud,
2014; Ratinaud & Marchand, 2012), la taille de ces segments est fixée par
défaut à 40 occurrences et leur découpage tient compte de la ponctuation. La
règle de découpage essaie donc de proposer des unités de taille homogène
(autour de 40 occurrences) et de respecter le découpage « naturel » des textes
marqué par la ponctuation. Une seconde originalité qu’il convient de préciser
est la distinction opérée entre formes pleines et mots outils. Dans ces
analyses, la plupart du temps, seules les formes pleines (verbes, adverbes,
adjectifs et substantifs) sont considérées. Les corpus peuvent alors être
représentés sous la forme de matrices qui croisent les segments de texte et les
formes pleines. Les cellules de ces matrices marquent la présence ou
l’absence des formes dans les segments en codant 1 la présence et 0 l’absence.
Le tableau 1 présente une telle matrice pour un corpus composé de 10
segments de texte (notés i1 à i10) et de 9 formes (notées j1 à j9).

618

JADT’ 18

Tableau 1 : Exemple d’une matrice croisant des segments de texte (en ligne) et les formes (en
colonne)
J1 J2
J3 J4 J5 J6 J7 J8 J9

I1

1

1

1

1

0

0

0

0

0

I2

0

0

0

0

1

1

1

1

1

I3

0

0

1

0

1

0

1

0

0

I4

1

0

1

0

1

0

0

0

1

I5

0

0

1

0

1

0

1

0

0

I6

1

1

1

1

0

0

0

0

1

I7
I8

0
1

0
0

0
1

0
0

1
1

1
0

1
0

1
0

0
0

I9

0

0

1

0

1

0

1

0

1

I10

0

0

1

0

1

0

1

0

0

La matrice présentée dans le tableau 1 est un exemple très simplifié de ce
qu’il se passe dans la réalité. Les matrices générées sur des corpus textuels
sont beaucoup plus grandes et beaucoup plus « creuses » (la proportion de 1
est très faible dans la matrice). Nous noterons N le nombre total de 1 dans la
matrice. L’objectif de la classification est de proposer une réorganisation de
cette matrice en sous-groupes de segments qui maximisent les propriétés
suivantes :
n) Les segments regroupés doivent être homogènes entre eux : la
méthode doit réunir les segments de texte qui se ressemblent, c’est-àdire les segments qui ont tendance à contenir les mêmes mots.
o) Les ensembles doivent être hétérogènes entre eux : les groupes de
segments constitués doivent être les plus différents possibles.
L’illustration 1 propose un découpage de la matrice présentée dans le
Tableau 1 en 4 classes qui respectent ces critères.

I1
I6

J1 J2 J3 J4 J5 J6 J7 J8 J9
1 1 1 1 0 0 0 0 0

J1 J2 J3 J4 J5 J6 J7 J8 J9

I8
1 0 1 0 1 0 0 0 0
1 0 1 0 1 0 0 0 1
I4
1 1 1 1 0 0 0 0 1
Illustration 1 : Découpage de la matrice du Tableau 1 en 4 classes

La « qualité » de cette solution peut être déterminée par le calcul du chi2/N
du tableau réduit (Reinert, 1983).
Dans cet exemple, la solution optimale serait obtenue en séparant les lignes
i6, i4, i2 et i9 de leur classe d’appartenance pour les laisser former leur propre
classe. La solution à 8 classes obtenue résumerait alors l’intégralité de
l’information contenue dans la matrice du Tableau 1.

JADT’ 18

619
Tableau 2 : Tableau réduit de la classification de l’illustration 1

J1 J2 J3 J4 J5 J6 J7 J8 J9
Σ [i1,i6]

2

2

2

2

0

0

0

0

1

Σ [i4,i8]

2

0

2

0

2

0

0

0

1

Σ [i9,i3,i5,i10]

0

0

4

0

4

0

4

0

1

Σ [i2,i7]

0

0

0

0

2

2

2

2

1

3. La CHD de la méthode Reinert
Rappelons que la méthode permettant de construire automatiquement ces
classes s’appuie sur une série de bi-partitions reposant chacune sur une
analyse factorielle des correspondances (AFC). La première coupure est
obtenue en cherchant le long du premier facteur de cette AFC les deux sousmatrices qui maximisent le chi2/N du tableau réduit. La partition produite
est améliorée en inversant chacune des lignes du tableau d’une classe à
l’autre et en recalculant le chi2/N du tableau réduit. Toutes les inversions qui
augmentent la valeur du chi2/N sont conservées. Cette étape boucle jusqu’à
ce que plus aucune inversion n’augmente cette valeur. Une dernière étape
consiste à retirer les formes (les colonnes) statistiquement sous-représentées
dans les matrices (sur la base d’un chi2).
Cette procédure (bi-partition de la matrice, inversion des lignes, suppression
des colonnes) constitue une des partitions de la CHD. La CHD dans son
ensemble réalisera cette procédure autant de fois que nécessaire pour
atteindre le nombre de classes terminales paramétré. Il faut n-1 partition(s)
pour constituer n classe(s) terminale(s). Après chacune de ces partitions, dans
sa formulation d’origine, l’algorithme sélectionne la plus grande des classes
constituées (c’est-à-dire celle qui contient le plus de lignes) pour lui faire à
son tour subir une partition.
Le tableau 3 présente, de façon très caricaturale, une matrice pour laquelle
cette démarche ne conduit pas à un résultat satisfaisant. Si nous soumettions
cette matrice à la CHD précédemment décrite, la première partition
conduirait à la création d’une classe constituée des lignes i1, i2 et i3 (notée
[i1,i2,i3]) et d’une autre constituée des lignes i4 et i5 (notée [i4,i5]). La première
de ces classes étant la plus grande, elle serait sélectionnée pour, à son tour,
subir une partition. Or, il est évident ici qu’il n’y a plus aucune information à
extraire de cette matrice, les lignes étant toutes identiques. Seule la séparation
des lignes i4 et i5 est, dans cet exemple, susceptible d’augmenter la qualité du
résultat. Pour cela, il aurait donc fallu sélectionner la classe restante la plus
hétérogène ([i4,i5]) plutôt que de sélectionner la plus grande ([i1,i2,i3]).

620

JADT’ 18
Tableau 3 : Une matrice problématique

J1 J2 J3 J 4 J5
i1

1

i2

1 1 0 0 0

1 0

0 0

i3

1 1 0 0 0

i4

0 0 1 1 0

i5

0 0 0 1 1

Il convient donc de percevoir que dans la version actuellement disponible de
cette méthode, l’algorithme de classification fait l’hypothèse que la matrice la
plus grande est également la plus hétérogène.
Nous pensons que certains corpus ne respectent pas cette propriété et qu’il
est tout à fait possible qu’à différents moments d’une classification, la plus
grande des matrices restantes ne soit pas la plus hétérogène.
4. Une nouvelle solution pour l’enchaînement des partitions
Il apparaît alors pertinent de pouvoir tester, après chacune des phases de
partition, l’homogénéité des matrices restantes de façon à sélectionner la plus
hétérogène. Comme le calcul de l’analyse factorielle des correspondances
nécessaire à chaque partition permet de déterminer le chi2 de la matrice dans
son ensemble, nous avons utilisé cette propriété pour revoir le déroulement
de l’algorithme. Dans cette nouvelle version, après chaque partition, l’AFC et
le chi2 des deux matrices générées sont calculés a priori. Pour chacune de ces
matrices, nous déterminons un indice d’homogénéité qui tient compte du
chi2 de la matrice, de sa taille et du nombre total de formes. Ce critère relève
de la formule suivante :

Il s’agit donc de multiplier le chi2 de la matrice par le ratio de 1 qu’elle
contient.
Cette méthode permet de ne plus supposer que la matrice la plus grande est
la plus hétérogène mais de tester cette hétérogénéité. Elle a pour désavantage
de nécessiter le calcul systématique de l’AFC sur pratiquement toutes les
matrices produites. Sans autre modification, cette procédure serait beaucoup
plus lente que la version précédente de l’algorithme. Dans l’objectif
d’accélérer ces analyses, la ré-écriture théorique de l’algorithme s’est
accompagnée d’une recherche de gain de performances qui a ici suivi deux
directions :

JADT’ 18

621



Les parties les plus gourmandes en calcul ont été ré-écrites en C++
par l’intermédiaire des packages Rccp (Eddelbuettel et al., 2017) et
RcppEigen (Bates, Eddelbuettel, Francois, & Yixuan, 2017) de R. Les
parties concernées sont la recherche de la partition qui maximise le
Chi2/N après l’AFC et le reclassement des lignes.
 Ces deux parties étant une suite de calculs de chi2 sur la base d’une
seule matrice, il a été possible de les paralléliser pour profiter de la
nature multi-coeur de la plupart des processeurs modernes. Les
calculs sont donc potentiellement distribués aux différents
cœurs/threads de la machine par l’intermédiaire des packages
Parallel et DoParallel (Calaway, Microsoft Corporation, Weston, &
Tenenbaum, 2017) de R.
Ces changements ont en fait nécessité la réécriture complète de l’algorithme
de la méthode Reinert dans IRaMuTeQ.
5. Expérimentation
De façon à tester les bénéfices apportés par cette nouvelle procédure, en
termes de précision et de rapidité, une expérimentation sur 6 corpus
différents a été réalisée. Nous avons associé à des corpus de grandes tailles
(les plus susceptibles de présenter des disproportions dans les thématiques
qu’ils contiennent) un corpus de taille plus réduite. Les caractéristiques de
ces corpus sont présentées dans le Tableau 4.
Tableau 4 : description des corpus utilisés dans l’expérimentation

Le corpus dataconf correspond à des titres et à des résumés de conférences
du domaine de l’informatique, il est uniquement en anglais. 20Newsgroup1
est un corpus également en anglais qui réunit 20 listes de discussions sur des
thématiques très diverses (Lang, 1995). lemondefr est un corpus d’articles du

1

http://qwone.com/~jason/20Newsgroups/

622

JADT’ 18

site web du monde en ligne2, il est en français. Ssm, pour « same sex
marriage », est un corpus d’articles de presse américaine et anglaise sur la
thématique du mariage entre personnes de même sexe. Il a été constitué par
Nathalie Paton. AN2011 correspond à l’année 2011 de la retranscription des
débats à l’assemblée nationale française (Ratinaud & Marchand, 2015). Enfin,
le corpus noté LRU regroupe 100 articles de la presse quotidienne française
sur la thématique de la loi liberté et responsabilité des universités.
L’expérimentation consiste donc à faire subir les deux versions de
l’algorithme de classification aux matrices extraites de ces corpus et à
comparer la qualité des résultats obtenus. Le nombre de classes terminales a
été fixé à 100 pour les “gros” corpus et à 30 pour le “petit”. Dans un cas,
l’algorithme utilisera le critère de taille pour sélectionner les matrices à
partitionner et dans l’autre il utilisera le critère d’homogénéité. Les résultats
se présentent sous la forme de graphiques qui montrent l’évolution de la
quantité d’information extraite après chacune des partitions. La valeur
renvoyée est celle du Chi2/N du tableau réduit des classes. Dans les
graphiques de l’illustration 2, les courbes rouges représentent les valeurs
obtenues avec l’ancienne version de l’algorithme (notée Reinert) et les
courbes bleues les valeurs obtenues avec la nouvelle version (notée
Reinert++). Une valeur supérieure correspond à une meilleure qualité de la
partition. Le graphique en barres présente le pourcentage d’augmentation ou
de diminution de la qualité de la partition du nouvel algorithme en prenant
l’ancien comme référence. Les barres vertes signalent une augmentation de la
qualité et les barres rouges une diminution. Pour la nouvelle version de
l’algorithme, 6 cœurs ont été alloués à la procédure3.
Ces résultats montrent assez clairement que la nouvelle version de
l’algorithme augmente dans la majorité des cas la précision de la
classification. Ils permettent également de percevoir que ce gain de qualité
est lié à la distribution des thématiques dans les corpus. Tous les corpus ne
profitent donc pas de cette évolution de la même façon. Il faut également
noter que sur le corpus LRU, il n’y a pratiquement pas de différences entre
les deux méthodes. La perte de précision de 1 à 3 % à différents moments de
la classification sur ce corpus est tout à fait négligeable et doit être attribuée à
des différences d’arrondis entre le code en R et le code en C++. À l’opposé,
certains corpus, comme 20newsgroup, présentent des gains de précision qui
peuvent atteindre 15 %.

http://www.lemonde.fr
Ces tests ont été réalisés sur un macbook pro 11,3 équipé d’un processeur
intel i7-4960HQ
2
3

JADT’ 18

623

Illustration 2 : Comparaison des résultats entre l’ancienne version (Reinert) et la nouvelle
version (Reinert++) de l’algorithme de classification

L’illustration 3 montre que sur les corpus conséquents, le gain de
performances introduit par le passage au C++ et à la parallélisation est
compris entre un facteur 4 et un facteur 6. Autrement dit, ce nouvel
algorithme est jusqu’à 6 fois plus rapide sur la machine sur laquelle ces
calculs ont été réalisés.

624

JADT’ 18

12000

7,0

4,3

4,9

4,8

6,0
5,0
4,0

6000

3,0

4000

2,0

2000

1,0

Gain de performance

Temps en seconde

10000
8000

6,0

5,6

Reinert
Reinert++
Gain de performance

0,0

0
AN2011

dataconf 20newsgroup lemondefr

ssm

Illustration 3 : comparaison des temps d’analyse entre l’ancienne version (Reinert) et la
nouvelle version (Reinert++) de l’algorithme

6. Conclusion
Dans ce travail, nous proposons une nouvelle formalisation de la procédure
de classification hiérarchique descendante de la méthode Reinert. Partant de
l’hypothèse que dans certains corpus et à certains moments de ces
classifications, la classe la plus hétérogène n’est pas forcément la plus grande,
nous proposons de substituer le critère du choix de l’enchaînement des
matrices d’un critère de taille à un critère d’homogénéité. Les résultats d’une
expérimentation sur 6 corpus montrent que les corpus volumineux profitent
effectivement de ce changement. Ces résultats sont aussi une invitation à
continuer les investigations sur cette méthode. Cette procédure sera
implémentée dans la prochaine version du logiciel IRaMuTeQ. L’utilisation
du critère d’homogénéité sera optionnelle, de façon à permettre aux
utilisateurs de revenir à l’ancienne version.
Bibliographie
Bates, D., Eddelbuettel, D., Francois, R., and Yixuan, Q. (2017). RcppEigen:
« Rcpp » Integration for the « Eigen » Templated Linear Algebra Library
(Version 0.3.3.3.1). Consulté à l’adresse https://cran.rproject.org/web/packages/RcppEigen/index.html
Calaway, R., Microsoft Corporation, Weston, S., and Tenenbaum, D. (2017).
doParallel: Foreach Parallel Adaptor for the « parallel » Package
(Version 1.0.11). Consulté à l’adresse https://cran.rproject.org/web/packages/doParallel/index.html
Eddelbuettel, D., Francois, R., Allaire, J. J., Ushey, K., Kou, Q., Russell, N., …
Chambers, J. (2017). Rcpp: Seamless R and C++ Integration (Version
0.12.14). Consulté à l’adresse https://cran.r-

JADT’ 18

625

project.org/web/packages/Rcpp/index.html
Lang, K. (1995). Newsweeder: Learning to filter netnews. In Proceedings of the
Twelfth International Conference on Machine Learning (p. 331-339).
Ratinaud, P. (2014). IRaMuTeQ : Interface de R pour les Analyses
Multidimensionnelles de Textes et de Questionnaires (Version 0.7 alpha
2) [Windows, GNU/Linux, Mac OS X]. Consulté à l’adresse
http://www.iramuteq.org
Ratinaud, P., and Marchand, P. (2012). Application de la méthode ALCESTE
à de « gros » corpus et stabilité des « mondes lexicaux » : analyse du
« CableGate » avec IRaMuTeQ. In Actes des 11eme Journées internationales
d’Analyse statistique des Données Textuelles (JADT 2012) (p. 835-844).
Liège, Belgique. Consulté à l’adresse http://lexicometrica.univparis3.fr/jadt/jadt2012/Communications/Ratinaud,%20Pierre%20et%20al
.%20-%20Application%20de%20la%20methode%20Alceste.pdf
Ratinaud, P., and Marchand, P. (2015). Des mondes lexicaux aux
représentations sociales. Une première approche des thématiques dans
les débats à l’Assemblée nationale (1998-2014). Mots. Les langages du
politique, 2015(108), 57-77.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique :
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, VIII(2), 187-198.
Reinert, M. (1990). ALCESTE : Une méthodologie d’analyse des données
textuelles et une application : Aurélia de Gérard de Nerval. Bulletin de
méthodologie sociologique, (26), 24-54.

626

JADT’ 18

Il parametro della frequenza tra paradossi e
antinomie: il caso dell’italiano scolastico
Luisa Revelli
Università della Valle d’Aosta– l.revelli@univda.it

Abstract
Emblem of a formal register, the linguistic variety proposed as a model in the
Italian school system ever since National Unity is characterized by a lasting
artificiality and a strong unwillingness to innovate, even within a frame of
progressive slow changes along its historical development. That's why lexical
frequencies recorded for “Scholastic Italian” can appear as inherently
inconsistent, contrasting with basic vocabulary, even contradictory compared
with other apparently similar Italian varieties. Consequently, to study their
configuration it's necessary to adopt analysis models capable to interpret
quantitative data (volume figures) in the light of the complexity of
paradigmatic relations between concurring solutions and of the composite
connections between number and type of meanings exhibited in current use.
By taking in consideration as a case study Scholastic Italian used by teachers
during the first 150 years of the national school system, and starting from the
data collected by the diachronic corpus of CoDiSV, the contribution aims at
verifying opportunities and criticalities of lexicometric analysis applied to
such a linguistic variety, that is addressed to an unsophisticated audience,
yet characterized by a specialized point of view; of high aspirations, but
influenced by educational needs; constantly evolving and yet always
recalcitrant to the solicitations of the contemporary language.
Riassunto
Emblema di un canone ‘antiparlato’, la varietà linguistica proposta a modello
nella scuola italiana a partire dall’Unità nazionale, pur presentando in
diacronia evidenti tratti evolutivi, si caratterizza per una duratura tendenza
all’artificiosità e per una marcata refrattarietà all’innovazione. Le frequenze
lessicali documentate nell’italiano scolastico possono, per queste ragioni,
risultare discordanti in rapporto a quelle del vocabolario di base, presentarsi
come intrinsecamente poco coerenti, contraddittorie rispetto alle evidenze
rintracciabili in varietà d’italiano apparentemente affini: lo studio delle loro
configurazioni richiede, pertanto, modelli di analisi capaci di interpretare i
dati quantitativi alla luce della complessità delle relazioni paradigmatiche tra
le potenziali soluzioni concorrenti nonché dei compositi rapporti tra numero

JADT’ 18

627

e tipologia delle accezioni testimoniate nei concreti impieghi contestuali.
Assumendo l’italiano scolastico proposto dagli insegnanti nei primi
centocinquant’anni di scuola nazionale a caso di studio, a partire dai dati
ricavati dal corpus diacronico del CoDiSV, il contributo si prefigge allora di
verificare opportunità e criticità poste dall’applicazione di parametri
lessicometrici a una varietà linguistica al contempo rivolta a un pubblico
ingenuo e connotata in prospettiva specialistica, di aspirazione elevata ma
condizionata da esigenze didascaliche, in costante evoluzione e ciò malgrado
costantemente recalcitrante rispetto alle sollecitazioni della lingua viva e
coeva.
Parole-chiave: italiano
vocabolario di base.

scolastico;

frequenza

lessicale;

lessicometria;

1. Introduzione
Al contempo rivolto a un pubblico ingenuo e connotato in prospettiva
specialistica, di aspirazione elevata ma condizionato da esigenze
didascaliche, in costante evoluzione e ciò malgrado costantemente
recalcitrante rispetto alle sollecitazioni della lingua viva e coeva, l’italiano
scolastico (d’ora in poi IS) proposto dagli insegnanti nei primi
centocinquant’anni di scuola nazionale sembra costituire un buon banco di
prova per far emergere le zone di criticità derivanti dall’applicazione di
parametri lessicometrici a varietà linguistiche poligenetiche e
costituzionalmente disomogenee1. Nell’IS, in effetti, un ideale di ricchezza
espressiva perseguito attraverso una marcata ostilità nei confronti di ogni
forma di ridondanza, ripetizione o generalità delle espressioni spinge verso
un’ostentata e ricercata variatio, ma la contemporanea esigenza di
alfabetizzare i giovani allievi orientandoli a privilegiare specifici membri di
serie sinonimiche ritenuti maggiormente corretti, appropriati o esornativi
tende, di fatto e in opposta direzione, a ridurre la gamma delle possibilità
espressive disponibili. La necessità di veicolare attraverso la lingua i saperi
disciplinari rende, d’altra parte, necessario l’uso di metalinguaggi, tecnicismi
e accezioni semantiche che sembrano destabilizzare ulteriormente il serbatoio
lessicale di riferimento dell’IS allontanandolo significativamente dal
vocabolario di base della lingua italiana. In che misura e in che termini
questo avvenga realmente è quanto ci si propone di verificare qui di seguito,
integrando i dati lessicometrici e quantitativi disponibili con alcune

1 Per un inquadramento delle caratteristiche, stabili ed evolutive, dell’IS si
rimanda a De Blasi 1993, Cortelazzo 1995, Benedetti G. e Serianni L. (2009), Revelli
2013.

628

JADT’ 18

riflessioni di natura qualitativa. Relativamente all’IS, la base lessicale presa a
riferimento è costituita da un lessico di frequenza elaborato da chi scrive
(Revelli 2013) a partire da un corpus iniziale di 830 quaderni di scuola
elementare redatti in area valdostana nel periodo compreso tra la fine del XIX
e i primi anni del XXI secolo. I 2.022 termini che compongono il vocabolario
di base sono stati individuati dopo che una selezione bilanciata dei
documenti, ripartiti in subcorpora cronologici ventennali, è stata sottoposta a
trattamento computazionale con lo scopo di identificare la dimensione della
variazione diacronica nei canoni linguistici proposti a modello da parte degli
insegnanti2. A fianco delle concordanze, è stato così ricavato in prima battuta
un vocabolario composto da 152.151 occorrenze (tokens), ricondotte a 18.898
forme (types) e 11.751 lemmi3. Un’ulteriore selezione ha poi dato luogo
all’identificazione dei 2.022 sostantivi, aggettivi e verbi considerati pancronici
perché stabilmente assestati nel vocabolario di base dell’italiano scolastico (d’ora
in poi VoBIS), in quanto testimoniati con più di cinque occorrenze in almeno
quattro dei sei repertori cronologici o in tre non consecutivi. Il termine di
paragone è costituito dall’edizione 2016 del Nuovo Vocabolario di base della
lingua italiana (d’ora in poi NVdB) di Isabella Chiari e Tullio de Mauro4, che
ripartisce le circa 7.000 parole statisticamente più frequenti e accessibili ai
parlanti italiani del XXI secolo nei tre serbatoi del lessico fondamentale (FO,
circa 2000 parole ad altissima frequenza usate nell’86% dei discorsi e dei
testi), del lessico ad alto uso (AU, circa 3.000 parole di uso frequente che
coprono il 6% delle occorrenze) e del lessico di alta disponibilità (AD, circa 2000
parole “usate solo in alcuni contesti ma comprensibili da tutti i parlanti e
percepite come aventi una disponibilità pari o perfino superiore alle parole di
maggior uso”). La scelta di fare riferimento a tale base, che comprende al
proprio interno anche le frequenze relative alle varietà parlate e si colloca
temporalmente in un periodo successivo a quello considerato per il lessico
scolastico, risponde all’esigenza di verificare se e in che misura il modello
scritto offerto da quest’ultimo possa aver inciso sulla configurazione dei
successivi usi concreti.

Le tipologie testuali prese in considerazione sono costituite dalle consegne
degli esercizi, dai titoli dei componimenti, da dettati, interventi correttivi, valutazioni
e giudizi documentati nei quaderni degli alunni.
3 Il vocabolario e le concordanze del corpus sono stati ricavati, previa
annotazione e lemmatizzazione, tramite il software T-LAB, ideato e distribuito da
Franco Lancia. Per un approfondimento a proposito dei principi adottati e la
metodologia seguita si rimanda a Revelli 2013.
4
https://www.internazionale.it/opinione/tullio-de-mauro/2016/12/23/il-nuovovocabolario-di-base-della-lingua-italiana.
2

JADT’ 18

629

2. Vocabolari di base a confronto: le frequenze nel NVdB e nel VoBIS
La comparazione del serbatoio lessicale dei due repertori presi a confronto
consente di compiere, in prima battuta, alcune osservazioni generali: dei 2022
lemmi del VoBIS, 1784 trovano riscontro nel NVdB, spartendosi per il 53%
nel serbatoio del lessico FO, per il 26% in quello di AU e per il 9% in quello di
AD. Senza entrare qui nel merito delle convergenze che accomunano i due
vocabolari, sembra comunque opportuno segnalare che dietro molti esempi
di apparente coincidenza delle distribuzioni di frequenza si celano in realtà
difformità significative, prevalentemente indotte dalla tendenza dell’IS al
restringimento o in alcuni casi anche alla rideterminazione semantica: fra le
molte parole che assumono specifici sensi scolastici (ad es. diario,
interrogazione, nota, pensierino, voto), alcune perdono del tutto l’ancoramento
ai significati di cui sono dotati nella lingua comune, come accaduto a tema,
passato a identificare non più un soggetto o argomento da trattare, ma invece
il prodotto di una specifica tipologia testuale.
Per ciò che concerne le 238 parole assenti nel NVdB (12%), esse possono
essere raggruppate in categorie utili a mettere fuoco diverse criticità relative
all’applicazione del parametro della frequenza comparativamente applicato.
Un primo, corposo gruppo che risulta esclusivo dell’IS è costituito da
logonimi caratteristici della nomenclatura metalinguistica dell’apparato
scolastico, del tipo alfabetico, apostrofo, coniugazione, preposizione, ecc.
Osserviamo che, malgrado il loro potenziale polisemico, molti di questi –
come coniugare, derivato, imperfetto, possessivo, primitivo - raggiungono
nell’ambito dell’IS frequenze molto elevate nel loro esclusivo ruolo di
etichette destinate alla riflessione metalinguistica5: la rappresentatività
quantitativa non implica quindi un contatto degli allievi con le diverse
accezioni di cui quegli stessi termini possono essere portatori, ma
corrisponde invece a un’insistita specializzazione motivata da esigenze
didascaliche. Un secondo gruppo è costituito da termini tipici dei contesti
d’insegnamento della letto-scrittura: si tratta principalmente di sostantivi che
fanno riferimento a referenti concreti ma di scarsa prominenza nella
quotidianità, la cui forma scritta guida e richiede la conoscenza di
convenzioni controintuitive eppure fondamentali per la corretta codifica e
decodifica ortografica. Citiamo a titolo di esempio parole come acquaio,
acquavite e acqueo, evidentemente introdotte non per stringente necessità
tematica quanto invece con scopo di consolidamento delle corrette
rappresentazioni grafematiche.
A scopi didattici legati agli insegnamenti disciplinari o più genericamente a

5 Ad esempio, dimostrativo - sempre preceduto da aggettivo o pronome - non entra
mai in combinazione con atto, gesto, ecc.

630

JADT’ 18

scelte tematiche caratteristiche del contesto educativo sono da imputare le
alte frequenze di diversi termini relativi all’ambito storico-geografico (legione,
vetta), di voci descrittive dell’universo naturale (arto, astro) e della vita rurale
(semina, vendemmia); di serie di verbi (castigare, disobbedire) di aggettivi
(diligente, ordinato) e di sostantivi astratti (umiltà, penitenza) appartenenti al
formulario tipico dell’educazione civica o morale e a quello della valutazione
scolastica. A differenza del NVdB, per la sua impostazione diacronica il
lemmario del VoBIS trova, d’altra parte, rappresentati numerosi arcaismi: si
tratta in alcuni casi di varianti formali oggi dismesse (ad es. annunziare per
annunciare,) o dispreferite (ubbidire per obbedire); di termini relativi a referenti
che i cambiamenti sociali dell’ultimo cinquantennio hanno reso superflui o
anacronistici (manto, ricamatrice); di membri di coppie o serie sinonimiche
superati o formali, che soltanto in ambito scolastico sono o sono stati più a
lungo privilegiati rispetto a concorrenti avvertiti dai parlanti come più attuali
(persuadere per convincere)6.
Proseguendo con le mancate corrispondenze nei due repertori, se l’assenza
nel NVdB di voci scolastiche un po’ leziose come diletto, garbato, vezzo e soave
risulta scontata, stupisce invece la mancata inclusione di termini che
appaiono stabili nel tempo e di diffusione panitaliana: è il caso di zoonimi
come bue, elefante, formica; di nomi di frutti usualmente presenti sulle tavole
degli italiani come fragola, noce e uva; di nomi concreti d’uso comune come
carezza, martello, ombrello. La mancanza di riscontri nel NVdB per termini di
questo tipo può essere solo in parte interpretato in una dimensione
propriamente sociolinguistica: pur essendo vero che - dato il pubblico cui si
orienta - l’IS fa più frequente riferimento a temi e referenti della cultura
materiale ed esperienziale di quanto non accada nelle varietà linguistiche
rivolte a e prodotte da parlanti adulti, è altrettanto vero che in linea teorica
tutti i vocaboli, a maggior ragione se accolti e veicolati dalla scuola,
dovrebbero rientrare in quel patrimonio di «parole che può accaderci di non
dire né tanto meno di scrivere mai o quasi mai, ma legate a oggetti, fatti,
esperienze ben noti a tutte le persone adulte nella vita quotidiana» (De
Mauro 1980: 148). Ci aspetteremmo quindi di trovare riscontri almeno
all’interno di quel serbatoio di parole di AD di cui tuttavia lo stesso De
Mauro ha in più occasioni dichiarato la natura sfuggente, non statistica ma

Ad es. bambagia, cagionare, figliolo, focolare, garzone, uscio. Proprio nell’ambito di
quest’ultima categoria il serbatoio dell’IS si differenzia d’altra parte in modo evidente
da quello del lessico corrente, privilegiando sistematicamente soluzioni assenti nel
NVdB, a scapito di quelle invece lì documentate e in molti casi dotate di marca d’uso
FO (ad es. appetito per fame, ardere per bruciare, sciupare per rovinare, ecc.).
6

JADT’ 18

631

congetturale7. E, in effetti, probabilmente neppure le analisi quantitative più
imponenti e minuziose possono aspirare ad azzerare inevitabili fattori di
imprevedibilità e accidentalità della frequenza. Nel caso qui preso a
campione, che relativamente all’IS non dispone di un corpus di partenza di
dimensioni del tutto soddisfacenti, lacune relative a termini rispetto ai quali
ci si aspetterebbe di avere riscontri si verificano anche capovolgendo la
prospettiva e quindi partendo dal lemmario del NVdB: pure ampliando
l’orizzonte all’intero vocabolario del corpus, a risultare mancanti non sono
soltanto termini marcati come AD, ma anche parole fondamentali che sono,
sì, probabilmente note ai bambini, ma non compaiono nel campione preso in
esame per ragioni meramente accidentali.8 Certamente motivate e
intenzionali sono invece specifiche tipologie di omissioni facilmente
identificabili come specifiche dell’IS: si tratta di neologismi e prestiti di lusso,
che i modelli dei maestri – forse in alcuni casi anche per ragioni ortografiche tendono a respingere quand’anche ormai stabilmente acclimatati nell’italiano
standard (jeans, quiz, smog); di termini riferiti a concetti ritenuti sconvenienti
per un pubblico acerbo (aborto, droga, sesso); di voci gergali, espressioni
volgari, insulti, improperi (coglione, culo, ruttare); di appellativi discriminatori
(ebreo, nano, negro) ma anche di parole prudenzialmente evitate perché
avvertite come potenzialmente faziose, propagandistiche o almeno
ideologicamente e politicamente orientate: su quest’ultimo aspetto, che
incarna l’intimità dei rapporti tra lessico, scuola, clima sociale e temperie
culturale non è tuttavia possibile compiere generalizzazioni, perché gli indizi
relativi alle diverse caratterizzazioni assunte dal fenomeno nel corso dei
tempi, anche molto recenti, richiedono di essere intercettati sulle frequenze
basse o inesistenti, piuttosto che su quelle elevate del lessico di base.
3. Conclusioni e prospettive
Come ci si è proposti di evidenziare, l’esame quali-quantitativo dell’IS
conferma che, pur presentando in diacronia tratti di ammodernamento, il
modello linguistico proposto dagli insegnanti risulta caratterizzato dallo

Nella Prefazione al NVdB è specificato che le parole di AD “sono state ricavate
partendo dalla lista di 2.300 parole di alta disponibilità del vecchio VdB e
sottoponendola a gruppi di studenti e studentesse universitari per eliminare le parole
non più avvertite come di maggior uso e per accogliere invece nuove parole avvertite
come di alta disponibilità”.
8 Esemplificativo dei margini di casualità può essere il caso degli etnici, che
mancano in alcuni casi al CoDiSV (ad es. cinese, iugoslavo) che pure ne documenta
moltissimi altri almeno apparentemente di analoga diffusione (ad es. giapponese,
inglese).
7

632

JADT’ 18

stabile impiego di termini estranei al vocabolario di base e dal parallelo
evitamento di termini correnti, ritenuti inadeguati o sconvenienti o più
semplicemente logorati da un uso reputato eccessivo. Lo studio dei dati
consente, poi, di rilevare un’abbondante presenza di logonimi ed etichette
tipici o esclusivi del metalinguaggio didattico e grammaticale, l’uso di hapax
spesso confinati nell’ambito di occasionali specifiche tipologie esercitative ma
per il loro ruolo strategico didatticamente irrinunciabili nonché per il ricorso
a un formulario al cui interno termini correnti assumono tramite fenomeni di
rideterminazione semantica accezioni differenti da quelle consuete,
specializzandosi in relazione a compiti e routines comunicative tipici del
contesto educativo. Le frequenze lessicali documentate nelle varietà dell’IS si
presentano in parte, per queste ragioni, come intrinsecamente poco coerenti,
discordanti in rapporto a quelle del vocabolario di base, contraddittorie
rispetto alle evidenze rintracciabili in varietà d’italiano apparentemente
affini: lo studio delle loro configurazioni richiede, pertanto, modelli di analisi
capaci di interpretare i dati quantitativi alla luce della complessità delle
relazioni paradigmatiche tra le potenziali soluzioni concorrenti nonché dei
compositi rapporti tra numero e tipologia delle accezioni testimoniate nei
concreti impieghi contestuali. In questa direzione, in parte già esplorata in
particolare negli studi di taglio psicolinguistico e glottodidattico dedicati ai
processi della comprensione e alla leggibilità dei testi, sembra che un
raffronto comparativo tra il lessico dell’IS e quello del VdB condotto in modo
sistematico su corpora cronologicamente armonizzati possa fornire ulteriori
linee di ricerca in almeno due specifici ambiti d’indagine.
Un primo, di prospettiva più propriamente acquisizionale, andrebbe
finalizzato a verificare gli effettivi esiti della protratta esposizione in età
scolare alla percentuale di parole dell’IS che risulta estranea al vocabolario di
base: in questa direzione, tenuto conto della natura incrementale e adattiva
degli apprendimenti lessicali ma anche dell’effetto di evanescenza che la
mancata pratica può esercitare sulle competenze possedute, si potrebbe
tentare di rispondere a domande del tipo: quanto incide effettivamente
l’insistenza con cui un termine è presente nell’input offerto nell’ambito
dell’IS sul suo effettivo impiego nei domini da questo distinti e
successivamente sperimentati? In che misura la concettualizzazione relativa
una determinata accezione di un termine veicolata dall’insegnamento può
condizionare (positivamente o negativamente) la successiva acquisizione di
significati ulteriori e diversi per quello stesso termine? In che termini le
soluzioni preferenziali e le scelte paradigmatiche proposte dall’IS risultano
vincenti, almeno a livello di competenza ricettiva, nella concorrenza con le
analisi statistiche che i parlanti sperimentano su altre varietà e in contesti
potenzialmente più pregnanti? E in questo senso, quanto può essere

JADT’ 18

633

percepito come autorevole, significativo, dotato di rilevanza comunicativa il
modello lessicale scolastico in un Paese in cui l’italiano è diventato lingua
materna per la gran parte dei cittadini e la concorrenza di input – non
soltanto lessicale - proveniente da fonti alternative alla scuola appare
quantitativamente strabordante?
Un secondo ambito d’indagine, al precedente correlato ma di prospettiva
principalmente lessicografica, potrebbe invece essere indirizzato ad esplorare
l’ipotesi che una parte del vocabolario scolastico di base possa essere considerata
denominatore comune delle competenze lessicali possedute dai parlanti
adulti alfabetizzati, e venire impiegata soprattutto come punto di riferimento
per la definizione del vocabolario di alta disponibilità. In questo senso, le
oggettive difficoltà di identificazione di quelle “parole che riteniamo più
comuni, psicologicamente e culturalmente, ma che poi hanno in realtà una
frequenza minima, vicina a zero, soprattutto nell’uso scritto” (De Mauro
2004: 142) potrebbero essere in parte superate facendo riferimento a quella
porzione di bagaglio lessicale condiviso e acquisito, se non attraverso altri
canali, per il tramite dell’IS: seppure statisticamente poco rilevanti nelle
produzioni adulte, i termini a chiunque familiari perché proposti con
frequenze elevate e funzioni significative nell’italiano per i bambini – ad
esempio i termini tipicamente indicati sugli alfabetieri (oca), usualmente
utilizzati per l’insegnamento delle particolarità ortografiche (camoscio),
presenti nelle denominazioni più diffuse di giochi e tipologie esercitative
(cruciverba), in fiabe e racconti (carrozza), corrispondenti a discipline
(geografia) o routines scolastiche (giustificazione) – potrebbero probabilmente
superare qualunque prova di elicitazione sui parlanti e quindi, seppur
difficilmente rintracciabili nel lessico adulto, essere selezionate per entrare
nel vocabolario di base con attribuzione della marca AD.
Anche in questo caso, certamente, per evitare insidie e ambiguità semantiche
andrebbero individuati dispositivi utili ad accertare la fenomenologia delle
accezioni effettivamente attive nonché a verificare e interpretare criticamente
le relazioni intercorrenti tra la frequenza dell’input lessicale (e semantico) in
ingresso e la frequenza dell’output lessicale (e semantico) fattuale ma anche
potenziale, in un modello descrittivo che – nel contemplare un’interazione
dialettica, dinamica e comparativa tra le dimensioni della ricettività,
produttività e disponibilità e attribuendo i giusti pesi a quella delicata e
complessa combinazione di quantità e qualità che De Mauro (1994: 97)
felicemente ebbe modo di battezzare binomio indispensabile – consenta di
distinguere gli autentici dai solo apparenti paradossi della frequenza.

634

JADT’ 18

Riferimenti bibliografici
Benedetti G. e Serianni L. (2009). Scritti sui banchi. L'italiano a scuola fra alunni
e insegnanti. Roma, Carocci.
Chiari I. e De Mauro T. (2012). The new basic vocabulary of Italian: problems
and methods. Rivista di statistica applicata / Italian Journal of Applied
Statistics, vol. 22 (1): 21-35.
Cortelazzo M. (1995). Un'ipotesi per la storia dell'italiano scolastico. In Antonelli,
Q. & Becchi E. curatori, Scritture bambine, Roma-Bari, Laterza: 237-252.
De Blasi N. (1993). L’italiano nella scuola. In Serianni, L. e Trifone, P. curatori,
Storia della lingua italiana, vol. I “I luoghi della codificazione”. Torino,
Einaudi: 383–423.
De Mauro T. (1980). Guida all'uso delle parole. Editori Riuniti, Roma 1980.
De Mauro T. (2004). La cultura degli italiani. A cura di Francesco Erbani.
Roma-Bari, Laterza.
De Mauro T. (2005). La fabbrica delle parole. Torino, Utet Libreria.
Revelli L. (2013). Diacronia dell’italiano scolastico. Roma, Aracne.

JADT’ 18

635

How Twitter emotional sentiments mirror on the
Bitcoin transaction network
Piergiorgio Ricci
Tor Vergata University – piergiorgio.ricci@gmail.com

Abstract
Bitcoin represents the first and most popular decentralized cryptocurrency. It
was launched in 2008 by Satoshi Nakamoto, the name used by the unknown
person or people who designed Bitcoin system and created its original
reference implementation. It is based on Blockchain technology that is
considered one of the most promising technologies for the future. It is
more than an instrument of finance and will likely disrupt many industries
from banking to governance in the next years. This research explores a
geolocalized subset of Bitcoin blockchain and compares it with Twitter
communication related to the topic in order to discover what people living in
different geographical areas think about Bitcoin cryptocurrency and to assess
potential relationship between characteristics of language adopted by Twitter
users in posts containing the key word Bitcoin and the structure of
geolocalized blockchain. It also answers a variety of interesting questions
about the national use of Bitcoin.
Keywords: Bitcoin, Blockchain, Cryptocurrency, Social Network Analysis,
Semantic Analysis.
1. Introduction
Bitcoin cryptocurrency is based on blockchain technology that consists in an
open and distributed ledger where all transactions occuring in the system are
recorded
in
a
verifiable
and
permanent
way.
(Narayanan A., Bonneau J., Felten E., Miller A., and Goldfeder S., 2016) They
are organized in blocks which are generated periodically and linked by using
cryptography techniques (SHA256)( Drainville D., 2012). Each of them needs
to be validated by a peer to peer network respecting a specific protocol for
validating new blocks. Once stored, data can not be tampered without
tampering all subsequent blocks, activity that requires collusion of the
network majority. (Nakamoto, 2008) This approach complies with consensus
theory, a social theory which holds that social changes and innovation can be
reached without conflicts and the social system is fair. In fact, Bitcoin's
protocol relies on a strong social consensus among all partecipants of the

636

JADT’ 18

system that represent a node of the network and run a software with the aim
to improve enforcement of rules they agree on. Bitcoin network is
decentralized and it does not require trusting in a third party, such as a bank
or a government institution. For sure, it represents a new concept of money
(Evans, 2014) and the main purpose of this work is to find out what people
living in different geographical areas think about Bitcoin cryptocurrency and
to assess potential relationship between characteristics of language used on
Twitter posts related to the topic and the structure of geolocalized Bitcoin
Blockchain. Research has been conducted to analyze correlations and
causalities between social network metrics performed on the geolocalized
Bitcoin Transaction Network and Bitcoin emotional signals interecepted by
analyzing Twitter users posts grouped by country. In particular, it has been
considered important to discover wheter there is a specific kind of
communication adopted by Twitter users belonging to a specific country that
holds certain transaction network centrality measures. In other words, the
core question to be answered has regarded the analysis on existence of
correlation between the centrality in the Bitcoin transactions network of a
country and characteristics of language used on Twitter Bitcoin posts by their
citizens. To achieve this purpose, two datasets reperesenting Bitcoin
transactions and Twitter communication related to Bitcoin, have been
collected and classified on the basis of geography. Prior research has been
focused on economic aspects (Ron D. and Shamir A., 2012) and structural
proprieties of Bitcoin transaction network (Lischke et Fabian, 2016) (Fleder
M., Kester M. and Pillai S. 2015), but it has rarely considered the existing
relationship between transactions and social media communication. This
study also answers a variety of interesting questions about the national use of
Bitcoin and how Twitter users perceive it through communication signals
posted on Twitter microblogging platform. One of the most widely accepted
use cases for Bitcoin has to do with payments for digital content (Grinberg R.,
2012) and, at present, Bitcoin system is used only by early adopters and
innovators among population.
2. Data set
2.1 Bitcoin dataset
In order to analyze and compare the network of Bitcoin transactions and the
relative user sentiment on Twitter, two differents dataset have been built by
using a serious of Application Program Interface (API) available on the web.
The first dataset to be extracted has been the Bitcoin transaction network that
is publicly available from many free web services (such as Blockchain.info) or
by using a Bitcoin client that requires and stores the whole transaction
history, known as blockchain (Moser M., 2013). In order to reduce and

JADT’ 18

637

manage its complexity, a subset of blockchain, composed by more than 2
million transactions from July
2013 to July 2017 has been collected. They have been imported through the
Blockchain Data API service that allows Bitcoin block and transaction
payments data query functionality, providing requests for data regarding
single block, single transaction and block heights.

Fig. 1 - Example of transaction with multiple inputs and outputs (www.blockchain.info)

Fig.2 - Word Cloud related to USA Twitter dataset

These transactions have been geolocalized by using IPInfo.it web service and
stored in a NoSQL database. Geolocalization activity has regarded the
discovery of the countries involved in each transaction and it has been
carried out by scraping transactions IP addresses (Kaminsky, 2011). Each
transaction block contains a set of transactions and is characterized by
following attributes: flow identifier, hash transaction, timestamp, origin
country, destination country, sender, recevier and total amount (Ober M.,
Katzenbeisser S. et Hamacher K. 2013). Since each transaction can allow
multiple input and output addresses (Reid F. et Harrigan, 2012), they have
been decomposed in transaction flows. In order to attach geographical
informations to each transaction, the service provides by ipinfo.io website
has been used. It offers a web interface where is possible to retrive the origin
country of an IP Address provided as input.

638

JADT’ 18

2.2 Twitter dataset
A set of tweets from 10 different countries containing the word "Bitcoin" have
been collected in order to be analyzed. Sentiment analysis have been
conducted using the Software Condor (MIT Center for Collective Intelligence)
that automatically recognizes sentiment in English, Spanish, German, French,
Italian and Portuguese and allows tweets fetching restricted to a given
geolocation. It also allows to calculate sentiment of posts by using semantic
analysis techinques. This dataset is partially misaligned with the first one for
technical reasons.
3. Research methodology
Research has been conducted combining social network analysis (SNA) and
semantic analysis methodology with a particular focus on the relationship
among main indicators related to these two fields calculated on the dataset.
3.1. Social Network Analysis
Using a Social Network Analysis approach, several strategies are possible to
examine the structure of the Bitcoin transaction network. In order to
counduct the analysis some of the most common measures of centrality have
been identified. Most of them have been proposed by Freeman (1979) and
also analyzed in other Social Network Analysis articles (Batagelj, 2011). In the
following subsections they are briefly described.
3.1.1 Degree centrality
This measure is based on the degree that indicates the number of nodes
attached directly to a specific node for which it is computed. In the case of
directed networks, two different measures of degree centrality can be
calculated, defined as indegree and outdegree. The first one is given by the
number of ties directed to the node, while outdegree is the number of ties
that the node directs to others. In such cases, the degree is the sum of
indegree and outdegree. The (weighted) all-degree for the generic node
a directed graph is represented by the following equation:

in

=
counts the number of incoming ties and
represent the number
where
of outgoing ties. A node with an high degree centrality is central in the
network structure and tend to influence the others.

JADT’ 18

639

3.1.2 Closeness Centrality
Closeness centrality indicates the inverse of the distance of a node from all
the others in the graph. It is based on the shortest paths that between each
couple of nodes in the network. Closeness centrality of node
with N nodes, is defined as following:

, in a graph

=
is

where,
linking

and

the

number

of

edges

in

the

shorterst

path

. Closeness centrality is normalized as shown below:
= (N - 1)

This measure can be considered as a proxy of the speed by which a social
actor can reach the others.
3.1.3 Betweenness Centrality
This variable considers the shortest paths that connect every other couple of
nodes and is higher when a node is more frequently in this subset. For a
network with N nodes, the betweenness centrality for node:

=
where,

is the number of the shortest paths linking two nodes in the

network and

is the number of shortest path linking two nodes that

go through the node . Social network indicators described above can be
used to analyze the structure and the dynamics of the Geographical Bitcoin
Network. In particular, once collected the target set of transactions and
enriched them with geographical informations, two directed graphs has been
modeled. In the first one, identified as Generic Network, each node
represents a Bitcoin address owned by a user belonging to a specific country,
while each link indicates a transaction of a certain amount (weight of link)
occuring between two different addresses, while in the second one, defined
as Geographical Network, each node symbolize a country and links act for
transactions that can involve single or different countries. All the network
metrics used in this study will be explained in the next chapter. They have
been performed on the Geographical Network, obtained by merging General
Network transactions on geographical basis.

640

JADT’ 18

3.2 Semantic Analysis
Semantic analysis of textual data allows to turn text into data for analysis.
This is possible applying natural language processing techinques and
analytical methods. (Hu X., Tang L., Tang J. and Liu H., 2013). In the
following subsections a set of communication indicators will be briefly
described.
3.2.1 Sentiment
This indicator describes wheter messages are positive or not. Its value is
between 0 in the case of very negative messages and 1 viceversa. It is
computed as the average score for the whole text in a message.
3.2.2 Emotionality
This variable expresses the degree of emotion of an individual text fragment
and it is involved in sentiment elaboration.
3.2.3 Complexity
It measures the rarity of a word, or the likelihood that a single word will
occur in a text. It is higher when a text contains many rare words.
4. Results
The aim of this study has been to find out whether characteristics of Twitter
communication related to Bitcoin reflects the GeoBlockchain network
structure. Analysis has been conducted combining most important social
network centrality metrics, such as Degree Centrality, Closeness Centrality
and Betweenness Centrality with some other language indicators measuring
the characteristics of the textual data used in Twitter communication. On the
one hand, centrality metrics measures the importance, influence or power of
a node in the network and are widely applied in social network analysis, on
the other, language indicators allow to identify whether communication
referred to Bitcoin is positive or not, its emotionality and the complexity of
word usage. During analysis, country rankings for each social network
indicator has been calculated in order to be correlated with Twitter
Sentiment, Complexity and Emotionality national rankings performed on
Tweets containing the key word "Bitcoin". Spearman's correlation, computed
considering a set of 10 different countries with a high number of transactions
and tweets, shows a significative correlation between centrality measures
computated in the Geographical blockchain netowork and language on
microblogging platform Twitter. In particular, communication of people
belonging to most central countries in the Bitcoin network, e.g. Germany and
USA, is more complex and less emotional than the one of peripheral country
nodes. This is probably due to a more depth knowledge of Bitcoin
phenomena in the most innovative countries as shown by their Word clouds.
In fact, they tweet more and with a quite technical language (e.g. they speak

JADT’ 18

641

about technical aspects such as fork of blockchain), while the others one, for
example Spain, appear frightened of the new cryptocurrency's diffusion.
Emotionality

Degree
Centrality

1,000

-,638*

Correlation Coefficient
Emotionality Sig. (2-tailed)
N

Spearman's Rho

Correlation Coefficient

Degree
Centrality

Sig. (2-tailed)
N

.

,047

10

10

-,638*

1,000

,047

.

10

10

Complexity

Degree
Centrality

1,000

-,693*

.

,026

*. Correlation is significant at the 0.05 level (2-tailed).

Correlation Coefficient
Complexity
Spearman's
Rho

Sig. (2-tailed)
N

Degree
Centrality

Correlation Coefficient
Sig. (2-tailed)
N

10

10

-,693*

1,000

,026

.

10

10

*. Correlation is significant at the 0.05 level (2-tailed).
Fig.3 - Spearman's correlations calculated on national rankings of Complexity - Degree
Centrality and Emotionality - Degree Centrality

5. Conclusion and future works
The analysis highlights the Bitcoin transactions geographical distribution and
shows national differences in its adoption, revealing the major businesses
and markets. In particular, the most central countries in Bitcoin transaction
network are characterized by a positive and quite complex language, while
peripheral countries use a more emotional language and the sentiment of
their people about it is fairly variable. This result leads to the interpretation
that Twitter emotional sentiments mirror the Bitcoin transaction network and
this could be seen as an interesting signal for investors and entrepreneurs
interested in the development of new payment systems based on Bitcoin
technology and in the choice of the start up country. Main findings of the
study could be applied to crypto-payments national regulation as well as to
the economic and financial impact assessment of cryptocurrencies and future

642

JADT’ 18

works include investigation on the principle barriers to mass adoption of
Bitcoin cryptocurrency.
References
De Nooy W.,Mrvar A. and Batagelj V. (2011). Exploratory social network
analysis with pajek (2nd Ed.). Cambridge University Press.
Freeman L.C. (1979). Centrality in social networks conceptual clarification. Social
Networks, 1 , 215–239.
Lischke M. and Fabian B. (2016). Analyzing the Bitcoin Network: The First Four
Years. MDPI AG.
Nakamoto S. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System.
Reid F. and Harrigan M. (2012). An Analysis of Anonymity in the Bitcoin
System. Springer.
Ober M., Katzenbeisser S. and Hamacher, K. (2013) Structure and Anonymity
of the Bitcoin Transaction Graph. Future Internet. MDPI.
Kaminsky D. (2011). Black Ops of TCP/IP. Black Hat & Chaos Communication
Camp
Drainville D. (2012). An Analysis of the Bitcoin Electronic Cash System.
University of Waterloo
Ron D. and Shamir A. (2012). Quantitative Analysis of the Full Bitcoin
Transaction Graph. IACR Cryptology ePrint Archive
Fleder M., Kester M. and Pillai S. (2015) Bitcoin Transaction Graph Analysis
Moser M. (2013) Anonymity of Bitcoin Transactions. Munster Bitcoin
Conference
Grinberg R. (2012). Bitcoin: An Innovative Alternative Digital Currency.
Hastings Sci. & Tech
Hu X., Tang L., Tang J. and Liu H. (2013). Exploiting social relations for
sentiment analysis in microblogging. In Proceedings of the sixth ACM
international conference on Web search and data mining. ACM.
Narayanan A., Bonneau J., Felten E., Miller A., and Goldfeder S. (2016).
Bitcoin and Cryptocurrency Technologies: A Comprehensive Introduction.
Princeton University Press
Evans D. (2014) Economic Aspects of Bitcoin and Other Decentralized PublicLedger Currency Platforms. University of Chicago Coase-Sandor Institute
for Law & Economics Research Paper No. 685

JADT’ 18

643

Analyse de contenu versus méthode Reinert :
l’analyse comparée d’un corpus bilingue de discours
acadiens et loyalistes du N.-B., Canada
Chantal Richard1, Sylvia Kasparian2
Université du Nouveau-Brunswick, Canada – chantal.richard@unb.ca
2Université de Moncton, Nouveau-Brunswick, Canada – sylvia.kasparian@umoncton.ca
1

Abstract
In this paper we compare two methods of thematic analysis by applying
them to the same corpus. Specifically, we will compare the results of the
classification of units of contexts using the Reinert method in IRAMUTEQ,
with a content analysis (manually coded themes) analyzed using SPHINX in
2012. The bilingual corpus consists of two sub-corpora: speeches at the
Conventions nationales acadiennes (in French) and centennial commemorative
speeches by Loyalists (in English). Our goal is to determine whether the
Reinert method of distribution by class confirms, contradicts, or enhances a
traditional content or thematic analysis.
Résumé
Cet article compare deux méthodes d’analyse thématique de données
textuelles appliquées à un corpus bilingue. Notamment, nous comparons la
répartition par classes selon la méthode Reinert, intégrée dans IRAMUTEQ,
avec les résultats d’une analyse de contenu (codification manuelle des
thèmes) analysés par SPHINX en 2012. Le corpus est constitué de discours
acadiens (en français) et de discours loyalistes (en anglais). Cette étude
permet de voir dans quelle mesure la méthode Reinert confirme, contredit,
ou bonifie l’analyse de contenu traditionnelle pour étudier les mondes
lexicaux ou univers de discours de ces deux sous-corpus.
Mots-clés : analyse de contenu, IRAMUTEQ, méthode Reinert, classification
hiérarchique descendante.
1. Introduction
Aux JADT 2012, nous avions présenté une analyse de contenu des thèmes
principaux d’un corpus bilingue tiré de la base de données Vocabulaires
identitaires. Cette base regroupe des discours en français et en anglais qui
traitent de l’identité collective de deux peuples diasporiques au NouveauBrunswick, Canada : les Acadiens et les Loyalistes. Depuis 2012, la base de

644

JADT’ 18

données est passée de 74 à 1525 textes. S’imposait alors une démarche plus
efficace – pour cela nous avons choisi la méthode Reinert de classification
hiérarchique descendante. Avant d’entamer l’analyse du corpus plus large
nous avons voulu comparer la méthode Reinert aux résultats de l’analyse de
contenu de 2012 en l’appliquant au corpus original de 74 textes. Cet article
permet de voir dans quelle mesure la méthode Reinert bonifie l’analyse de
contenu traditionnelle pour étudier les mondes lexicaux de ce corpus.
2. Analyse de contenu et méthode Reinert
Avant de procéder à l’analyse, nous définirons brièvement les deux types
d’analyse tout en expliquant notre démarche méthodologique.
2.1 Analyse de contenu
Nous entendons par analyse de contenu une « méthode de classification ou
de codification dans diverses catégories des éléments du document analysé
pour en faire ressortir les différentes caractéristiques en vue d’en mieux
comprendre le sens exact et précis » (L’Écuyer 50). En d’autres mots, une
lecture exhaustive du corpus permet de choisir des unités de classification,
de générer une catégorisation sous forme de tableaux à être traités
statistiquement, et l’interprétation des résultats de l’analyse statistique
permet une description des thèmes relevés. C’est la méthodologie utilisée
dans notre première étude du corpus à l’aide des logiciels SPHINX et
HYPERBASE afin d’extraire les mots-clés des sous-corpus. Ci-dessous
(Tableaux 1 et 2) se trouvent les thèmes et quelques mots-clés qui les
constituent.
2.2 Méthode Reinert
La méthode Reinert de la classification hiérarchique descendante a été
adaptée pour le logiciel IRAMUTEQ et appliquée à notre corpus selon les
modalités décrites par Ratinaud et Marchand (2012). Cette méthode consiste
à identifier les unités de contexte élémentaires selon l’organisation interne du
texte qui a été lemmatisé pour ensuite être réparti par classes en procédant
par bipartitions successives. Comme pour l’analyse de contenu, nous avons
analysé séparément les sous-corpus par langue. Les classifications obtenues
ainsi ont été contrastées avec les premiers résultats obtenus à l’aide de
l’analyse de contenu.
3. Corpus
Les 34 discours des conventions nationales acadiennes, prononcés de 1881 à
1890, constituent le corpus acadien de langue française, qui compte 56 368
mots. À cette époque, les Acadiens procédaient à une réorganisation sociale

JADT’ 18

645

par le choix de symboles nationaux. Les Loyalistes du Nouveau-Brunswick,
pour leur part, sont un groupe d’Américains royalistes ayant fui le pays suite
à l’Indépendance pour s’établir au Nouveau-Brunswick où ils fêtent leur
centenaire en 1883. Les 40 discours du centenaire des Loyalistes, publiés
entre 1882 et 1887, forment le corpus de langue anglaise qui compte 69 610
mots.
4. Analyse
L’analyse contrastive des résultats obtenus par ces deux méthodes d’analyse
sont présentés par sous-corpus en affichant en premier le tableau thématique
accompagné de quelques mots-clefs générés par l’analyse de contenu, suivi
du dendrogramme produit par IRAMUTEQ.
4.1 Corpus des Conventions nationales acadiennes (français)
Tableau 1 : Thèmes et mots-clés extraits du sous-corpus acadien par l’analyse de contenu
Événement Progrès et
rassembleur avenir
(symboles)
fête
convention
drapeau
adopter
distinct
monument
assemblée
tricolore
légitime
étoile…

avancement
intérêts
droits
développement
sauvegarde
surmonter
triomphant
amélioration
combattre…

Références
au passé

Relations
(inter)nationales

colonie
histoire
perdu
ancêtres
origine
persécutés
misère
pères mort
larmes
souvenir
infortune
ruine…

compatriotes
anglais
union
sympathie
ennemi
confédération
américains
fusion
puissance
Louisiane
préjugés…

Caractéristiques
associées au
peuple
grand
bonheur
malheur
honneur
noble, digne
devoir, petit
courage
difficultés
persévérance
faible. pauvre
humble…

Race,
ethnie et
culture
peuple
nation
race
patriotisme
sang
Acadie
patrie
âmes
usages
traits…

Religion

saint
religieuses
frères
foi
patron
Dieu Marie
Église
Assomption
chrétien…

La répartition par classes selon la méthode Reinert effectuée par IRAMUTEQ
sépare en premier la classe 6 des autres classes. Cette classe est représentée
par un lexique autour du choix d’une fête nationale acadienne, premier
objectif de ce grand rassemblement patriotique. Une deuxième partition se
fait entre les classes 3 et 4 et les classes 2, 1 et 5. La classe 4 est caractérisée par
un lexique de valeurs associées à la religion alors que la classe 3 illustre des
valeurs associées à un style de vie traditionnel attaché au passé. Le lien entre
les deux est révélateur du fait que pour les Acadiens de l’époque, le style de
vie traditionnel est fortement lié à la religion catholique. Si les classes 3 et 4 se
réfèrent au passé, les classes 2, 1 et 5 suggèrent plutôt un regard tourné vers
l’avenir, notamment dans les domaines des progrès matériel et intellectuel

646

JADT’ 18

(classe 2), de la presse francophone (1) et de l’éducation (5).

Figure 1 : Dendrogramme CHD1 – phylogram produit par IRAMUTEQ : classification
hiérarchique descendante par la méthode Reinert pour le corpus acadien

Quant à la comparaison aux thèmes relevés par l’analyse de contenu
traditionnelle (Tableau 1), certains rapprochements sont possibles. La classe 6
partage une quantité importante de formes avec le thème « événement
rassembleur » dans l’analyse de contenu, notamment les mots-clés communs
aux deux méthodologies : fête, adopter, drapeau, tricolore et distinct. Il est
également possible de rapprocher les classes 3 et 4 des thèmes « Religion » et
« Références au passé » du Tableau 1. Ces deux classes contiennent quelques
mots présents sous le thème « Caractéristiques associées au peuple » du
Tableau 1. Ces classes (2, 1 et 5) partagent une certaine partie de leur lexique
avec le thème « Progrès et avenir » du Tableau 1.
Quel est l’apport de la méthode Reinert à notre analyse? Dans ce cas, il est
pertinent de s’interroger sur ce qu’elle ne relève pas. Notamment, les
catégories de l’analyse de contenu « Relations nationales et internationales »
et « Race, ethnie et culture » (bien que certaines formes telles que « sang » et
« Acadie » se retrouvent dans les classes 3 et 4). Ces deux thèmes se
rapprochent le plus des axes d’intérêt des chercheurs, ce qui suggère une
interférence humaine probable. De plus, l’ordre des partitions proposé par
IRAMUTEQ, qui sépare la classe 6 et répartit les 5 autres classes entre le

JADT’ 18

647

passé et l’avenir, est très révélateur d’un discours paradoxal juxtaposant le
progrès social à la préservation d’une identité ancrée dans le passé, ce qui
n’était pas ressorti lors de l’analyse de contenu traditionnelle par thèmes.
4.2 Corpus des commémorations centenaires des Loyalistes du N.-B.
Tableau 2 : Thèmes et mots-clés extraits du sous-corpus loyaliste par l’analyse de contenu
(HYPERBASE et SPHINX)
Événement
Progrès et
rassembleur
avenir
(commémoration)

Références Relations
Caractéristiques Race,
au passé
nationales et
associées au
ethnie
internationales peuple
et
culture

anniversary
commemorate War,
1783 forefathers
memorial Parrtown
Victoria
1883, 18th Institute
Regiment…

abandoned
bitterness
choice
confiscated
defence
hardship
heroes, duty
Israelites
rugged
struggle…

advancement
building
cities
commerce
development
establishment
factories
harbour
hotels
industrial…

alliance
annexation
commonwealth
constitution
Independence
monarchy
government
King
Mother
protection…

active
brave brotherhood
conservative
determination
intelligent
deserving
strength…

civil
civilized
humanity
race
superior
anglosaxon
yanks
elevate
blood…

Religion

God
bibles
bless
Christian
churches
devotion
Faith
morality
temperance
…

Sept classes sont proposées dans le dendrogramme produit par IRAMUTEQ
pour le corpus loyaliste en anglais. Une première répartition sépare les
classes 3 et 2 de toutes les autres classes. La classe 3 est composée de
références militaires à des personnages, des lieux et des dates, et la classe 2
rassemble un lexique désignant des structures associatives responsables de
préserver la mémoire. Les deux classes sont caractérisées par un grand
nombre de noms propres. La classe 7 se distingue ensuite par ses termes
juridiques rattachés à l’empire britannique et ses colonies. Pour sa part, la
classe 6 est constituée d’un lexique autour des ressources naturelles et du
progrès matériel ou commercial, ce qui suggère une vision de domination de
la nature par l’être humain. La classe 1 traite des valeurs morales et
religieuses prisées par les Loyalistes. Finalement, les classes 4 et 5 sont très
proches, et désignent respectivement les circonstances du départ des
Loyalistes des États-Unis par loyauté à la couronne britannique, et la
célébration de leur succès en tant que fondateurs d’une nouvelle province (le
Nouveau-Brunswick) cent ans plus tard.

648

JADT’ 18

Figure 2 : Dendrogramme CHD1 – phylogram produit par IRAMUTEQ : classification
hiérarchique descendante par la méthode Reinert pour le corpus loyaliste

Les classes ainsi obtenues peuvent être comparées aux thèmes du Tableau 2.
Par exemple, la classe 1 (valeurs morales et religieuses) partage son lexique
avec les thèmes « Religion » et « Caractéristiques associées au peuple ». La
classe 4 (circonstances du départ) est très semblable au thème « Références au
passé » et la classe 5 (célébration du succès) pourrait également être mise en
parallèle avec « Événement rassembleur : commémoration » ainsi que le
thème « Race, ethnie et culture », extrait par l’analyse de contenu. Les classes
2 (structures associatives) et 3 (références militaires) peuvent être
rapprochées du thème désigné dans le Tableau 2 sous « Événement
rassembleur : commémoration ». La classe 7 (empire britannique et ses
colonies) se rapproche du thème « Relations nationales et internationales »
sans toutefois être identique, et la classe 6 (ressources naturelles et progrès)
ressemble au thème « Progrès et avenir », mais avec certaines distinctions,
notamment, l’inclusion des mots se référant à la nature dans le thème du
progrès matériel. L’originalité de la répartition par classes par IRAMUTEQ se
trouve en partie dans la juxtaposition du passé et du présent dans les classes
3 (références militaires du passé) et 2 (associations pour la préservation de la
mémoire par des activités commémoratives), ainsi que les classes 5
(célébration du succès) et 4 (circonstances du départ) qui en sont, en quelque
sorte, l’écho. De plus, les catégories établies dans l’analyse de contenu se sont
avérées incomplètes, et le lexique est réorganisé par la classification
hiérarchique descendante. Selon les répartitions de la méthode Reinert, les

JADT’ 18

649

termes juridiques (parliament, act, law, etc.) se retrouvent avec les termes se
référent à la couronne britannique et ses colonies alors qu’ils n’avaient pas
été relevés dans notre étude de 2012. De même, les mots désignant le monde
naturel (forest, ocean, tree, etc.) côtoient le lexique du progrès matériel et
commercial dans le dendrogramme, ce qui n’était pas intuitif à la lecture
humaine, mais fort révélateur. C’est précisément dans ces apparentes
contradictions qu’apparaissent les interprétations les plus nuancées, et donc
les plus judicieuses d’un corpus textuel.
5. Conclusion
Outre le fait de pouvoir traiter de corpus plus volumineux dans plusieurs
langues, quels sont donc les avantages de l’application de la méthode Reinert
à notre corpus bilingue? En somme, la répartition par classes nous a amené à
réviser et nuancer les résultats de l’analyse de contenu originale. Si les
partitions ressemblent parfois aux thèmes relevés en 2012, la méthode
Reinert a l’avantage de dévoiler les liens entre les classes par ses partitions
graduelles sans égard à la langue, ce qui nous a permis d’observer une
répartition temporelle passé/avenir dans le sous-corpus acadien et
passé/présent dans le sous-corpus loyaliste. De plus, les unités de contexte ne
reposent pas sur des préconçus ou des dictionnaires internes, mais sur une
répartition des mondes lexicaux qui respecte l’organisation interne des
corpus, ce qui a donné une réorganisation du lexique et l’inclusion de mots
qui ne figuraient pas dans l’analyse originale.
C’est justement l’inclusion de ce lexique apparemment paradoxal qui mène à
une analyse plus objective et plus fine. Par exemple, le côtoiement de la
nature et du progrès matériel dans les discours loyalistes suggère une vision
de la domination de la nature par l’être humain et les discours acadiens
visent un progrès social, économique et commercial tout en souhaitant
préserver une identité ancrée dans le passé. Ainsi, nos observations sur les
discours patriotiques des Loyalistes et des Acadiens à la fin du 19e siècle se
trouvent considérablement enrichies par la méthode Reinert telle qu’intégrée
dans le logiciel IRAMUTEQ.
Note : Cet article a bénéficié d’une subvention Savoir du Conseil de recherches en
sciences humaines du Canada. Nous remercions aussi Marc-André Bouchard pour
son aide technique.
Bibliographie
Baulac Y. et Moscarola J. SPHINX Solutions d’enquêtes et d’analyses de
données. www.lesphinx-developpement.fr.
Brunet É. HYPERBASE Laboratoire UMR 6039 Bases Corpus Langage,
Université de NICE-Sophia Antipolis.

650

JADT’ 18

http://ancilla.unice.fr/~brunet/pub/logiciels.html.
L'Écuyer R. (1987). L'analyse de contenu : notion et étapes. In Deslauriers, J.P., editor. Les méthodes de la recherche qualitative. Presses de
l'Université du Québec, pp. 49-64.
Ratinaud P. et Marchand P. (2012) Application de la méthode ALCESTE aux
« gros » corpus
et stabilité des « mondes lexicaux » : analyse du « CableGate » avec
IRAMUTEQ. Dister A., Longrée D., Purnelle G., editors,
Actes/Proceedings of JADT 2012. (11e journées internationales d’Analyse
statistique de Données Textuelles, pp. 845-857.
Ratinaud P. (2009). IRAMUTEQ: Interface de R pour les Analyses
Multidimensionnelles de Textes et de Questionnaires.
http://www.iramuteq.org.
Richard C. et Kasparian S. (2012). Vocabulaire de l’identité nationaliste :
analyse lexicale et morphosyntaxique des discours acadiens et loyalistes
entre 1881 et 1890 au N.-B., Canada. Dister A., Longrée D., Purnelle G.
editors, Actes/Proceedings of JADT 2012. (11e journées internationales
d’Analyse statistique de Données Textuelles), pp. 845-857.
Richard C., Bourque D., Brown A., Conrad M., Davies G., Francis C., Huskins
B., Kasparian S., Marquis G., Mullally, S. Base de données : Vocabulaires
identitaire/Vocabularies of Identity. https://voi.lib.unb.ca

JADT’ 18

651

Bridge over the ocean: Histories of social psychology
in Europe and North America. An analysis of
chronological corpora1
Valentina Rizzoli, Arjuna Tuzzi
University of Padova – valentina.rizzoli@phd.unipd.it; arjuna.tuzzi@unipd.it

Abstract
Since the European Association of Social Psychology (EASP - initially called
European Association of Experimental Social Psychology) has been
established in 1966, what was then considered “European” social psychology
has been working to affirm its own identity by presenting a distinctive brand
to the rest of the world in general and to North America in particular. This
study aims to compare European and U.S. social psychology through the
analysis of the papers published by two of the main journals in their field:
The Journal of Personality and Social Psychology and the European Journal
of Social Psychology. All the abstracts (from the first publication to the last
one in 2016) of the two journals papers have been collected. By means of a
(lexical) correspondence analysis (SPAD software), the existence of a latent
temporal pattern in keywords’ occurrences was explored. Furthermore, in
order to detect, retrieve and compare the main topics the journals dealt with
over time, an analysis implemented by means of Reinert’s method was
conducted (IRaMuTeQ and R software). Results show that even if there are
some typical features that distinguish the “European” from the “American”
social psychology some publication trends seem to converge. Results will be
discussed also reflecting on the contribution of these methods in studying the
history/ies of a discipline.
Keywords: diachronic corpora, chronological textual data, text clustering,
correspondence analysis, Reinert’s method, history of social psychology
1. Introduction
It is widely spread that what is called “the modern social psychology” came
from Europe with the migration of scholars during the second world war,

1This study is a new development of a an interdisciplinary research project
funded by the University of Padova, fund CPDA145940 (2014) “Tracing the History of
Words. A Portrait of a Discipline Through Analyses of Keyword Counts in Large
Corpora of Scientific Literature" (P.I. Arjuna Tuzzi).

652

JADT’ 18

and started to develop mainly in the United States. Moscovici and Markova
(2006) referred to an American indigenous tradition that compete with a
newer Euro-American tradition, not intending to argue that there was a
socio-psychological tradition born in Europe and brought to America; but a
genuinely American tradition that began with the work of the immigrant
Lewin and his new students. While there was a prosperous development of
social psychology in U.S., in Europe there were scholars working on social
psychology, but there was no European school (Moscovici, 1999). The
establishment of the European Association of (Experimental) Social
Psychology (EASP - initially EAESP) in 1966 has been fundamental in the
development of a “European” social psychology. EASP represented a
distinctive brand of the discipline to the rest of the world in general and to
North America in particular, by providing a voice for a more “social” social
psychology (http://www.easp.eu/about/?). To consider an "American" and a
"European" social psychology as two completely separated and
counterpoised entities would be wrong since there was a clear influence
between them. Moreover, the first EASP meeting, which fostered the birth of
EAESP, was an initiative of U.S. scholars (cf. Moscovici and Markova, 2006).
By saying “American” social psychology we usually refer to the indigenous
U.S. tradition explicated by Floyd Allport’s work in 1924, which considers
social psychology as part of general psychology and keeps more attention on
the “individual”. "European" social psychology usually refers to the EuroAmerican tradition, promoted by the EASP, that regards social psychology as
strictly connected to close disciplines such as sociology and anthropology
and accords a greater role to social and cultural aspects
(http://www.easp.eu/about/?). This contribute consists in an empirical
analysis that moves from the study of scientific production. Over time,
scientific journals shape the history of a discipline as they include objects,
fields of application and methods that contribute to delineate the trajectory of
a discipline. Thus, an in-depth understanding of the past and the temporal
evolution of a discipline can be achieved by analysing the scientific debate
inside relevant scientific journals (Trevisani and Tuzzi, 2015; 2018). We have
taken into account the European Journal of Social Psychology (EJSP) and the
Journal of Personality and Social Psychology (JPSP). The former is an official
publication of the EASP and worldwide represents the association's voice.
The JPSP belongs to the American Psychological Association, that represents
the most widespread community of psychologists in the United States, and
not only: It is an important scientific reference that provides guidelines also
in Europe. In terms of visibility and prestige, the JPSP is considered one of
the most relevant journals of the field. The main aim is to observe and
compare the trajectory of the two Journal publications and to reflect about

JADT’ 18

653

what contribution these methods can provide for the study of the history of a
discipline. We particularly intend: 1) to portray the temporal pattern of the
main concepts debated in the past and covered today by EJSP and JPSP; 2) to
detect, retrieve and compare the main topics these journals dealt with over
time.
2. Methods
All the available abstracts of the two journals have been included in two
corpora and collected from different acknowledged sources compared with
the website of the journals. As regards EJSP, a total of 2,559 items was
collected, for a period of 46 years, from the very first in 1971, Volume No. 1,
Issue No. 1 to the latest of 2016, No. 46, Issue No. 7. Regarding JPSP, an
amount of 9,568 item was downloaded, for a period of 52 years, from 1965,
Volume No. 1, Issue No. 1 to 2016, No. 111, Issue No. 6. Items without any
abstract have been deleted (e.g., editorials, master heads, errata,
acknowledgements). The EJSP corpus is composed of 2,195 abstracts, while
the JPSP one of 9,536 abstracts.
To improve the homogeneity of the corpora we decided to privilege the
British spelling (e.g., we replaced analyzed with analysed) in EJSP and those
American in JPSP. Our corpora have been normalised only replacing
uppercase with lowercase letters. The lexicometric measures showed that
there is a good redundancy, that is fundamental to work with frequencies
(Lebart, Salem, & Berry, 1998; Tuzzi, 2003; Bolasco, 2013).
Multi-words (MW) with frequencies ≥ 5 for the EJSP corpus and ≥ 10 for the
JPSP one (it is consistently larger than the former) have been recognised,
selected and considered as textual units. We resort to a procedure for
automatic information retrieval that permits to recognise repeated
informative sequences, e.g., an adjective followed by a noun as in “social
psychology”, that produce a MW (Pavone, 2010). Two encyclopaedias of
social psychology (Manstead et al., 1995; Baumeister & Vohs, 2007) and index
of keywords available in the downloading process provided further MWs.
In order to depict the structure of the association between years and words
and to establish the existence of a chronological dimension, a (lexical)
correspondence analysis (CA) has been conducted on two matrices: 5,784
words over 46 years (rows per columns) for EJSP corpus and 8,349 x 52 for
JPSP. To detect a set of relevant topics included in the journals and observe
their temporal development, an analysis implemented by means of Reinert's
method (1986) has been conducted. Topics can be defined as “lexical worlds”
(Reinert, 1993), that are groups of words referring to a class of meaning. The
result, performed with a hierarchical descending classification, is a
dendrogram that groups units into classes that mirror a similar lexical

654

JADT’ 18

context. Textual data were processed with the Taltac2 dedicated software and
statistical analyses were conducted with SPAD, Iramuteq and R software
packages.
3. Results
By means of CA we can observe the existence of clear-cut temporal
dimension in both the corpora (Figure 1). The keywords which mainly
contributed to the factorial solution show which concepts typifies each timespan.

Figure 1 - First factorial plan of Correspondence Analysis of EJSP (left side) and JPSP (right
side). Projection of years

In the EJSP (Figure 1, left side) the first period (1971-1990) is strongly
characterised by words that refer to the experimental design. This is the
period mainly concerned with the study of aggression, risk taking,
dissonance, and attribution theory. The keywords of the subsequent period
(nineties) seem to be related to social change, which is characterised by the
study of social influence, categorization, and words referring to Moscovici
and Tajfel's theories (that marked the European production: social
representations, minority influence and minimal group paradigm). In the
following years (2000s) we can observe that the attention has turned on the
self, ingroup/outgroup relations and the social cognition with the study of
stereotypes, emotions, motivation, agency/communion, and so on. In recent
years (2011-2016) mainly social issues (e.g., gender, migration, environment,
religion) and everyday life concerns are highlighted.
As regards the JPSP (Figure 1, right side), in the first decade considered
(1965-1976) the main contribution is given by words as reinforcement, verbal
reinforcement, conditioning, and so on, that together refer to behaviourism.
At the same time, we can observe the occurrence of words pertain to game’s
theories, conflict/cooperation as well as aggression and dissonance theory.

JADT’ 18

655

Also physiological measurements (e.g. heart rate) and experiments
(experimental) are visible. The second period includes the last Seventies until
the last Eighties. Its distinctive words are masculinity/femininity, and other
terms that remind to motivational theories. Moreover, the presence of words
related to personality is evident and becomes stronger in the following
period, that includes the Nineties. In this period mood, personality,
individual differences, memory and the self represent the main contribution.
At the same time also issues about gender and women are noteworthy. The
last period starts from the 2000s and shows many references to
explicit/implicit, and intimate relationships. Moreover, further specific words
about positive psychology (life satisfaction, goal pursuit, and so on) and
culture (cultural, culture) are relevant.
The analysis conducted by means of Reinert’s method enlightens the
presence of nine different lexical worlds (79.64% of the abstracts have been
classified) in EJSP (Figure 2).

Figure 2 - EJSP classes and their distributions over years – Unsupervised clustering method

Following the classes order from the bottom to the top of Figure 2, a brief
outline of their contents is provided below. Class 1 (red) concerns attribution
and methodological issues (e.g., method, statistical, model). Class 9 (fuchsia)
contains words related to impression formation, categorisation and
stereotype. Both these classes show decreasing trends without disappearing.
Class 6 (light blue) includes mainly words related to gender studies and
implicit measures (e.g., prime, IAT). Class 5 (water blue) concerns moods and
regulatory focus theory. These two classes show increasing trends. Class 8
(purple) concerns studies on aggression (in which mainly male/female as
subjects involved in an experiment were compared). This class was initially
hegemonic in the field and then disappeared along time. Class 7 (blue)
includes game theories and studies on cooperation competition and shows a
decreasing trend. Class 2 (orange) concerns politics and culture (mainly cross
cultural studies) and it is an ever-present topic, as well as Class 4 (green) that
concerns the social identity theory and ingroup/outgroup dynamics. Class 3,
that concerns the applications of that theory (e.g., migration), shows a clear

656

JADT’ 18

increasing trend. As regards JPSP, the analysis shows the presence of eleven
clusters (76,08% of the abstracts have been classified - Figure 3). Following
the order of the classes from the bottom to the top of Figure 3: Class 7 (light
blue) concerns consensus formations and attribution, and seems to be an
ever-present topic. Class 6 (water blue) contains processes regarding
memory, stereotypes and categorisation and it is particularly recurrent in the
nineties and 2000s. Class 3 (grey) contains studies on self, emotion and
motivation and shows a clear increasing trend, becoming one of the most
relevant topics nowadays. Classes 11 (fuchsia), 10 (lilac), and 1 (red) concern,
respectively, studies on aggression and physical measurements, on
dissonance and opinion changes, and male and female involved in
experimental studies. They were predominant in the first years considered
and then disappeared. Class 9 (purple) concerns culture (mainly comparing
west and east ones) and politics. It shows an increasing trend although it is
not among main topics nowadays. Class 2 (orange) includes words regarding
the measurements and their validity (e.g., scale, reliability, test retest) and
shows a stable trend. Class 8 (blue) contains words relate to interpersonally
differences (based on gender or studied with twin studies). It seems to
remain constant even if with a slight decreasing trend. Class 5 (water green)
is represented by words concerning health (mental and physical) and how to
cope with related problems. Class 4 (green) concerns romantic and couple
relationships. Both those classes show increasing trends.

Figure 3 - JPSP classes and their distributions over years – Unsupervised clustering method

4. Discussion and conclusions
The aim of the present study is to compare American and European social
psychology offering food for thought on the contribution of the methods
used in studying the histories of a discipline. Thanks to these preliminary
results we succeeded in highlighting the history of a discipline from the
particular point of view of its effective scientific production.
In the first years considered, some similarities among the contents tackled in
the two journals can be noticed (e.g., dissonance theory and aggression). The
main differentiation that emerged concerns the stronger attention on

JADT’ 18

657

individual and personality in JPSP, on the one hand, and the different impact
of Tajfel and Moscovici's contributions on the psychology of groups and
Moscovici's works on social representations, on the other. This emerged as
particularly evident in ‘80s and ‘90s. The predominant approach of social
cognition seems to be a common feature, as well as methods and research
design that mainly refer to the experimental method and topics concerning
cross cultural studies and politics. As regards the topics identified, some
common trajectories of publication were enlightened. They are, for example,
Class 11 in EJSP and 8 in JPSP, concerning studies on aggression that were
predominant in the first decades and later decline. Class 1 in EJSP and 7 in
JPSP, as regards, studies on attribution. Also, class 2 in EJSP and 9 in JPSP,
that are related to culture and politics. Similar contents but different
trajectories are shown by Class 9 in EJSP and 6 in JPSP. The main difference
between the journals is observed in JPSP Classes concerning personality,
health, cope, and romantic and couple relationships (8, 5, 4), and EJSP
Classes concerning ingroup/outgroup processes, and intergroup contact and
applied concerns (4, 3).
It is worth mentioning the core of the difference between American and
European social psychology: the attention on the individual in the American
and on the social in the European one. That difference finds its way as a
greater attention on social issues in EJSP and individual related studies (e.g.
interpersonal relations, personality) in JPSP. Two histories of publications in
social psychology have been traced, one North American and the other
European. Their typical differentiation is historically well known in the
community, but the empirical works that contributed to that debate are less.
This is an example of the contribution that quantitative analysis of textual
data can provide to the study of the history of a discipline, also known as
digital history.
References
Allport, F. (1924). Social Psychology. Boston, MA: Houghton Mifflin.
Baumeister, R. F., & Vohs, K. D. (2007). Encyclopedia of social psychology.
Thousand Oaks, CA: Sage.
Lebart, L., Salem, A., & Berry, L. (1998). Exploring textual data. Netherlands:
Springer. doi:10.1007/978-94-017-1525-6
Manstead, A. S., Hewstone, M. E., Fiske, S. T., Hogg, M. A., Reis, H. T., &
Semin, G. R. (1995). The Blackwell Encyclopedia of Social Psychology.
Blackwell Reference/Blackwell Publishers.
Moscovici, S. (1999). Ringraziamento. In Laurea Honoris Causa in Psicologia a
Serge Moscovici. Università degli studi di Roma “La Sapienza”: Centro
Stampa d’Ateneo.

658

JADT’ 18

Moscovici, S., & Markova, I. (2006). The making of modern social psychology.
Cambridge: Polity.
Pavone, P. (2010). Sintagmazione del testo: una scelta per disambiguare la
terminologia e ridurre le variabili di un’analisi del contenuto di un
corpus. In S. Bolasco, I. Chiari, & L. Giuliano (Eds.) Statistical Analysis of
Textual Data: Proceedings of 10th International Conference Journées d’Analyse
statistique des Données Textuelles, 9-11 June 2010, Sapienza University of
Rome, pp. 131-140. LED.
Ratinaud, P. (2014). Visualisation chronologique des analyses ALCESTE:
application à Twitter avec l’exemple du hashtag# mariagepourtous. Actes
des 12es Journées internationales d’Analyse statistique des Données Textuelles.
Paris Sorbonne Nouvelle–Inalco.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique:
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, 8(2), 187-198.
Reinert, M. (1993). Les «mondes lexicaux» et leur «logique» àtravers l’analyse
statistique d’un corpus de récits de cauchemars. Langage & Société, 66, 5–
39.
Trevisani, M., & Tuzzi, A. (2015). A portrait of JASA: The history of Statistics
through analysis of keyword counts in an early scientific journal. Quality
and Quantity, 49, 1287-1304.
Trevisani, M., & Tuzzi, A. (2018). Learning the evolution of disciplines from
scientific literature. A functional clustering approach to normalized
keyword count trajectories. Knowledge-Based System, 146, 29-141

JADT’ 18

659

Les « itemsets fréquents »
comme descripteurs de documents textuels
Louis Rompré1, Ismaïl Biskri2
1

Université du Québec à Trois-Rivières – rompre.louis@courrier.uqam.ca
2 Université du Québec à Trois-Rivières – ismail.biskri@uqtr.ca

Abstract
Automated classification is one of the preferred approaches applied to the
problem of organizing information. The classification process is based on
identification and evaluation of descriptors which characterize the
information. It’s usually necessary to discover them following a raw data
analysis. Generally, words are considered during this analysis. In this paper,
we propose to use frequent itemsets as descriptors. We present how they can
be identified and used to define a level of similarity between several texts.
The experiments conducted demonstrate the potential of the proposed
approach for defining similarity between texts and linking news broadcast on
the web.
Résumé
La classification automatisée est une des principales approches appliquées au
problème d’organisation de l’information. Le processus de classification
repose sur l’identification et l’évaluation de descripteurs qui caractérisent
l’information. Il est souvent nécessaire de déduire ces descripteurs à partir
d’une analyse des données brutes. Généralement, les mots sont considérés
pour mener cette analyse. Dans cet article, nous proposons d’utiliser des
itemsets fréquents comme descripteurs. Les expérimentations effectuées
démontrent le potentiel de cette approche pour établir un degré de similarité
entre différents textes et lier des nouvelles diffusées sur le web.
Keywords: Classification, Itemset fréquent, Descripteur, Document, Texte.
1. Introduction
La digitalisation des documents a facilité la diffusion de l’information. Dès
qu’un événement se produit de multiples articles sont rédigés et diffusés sur
les différentes plateformes numériques. Plusieurs documents textuels
diffusés sur le web sont composés uniquement de quelques centaines de
mots. C’est en consultant différents documents, qu’une description riche peut
être obtenue. Différents documents peuvent aborder un même sujet et
chacun de ces documents est susceptible de contenir de l’information

660

JADT’ 18

complémentaire. Toutefois, la quantité de données disponibles et leur
manque de structure limitent notre capacité à capturer ces informations d’où
la nécessité d’avoir recours à des outils facilitant l’accès à l’information. La
classification automatique est l’une des stratégies appliquées au problème
d’organisation de l’information. Un processus classificatoire appliqué à des
documents textuels, qu’il soit automatisé ou non, organise les documents de
sorte que ceux qui partagent des similarités soient regroupés. L’organisation
qui en découle peut être utilisée pour orienter, par exemple, la recherche
d’information, l’extraction de connaissances, l’aide au résumé, etc.
Plusieurs classifieurs automatiques ont fait l’objet de publications. Comparer
ces classifieurs pour déterminer leur performance est une tâche complexe et,
surtout, subjective. Un classifieur peut performer avec un ensemble
particulier de documents et engendrer des classes bruitées avec un autre
ensemble. La pertinence d’une classification est jugée en fonction de
l’homogénéité des classes qui en résultent. Ce critère est toutefois relatif.
L’examen d’une classe par un intervenant est accompli à partir de ses
objectifs de recherche et de ses connaissances du domaine abordé. La qualité
recherchée pour un système de classification automatisée est d’être capable
de cibler les informations pertinentes à l’intérieur des documents visés et de
déterminer comment ces informations peuvent être utilisées pour établir un
niveau de similarité entre ces documents. La classification numérique repose
sur l’identification et l’évaluation de descripteurs qui permettent de
différencier une classe d’une autre. Le choix d’un descripteur aux dépens
d’un autre revient à prendre position sur la nature des résultats générés. Il
influence le comportement du classifieur car la présence ou l’absence d’un
descripteur est un indice permettant de cibler la classe à laquelle appartient
un document. Pour la classification textuelle, le mot est souvent utilisé
comme descripteur discriminant (McCallum et Nigam, 1998). Lorsque
plusieurs mots apparaissent à des fréquences comparables dans deux
documents alors ces documents sont considérés comme étant similaires.
Toutefois, il est courant que des documents partagent un nombre important
de mots et ce même si ces documents traitent de sujets différents. La présence
seule de ces mots est donc peu porteuse d’information et son utilité pour
établir le niveau de similarité entre des documents est limitée. Néanmoins,
les relations qu’entretiennent ces mots avec d’autres peuvent mettre en
lumière des particularités propres à certains documents. Il est possible
d’utiliser ces relations pour établir le niveau de similarité entre documents.
2. Les règles d’associations
Le développement récent des règles d’association découle des travaux
d’Agrawal sur l’extraction de connaissances à partir de données

JADT’ 18

661

transactionnelles (Agrawal et al., 1993). Agrawal proposait de dégager des
relations entre des items qui cooccurrent dans des transactions commerciales.
Par exemple, les clients qui achètent les items x et y achètent également l’item
z. Depuis, l’approche a été transposée à d’autres domaines, les règles
d’association pouvant être appliquées à divers domaines dans la mesure où
le concept de transaction peut y être défini.
Soit T un ensemble de transactions tel que

, les

sont appelés des items. Un
éléments qui composent les transactions
item est une donnée dont la nature dépend du domaine abordé. Par exemple,
les items peuvent correspondre à des descripteurs extraits d’une musique
(Rompré et al., 2017), à des descripteurs extraits d’une image (Alghamdi et
al., 2014) ou simplement à des mots extraits d’un texte (Zaïane et Antoine,
2002). Ainsi, une transaction peut être définie simplement comme un sousensemble de descripteurs.
Soit

un ensemble de d items distincts, chaque sous-

ensemble qu’il est possible de générer à partir des items

est appelé un

itemset. Pour un ensemble I de taille d, le nombre d’itemsets possibles est
(Tan et al., 2002). Le nombre d’itemsets potentiels est exponentiel, en fonction
de la taille de I. L’objectif à atteindre lors du processus d’extraction des règles
d’association étant de découvrir des relations cachées, il n’y a pas d’indice
permettant de cibler les items à considérer. Ainsi, l’espace de recherche
équivaut à l’ensemble des itemsets possibles. Même s’il est théoriquement
possible de créer
itemsets à partir d’un ensemble de taille d, en pratique
plusieurs combinaisons apparaissent peu ou tout simplement pas dans les
transactions. Ces combinaisons peuvent, donc, être ignorées. Le support est
une mesure qui permet de cibler les itemsets à ignorer. Le support d’un
itemset X représente le pourcentage des transactions de
Il est noté

qui contiennent X.

, et donné par l’équation 3.1 où n équivaut au nombre total de

transactions contenues dans T et

au support brut. Le support brut d’un

itemset représente le nombre de transactions de
donné par l’équation 3.2.

qui contiennent . Il est
(3.1)

662

JADT’ 18

(3.2)
Un itemset est considéré fréquent lorsque son support est supérieur ou égal
à un seuil prédéterminé. Soit X et Y deux itemsets fréquents tel que
, une règle d’association notée
traduit une relation de
cooccurrence entre ces itemsets. Par convention, le premier terme est appelé
l’antécédent tandis que le second est appelé le conséquent. Une règle
d’association est jugée de qualité selon une mesure
préalablement fixé. Ainsi, une règle d’association

et un seuil

est jugée de qualité

. La quantité de règles générées, leur pertinence de
si
même que leur utilité dépendent fortement des mesures et des seuils
minimaux fixés. L’évaluation des mesures d’intérêt des règles d’association a
fait l’objet de plusieurs travaux (Le Bras et al., 2010 ; Geng et Hamilton, 2006;
Tan et al., 2002). Même s’il existe plusieurs variantes, l’extraction des règles
d’association est généralement effectuée à l’aide de l’algorithme Appriori
(Agrawal et Srikant, 1994) ou FP-Growth (Han et al., 2000). D’autres
algorithmes sont présentés dans (Fournier-Viger et al., 2017). Les deux
principales difficultés liées à l’extraction des règles d’association sont la
gestion de la mémoire et l’effort computationnel nécessaire à la recherche des
itemsets fréquents. Contrôler le nombre d’items à considérer demeure le
meilleur moyen de traiter ces difficultés. Depuis deux décennies plusieurs
travaux portent sur l’application des règles d’association à des fins de
classification (Liu et al., 1998; Zaïane et Antoine, 2002 ; Bahri et Lallich, 2010).
Les différents classifieurs qui découlent de ces travaux produisent des
résultats qui sont en mesure de rivaliser avec ceux obtenus à l’aide d’autre
approches comme les arbres de décision (Mittal et al., 2017). Le principal
avantage des classifieurs à base de règles d’association est que les
connaissances qu’ils exploitent pour guider le processus classificatoire
peuvent être facilement interprétées. Ainsi, un classifieur qui exploite des
règles d’association peut être utilisé pour identifier les descripteurs
pertinents. Les différentes approches proposées dans la littérature impliquent
généralement des règles de la forme

où

correspond à un ensemble

de descripteurs et à une classe de similarité. Les documents sont considérés
comme étant les transactions tandis que les descripteurs (mots clés, fréquence
d’apparition des mots, etc.) et les classes sont considérés comme étant les

JADT’ 18

items. Soit un ensemble de descripteurs

663

, et un

représentant différentes
ensemble d’étiquettes
classes, alors un ensemble de documents peut être représenté de la manière
suivante :

Cette forme de représentation implique que les classes de similarité
auxquelles appartiennent les documents soient préalablement connues. Un
ensemble d’apprentissage est constitué et utilisé pour entraîner le classifieur.
Les règles d’association dégagées lors de la phase d’entraînement sont
utilisées pour prédire la classe de nouveaux documents. Ce processus
demande généralement un effort considérable et les résultats générés
dépendent de l’ensemble utilisé pour entraîner le classifieur.
3. Méthodologie
À l’instar des classifieurs à base de règles d’association, notre approche
exploite des itemsets fréquents pour décrire les documents. Toutefois, elle ne
nécessite pas de phase d’entraînement. Des itemsets fréquents sont extraits
de chacun des documents et comparés. Le degré de similarité entre deux
documents est fonction du nombre d’itemsets fréquents qu’ils partagent.
L’hypothèse derrière cette approche est que lorsque des mots co-occurrent
fréquemment au sein des phrases qui composent un texte, alors ces mots sont
représentatifs de ce texte. Ainsi, en considérant quelques itemsets fréquents,
il est possible de dégager les thèmes spécifiques traités dans les documents.
L’approche proposée comporte 4 étapes.
La première étape consiste à segmenter les documents afin de les préparer à
l’extraction des itemsets fréquents. Les documents sont traités comme des
ensembles de transactions où les phrases constituent les transactions et les
mots les items. Le nombre de mots différents susceptibles d’apparaître dans
un ensemble de documents textuels est théoriquement de l’ordre de la taille
du vocabulaire de la langue d’écriture de ces documents. Le nombre de mots
qui composent le français est estimé par l’Office Québécois de la Langue
Française à plus de 500 000. Considérant qu’à partir de 500 000 mots il est
possible de générer
itemsets, il est nécessaire d’imposer certaines
conditions aux textes en entrée afin de contrôler le nombre de mots. La

664

JADT’ 18

diversité d’un lexique augmentant avec la taille d’un texte, nous devons
limiter les textes en entrée à quelques milliers de mots.
La deuxième étape consacre la réduction du nombre d’items et donc de
l’espace de recherche lors de l’extraction des itemsets fréquents. Certains
mots jugés peu porteur d’information sont supprimés des transactions. Une
liste de 502 mots vides est utilisée. Les chiffres et les caractères de
ponctuation sont également supprimés.
La troisième étape vise à extraire les itemsets fréquents. Cette étape est
réalisée à l’aide de l’algorithme Apriori. Un effort est porté afin de dégager
un nombre restreint d’itemsets fréquents. La recherche des itemsets fréquents
est effectuée de manière itérative. Lors de la première itération, le support
minimum est fixé à une valeur élevée. Lorsque le nombre d’itemsets
fréquents extraits est inférieur à 10 alors le support minimum est diminué de
0.1. Le processus cesse lorsque le nombre d’itemsets obtenus est supérieur à
10 ou que le support minimum est inférieur à 0.1.
La dernière étape établit le degré de similarité entre les documents. Les
itemsets fréquents utilisés pour décrire les documents sont comparés. Plus le
nombre d’itemsets partagés par deux documents est grand, plus ces
documents sont jugés comme étant similaire.
4. Expérimentation et discussion
Afin d’évaluer l’approche proposée, plusieurs expérimentations ont été
effectuées avec une application que nous avons développée en Python. Un
corpus formé d’une centaine d’articles tirés de l’actualité et diffusés sur le
web a été utilisé. Ce corpus se distingue par le fait qu’il présente les mêmes
nouvelles sous l’angle de différentes agences de presse. Il regroupe des
articles diffusés sur le web provenant de 6 sources différentes et contenant
entre 500 et 1500 mots. Ces articles sont parfaitement adaptés aux conditions
de l’approche proposée.
Lors de nos expérimentations, nous avons mesuré le pouvoir discriminant
des itemsets fréquents. Nous avons effectué une comparaison entre les
classifications produites lorsque les descripteurs sont les itemsets fréquents et
les classifications produites lorsque les mots sont les descripteurs. La nature
des résultats obtenus suggère que les itemsets fréquents peuvent servir à
raffiner la description d’une classe. À titre d’exemple, le mot {avions} est
utilisé pour décrire 15% des articles du corpus. Même si ces articles sont
associés à l’aviation, ils traitent néanmoins de 4 sujets différents. Nos
expérimentations démontrent que l’utilisation des itemsets fréquents comme
descripteurs peut servir à décrire plus précisément le contenu de ces articles.
Les figures 1 et 2 illustrent respectivement la précision obtenue en
considérant des itemsets fréquents et celle obtenue en considérant

JADT’ 18

665

uniquement des mots. Il est à noter que lorsque seuls les mots sont
considérés, les classes de similarité générées sont moins homogènes. En effet,
des articles qui traitent de sujets autres que l’aviation y sont inclus.

Figure 1 : Précision avec les itemsets fréquents

Figure 2 : Précision avec les mots

La figure 3 illustre la matrice de similarité produite pour des articles traitant
de la crise nord-coréenne. La première colonne contient l’identifiant de
l’article, la seconde indique le sujet abordé tandis que les colonnes suivantes
donnent le nombre d’itemsets fréquents partagés par les articles. La
diagonale équivaut aux nombres d’itemsets fréquents extraits pour un article.
La figure 2 est représentative des résultats observés. Moins de 10 itemsets
fréquents ont été extraits pour la moitié de ces articles. Néanmoins, ils ont
tous été associés à la même classe.

Figure 3 : Matrice de similarité des documents traitant de la crise Nord-coréenne.

Malgré le fait qu’ils traitent du même sujet, certains articles partagent peu
d’itemsets fréquents avec les autres articles qui forment la classe. Ceci
s’explique par le lexique employé. Il est possible que les performances
puissent être améliorées en ajoutant une étape de lemmatisation. Toutefois,
certaines relations demeurent difficiles à établir automatiquement. Par
exemple, le document 45 contient les itemsets {nucléaire, pyongyang} et
{nucléaire, washington} tandis que le document 46 contient les itemsets
{nucléaire, corée} et {nucléaire, américaine}. Les résultats présentés
constituent uniquement un échantillon des connaissances extraites à l’aide de

666

JADT’ 18

l’approche proposée. En plus d’être faciles à interpréter, les itemsets
fréquents permettent de décrire plus précisément le contenu des documents
que les mots seuls.
5. Conclusion
Nous avons proposé une approche non supervisée pour établir des relations
entre des documents textuels. L’approche proposée repose sur l’utilisation
d’itemsets fréquents. Ces descripteurs expriment la cooccurrence de mots au
sein des phrases qui composent un texte. Les itemsets fréquents ont tendance
à être plus discriminant que les mots seuls. Par conséquent, ils peuvent aider
à rehausser la description d’une classe. L’un des avantages de la méthode
proposée est que les résultats produits sont faciles à interpréter. Les
expérimentations effectuées suggèrent que les itemsets fréquents, tels que
définis, sont suffisamment informatifs pour servir à établir des liens
cohérents entre des documents. Plusieurs débouchés sont envisageables.
Entre autres, l’approche proposée pourrait servir comme prétraitement à la
navigation entre différents documents, à l’annotation, au filtrage de
l’information, etc.
Références
Agrawal, R., Imielinski T., et Swami, A. (1993). Minning association rules
between sets of items in large databases, In Proc. of the SIGMOD
Conference on Management of Data, pp 207-216.
Agrawal, R., Srikant, R. (1994). Fast Algorithms for Mining Association Rules,
In Proc. of the 20th International Conference on Very Large Database, pp.
487-499
Alghamdi, R. A., Taileb, M., et Ameen, M. (2014). A new multimodal fusion
method based on association rules mining for image retrieval. In
Mediterranean Electrotechnical Conference (MELECON), 2014 17th IEEE
(pp. 493-499). IEEE.
Bahri, E., et Lallich, S. (2010). Proposition d'une méthode de classification
associative adaptative. 10eme journées Francophones d'Extraction et
Gestion des Connaissances, EGC 2010, pp. 501-512.
Fournier-Viger, P., Lin, J. C. W, Vo, B., Chi, T. T., Zhang, J. et Le, H. B. (2017).
A survey of itemset mining. Wiley Interdisciplinary Reviews: Data
Mining and Knowledge Discovery.
Geng, L., et Hamilton, H. J. (2006). Interestingness measures for data mining:
A survey. ACM Computing Surveys (CSUR), vol. 38, no 3, p. 9.
Han, J., Pei, J., et Yin, Y. (2000). Mining frequent patterns without candidate
generation. In ACM sigmod record (Vol. 29, No. 2, pp. 1-12). ACM.
Le Bras, Y., Meyer, P., Lenca, P., et Lallich, S. (2010). Mesure de la robustesse
de règles d’association. QDC 2010.

JADT’ 18

667

Liu, B., W. Hsu, et Y. Ma (1998). Integrating classification and association rule
mining. In Knowledge Discovery and Data Mining, pp. 80–86.
McCallum, A., et Nigam, K. (1998). A comparison of event models for naive
bayes text classification. In AAAI-98 workshop on learning for text
categorization (Vol. 752, pp. 41-48).
Mittal, K., Aggarwal, G., et Mahajan, P. (2017). A comparative study of
association rule mining techniques and predictive mining approaches for
association classification. International Journal of Advanced Research in
Computer Science, 8(9).
Rompré, L, Biskri, I et Meunier, J-G (2017). Using Association Rules Mining
for Retrieving Genre-Specific Music Files, In Proc. of FLAIRS 2017, pp.
706-711.
Tan, P. N., Kumar, V., et Srivastava, J. (2002). Selecting the right
interestingness measure for association patterns. In Proceedings of the
eighth ACM SIGKDD international conference on Knowledge discovery
and data mining (pp. 32-41). ACM.
Zaïane, O. R., et Antonie, M. L. (2002). Classifying text documents by
associating terms with text categories. In Australian computer Science
communications (Vol. 24, No. 2, pp. 215-222).

668

JADT’ 18

Discursive Functions of French Epistemic Adverbs:
What can Correspondence Analysis tell us about
Genre and Diachronic Variation?
Corinne Rossari, Ljiljana Dolamic, Annalena Hütsch,
Claudia Ricci, Dennis Wandel
University of Neuchâtel – corinne.rossari@unine.ch

Abstract
Our aim is to describe discursive functions of a set of French epistemic
adverbs by establishing their combinatory profiles on the basis of their cooccurrence with different connectors. We then compare these profiles using
correspondence analysis in order to find evidence of genre and diachronic
variation. The use of these adverbs is explored in contexts of informative
discourse within two distinctly different genres – contemporary written press
and encyclopedic discourse – as well as within two diachronic spans.
Keywords: epistemic adverbs, connectors, co-occurrences, correspondence
analysis, genre variation, diachronic variation
1. Introduction
Our aim is to analyze the genre and diachronic variation of discursive
functions of French epistemic adverbs (E-ADV). By discursive function we
mean the rhetorical aim of the utterance in which the adverb occurs: counterargument, argument, or conclusion (cf. Roulet et al., 1991). Our paradigm of
E-ADVs consists of the following items: certainement, certes, peut-être,
probablement, sans doute and sûrement1. The functions of these adverbs are
explored in contexts of informative discourse within two distinctly different
genres: contemporary written press and encyclopedic discourse. The former
is represented by three daily newspapers: Le Monde (2008, 20 410 766 tokens),
Le Figaro (2008, 10 795 373 tokens) and Sud-Ouest (2002, 29 763 988 tokens). In
the latter, we consider two diachronic spans: the 18th century, represented by
Diderot & d’Alembert’s Encyclopédie (DDA, 29 940 181 tokens) and the 21st
century, represented by the 2005 edition of Encyclopédie Universalis (UNI, 49
859 864 tokens) and by a random sample of the 2015 version of Wikipédia

1

Selection based on Roulet’s (1979) paradigm of epistemic assertive adverbs.

JADT’ 18

669

(WIKI, 50 396 345 tokens).2
We first proceed to an analysis based on the combinatory profile of each EADV (section 2) in our corpus of contemporary written press, and then, after
having pinpointed what such an analysis can and cannot show, we use a
more holistic approach based on correspondence analysis (section 3).
2. Analysis of Combinatory Profiles
In order to identify the discursive functions of the E-ADVs considered here,
we searched connectors (C) specifically co-occurring with each of these EADVs within a 20-token span. We have chosen a 20-token span rather than a
sentence span, because a connector’s combinatory profile can go beyond the
sentence boundaries. We define connectors as linguistic forms linking
segments of discourse. Such a functional category is not part of the tagset of
the platform we used. We therefore made our query by searching for three
different categories: adverbs, subordinating conjunctions and coordinating
conjunctions. We then manually filtered the resulting forms by keeping those
which proved to function as a connector.
For all our sub-corpora, each of these adverbs is thus specifically assigned a
series of connectors within constructions of the type “E-ADV…C1/C2/Cn”
and “C1/C2/Cn…E-ADV”, which represent their discursive combinatory
profile3. We call each sequence within a combinatory profile a discourse
movement as we consider it to have specific, rhetorically motivated discursive
aims. These aims (mentioned in section 1) are signaled by the connectors cooccurring specifically with an E-ADV: néanmoins and mais signal that the
utterance preceding them is a counter-argument to the utterance they
introduce; donc and finalement signal that the utterance they introduce is a
conclusion; car and parce que signal that the utterance they introduce is an
argument in favor of the utterance preceding them.
The tables below show the discursive combinatory profiles in three subcorpora of contemporary press (Le Monde 2008 ; Le Figaro 2008 ; Sud-Ouest
2002). The significance of each co-occurrence of a connector with an E-ADV is
calculated using log-likelihood (LL).4

All the corpora used were supplied by the platform BTLC (Base Textuelle Lexicostatistique de Cologne), conceived by Sascha Diwersy (Diwersy, 2014), and were
constituted within the French-German projects Presto (http://presto.ens-lyon.fr) and
Emolex (http://emolex.u-grenoble3.fr).
3 We adapt the term combinatory profile used by Blumenthal et al. (2005) and
Blumenthal (2008; 2012).
4 Although LL can be directly calculated on the BTLC platform, we used the
platform to extract the corresponding frequencies, and calculated the LL by using R.
2

670

JADT’ 18

Tables 1-3. Log-likelihood scores (threshold: 10.83; all scores equal or above are marked in
bold).
Le Monde
(2008)
car
(6 706)
donc
(8 276)
finalement
(1 559)
mais
(51 544)
néanmoins
(968)
parce que
(2 514)

certainement
(385)
L
R
0.08
0.08
(3)
(3)
2.09
2.09
(6)
(6)
-1.18
-0.01
(0)
(1)

certes
(1943)
L
-0.27
(11)
-2.47
(10)
-1.77
(1)

29.83
(47)

27.94
(46)

3248
(979)

-0.73
(0)
0.88
(2)

-0.73
(0)
2.81
(3)

5.84
(6)
1.78
(8)

R
0
(13)
2.9
(23)
1.14
(5)
22.55
(52)
14.22
(9)
-4.47
(1)

Le
Figaro
(2008)
car
(3 922)

certainement
(268)
L
R
-0.57
(1)

-0.57
(1)

3.89
(14)

donc
(4 763)
finalement
(1 150)

-1.02
(1)

-1.02
(1)

-0.03
(9)

-1.14
(0)

-1.14
(0)

-0.04
(2)

36.23
(41)

14.25
(30)

1757.55
(545)

1.07
(1)

-0.58
(0)

10.02
(6)

0.10
(1)

-1.43
(0)

-1.65
(1)

mais
(28 552)
néanmoins
(580)
parce que
(1 435)

certes
(1084)
L

peut-être
(2900)
L
R
49.30
-0.37
(57)
(12)
-1.45
24.09
(18)
(51)
3.60
15.43
(9)
(15)

probablement
(723)
L
R
18.99
0.92
(17)
(7)
-0.68
1.43
(4)
(9)
-0.01
0.58
(1)
(2)

sans doute
(2482)
L
R
16.09
0.02
(35)
(17)
0.03
4.18
(21)
(30)
0.01
6.96
(4)
(10)

sûrement
(307)
L
R
1.51
-0.64
(4)
(1)
0.10
0.10
(3)
(3)
-0.94
6.08
(0)
(3)

371.45
(423)

107.28
(284)

10.30
(57)

10.30
(57)

205.88
(310)

32.85
(193)

30.90
(41)

65.68
(55)

0.02
(3)
62.15
(37)

-0.23
(2)
9.75
(17)

0.12
(1)
2.03
(4)

-1.38
(0)
-3.58
(0)

-0.06
(2)
86.58
(41)

-0.06
(2)
0.12
(7)

-0.58
(0)
3.78
(3)

-0.58
(0)
0.07
(1)

peut-être
(1851)
L
R

probablement
(441)
L
R

sans doute
(1240)
L
R

sûrement
(211)
L
R

28.08
(37)

-0.48
(11)

14.27
(12)

1.95
(6)

8.45
(19)

4.43
(16)

7.53
(6)

-3.08
(0)

-20.39
(2)

3.16
(24)

-0.22
(3)

6.74
(10)

-0.37
(9)

0.37
(13)

-3.74
(0)

-0.48
(1)

-3.15
(1)

6.52
(10)

0
(1)

0
(1)

1.67
(5)

-1.34
(1)

-0.90
(0)

-0.90
(0)

245.20
(281)

93.30
(204)

0.56
(27)

2.38
(31)

86.88
(151)

34.59
(117)

17.23
(27)

87.11
(52)

0.49
(2)

0.44
(3)

-3.84
(0)

-0.95
(0)

-2.67
(09

-2.67
(0)

-0.45
(0)

-0.45
(0)

0.30
(2)

31.90
(22)

-2.25
(2)

6.88
(5)

4.79
(8)

0.14
(4)

0.28
(1)

2.22
(2)

R
2.35
(4)
1.81
(14)
0.95
(1)
1.39
(49)

0.95
(0)
0.03
(1)

JADT’ 18
Ouest
Sud
(2002)
car
(12 434)
donc
(19 185)
finalement
(2 858)
mais
(77 108)
néanmoins
(1 698)
parce que
(5 981)

671

certainement
(1277)
L
R

certes
(2795)
L

R

peut-être
(4950)
L
R

probablement
(812)
L
R
1.34
1.35
(10)
(4)

10.78
(23)

6.53
(20)

2.89
(32)

5.92
(36)

44.71
(91)

-0.29
(38)

-3.00
(10)

5.72
(27)

-6.45
(22)

0.10
(38)

-1.98
(53)

1.55
(74)

-1.32
(7)

-0.09
(2)

0.11
(3)

3.19
(10)

0.07
(6)

7.35
(19)

3.68
(16)

-0.23
(1)

28.87
(113)

64.77
(139)

6962.42
(1778)

-5.72
(118)

520.54
(682)

211.15
(513)

10.35
(64)

-0.16
(1)

-0.16
(1)

9.25
(10)

-0.01
(3)

5.39
(12)

-0.54
(4)

0.93
(2)

-4.77
(11)

13.45
(6)

12.51
(15)

8.48
(13)

814.50
(135)

104.89
(33)

16.58
(13)

9.77
(22)
0.23
(1)
9.48
(63)
1.85
(0)
0.57
(2)

sans doute
(3930)
L
R

sûrement
(684)
L
R

45.13
(78)

-0.26
(30)

6.87
(13)

-1.61
(42)

9.50
(74)

2.23
(12)

0.72
(10)

209.38
(434)

-0.09
(5)

123.59
(376)

0.41
(7)
0.08
(1)
7.09
(52)

162.80
(130)

6.72
(11)

1.20
(7)

0.06
(1)

0.06
(1)

233.04
(108)

0.00
(16)

1.04
(8)

-0.48
(9)

The data lead to the following observations: (i) Although the E-ADVs belong
to the same semantic class, each has its own specific combinatory profile. (ii)
Certain E-ADVs share comparable combinatory profiles: sans doute and peutêtre share an almost identical set of specific connectors; more frequently,
several E-ADVs essentially only share one or more specific connectors (for
instance the connector mais for certainement, sûrement, peut-être and sans
doute). (iii) Certain E-ADVs stand out for their unique combinatory features:
certes is almost exclusively associated with mais, but only with mais_R, and
with a notably higher log-likelihood score than the other E-ADVs.
Probablement is also associated with only a few connectors, but with a low
log-likelihood score, close to the threshold of 10.83. (iv) There is homogeneity
in the significant association for each E-ADV in the three sub-corpora of
contemporary press. However, preceding studies – Rossari et al. (2016) and
Rossari & Salsmann (2017) – show that the E-ADVs’ combinatory profile
varies depending on different genres and diachronic periods: contrary to
what is observed for the press genre, in DDA and UNI the association peutêtre…mais is less significant than the association mais…peut-être. For instance,
in DDA, no significant association certes…mais is observed, while the
association sans doute…mais in the same corpus proves to be highly
significant. The analysis of combinatory profiles (based on the significance
measure log-likelihood; cf. Blumenthal et al., 2005) allows for one-to-one
comparison of the different sequences of the type E-ADV...C and C...E-ADV.
Thus, the associations of each E-ADV with each connector can easily be
compared across corpora representing different newspapers, but also across
different genres and diachronic periods. It is also possible to compare the

0.15
(10)
0.31
(2)

672

JADT’ 18

associations of different E-ADVs with one or a few connectors. However, this
method has certain insufficiencies when it comes to simultaneously
comparing all of these variables in a holistic view. This type of analysis of
combinatory profiles never takes into account all variables at the same time
(e.g. frequencies, log-likelihood scores, paradigm of E-ADVs, paradigm of
connectors). Moreover, using a threshold (in our case 10.83) in order to
decide whether an association is significant is useful for traditional
collocation analysis. But our goal is to also represent the use of each E-ADV
in its typical discourse movements in contrast to its non-typical discourse
movements. It thus seems counterproductive that all sequences (EADV...C/C...E-ADV) which are not statistically significant for certain E-ADVs
as such are not taken into account when establishing their combinatory
profiles, since these nonsignificant cases play an important role in
characterizing the overall use of the E-ADVs and connectors. In order to
allow for a holistic approach, we propose to use correspondence analysis
(CA) (Greenacre, 2017).
3. Correspondence Analysis (CA)
The correspondence analysis presented in this section was performed using
the R software and the package “CA” (Nenadić & Greenacre, 2007). (1) In
DDA, representing the 18th century, certes has a use which stands out. Certes
left and right of mais differ clearly from all other E-ADVs as to their
associations with the connectors. Certes is not typically used with any other
connector analyzed and, most importantly, its association is not stronger
with mais on its right than it is with mais on its left. Conversely, in all other
five sub-corpora (encyclopedic and press corpora), which represent the 21st
century, there is an important difference between the use of certes right and
left of mais: while certes_L is strongly linked to mais, certes_R is not. (2) In all
six sub-corpora, mais appears to be opposed to all other connectors when it
comes to its associations with E-ADVs. Its central position appears to be
linked to its high frequency, indicating its high contribution to the horizontal
axis, this being confirmed by the analysis of the correspondence analysis
indicators. (3) An association between sans doute_L and parce que can be
observed in DDA and WIKI, whereas in UNI, the adverb and the connector
appear to be in the opposite relation. This behavior indicates variation has to
be expected even within the encyclopedic sub-corpus, based on at least two
parameters: on the one hand, the diachronic parameter is involved in some
discursive uses of E-ADVs, like certes_L and certes_R showing no difference
as to their association with mais in DDA, consistently with its different
meaning at that time, whereas only certes_L is associated with mais in all
other sub-corpora; on the other hand, some convergence between DDA and

JADT’ 18

673

WIKI could be interpreted as showing similarities in writing style. (4) The
results of the correspondence analysis show that in all sub-corpora of one
particular genre, in most cases, the same E-ADVs are strongly associated
with the same connector or group of connectors (donc and finalement ; car and
parce que ; mais); this phenomenon is particularly pronounced in the subcorpora representing written press. The connector mais differs the most from
the other connectors in what concerns the strength of its associations.
Although mais is associated with most E-ADVs, its association appears to be
strong with only a few of them in all sub-corpora (certes_L being the only
constant), while most other connectors have a higher number of strong
associations. This indicates that certain discourse movements (such as EADV...car / parce que) seem to be rather regular, whereas certes...mais proves to
be a special association, although only in the 21st century corpora. (5) The
behavior of néanmoins in the Figaro 2008 corpus should be interpreted with
caution since the two axes describe only 10% of its variation.
4. Perspectives
Our first attempt to use correspondence analysis to study different discursive
movements has provided promising results regarding the genre and
diachronic variation of discursive functions of French epistemic adverbs in
these cases. We intend to further extend our analysis in three directions: First,
we would like to enlarge our corpora to see if this allows to extend the
paradigm of connectors, so as to give a better overview of the different
discursive movements that exist and to better represent the different
discursive functions of the E-ADVs that we have found. It would be
especially interesting to cover different diachronic spans of press, allowing
for a study of possible changes within this specific genre. Likewise, other text
types may be considered in order to better represent possible variation
between genres. Second, through the comparative analysis of the discursive
combinatory profiles of each E-ADV, we aim to identify regularities
concerning the rhetorical purpose of the sequence in which the E-ADV
typically occurs by understanding its motivation. For instance, beyond the
difference between a counter-argument, an argument, and a conclusion, there
is a more fundamental difference between a discourse movement used with
the rhetorical aim (i) to present a content as being in the discursive
background (when the E-ADV is followed by mais), (ii) to introduce a content
which the speaker considers to be most relevant (when the E-ADV is
introduced by mais or donc), and (iii) to add evidence to a relevant content
(when the E-ADV follows car or parce que). Third, in order to confirm the
reliability and precision of the positions on the correspondence analysis
planes, our intention is to apply bootstrap validation (Lebart, 2010).

674

JADT’ 18

Figures 1-6. Correspondence analysis scatter plots for the six corpora.

JADT’ 18

675

References
Blumenthal P. (2008). Combinatoire des prépositions : approche quantitative.
Langue française, 157: 37-51.
Blumenthal P. (2012). Particularités combinatoires du français en Afrique :
essai méthodologique. Le français en Afrique, 27: 55-74.
Blumenthal P., Diwersy S. and Mielebacher, J. (2005). Kombinatorische
Wortprofile und Profilkontraste. Berechnungsverfahren und
Anwendungen. Zeitschrift für romanische Philologie, 121: 49-83.
Diwersy S. (2014). Corpus diachronique de la presse française : base textuelle
créée dans le cadre du projet ANR-DFG PRESTO. Institut des Langues
Romanes, Université de Cologne.
Greenacre M. J. (2017). Correspondence analysis in practice. 3rd ed. Boca Raton:
Chapman.
Lebart L. (2010). Validation techniques for textual data analysis. Statistica
Applicata - Italian Journal of Applied Statistics, 22(1): 37-51.
Nenadić O. and Greenacre M. J. (2007). Correspondence Analysis in R, with
two- and three-dimensional graphics: The ca package. Journal of Statistical
Software, 20(3): 1-13.
Rossari C., Hütsch A., Ricci C., Salsmann M. and Wandel, D. (2016). Le
pouvoir attracteur de mais sur le paradigme des adverbes épistémiques :
du quantitatif au qualitatif. In Mayaffre D. et al. (eds), Proceedings of 13th
International Conference on Statistical Analysis of Textual Data, II: 819-823.
Rossari C. and Salsmann M. (2017). Étude quantitative des propriétés
dialogiques des adverbes épistémiques. Actes des 9èmes Journées
Internationales de la Linguistique de corpus: 87-93.
Roulet E. (1979). Des modalités implicites intégrées en français contemporain.
Cahiers Ferdinand de Saussure, 33: 41-76.
Roulet E., Auchlin A., Moeschler J., Schelling M. and Rubattel C. (1991).
L'articulation du discours en français contemporain. 3rd ed. Bern: Lang.

676

JADT’ 18

Misleading information
in online propaganda networks
Vanessa Russo1, Mara Maretti2,
Lara Fontanella3, Alice Tontodimamma4
D’Annunzio University of Chieti-Pescara – russov1983@gmail.com
D’Annunzio University of Chieti-Pescara – mara.maretti@unich.it
3D’Annunzio University of Chieti-Pescara – lara.fontanella@unich.it
4D’Annunzio University of Chieti-Pescara – alicetontodimamma@gmail.com
1

2

Abstract 1
Nowadays, the spreading of inaccurate, false or misleading information over
the digital space is amplified by the increasing use of social networks and
social media. In different cases, misleading information can be linked to a
propaganda activity aimed at supporting offline organizations. In fact, in
such cases, online pages, conveying unintentionally (misinformation) or
intentionally (disinformation) inaccurate information, are embedded into a
network system composed by political and ideological advertise. In this
paper, we discuss the different structures of online networks linked to some
official pages of different political parties. The analyzed networks were
identified through Social Network Analysis.
Abstract 2
La diffusione di informazioni inesatte, false o fuorvianti nello spazio digitale
è amplificata dal crescente uso di social network e social media. In diversi
casi, tali informazioni approssimative e/o fuorvianti possono essere collegate
ad un'attività di propaganda volta a supportare organizzazioni offline.
Infatti, in questi casi, le pagine online, che trasmettono informazioni non
intenzionalmente (misinformation) o intenzionalmente (disinformation) errate,
sono incorporate in un sistema di rete composto da pubblicità politica e
ideologica. In questo articolo, discutiamo le diverse strutture delle reti online.
Le reti analizzate sono state identificate attraverso la Social Network
Analysis.
Keywords: misinformation, disinformation, propaganda activity, Social
Network Analysis
1. Background: misinformation and disinformation online
The development of the digital space relates to a new form of web-mediated
communication, which can be defined according to the following main
features. Web-communication can be thought of as a participative act and is

JADT’ 18

677

not part of a broadcast system (McLuhan, 1962) but is a networkcast system.
In fact, a web content generates connections, denoted as “Affinity networks”
(Rainie and Wellman, 2012; Castells, 2000), based on the sharing of a given
content. In this network system, Web-communication yields temporary
consensus areas based on alliances between users with respect to the shared
contents. Moreover, Web-communication favors a mobilization of skills that
generates new paths of social action and collective projects (Levy, 2002). In
the digital space, content validity relies on activism and interest of digital
users and every opinion “has citizenship rights” (Quattrociocchi and Vicini,
2016; Mocanu, 2015). In this framework, misinformation and disinformation
processes share the previous characteristics. Furthermore, the accidental or
deliberate propagation of false information is strictly linked to a “loss of
disintermediation” (Jenkins, 2006). According to this theory, one of the most
important effects of webmediated communication is the loss of traceability of
official information sources. In fact, phenomena like Wikipedia, Social Media
sites or Blog news produce the culture of unofficial knowledge, creating a
virtuous circle of free sources, on the one hand, and a vicious circle of
misleading information, on the other hand. Disinformation and
misinformation processes can be both related to Fake news and Hate
Speeches. “Fake news” or “Junk news” refers to web sources completely
invented or simply distorted. In fact, in the digital space, anyone gain access
at different information sources and can, also, create information content
with low costs and high distribution potential. Furthermore, the fake new
propagation process can develop into a viral system, dominated by the high
sharing power of different recurring themes. Usually, Hate Speech
phenomenon is linked to sharing and commenting fake news. Web 3.0 era is
permeated by hatred, mainly directed to immigrants, political parties and
homosexual people. Although hater activity concerns specific themes, it
becomes a fundamental part in redefining the digital public sphere (Lévy,
2002).
2. Research Design and Methodology
The disinformation and misinformation online phenomena have become a
propaganda activity to support offline organizations. In fact, in many cases
online fake news and hate speeches are contained within a network system
consisting of political and ideological advertising. In particular, this tendency
gained attention during Trump’s election campaign (Ott, 2017). The
Computational Propaganda Research Project, promoted by Oxford
University, aims at investigating «how tools like social media bots are used
to manipulate public opinion by amplifying or repressing political content,
disinformation, hate speech, and junk news». Woolley and Howard (2017),

678

JADT’ 18

mapping the computational propaganda in different countries, analyzed tens
of millions of posts on seven different social media platforms, referring to
elections, political crises and national security incidents. Each case study
takes into account qualitative, quantitative and computational evidences
collected between 2015 and 2017. In this framework, following a
computational approach (Lazer et al., 2009), our research aims at identifying
and comparing propaganda policy networks. For this purpose, we
investigated the networks in which different political Facebook Like pages
are embedded. More specifically, we selected the following Facebook Like
pages related to political institutional information: “Ricostruiamo il centro
destra” (Centre-Right wing), “Di Battista Alessandro” (Five Star Movement) e
“Partito Democratico” (Centre-Left wing). Exploiting Social Network Analysis
and focusing the attention on each of the chosen pages, we detected the
online networks. The analyzed adjacency matrices were built considering as
link the “likes”. The analysis was implemented using the free and open
source NodeXL extension of the Microsoft Excel spreadsheet (Hansen et al.,
2011). For each network, we present the centrality measures, which describe
how a particular vertex can be said to be in the “middle” of the network. In
particular, betweenness centrality measures how often a given vertex lies on
the shortest path between two other vertices. Vertices with high betweenness
may have considerable influence within a network by virtue of their control
over information passing between others. As pointed out by Hansen et al.
(2011), these measures can be thought of as a kind of “bridge” score, a
measure of how much removing a node would disrupt the connections
between other vertices in the network. Closeness centrality captures the
average distance between a vertices and every other vertex in the network. In
NodeXL the inverse of the average distance is implemented so that higher
closeness values indicate more central vertices. The Eigenvector Centrality
network metric takes into consideration not only how many connections a
vertex has (i.e., its degree), but also the degree of the vertices that it is
connected to. A node with few connections could have a very high
eigenvector centrality if those few connections were themselves very well
connected. These centrality measures allowed to identify the most relevant
nodes of each network. The identified Facebook Like Pages were classified in
“official pages” and “junk pages” according to their contents. Junk
information is strictly linked to the so-called post-truth politics, meaning a
political culture in which truth is no longer significant or relevant and
«objective facts are less influential in shaping public opinion than appeals to
emotion and personal belief» (Oxford Dictionaries, 2016). In this context, the
term junk information refers to fake news, conspiracy theories, hate speeches,
misinformation and deliberately misleading disinformation. Accordingly,

JADT’ 18

679

Facebook Like pages containing posts, comments or images conveying this
kind of information were classified as “junk pages”. It is worth noticing how
in the identified networks we did not retrieve hybrid forms, that is pages
composed of both official and junk contents.
3. Preliminary results
The network built by considering the Facebook Like page “Ricostruiamo il
centro destra” is depicted in Figure 1. This social media network, linked to a
Centre-Right political view, is composed by 159 nodes, comprising both
institutional
and
junk
pages
(e.g.
“unitaliasenzacomunisti”,
“SapereEundovere”). Centrality values, provided in Table 1 for the six pages
with higher levels of betweenneess centrality, highlight a connection between
junk and institutional nodes; furthermore, the influence of junk pages in the
network is very outstanding.

Figure 1. NodeXL social media network diagram of relationships derived from the Facebook
Like page “Ricostruiamo il centro destra”.
Table 1: Social media network of relationships derived from the Facebook Like page
“Ricostruiamo il centro destra”: centrality measures for the vertex pages with higher levels of
betweenness
Vertex
ricostruiamocentrodestra
unitaliasenzacomunisti
SapereEundovere
radionewsinformazionelibera
italianinonsonorazzistisonostanchi
diquestainvasione

Betweenness
Centrality
22644.000
10986.000
10044.000
1087.000

Closeness
Centrality
0.004
0.003
0.003
0.002

Eigenvector
Centrality
0.009
0.009
0.000
0.000

777.000

0.002

0.000

A similar situation was detected for the Five Star Movement. This network,
represented in Figure 2, is composed by 664 nodes comprising again both
institutional and junk pages. In this case, the junk pages are specifically of the
Five Star Movement and institutional pages are personal pages of political
candidates. The Five Star Movement network shows three big cluster in
which the central node (WIlM5s) is a junk page.

680

JADT’ 18

Figure 2. NodeXL social media network diagram of relationships derived from the Facebook
Like page “Di Battista Alessandro”.
Table 2: Social media network of relationships derived from the Facebook Like page “Di
Battista Alessandro”: centrality measures for the vertex pages with higher levels of
betweenness.
Vertex
Betweenness Centrality Closeness Centrality EigenVector Centrality
MassimoEnricoBaroni
281353.000
0.001
0.032
WIlM5s
172430.333
0.001
0.024
sorial.giorgio
143457.000
0.001
0.013
dibattista.alessandro
3405.667
0.001
0.006
pierrecantagallo89
1324.000
0.001
0.001
perchevotarem5s
702.000
0.001
0.003

The social media network of relationships derived from the Facebook Like
page “Partito Democratico” does not show the features found out for the
previous networks. In fact, the network related to the Centre-Left political
party is composed by only institutional propaganda pages.

Figure 3. Centrality measures for the social media network of relationships derived from the
Facebook Like page “Partito Democratico”.

4. Community clusters
The mapping process of propaganda pages resulted into different structures
of network. For the classification of these structures, we make use of the
model elaborated by Smith et al. (2014) in order to define a taxonomy of
social networks derived from conversations within Twitter. The authors
defined six types of Networks: polarized crowds, tight crowds, community
cluster, brand cluster, broadcast network and support network (see Figure 5).

JADT’ 18

681

Table 3: Social media network of relationships derived from the Facebook Like page “Partito
Democratico”: centrality measures for the vertex pages with a higher level of betweenness
Vertex
Betweenness Centrality Closeness Centrality EigenVector Centrality
partitodemocratico.it
46486.100
0.002
0.024
enricoletta.it
28853.657
0.002
0.047
scalfarotto
24167.162
0.002
0.038
giannipittella
23136.533
0.001
0.018
giovanidem
19798.000
0.001
0.011
palazzochigi.it
12633.519
0.001
0.009

Figure 5: Diagrams of the differences in the six types of social media networks (Smith et al
2014).

In this framework, we can recognize how the Centre-Right wing social media
network shows a conformation similar to a mixture of Polarized Crowd and
Support Network. On the one hand, the Polarized Crowd model is
characterized by two groups, polarized on specific opinions and sharing few
connections. On the other hand, the Support Network model consists of a
central node that sends information to the peripheral nodes. The Five Star
Movement social network adheres more closely to Tight Crowd and Support
network structures. The Tight Crowds is composed by highly connected
nodes and specific shared themes. Finally, the Democratic Party network
reflects the structures of a Community Cluster, which is organized in many
cliques that share specific topics of conversation.

682

JADT’ 18

4. Conclusions and future works
In this preliminary phase of our research, we considered the network
structures related to the online propaganda linked to different political areas.
Our analysis allowed to highlight the differences in the networks and to cast
the reconstructed networks into the taxonomy proposed by Smith et al.
(2014). In addition, in two out of the three analyzed social networks we
found out the presence of junk pages contributing to the disinformation and
misinformation processes by spreading out fake news and indulging in hate
speeches. The cluster structures of those two networks, leading to closed
circle of highly polarized information, facilitates the diffusion process of
misleading information. Based on these preliminary results, future works
will focus on the textual analysis of posts and comments shared on the
retrieved junk pages, in order to identify the main discussed topics. To this
end, Text mining and machine learning techniques will be exploited.
References
Castells M. (2000). The Rise of the Network Society, Blackwell Publishers Oxford
Hansen D. L., Schneiderman B., Smith M. A. (2011). Analyzing social media
networks with NodeXL: insights from a connected world, Morgan Kaufmann
Jenkins H., (2006). Fans, Bloggers and Gamers: Exploring partecipatory culture,
New York University Press.
Lazer D., Pentland A., Adamic L., Aral S., Barabási A.L., Brewer D.,
Christakis N., Contractor N., Fowler J., Gutmann M., Jebara T., King G.,
Macy M., Roy D., Van Alstyne M., (2009). Life in the network: the coming
of computational social science, Science 323(5915): 721–723
Lévy P. (2002). Cyberdémocratie. Essai de philosophie politique, Paris: O. Jacob
McLuhan, M. (1962). The Gutenberg Galaxy: the making of typographic man,
University of Toronto Press.
Mocanu, D.; Rossi L., Zhang Q., Karsai M., Quattrociocchi W. (2015)
Collective attention in the age of (mis)information. Computers In Human
Behavior, 51, 1198-1204
Ott B. L. (2017). The age of Twitter: Donald J. Trump and the politics of
debasement, Critical Studies in Media Communication, 34, (1): 59-68
Oxford Dictionaries (2016). Word of the Year 2016 Is...,
https://en.oxforddictionaries.com/word-of-the-year/word-of-theyear-2016.
Quattrociocchi W., Vicini A. (2016). Misinformation. Guida alla società
dell’informazione e della credulità, Franco Angeli.
Rainie L., Wellman B. (2012). Networked: The New Social Operating System, MIT
Press.
Smith M., Raine L., Shneiderman B., Himelboim I. (2014). Mapping Twitter
Topic Network: From polarized Crowds to community Cluster, Pew

JADT’ 18

683

Research Internet Project, February 20,
http://www.pewinternet.org/2014/02/20/mapping-twitter-topic-networks-frompolarized-crowds-to-community-clusters/#
Woolley S. C, Howard P. N. (2017). Computational Propaganda Worldwide:
Executive Summar,. Working Paper 2017.11. Oxford, UK: Project on
Computational Propaganda. comprop.oii.ox.ac.uk. 14 pp.

684

JADT’ 18

Topic modeling of Twitter conversations
Eliana Sanandres1, Camilo Madariaga2, Raimundo Abello3
1

Universidad del Norte – esanandres@uninorte.edu.co
2Universidad del Norte – cmadaria@uninorte.edu.co
3Universidad del Norte – rabello@uninorte.edu.co

Abstract
Topic modeling provides a useful method of finding symbolic
representations of ongoing social events. It has received special attention
from social researchers, particularly among cultural sociologists, in the last
decade (DiMaggio et al., 2013; Sanandres and Otalora, 2015). During this
time, Twitter has acted as the most common platform for people to share
narratives about social events (Himelboim et al., 2013). This study proposes
LDA (Latent Dirichlet Allocation) based topic modeling of Twitter
conversations to determine what topics are shared on Twitter in relation to
social events. The dataset for this study was constructed from public
messages posted on Twitter related to the financial crisis of the National
University of Colombia. Over an eight-week period, we downloaded all
tweets that included the hashtag #crisisUNAL (UNAL is the Spanish
acronym of the university) using the Twitter API interface. We analyzed over
45,000 tweets published between 2011 and 2015 using the R package
topicmodels to fit the LDA Model in five steps: first, we transformed the
tweets into a corpus, which we exported into a document-term matrix; the
terms were stemmed and the stop words, punctuation marks, numbers, and
terms shorter than three letters were removed. Second, we used the mean
term frequency-inverse document frequency (tf-idf) over documents
containing this term to select the vocabulary. We only included terms with a
tf-idf value of at least 0.1, which is a bit less than the median, to ensure that
the most frequent terms were omitted. Third, we defined the number of
topics k by estimating the log-likelihood of the model for each topic number
starting with 1 though to 300 topics and selected k = 12 because it had the
highest log-likelihood value (LL = -198000). Fourth, we run the LDA Model
for k = 12 topics. Fifth, we labeled the k = 12 topics previously identified by
choosing the top N terms ranked based on the probability of that topic. This
article illustrates the strength of topic modeling for analyzing large text
corpora and provides a way to study the narratives that people share on
Twitter.
Keywords: Topic modeling, LDA, Twitter.

JADT’ 18

685

1. Introduction
This article presents a way to analyze large amounts of textual data from
Twitter conversations in an efficient and effective way. Specifically, we
explain how to capture the narratives that people share on Twitter about
social events, reduce their complexity, and provide plausible explanations.
This is a research concern that has received special attention among social
researchers (Kovanović et al., 2015; Yann et al., 2011; Newman and Block,
2006; Griffiths and Steyvers, 2004), particularly among cultural sociologists,
who face the methodological challenge of working qualitatively with large
amounts of data (Sanandres and Otalora, 2015; Eyerman et al., 2011;
Alexander, 2004). In this paper we propose an LDA (Latent Dirichlet
Allocation) based topic model to address this challenge. Topic modeling is a
useful approach because the set of terms found within topics index
discursive environments or frames that define patterns of association
between a focal issue and other constructs (DiMaggio et al., 2013). These
patterns of association are to be interpreted as symbolic representations of
ongoing social events, which represent claims about the shape of social
reality, its causes, and the responsibility for action such causes imply
(Alexander, 2004). We applied an LDA-based model to Twitter conversations
about the financial crisis of the National University of Colombia to examine
how the debate over this crisis was framed on Twitter, from 2011 when it
emerged, until 2015. We analyzed over 45,000 tweets and illustrated the
strength of topic modeling for the analysis of large text corpora as a way to
study narratives shared on Twitter.
2. Background: The financial crisis of the National University of
Colombia
Over the last decade, Colombian academics and representatives of the
government have recognized that the limitations of their budgets are the
major limitation in the response of public universities to the increasing
demands of society. To face this problem, the government proposed to
reform the entire system of higher education (Ministry of National
Education, 2010). The intention was to find new sources of money for higher
education, enable more people to attend college, encourage transparency and
good governance in the education sector, and improve the quality of higher
education. One of the most controversial proposed changes was the opening
of the education sector to private investment by for-profit companies (El
Espectador, 2011). This was immediately rejected by public universities, who
claimed that the proposed reform would lead to a full-scale privatization of
the system of higher education (Semana, 2011).
At the public National University of Colombia, the largest higher education

686

JADT’ 18

institution in Colombia, some students and professors claimed that the
reform offered no clear solution to the financial crisis of the university. They
explained that the university had been using a funding model with its
sources of support mixed between the state and external resources, claiming
that since 2004 this model had borne dwindling state support and everincreasing costs to be covered by external resources. They showed that
government transfers had decreased from 70% in 2004 to 64% in 2013, while
the external resources produced from activities such as tuition fees, nonformal education courses, and academic extension services, among others,
had increased from 30% to 36% in the same period (National University of
Colombia, 2014). This statement reopened the debate on the financial crisis of
the National University of Colombia and became a Twitter trending topic
with the hashtag #CrisisUnal (UNAL is the Spanish acronym for the name of
the university).
3. The financial crisis of the National University of Colombia on Twitter
Here, we investigate how the financial crisis in the National University of
Colombia was framed on Twitter. It may be asked why we should care about
Twitter conversations on this topic? However, it should be considered that
Twitter conversations can offer clues to what the university is thinking and
doing about the crisis. A central advantage of using Twitter for analyses is
that it covers topics in real time, producing a large amount of data that can be
used to look at people’s perceptions and narratives of particular events.
Twitter also provides a practical way to examine collective experience related
to a topical event, to study behaviors and attitudes where social desirability
bias may occur in official surveys, and to collect large amounts of data with a
limited budget (Himelboim et al., 2013). Twitter conversations also illustrate
the views of the reading public and show dominant viewpoints, which
emerge quickly and are difficult to change (Xiong et Liu, 2014).
We collected every tweet published between 2011 and 2015 that contained
any reference to the financial crisis in the National University of Colombia
with the hashtag #CrisisUNAL. We chose this period to track Twitter
conversations around this topic, from the time it became a Twitter trend in
2011 through 2015 (the last year in which we collected data). Our collection
formed a corpus of over 45,000 tweets. In the next section we describe how
we used topic modeling.
4. Method
Topic modeling is a machine-learning method used to discover hidden
thematic structures in large collections of documents. In this work we used
LDA, a widely used method in topic modeling (Jelodar et al., 2017; Fligstein

JADT’ 18

687

et al., 2014), which assumes that there is a set of topics to be found in a
collection of documents. The intuition behind LDA is that documents exhibit
multiple topics. A topic is formally defined as a distribution of words over a
fixed vocabulary (Blei, 2012). For LDA, topics must be specified before any
data are generated. For each document in the collection, this method
generates the words in a two-stage process. During the first stage, it
randomly chooses a distribution over topics (step 1). In the second stage, for
each word in the document, it randomly chooses a topic from the distribution
over topics in step 1 (step 2a), and a word from the corresponding
distribution over the vocabulary (step 2b). At the end, each document
exhibits topics in different proportions (step 1) and each word in each
document is drawn from one of the topics (step 2b), where the selected topic
is chosen from the per-document distribution over topics (step 2a) (Blei,
2012). To run the LDA model, we followed five steps. First, we transformed
the tweets into a corpus and exported this corpus to a document-term matrix;
the terms were stemmed and the stop words, punctuation, numbers and
terms shorter than three letters were removed. Second, we used the mean
term frequency-inverse document frequency (tf-idf) to select the vocabulary.
We only included terms with a tf-idf value of at least 0.1, which is a bit less
than the median, to make sure that the most frequent terms were omitted.
Third, we defined the number of topics k by estimating the log-likelihood of
the model for each topic number, from 1 to 300 topics; we selected k = 12 as
having the highest log-likelihood value (LL = -198000). Fourth, we run the
LDA model for k = 12 topics. Fifth, we labeled the k = 12 topics previously
identified by choosing the top N terms, ranked according to the probability
of that topic. For this we used the R package topicmodels.
5. Results
Table 1 displays the 12-topic solution and lists the 10 highest-ranking terms
for each topic. We call attention to four sets of topics: six topics concerned
with social protest (dark shading), three topics on educational reform
(medium shading), two topics calling for investment (light shading), and one
topic emphasizing the role of the National University of Colombia in the
Colombian peace process (no shading). To more easily interpret the topics,
after reviewing the list of terms we examined those tweets that exhibited
each topic with the highest probability.
5.1 Protest topics
Protest topics are the focus of the Twitter conversations on the financial crisis
in the National University of Colombia. Topic 1 covers the protests of the
education workers. The most highly ranked terms were sintraunal (the labor

688

JADT’ 18

union covering all workers at public universities), protest, strike, campus, riot,
gas, blocked, and wall. The tweets in which this topic was strongly represented
locate protests in national and international contexts with terms like nation
and clacso (Latin American Council of Social Sciences), indicating that the
protests were a matter of concern in Colombia and in Latin America. Topic 3
also refers to the protests of the education workers. Some of the top words
are sintraunal, gases, wall, and block. This topic frequently exhibits tweets that
show negative aspects of protests, such as confrontation, death, and bombs.
Table 1: 12-topic solution
Topic 1
sintraunal
protest
strike
campus
riot
gas
blocked
wall
nation
clacso

Topic 2
agricultural
strike
graffiti
hate
block
bombs
terrorists
crash
delinquents
guevara

Topic 3
sintraunal
gases
wall
block
undefined
bombs
hood
criticism
death
confrontation

Topic 4
agrarian
protest
movement
mobilization
participation
people
bombs
poor
assembly
disturbance

Topic 7
defend
university
improvement
campus
crisis
infrastructure
cement
hospital
architecture
sociology

Topic 8
no to the
reform
propose
threat
oblivion
save
closed
blocked
abnormality
upedagogica
uncertainty

Topic 9
Stamp
demand
support
public
university
strike
resources
deserve
financial
pride

Topic 10
intimidation
blocked
abandoned
public
eviction
strike
che
graffiti
protest
worker

Topic 5
solidarity
no to the
reform
justice
march
respect
charge
help
block
upedagogica
studying
Topic 11
peace
process
mobilization
research
studying
participation
talks
intellectuals
solidarity
civil

Topic 6
no to the
reform
universities
listen
sciences
confrontation
media
classrooms
abandoned
mobilization
block
Topic 12
revolutionary
victory
popular
campus
strike
eviction
denounce
deserve
abandonment
took

Topics 2 and 4 refer to the agricultural sector protests. While Topic 4 is
related to the mobilization of people to take part in these protests, Topic 2
emphasizes the participation of terrorists and delinquents in agricultural strikes.
In this context, social protest is associated with the Argentine Marxist
revolutionary Ernesto Che Guevara. Che is also mentioned in Topic 10,
which deals with the protests of the working class and the intimidation of
protesters. The most highly ranked terms in this topic are intimidation, blocked,
abandoned, public, eviction, worker, strike, che, graffiti, and protest. Finally, Topic
12 covers the revolutionary cause of social protest and includes the words
revolutionary, victory, popular, campus, and strike.

JADT’ 18

689

5.2 Anti-reform topics
Five topics deal with the reforms of higher education proposed by the
government. According to the terms included in Topic 5, public universities
reject this reform and called for justice and respect; terms in this topic include
solidarity, no to the reform, justice, march, and respect; tweets representing this
topic show strong solidarity among public universities, specially from the
Universidad Pedagógica (upedagogica). Topic 8 is also related to the rejection
of the planned educational reform to save public education; this includes
terms like no to the reform, propose, threat, oblivion, and save; Universidad
Pedagógica (upedagogica) is mentioned as well. In the same way, Topic 6
indicates that public universities reject the reform of higher education,
mobilize to denounce the government’s abandonment, and demand to be
listened to; some of the words in this topic are: no to the reform, universities,
listen, sciences, confrontation, media, classrooms, abandoned, mobilization, and
block.
5.3 Investment topics
Topics 7 and 9 cover demands for investment to face the crisis. Topic 7 calls
for infrastructure investment. Many tweets in which this topic is prominent
focus on the infrastructure crisis of the campus buildings, in particular the
sociology and architecture buildings and the university’s hospital. The top terms
in this topic include defend, university, improvement, campus, crisis,
infrastructure, cement, hospital, architecture, and sociology. Topic 9 plays a
similar role in investment demands focusing on the pro-National University
of Colombia stamp, created to acquire financial resources to improve the
university facilities. Some tweets containing this topic highlight the role of
the University as a national pride. The top ranked terms include stamp,
demand, support, public, university, resources, financial, strike, deserve, and pride.
5.4 Peace topic
Topic 12 represents the integration of the crisis in the National University of
Colombia into a broader frame of national concern associated with the
Colombian peace process. The top-ranked terms are peace, process,
mobilization, research, studying, participation, talks, intellectuals, solidarity, and
civil. Tweets in which this topic was strongly represented are related to the
role of the university as facilitator in peace talks among the government,
rebel groups involved in the Colombia’s internal armed conflict (which
began in the mid-1960s and is currently in negotiation, in a process known as
the Colombian peace process), intellectuals, and representatives of civil
society.

690

JADT’ 18

6. Conclusions
Producing an interpretable way to study Twitter conversations efficiently
and effectively is only the beginning. The solution of this issue presents
meaningful categories to address the analytic question that motivated the
study: how was the financial crisis in the National University of Colombia framed
on Twitter? The 12-topic solution showed that it was framed through four
categories: protest, anti-reform, investment, and peace.
Each topic constitutes a frame, in that it includes terms calling attention to
particular ways in which the crisis under study may arouse controversy:
protest frames emphasize public displays, demonstrations and the civil
disobedience of the working class; anti-reform frames refer to the rejection of
the reform of higher education by public universities; investment frames
focus on investment demands to face the crisis; and the peace frame draws
attention to the role the National University of Colombia played in acting as
a facilitator in the Colombian peace process. Each of these frames represents
a discursive environment for the financial crisis, which broadcasts not just
the structural characteristics of the crisis (investment demands and education
reform), but also symbolic representations of ongoing social events (workers
protests and peace process), which can be seen as claims about ongoing social
processes and demands of reparation.
These results provide substantive insight into Twitter conversations about
the financial crisis in the National University of Colombia. Using LDA to
discover topics allowed us to locate two narratives: one focused on the
structural characteristics of the crisis and the other concerned with symbolic
representations of ongoing social events surrounding that crisis. For cultural
sociologists, this is only the beginning of the analysis. A topic model allows a
starting point to be found, which in this case is the structure of Twitter data.
Used properly, with appropriate validation, topic models are valuable
complements to other interpretive approaches, offering new ways to extract
topics and make sense of online data.
References
Alexander, J. (2004). Toward a theory of cultural trauma. In Alexander, J.,
Eyerman, R., Giesen, B., Smelser, N. and Sztompka, P. Cultural trauma and
collective identity. Univ of California Press.
Blei, D. (2012). Probabilistic topic models. Communications of the ACM, 55(4):
77–84.
DiMaggio, P., Nag, M., and Blei, D. (2013). Exploiting affinities between topic
modeling and the sociological perspective on culture: Application to
newspaper coverage of US government arts funding. Poetics, 41(6): 570–
606.

JADT’ 18

691

El Espectador (2011). Universidades con ánimo de lucro, apuesta del
gobierno. March 10.
Eyerman, R., Alexander, J. C., and Breese, E. B. (2011). Narrating trauma: on
the impact of collective suffering. Routledge.
Fligstein, N., Brundage, J. S., and Schultz, M. (2014). Why the Federal Reserve
failed to see the financial crisis of 2008: The role of “Macroeconomics” as a
sense making and cultural frame. IRLE Working Paper No. 111–14.
Griffiths, T., and Steyvers, M. (2004). Finding scientific topics. Proceedings of
the National academy of Sciences, pp. 5228–5235.
Himelboim, I., McCreery, S., and Smith, M. (2013). Birds of a feather tweet
together: Integrating network and content analyses to examine crossideology
exposure
on
Twitter. Journal
of
Computer-Mediated
Communication, 18(2): 154–174.
Jelodar, H., Wang, Y., Yuan, C., and Feng, X. (2017). Latent Dirichlet
allocation (LDA) and Topic modeling: Models, applications, a
survey. arXiv preprint arXiv:1711.04305.
Kovanović, V., Joksimović, S., Gašević, D., Siemens, G., and Hatala, M. (2015).
What public media reveals about MOOCs: A systematic analysis of news
reports. British Journal of Educational Technology, 46(3): 510–527.
Ministry of National Education (2010). Proposal for the education reform in
Colombia. April 12.
National University of Colombia (2014). Estadísticas e indicadores de la
Universidad Nacional de Colombia. 19. ISSN 2357-5646.
Newman, D., and Block, S. (2006). Probabilistic topic decomposition of an
eighteenth-century American newspaper. Journal of the Association for
Information Science and Technology, 57(6): 753–767.
Sanandres, E., and Otálora, J. (2015). Application of topic modeling for
Trauma Studies: The case of Chevron in Ecuador. Investigación &
Desarrollo, 23(2): 228–255.
Semana (2011). Reforma a la Ley 30: por qué sí, por qué no. April 1.
Yang, T., Torget, A., and Mihalcea, R. (2011). Topic modeling on historical
newspapers. In K. Zervanou & P. Lendvai (Eds.), LaTeCH ’11 Proceedings
of the 5th ACL-HLT Workshop on Language Technology for Cultural Heritage,
Social Sciences, and Humanities, pp. 96–104.

692

JADT’ 18

What volunteers do? A textual analysis of voluntary
activities in the Italian context
Francesco Santelli, Giancarlo Ragozini, Marco Musella
University of Naples Federico II
francescosantelli@unina.it marcomusella@unina.it

Abstract
The complex phenomena of volunteering was mainly analyzed in economic
literature with respect to its “economic value added”, i.e the capability of this
kind of activities to increase the level of productivity of some specific gods or
services. In this paper, the point of view switches and voluntary
organizations are analyzed as place of job market innovation, where new jobs
arise and where people acquire new skills. Thus, volunteering can be thought
as “social innovation” factor. In order to analyze the contents of voluntary
works we use data coming from Istat survey “Multiscopo, Aspetti della vita
quotidiana” (Multi-purposes survey, daily life aspects), for the year 2013. In
our textual analysis, we use information included in the open answers given
by people about the description of the tasks performed individually as
volunteer. After stemming, lemmatization, and cleaning, data have been
analyzed by means of Community Detection based on Semantic Network
Analysis in order to discover patterns of jobs and through Correspondence
Analysis on Generalized Aggregated Lexical Tables (CA-GALT) in order to
discover profiles of volunteers. In particular, we look for differences given by
gender, age, educational level, region of residence and type of voluntary
association.
Keywords: Text Mining, Volunteers, Lexical Correspondence Analysis,
Semantic Network Analysis
1. Introduction
Volunteer work differs from the traditional forms of work for several
features. Nevertheless, most of the authors approaching the volunteering
phenomenon are interested mainly in the economic value that this sector is
able to add to the labour market (Ironmonger, 2000; Salamon et al., 2011)
considering it like a special case of job in the economic theory framework.
From this point of view, volunteering is assumed to be a peculiar sector of
the production with a considerable number of divergent rules and dynamics
compared to the standard work patterns, but still able to provide goods and
services to the community like all the other sectors. It will lead, of course, to
increase the overall economic value of the society.

JADT’ 18

693

In this work, the focus will be instead from a different perspective:
volunteering will be considered as a laboratory of social innovation
embedded in the labour market. The main concept behind it is that
volunteering is based on different guidelines and different principles
(Zamagni, 2005); therefore, it could develop new professional profiles and
modify pre-existent ones. Text Mining approach will be performed on openend questions given by volunteers, assuming that their self-concepts is a
consistent proxy of volunteering world. The empirical statistical analysis will
make use of two tools chosen for their capability to profile both groups of
words and cluster of volunteers. The latter, in the Italian context, will be
analyzed in parallel with the traditional categories applicable to the classic
labor theory. It will be shown that most of the determinants of the
segmentation of the professions (Colombo, 2003), such as gender, age or
geographic area of origin, can be adopted as well in this framework.
2. Data and statistical approach
Data are taken from the Istat Survey of 2013 “Multiscopo, Aspetti della vita
quotidiana” (Multi-purposes survey, daily life aspects) (Istat, 2013). It is a
large annual sample survey that covers the resident population in private
households, by interviewing a sample of about 20000 households and about
50000 people with P.A.P.I. technique. The main dimensions questionnaires
concern education, work, family and social life, spare time, political and
social participation, health, life style and access to the services.
From the whole sample, we selected about 5000 persons that declared to be
involved in volunteering and that answered to open-end questions about
their voluntary activities and if they carried out it within an organization or
by themselves. The main core of the statistical text mining procedure will be
focused on these brief descriptions of their own volunteering jobs. We
analyzed the descriptions along with the socio-demographic variables
available: gender, age, geographic macro-area and educational level. Given
the definition of volunteering (Istat, 2013; Wilson, 2000), several descriptions
were erased from the database as they do not belong to voluntary activities
(e.g., people donating blood to AVIS organization, or people that provides
help to family members).
Therefore, after this preliminary procedure in order to delete inappropriate
or missing answers, the valid number of volunteers are 4254 from the
original 5000. Before going through the analysis, we perform a preliminary
transformation of the original lexical data by removing punctuation and
stop-words, and by stemming the words, i.e. deleting all the derivational and
inflectional suffixes (Lovins, 1968; Willet, 2006). Therefore, all the words that
evolved from the same root will be considered to be the same after the

694

JADT’ 18

stemming. For this task we use the Porter Stemming Algorithm using
software R implemented in the package tm (Meyer et al., 2008).
After the preliminary analysis, in order to discover groups of activities that
can be described as jobs we apply a Semantic Network Analysis (van
Atteveldt, 2008; Drieger, 2013), and in order to profile of voluntary jobs with
respect to socio-demographic dimensions we use Correspondence Analysis
on Generalized Aggregated Lexical Tables (CA-GALT) (Kostov et al., 2015).
The former is an extension of Social Network Analysis that treats text as
graph structure: each word is defined as a node, and the ties between words
are undirected links weighted by the count of co-occurrences (how many
times do these words appear together in the same answers). Groups of terms
corresponding to semantic clusters can be found through community
detection algorithms (Fortunato, 2010). We use the Fast Greedy method that
is suited to deal with undirected and weighted edges (Clauset et al., 2004).
On the other hand, the CA-GALT method allows us to jointly analyze in a
multiple correspondence framework both the lexical table and sociodemographic profiles, combining the document-term matrix and the matrix
containing the individual characteristics.
3. Main findings of the analysis
After the preliminary transformations, the overall corpus shows a high
degree of heterogeneity with 1649 different words, and a high level of
sparsity, close to 100% due to the large number of documents and their
shortness. The term frequency distribution has a median equal to 2, and a p0.75
percentile equal to 4. Given the sparsity, we focus the analysis on the most
frequent words that profile and describe voluntary activities, taking into
account only words that are above the p0.90 percentile (frequency equal to 11),
and ending up in a vocabulary consisting of 175 words. The most used of
them are organizz (to organize, or organization) that appears 296 times,
assistent (assistent) with 225 occurrencies, attiv (activity) that occurs 215
times, then assoc (association), aiut (to help) and volontar (volunteer and
derived words). Those terms can be considered pretty generic, and could be
related to several aspects inside the volunteers’ community, without showing
additional informative power to profile volunteers. They are followed by
terms describing specific field of intervention: sport, fond (fund), event,
bambin (child/children), anzian (senior/old). Further, some of them are
expressing just one semantic meaning, and can be considered bi-grams
(Collins, 1996): croce rossa (red cross), croce verde (green cross), croce bianca
(white cross), protezione civile (civil protection/defense), vigili fuoco
(firefighters), capo scout (scoutmaster). We merge them in the following.
Applying the Semantic Network and the community detection algorithm to

JADT’ 18

695

these data, we found 7 groups/communities. In Fig. 1 we plot the semantic
network along with the communities, in which words are colored according
to the community. It is possible to identify a set of “jobs” related to the
typical charity organizations, mainly in a religious context: the care of old
people and hospitalized people -ospedal, malat, assistenz, ascolt, accud, cur,
sostegn- (orange), the education and animation od disadvantaged children,
mainly in religious organizations -insegn, parrocc, scuol, orator, cateches, anim(purple), the food and cloth drive and its distribution to the poor -cibo, vestiar,
caritas, raccolt, aliment, mens, pover- (green). Another large group is related to
the executives and officers of organizations and to the cultural events
organizers -organizz, event, cultural, membr, consigl, dirigent, reunion- (blue).
Related to this large group we found the musicians (black) characterized by
suon, band, musical. Finally, the last important area of the network is
associated to the organized volunteers on the territory -vigilefuoc,
protezionecivil, territor, croceross, soccors, ambul- (red). The coaches are mixed
with this group -squadr, allen, calc, pallavol- (brown). All these activities are
mainly done in nonreligious organizations and are not directly related to
charity aims.
Analyzing categories and lexical CA in (fig:2) is possible to profile
individuals according to their demographic status. In this context is not
performed a real clustering procedure, but as in classical Correspondence
Analysis the two spaces, units and variables, are linked taking into account
that words close to a specific categories are more likely to occur for people
belonging to the given category. It is clear that there is a gender gap: men are
related to sport activities, they play music in band, they are driver (mainly
ambulance) and they are involved in administration tasks. Women are more
involved in providing services to individuals (taking care of children and old
people), also carrying out food and cloth drive for the poor. Geographic
differences come up as well: volunteers from North-Est and North-Ovest
describe their activities as manutenzion, dirigent, addett, consigl, showing a
higher organization level. South and Islands are more related to a female
style of volunteering, with a predisposition for religious organization and
mainly aimed to assistance. Educational level and age have an impact: lowest
level of education, crossed with age information, profile a group of old and
less educated volunteers involved in religious volunteering. Highest
educated people carry out mainly administrative tasks. The central group of
age (35-64) shows, on the other hand, an average profile close to the origin of
axis, as well as people from Center Italy.

696

JADT’ 18

Figure 1: Semantic network: different colors for different communities identified by
FastGreedy algorithm. Size of words and width of the edges are proportional to the weights

4. Discussion and conclusion
As introduced in the first section, the aim of this work is to present a general
perspective about volunteering work in Italy under the assumption that is
possible to study it in an analogue way in which labour market is studied in
classic economic literature. Some authors already gave example how it
follows also the rule of supply and demand under given condition (i.e. Wolff
et. al, 1993) and also volunteering companies make use of marketing
strategies similarly to business companies (Dolnicar et Randle, 2007). The
two different statistical tools presented in previous section give to the
empirical analysis different hints, and are somehow complementary.
Communities in Semantic Network of (fig:1) are based on the connection
level between words, without taking into account other previously known
characteristics of individuals. Communities thus discovered are groups of
words that define several activities and so clusters of jobs in some specific
fields. In the second analysis, both spaces build in Ca-Galt, individuals and
categories, stress out how segmentation is clearly present in volunteering as
in labour market, and words used (and so activities done) change for gender,
education, age and macro-area, in an equivalent way as for standard jobs. It

JADT’ 18

697

gives so an overview about the relationships between words (as description
of activities) and categories (socio-demographic variables). Summing up,
both analysis highlight how volunteering is complex and heterogeneous; it
shows that people involved are in some cases highly skilled, often using
some of the competencies trained in their life. Generally, they are able to
describe their activities in a thorough way, explaining openly the aim of their
voluntary jobs. The Text Mining analysis presented in this work could lead to
figure out some needs of the population that are not adequately satisfied,
given the assumption that volunteers spend their time and use their skills to
give something to individuals that strongly ask for demands, in a framework
similar to supply and demand mechanism. Furthermore, to have a more
exhaustive overview for future policies to undertake, next step could be
likely to go on the other side; another survey should be done asking people
why do they ask help to volunteers. It will lead to better understand the real
needs of individuals that are not fully satisfied of what they get in terms of
assistance, especially from official institutions welfare.

698

JADT’ 18

Figure 3: Ca-Galt for both terms (blue) and categories (red). Overlapping both factor maps is
possible to profile cluster of individuals.

References
Amati, F., Musella, M. and Santoro, M. (2015). Per una teoria economica del
volontariato. (Vol. 1). G. Giappichelli Editore, Torino
Clauset, A., Newman, M. E., and Moore, C. (2004). Finding community
structure in very large networks. Physical review E, 70(6), 066111.
Collins, M. (1996). A new statistical parser based on bigram lexical
dependencies. In Proceedings of the 34th annual meeting on Association for
Computational Linguistics, 184-191, Association for Computational
Linguistics
Colombo, A. (2003). Razza, genere, classe. Le tre dimensioni del lavoro domestico
in Italia. Polis, 17(2), 317--344,
Dolnicar, S. and Randle, M. (2007). The international volunteering market:
Market segments and competitive relations. International Journal of
Nonprofit and Voluntary Sector Marketing, 12(4), 350-370.
Drieger, P. (2013) Semantic network analysis as a method for visual text
analytics, Procedia-social and behavioral sciences, 79, 4 – 17
Fortunato, S. (2010). Community detection in graphs. Physics reports, 486(35), 75-174.
Indagine Istat Multiscopo sulle famiglie: aspetti della vita quotidiana, (2013),
Retrieved from http://www.istat.it/it/archivio/91926
Ironmonger, D. (2000). Measuring volunteering in economic terms.
Volunteers and Volunteering, The Federation Press, Sydney, 56--72
Kostov, B., Bécue Bertaut, M. and Husson, F. (2015). Correspondence analysis
on generalised aggregated lexical tables (CA-GALT) in the FactoMineR
package, R Journal, 7(1), 109 -- 117,

JADT’ 18

699

Lovins, J. (1968). Development of a stemming algorithm. Mech. Translat.
Comp. Linguistics, 11(1-2), 22--31, (1968)
Meyer, D., Hornik, K., and Feinerer, I. (2008). Text mining infrastructure in R.
Journal of statistical software, 25(5), 1-54.
Salamon, L., Sokolowski and S., Haddock, M. (2011). Measuring the
economic value of volunteer work globally: Concepts, estimates, and a
roadmap to the future, Annals of Public and Cooperative Economics, 82(3),
217--252, (2011)
van Atteveldt, W. (2008). Semantic network analysis: Techniques for extracting,
representing, and querying media content, BookSurge Publishers, Charleston
SC
Willett, P. (2006). The Porter stemming algorithm: then and now. Program, Vol.
40 Issue: 3, 219--223, doi: https://doi.org/10.1108/00330330610681295
Wilson. J. (2000). Volunteering, Annual review of sociology, 26(1), 215—240
Wolff, N., Weisbrod, B. A., and Bird, E. J. (1993). The supply of volunteer
labor: The case of hospitals. Nonprofit Management and Leadership, 4(1),
23-45.
Zamagni, S. (2005). Gratuità e agire economico: il senso del volontariato. In
Working Paper presented at Aiccon meeting, Bologna

700

JADT’ 18

A longitudinal textual analysis of abstract presented
at Italian Association for Vocational guidance and
Career Counseling’ Conferences from 2002 to 2017
S. Santilli1., S. Sbalchiero2, L. Nota3, S. Soresi4
2

1 University of Padova – sara.santilli@unipd.it
University of Padova – stefano.sbalchiero@unipd.it
3 University of Padova – laura.nota@unipd.it
4 University of Padova – salvatore.soresi@unipd.it

Abstract
This new century is characterized by phenomena such as globalization,
internationalization, and rapid technological advances, that influence people
life and the ways in which they seek and do their jobs. Changing the shape
of organizations changes the shape of careers. To better account for the
complexities of work due to the least socio economic crisis, the Life Design
paradigm, a new paradigm for career theory in the 21st century (Savickas et
al., 2009) has been recently developed an it represent the third wave of
career theory and practice. The first wave emerged as the psychology of
occupations in the first half of the 20th century to match people to jobs. The
second wave comprised the psychology of careers ascending at mid-20th
century to manage worker and other life roles across the lifespan.
The main aims of the present study was illustrate the changes in theory,
technique e measure emerged in the Italian vocational guidance and career
counseling psychology by the analysis of the abstract presented at Italian
Association for Vocational guidance and Career Counseling’ Conferences.
The corpus was composed of 1,250 abstracts that have been collected from
2002 to 2017. In order to compare and contrast the main semantic areas over
time, a topic analysis by means of Reinert's method (1983) was conducted
(IRaMuTeQ and R software) to detect the clusters of words that
characterized the different orientations over time. The results show that
career counseling theories and technique evolved during the time to better
assist workers in adapting to fluid societies and flexible organization and to
better help clients design their lives in 21st century.
Keywords: longitudinal textual analysis, career counseling, vocational
psychology
1. Introduction
In Western countries the economic recession that characterized the years

JADT’ 18

701

2008–2009 lead to a dramatic loss of jobs throughout the Union’s private
sector. Furthermore fast moving global economy and phenomena such as
globalization, internationalization, and rapid technological advances,
influence people’s lives and the ways in which they seek and do their jobs.
The world of work is in general much less clearly defined or predictable, and
employees face greater challenges in coping with work transitions (Savickas
et al., 2009). Therefore, life in a 21st-century requires new models and
methods to deal with the new issues such as uncertainty, inequalities,
poverty, immigration precariousness in the labor market, and with the
worrying consequences also on individual and relational wellbeing. For these
reasons existing traditional career guidance assumptions have been swept
away, together with other certainties, by the sudden changes that have taken
place in the world of work and in the economic field. To better account for
the complexities of work, the Life Design paradigm, a new paradigm for
career theory and intervention in the 21st century (Savickas et al., 2009) has
been developed. The psychology of life design advances a contextualize
epistemology emphasizing human diversity, uniqueness, and purposiveness
in work and career to make a life of personal meaning and social
consequence. Rather than matching self to occupation, it reflects a third wave
of career theory and practice. The first wave emerged as the psychology of
occupations in the first half of the 20th century to match people to jobs. The
second wave comprised the psychology of careers ascending at mid-20th
century to manage worker and other life roles across the lifespan. The third
wave arose as the psychology of life design to make meaning through work
and relationships.
The main aims of the present study was illustrate the longitudinal changes
that emerge in the Italian context regarding the models and the theoretical
paradigms that drive vocational guidance and career counseling by the
analysis of the abstract presented at the Italian Association for Vocational
guidance and Career Counseling 'Conferences. Specifically, we analyzed
differences between the abstract presented before the economic recession
(from 2002 to 2008) and during/after the economic recession (form 2009 to
2017) in the topics related to research, theories, and practice. The corpus was
composed of 1,250 abstracts that have been collected from 2002 to 2017.
2. Corpus and method
All the abstracts have been collected by the Italian Association for Vocational
guidance and Career Counseling - SIO. SIO represents at the national and
international level a focal center in which the main scholars and practioners
converge, gather, share and compare the theories and practices in terms of
vocational guidance and career counseling. The Abstracts from the first

702

JADT’ 18

SIO’s Conference (2002) to the latest one (2017) were collected. No abstract
were collected during the year 2003, 2007, 2016, and 2014, becouse SIO has
not organize national conferences. The corpus is composed of 1,250
abstracts. The corpus was pre-processed by means of IRaMuTeQ and R
software (Ratinaud 2009; Sbalchiero e Santilli, 2017). The corpus was
normalized replacing uppercase with lowercase letters, and punctuation,
numbers and stop words have been removed because are not significant to
analyse the content of abstract. The pre-processing steps were useful to
reduce the redundancy and to provide homogeneity among forms. The
lexicometric measures (Tab.1) indicate that it is plausible to apply statistical
analysis of textual data to the corpus (Lebart et al., 1998). The corpus is
composed of 20,932 word-type and 462,034 word-tokens.
Tab. 1. Lexicometric Characteristics of the corpus

Number of texts
(V) Word-type
(N) Word-tokens
(V1) Hapax
(V/N)*100 = Type/Token Ratio
(V1/V)*100 = Percentage of hapax

1250
20932
462034
8902
4,53
42,53

Using the Reinert method (Reinert, 1983), we extracted a series of ‘lexical
worlds’. The texts was divided into elementary content units of similar
length, then, the algorithm provides reports on ‘words x units’ matrix. The
classification of units consents to identify and extract only parts of texts
relating to the same topic, so for each cluster the list of the most significant
words calculated using the chi square measurement, are identified (Reinert,
1993; Sbalchiero and Tuzzi 2016; Sbalchiero e Santili, 2017).
3. Results
The analysis conducted by means of Reinert’s method detected five different
lexical worlds, as the dendrogram shows (Fig. 1). The methods identify the
lexical worlds quite well because 98,42% of the abstracts have been classified
and the words in the same sematic area are semantically associated, i.e. they
refer to the same issue.
Specifically, the first class of the present corpus refers to career counselor’s
professional knowledge, skills, resources and training. The second class
refers to the principals variables and constructs related to vocational
guidance and career counseling, such as self-efficacy, personality, coping,
intelligence, emotions, satisfaction, optimism. The third class include the
statistic measure and instruments used in vocational guidance to assess

JADT’ 18

703

people career self and personality. The fourth class refers to context variables,
to the supports and barriers for inclusion, rights of people with
vulnerabilities (people with disabilities, psychological sidelines, etc.). The
fifth class includes the guidance services, projects, career guidance activities
that are provided by local centers (university, region, province).
As already mentioned, differences between the abstract presented before the
economic recession (pre-crisis: from 2002 to 2008) and during/after the
economic recession (post crisis: form 2009 to 2017) were analysed. These two
period in vocational guidance history are specific because the stable
employment and secure organizations of the pre-crisis have in post crisis
given way to a new social arrangement of flexible work and fluid
organization, causing people tremendous distress, making difficult to
comprehend career with theories that emphasize stability than mobility.
Furthermore, it seemed interesting to analyze whether differences could be
found in the theories and techniques presented in the abstract in pre e post
crisis. To differentiate between papers presented pre- and post-crisis, a
specific procedure was used based on the Chi2 association of semantic
classes (Ratinaud, 2014) over the two period of time (Fig. 2).
The classes related to the pre-crisis are three and five characterised by
statistic measure and instruments used in vocational guidance to assess
people and guidance services, projects, and career guidance activities. The
post-crisis period is characterized by the class four, that refers to context
variables, to the support and barriers for inclusion, rights of people with
vulnerabilities.

Fig. 1: Cluster Dendrogram and list of most relevant words for each lexical world (in
descending order according to the Chi2 value of each class).

704

JADT’ 18

Fig. 2: Comparison among pre-crisis and post-crisis papers

These results highlighted that the topics presented in the abstract related to
pre-crisis are more oriented towards “people” focusing on the assessment
and measure with a statistical background. In the post-crisis period, the
attention of counsellors is more oriented toward the “environment” in which
people live and the relation between people and their context, so the
uniqueness and the vulnerability of people are considers in relation to social
and work inclusion. Finally, in order to compare and contrast the main
semantic areas over time, the classes were analysed using the Chi2
association of semantic classes and their distribution over years (Fig. 3).

Fig. 3: Comparison among classes and their distributions over years

JADT’ 18

705

In addition to the classes already analyzed in the pre and post crisis periods,
the comparison among classes and their distributions over years, highlights
also class 1 and class 2, which can be considered as evergreen in the
vocational guidance and career counseling field because they are present
throughout almost the entire period considered. The class 1 refers to career
counselor’s professional knowledge, skills, and competences. The class 2
refers to variables and constructs related to vocational guidance and career
counseling such self-efficacy, coping, life satisfaction, and positive attitudes.
4. Conclusions and discussion
The aim of the present study were to highlight the changes in theory,
technique e measure emerged in the Italian vocational guidance and career
counseling psychology by the analysis of the abstract presented at Italian
Association for Vocational guidance and Career Counseling’ Conferences.
The results show five different lexical worlds classes, related to career
counselor’s professional knowledge, variables and constructs of vocational
guidance and career counseling, measure and instruments to assess people
career self and personality, context variables to support inclusion of people
with vulnerabilities, and career guidance services and center.
Differences between the abstract presented before the economic recession
(pre-crisis: from 2002 to 2008) and during/after the economic recession (post
crisis: form 2009 to 2017) were also analysed. The results shows that career
counseling theories and technique evolved during the time to better assist
workers in adapting to fluid societies and flexible organization and to better
help clients design their lives in 21st century. In fact, while in the abstracts
relented to the pre-crisis period, emphasis is given to all those guidance
activities that consider particularly important to allow the person to collect
information about their characteristics and needs before advancing decisionmaking hypotheses (measure and instrument for the assessment), in the
abstracts related to the post-crisis period attention is paid to the "contexts"
where people live. Career guidance practices that are limited to the analysis
of "attitudes" and "interests" are considered obsolete, while current policies,
challenges, socio-economic conditions, the way in which vulnerability is
conceptualised are inputs from the environment which act at various levels
and on which scholars should pay attention (Shogren, Luckasson, &
Schalock, 2014).The evolution of social sciences that revolve around
orientation is undoubtedly a very complex phenomenon. Career scholars and
practioners should support people's needs taking into account the
organizational and environmental context in which they develop and take
shape. Currently the career guidance theory and model are numerous and
not always denominated and defined in the same way by the various authors

706

JADT’ 18

and scholars. For these reasons is important to analyze and understand the
different model developed over the time in order to activate a continuous
comparison in the field of career counselor’s competences that produces
precise trajectory regard the constructs to develop in the people by program
and activity provided by career services. In fact, noteworthy is the result that
highlights how the classes that refer to vocational guidance and career
counseling are presented throughout the entire period considered.
Nevertheless, these are just some results and other analyzes will be useful for
examining the peculiarities that these specific classes assume during the
years considered, in order to identify the specific skills and constructs that
characterized different historical periods. It could also be important to
compare the results that emerged in the Italian context with those of other
European and North American contexts, to generalize the results obtained.
References
Lebart, L., Salem, A., & Berry, L. (1998). Exploring textual data. Kluwer
Academic Publishers: Dordrecht.
Ratinaud, P. (2014). Visualisation chronologique des analyses ALCESTE:
application à Twitter avec l’exemple du hashtag #mariagepourtous. Actes
des 12es Journées internationales d’Analyse statistique des Données Textuelles.
Paris Sorbonne Nouvelle–Inalco.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique:
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, 8(2), 187-198.
Reinert, M. (1993). Les «mondes lexicaux» et leur «logique» a` travers
l’analyses tatistique d’un corpus de re´cits de cauchemars. Langage &
Société, 66, 5–39.
Shogren, K. A., Luckasson, R. & Schalock, R. L. (2014). The fefinition of
“context” and its application in the field of intellectual disability. Journal
of Policy and Practice in Intellectual Disabilities, 11(2), 109-116.
Savickas, M. L., Nota, L., Rossier, J., Dauwalder, J. P., Duarte, M. E.,
Guichard, J., ... & Van Vianen, A. E. (2009). Life designing: A paradigm for
career construction in the 21st century. Journal of Vocational Behavior, 75,
239-250.
Sbalchiero, S. & Santilli, S. Some introductory methodological notes. In L.
Nota & S. Soresi (Eds.), For A manifesto in favor of Inclusion. Florence:
Hogrefe Editore
Sbalchiero, S., & Tuzzi, A. (2016). Scientists’ spirituality in scientists’ words.
Assessing and enriching the results of a qualitative analysis of in-depth
interviews by means of quantitative approaches. Quality & Quantity,
50(3), 1333-1348.

JADT’ 18

707

A la poursuite d’Elena Ferrante
Jacques Savoy
Université de Neuchâtel (Suisse) – Jacques.Savoy@unine.ch

Abstract
The objective of an authorship attribution model is to determine, as
accurately as possible, the true author of a document, literary excerpt,
threatening email, legal testimony, etc. Recently a tetralogy called My
Brilliant Friend has been published under the pen-name Elena Ferrante, first
in Italian and then translated into several languages. Various names have
been suggested as possible true author (e.g., Milone, Parrella, Prisco, etc.).
Based on a corpus of 150 contemporary Italian novels written by 40 authors,
two computer-based authorship attribution methods have been employed to
answer the question “Who is the secret hand behind Elena Ferrante?” To
achieve this objective, the nearest neighbor (k-NN) approach was applied on
the 100 to 2,000 most frequent tokens using the Delta model. As a conclusion,
we found that Domenico Starnone is the true author behind Elena Ferrante’s
pseudonym. As a second approach and using the entire vocabulary, Labbé’s
model confirms this finding.
Résumé
L’objectif d’un modèle d’attribution d’auteur consiste à identifier, de la
manière la plus fiable possible, le véritable auteur d’un document, extrait
d’une œuvre, d’un courriel menaçant ou d’un testament. Récemment, la
tétralogie débutant avec L’amica geniale (Une Amie Prodigieuse) a été publié
sous le nom de plume d’Elena Ferrante, d’abord en italien puis traduite dans
plusieurs langues. Plusieurs noms ont été proposés comme le possible
véritable écrivain (par exemple, Milone, Parrella, Prisco, etc.). En s’appuyant
sur un corpus composé de 150 romans contemporains italiens écrit par 40
auteurs, deux méthodes d’attribution d’auteur ont été utilisés pour
déterminer qui se cache derrière le pseudonyme Elena Ferrante. Dans ce but,
la technique du plus proche voisin a été appliquée sur la base des 100 à 2 000
vocables les plus fréquents avec le modèle Delta. Comme conclusion, on
aboutit au nom de Domenico Starnone comme la véritable identité de Elena
Ferrante. Comme deuxième approche basée sur l’ensemble du vocabulaire, le
modèle de Labbé confirme cette conclusion.
Keywords : Authorship attribution, corpus linguistics.
Mots-clés : Attribution d’auteur, linguistique de corpus.

708

JADT’ 18

1. Introduction
Avec la parution de L’amica geniale (2011) débute une tétralogie sur la vie à
Naples depuis les années 50. Cette série de romans rencontre un étonnant
succès, en particulier aux États-Unis. Toutefois, l’auteur indiquée, Elena
Ferrante, représente un pseudonyme dont la véritable identité n’a pas été
révélée. Des érudits et journalistes ont proposé plusieurs noms en tenant
compte de possibles similarités stylistiques ou en affirmant que l’auteur doit
connaître le Naples d’après-guerre, voire être une femme (par exemple, Erri
De Luca, Francesco Piccolo, Michele Prisco, Fabrizia Ramondino, …). Sur la
base des royalties versés, le journaliste C. Gatti (Gatti, 2016) affirme que la
plume de Ferrante est tenue par Anita Raja (femme de l’écrivain Domenico
Starnone). Aucune étude scientifique approfondie n’a abordé cette question,
mais une première ébauche indique que le véritable auteur serait Domenico
Starnone (Tuzzi et al., 2018). L’identification du véritable auteur de ces
romans nous rappelle les investigations sur les relations Gary-Ajar en France
dans les années 1970. Dans le monde anglo-saxon, la parution de The Cuckoo’s
Calling (2013) sous la signature de R. Galbraith correspond à une affaire
similaire puisque le véritable auteur était J. K. Rowling (Juola, 2016). La
découverte d’un poème inédit soulève également la question de son véritable
auteur (Thisted & Efron, 1987), (Craig & Kinney, 2009). Pour lever le voile sur
l’identité exacte de Ferrante, notre étude dispose d’un corpus de 150 romans
italiens contemporains. De plus, on s’appuiera sur deux méthodes
d’attribution d’auteur (Juola, 2006) reconnues et ayant fait l’objet de plusieurs
études. En effet, afin d’admettre une preuve devant un tribunal celle-ci doit
posséder plusieurs caractéristiques (Chaski, 2013) comme, par exemple,
correspondant aux meilleures pratiques dans le domaine, avoir été testée et
pouvant être vérifiée et répliquée. Enfin, nous faisons l’hypothèse que le
véritable auteur derrière la signature Ferrante est bien l’un des 39 écrivains
italiens présents dans notre corpus (attribution dans un ensemble fermé).
2. Travaux reliés
Afin de déterminer l’identité d’un écrivain, trois paradigmes principaux ont
été proposées (Juola, 2006), (Stamatatos, 2009). D’abord, on s’est appuyé sur
des mesures stylométriques admises comme invariantes pour chaque auteur,
à l’exemple de la longueur moyenne des phrases, la taille du vocabulaire par
rapport à la taille du document (TTR) (Rexha et al., 2016). Face à des textes de
tailles variables, ces mesures s’avèrent d’être instables (Baayen, 2008).
Deuxièmement, les choix lexicaux permettent de différencier les auteurs, tant
dans la sélection des mots que dans leur fréquence d’occurrences ; « Le style
c’est l’homme » disait Buffon en 1753). Dans ce but, Mosteller & Wallace
(1964) proposent de sélectionner semi-automatiquement les vocables les plus

JADT’ 18

709

pertinents. Burrows (2002) choisit les mots les plus fréquents et, en
particulier, les mots fonctionnels (déterminants, prépositions, conjonctions,
pronoms et verbes auxiliaires). Ces derniers possèdent l’avantage d’être plus
fortement reliés au style de l’auteur qu’à la sémantique. Cette liste
comprendra entre 50 à 1 000 vocables les plus fréquents (Hoover, 2007), voire
l’ensemble du vocabulaire (Labbé, 2007). D’autres auteurs proposent de
définir a priori une telle liste (Hughes et al., 2012). Sur cette base, chaque texte
est représenté par les fréquences relatives d’occurrence des vocables
sélectionnés. Ensuite, une mesure de distance (ou de similarité) permet
d’estimer la proximité de deux textes. L’attribution s’établit habituellement
selon la règle du plus proche voisin. Troisièmement, en recourant à des
modèles d’apprentissage automatique (Stamatatos, 2009) les attributs les plus
pertinents (mots, bigrammes de mots ou de lettres, partie du discours,
émoticons, etc.) peuvent être sélectionnés. Ensuite un classifieur est entraîné
pour générer les profils des auteurs retenus (Naïve Bayes, régression
logistique, SVM, apprentissage en profondeur (Kocher & Savoy, 2017), etc.).
Enfin, le texte d’attribution douteuse est représenté et le nom du profil le plus
similaire est retourné comme réponse.
3. Le corpus de romans italiens contemporains
Grâce aux efforts de A. Tuzzi et M. Cortelazzo (Université de Padoue), le
corpus PIC (Padova Italian Corpus) a été créé en 2017. Cette collection
contient 150 romans italiens couvrant la période de 1987 à 2016. Comme
l’indique le tableau 1, ce corpus contient des œuvres de 40 auteurs (dont
Elena Ferrante avec sept textes). Lors de sa création, les auteurs originaires de
Naples et de sa région ont été favorisés (10 noms indiqués en italique dans le
tableau 1), de même que les femmes (12, pour 27 hommes).
Ce corpus contient 9 609 234 formes, avec une moyenne de 64 061 mots par
œuvre (un seul écrit comprend moins de 10 000 formes). La longueur
moyenne des romans signés par Ferrante s’élève à 88 933 mots. Enfin, un
contrôle éditorial a été appliqué afin d’éliminer les éléments non-textuels
(titre courant, numérotation des pages, etc.) ainsi qu’une inspection de
l’orthographe. Ce corpus renferme donc des écrits de la même époque et
langue, du même genre littéraire et dont la qualité a été vérifiée. Le 7
septembre 2017, un workshop regroupant sept équipes de chercheurs s’est
tenu à l’Université de Padoue durant lequel le nom de Domenico Starnone a
été identifié unanimement comme l’auteur derrière les œuvres de Elena
Ferrante. Pour atteindre cette conclusion, notre approche s’appuie sur les
techniques suivantes.

710

JADT’ 18
Tableau 1 : Nom des écrivains inclus dans le corpus avec le nombre de romans

Nom
Affinati
Ammaniti
Bajani
Balzano
Baricco
Benni
Brizzi
Carofiglio
Covacich
De Luca
De Silva
Faletti
Ferrante
Fois

H/F
H
H
H
H
H
H
H
H
H
H
H
H
?
H

Nombre
2
4
3
2
4
3
3
9
2
4
5
5
7
3

Nom
Giodano
Lagiola
Maraini
Mazzantin
Mazzucco
Milone
Montesano
Morazzon
Murgia
Nesi
Nori
Parrella
Piccolo
Pincio

H/F Nombre
Nom
H
3
Prisco
H
3
Raimo
F
5
Ramondino
F
4
Rea
F
5
Scarpa
F
2
Sereni
H
2
Starnone
F
2
Tamaro
F
5
Valerio
H
3
Vasta
H
3
Veronesi
F
2
Vinci
H
7
H
3

H/F
H
H
F
H
H
F
H
F
F
H
H
F

Nombre
2
2
2
3
4
6
10
5
3
2
4
2

4. Identifier l’auteur derrière la signature Elena Ferrante
Notre étude débute par l’application du modèle Delta (Burrows, 2002) dans
lequel la sélection des attributs stylistiques correspond aux k vocables les
plus fréquents. Toutefois, aucune limite précise pour le paramètre k n’est
indiquée et des travaux précédents (Savoy, 2015) soulignent que des valeurs
entre 200 et 500 tendent à apporter les meilleures performances. Cette limite
fixée, la méthode Delta estime un Z score pour chaque vocable ti basé sur la
fréquence relative (dénotée rtfij pour le terme ti et dans le document Dj)
comme indiqué par l’équation 1 (avec meani indique la fréquence moyenne
du vocable et si son écart-type).
Z score(tij) = (rfrij – meani) / si
Pour chaque auteur, on concatène tous ses écrits pour générer son profil Aj.
Enfin, on calcule la distance entre la représentation du texte à attribuer
(dénotée Q) et les profils des auteurs Aj (voir équation 2). Ensuite, les
différents auteurs peuvent être triés avec la plus faible distance signalant
l’auteur le plus probable. Le tableau 2 redonne les trois premiers auteurs avec
des valeurs pour k = 200, 300 et 500. Dans la dernière colonne (Stopword), les
vocables choisis correspondent uniquement aux mots fonctionnels de l’italien
(k = 307).

Le tableau 2 nous renseigne sur l’attribution du roman L’amica geniale (2011).
En considérant les six autres ouvrages, le même nom apparaît au premier
rang. De même, si le nombre de vocables s’élève à 50, 100, 150, 250, 400,
1 000, 1 500 ou 2 000, nous retrouverons toujours Starnone en première place
et ceci pour toutes les œuvres de Ferrante.

JADT’ 18

711

Une analyse plus fine des distances du tableau 2 indique que la différence (en
pourcentage) entre les distances du premier et deuxième rang présente des
valeurs nettement supérieures à celles entre le deuxième et troisième rang.
Ainsi, si k = 200, la différence entre 0,524 et 0,686 s’élève à 30,9 % tandis que
celle entre 0,686 et 0,700 n’est que de 2,0 %. Le premier nom proposé se
détache clairement des autres.
Dans une deuxième série d’expériences, nous avons regroupé tous les
romans attribués à Elena Ferrante pour en former qu’un seul texte (ou profil).
En variant le nombre de vocables de 50, 100, 150, 200, 250, 300, 400, 500, 1 000,
1 500 à 2 000, Starnone se retrouve toujours au premier rang des auteurs
ayant la plus forte similarité avec le profil d’Elena Ferrante.
Tableau 2 : Listes triées des auteurs les plus probables pour L’amica geniale (méthode Delta)
k = 300
k = 500
Stopword
k = 200
Rang Distance Auteur Distance Auteur Distance Auteur Distance Auteur
1
0,524
Starnone
0,515 Starnone 0,505 Starnone 0,421 Starnone
2
0,686
Veronesi
0,684
Brizzi
0,686 Veronesi 0,640
Milone
3
0,700
Balzano
0,719 Veronesi 0,710
Brizzi
0,660 Veronesi

Comme second modèle d’attribution d’auteur, l’approche de Labbé (2007)
suggère de recourir à l’ensemble du vocabulaire. Dans ce cas, la distance
entre deux textes A et B (indiquée par D(A,B) dans l’équation 3) dépend des
fréquences absolues des vocables dans les deux textes (dénotées par tfiA,
respectivement tfiB, avec i = 1, 2, …, k). La variable nA (ou nB) signale la
longueur de l’écrit A (en nombre de formes). Comme les deux textes ne
possèdent pas des tailles identiques, les fréquences du plus long (B dans
l’équation 3) seront multipliées par le rapport des tailles (voir partie droite de
l’équation 3). Enfin, les valeurs D(A,B) seront comprises entre 0 (aucun mot
en commun) et 1 (mêmes mots avec des effectifs identiques).

avec
En appliquant cette méthode, une distance est calculée entre chaque roman et
la distance permet de trier les couples d’écrits, de la plus faible distance à la
plus grande. Le corpus PIC génère (150 x 149) / 2 = 11 175 couples. Un extrait
est repris dans le tableau 3.
Dans ce tableau, la première place correspond aux deux œuvres les plus
similaires, deux romans écrits par Ferrante dans notre cas, soit Storia di chi
fugge e di chi resta (Id : 51, (2013)) et Storia della bambina perduta (Id : 52, (2014)).
Les deux autres romans de la tétralogie suivent, du deuxième au quatrième

712

JADT’ 18

rang, soit avec Storia del nuovo cognome (Id : 50, 2012) et L’amica geniale (Id : 49,
(2011)). En cinquième position, on rencontre deux écrits de Faletti, soit Niente
di vero tranne gli occhi (Id : 42, 2004) et Io sono Dio (Id : 44, 2009), puis deux
romans de Veronesi (Id : 145, Caos calmo (2009) et Id : 147, Terre rare (2014)).
Avec des distances faibles, les appariements s’opèrent entre des œuvres
rédigés par le même auteur et dans un intervalle de temps assez court.
Tableau 3 : Liste triée des romans les plus similaires (méthode Labbé)

Rang
1
2
3
4
5
6
…
43
…
63

Distance
0,140
0,148
0,155
0,157
0,165
0,166
…
0,228
…
0,241

Id.
51
50
49
50
42
145
…
47
…
108

Auteur 1
Ferrante
Ferrante
Ferrante
Ferrante
Faletti
Veronesi
…
Ferrante
…
Raimo

Id.
52
51
50
52
44
147
…
127
…
147

Auteur 2
Ferrante
Ferrante
Ferrante
Ferrante
Faletti
Veronesi
…
Starnone
…
Veronesi

Lorsque la distance augmente, la probabilité de rencontrer le même auteur
pour les deux ouvrages reliés diminue. Le premier lien apparament incorrect
se situe au 43e rang avec un écrit de Ferrante (Id : 47, I giorni dell abbandono
(2002) apparié avec un de Starnone (Id : 127, Eccesso di zelo (1993)). Un
appariement entre ses deux auteurs apparaît également au rang 44, 53, et 54,
avant que l’on découvre un autre type d’erreur en position 63 reliant un
roman rédigé par Raimo (Id : 108, Il peso della grazia (2012)) et un autre de
Veronesi (Id : 147, Terre rare (2014)). Puis, on découvre à nouveau un
appariement entre Ferrante et Starnone aux rangs 65, 69, 71, 72, 73, 74, soit un
total de dix couples entre ces deux auteurs et seulement un seul avec des
autres écrivains. Sachant que Ferrante correspond à un pseudonyme, la forte
similarité de style avec celui Starnone fait de ce dernier un choix de premier
ordre.
5. Analyse
Les choix lexicaux ne sont pas le fruit du hasard et chaque auteur a ses
préférences qui sont détectables par les mesures stylistiques. Le
rapprochement entre Ferrante et Starnone s’explique également en analysant
quelques exemples. Dans notre corpus, les sept romans de Ferrante
correspondent à 6,5 % de la taille tandis que 6,4 % est constitué par les dix
œuvres de Starnone. Si les fréquences d’occurrences de certains mots
s’écartent de ces proportions et dans la même direction pour les deux
auteurs, nous pouvons rapprocher leur style.

JADT’ 18

713

Le nom padre apparaît 9 815 fois dans le corpus PIC. Dans les œuvres de
Ferrante, on en dénombre 833 (8,5 % du total) et 1 170 chez Starnone (11,9 %).
Ce mot est clairement employé plus fréquemment par ces “deux” auteurs. De
manière similaire, le mot madre possède une fréquence de 8 246 dans le
corpus pour 1 104 occurrences (13,4 %) sous la plume de Ferrante et 762
(9,2 %) avec Starnone. D’autres vocables fonctionnels possèdent des
distributions similaires. Ainsi le mot persino (même) apparaît 1 351 fois dans
la collection PIC et on en compte 266 (19,7 %) chez Ferrante et 205 (15,2 %)
chez Starnone. On notera également que ce terme peut également s’écrire
perfino (avec une fréquence d’occurrences de 20 avec Ferrante, 18 chez
Starnone). Pour Ferrante et Starnone, on voit une préférence pour une forme,
tandis que d’autres auteurs recourent uniquement à l’une des orthographes
(Baricco : uniquement perfino, Tamaro : seulement persino). Enfin certains
écrivains ignorent les deux mots (Covacich, Parrella) ou l’utilisent très
rarement (De Luca ou Balzano). Comme exemples complémentaires, certains
mots ne sont employés que par Ferrante et Starnone comme risatella
(gloussement, 16 occurrences chez Ferrante, 4 avec Starnone) ou
contraddittoriamente (contradictoirement, Ferrante : 6; Starnone : 9).
Pour un écrivain italien, le lexique peut inclure des formes provenant du
dialecte comme celui de Naples avec le terme strunz (stronzo en italien). Ce
terme apparaît 85 dans le corpus, avec 63 occurrences dans les romans de
Starnone et 18 chez Ferrante (et deux fois chez De Silva et Raimo).
Certains n-grammes de mots s’avèrent plus fréquents chez Ferrante et
Starnone comme no essere che (ne pas être ça) qui apparaît 23 fois (100 %)
dans le corpus mais 6 (26,1 %) sous la plume de Ferrante et 7 (30,4 %) sous
celle de Starnone. Ensemble ces deux auteurs apportent plus de 56 % des
occurrences de cette séquence.
6. Conclusion
Cette étude s’appuie sur deux méthodes d’attribution d’auteur reconnues
d’une part, et, d’autre part sur un corpus de 150 romans contemporains
rédigés par 40 auteurs. Comme attributs stylistiques, nous avons retenu les
100, 150, 200, 250, 300, 400, 500, 1 000, 1 500 et 2 000 mots les plus fréquents
pour la méthode Delta (Burrows, 2002). Avec ces différentes valeurs, le
premier nom retourné comme le probable auteur s’avère toujours Domenico
Starnone et ceci pour les sept romans parus sous le nom Ferrante. En
s’appuyant sur l’ensemble du vocabulaire et la méthode de Labbé (2007), la
même conclusion est obtenue.
En analysant quelques choix lexicaux, on découvre des relations étroites
entre Starnone et Ferrante. Par exemple, le mot persino est sur-employé dans
les romans des deux auteurs, et la second forme perfino n’apparaît que plus

714

JADT’ 18

rarement. Chez d’autres écrivains, on rencontre habituellement une
préférence pour l’un des deux termes ou l’absence de leur usage. Enfin, suite
à l’atelier qui s’est tenu à Padoue le 7 septembre 2017 aboutissant à désigner
Domenico Starnone comme l’écrivain derrière la signature Ferrante, celui-ci a
démenti en être le véritable auteur (Fontana, 2017).
Remerciements
Cette recherche a été possible grâce à A. Tuzzi et M. Cortelazzo qui nous ont
transmis le corpus PIC.
Références
Baayen, H.R. (2008). Analysis Linguistic Data: A Practical Introduction to
Statistics using R. Cambridge University Press, Cambridge.
Burrows, J.F. (2002). Delta: A measure of stylistic difference and a guide to
likely authorship. Literary and Linguistic Computing, 17(3):267-287.
Chaski, C. (2013). Best practices and admissibility of forensic author
identification. Journal of Law and Policy, 21(2):333-376.
Craig, H., & Kinney, A.F. (2009). Shakespeare, Computers, and the Mystery of
Authorship. Cambridge University Press, Cambridge.
Fontana, E. (2017). Lo scrittore Domenico Starnone: “Io non sono Elena
Ferrante”. Il Giornale, 9 sept.
Gatti, C. 2016. La véritable identité d’Elena Ferrante révélée. BublioObs, 2
octobre 2016.
Hoover, D.L. (2007). Corpus stylistics, and the styles of Henry James. Style,
41(2):160-189.
Hughes, J.M., Foti, N.J., Krakauer, D.C., & Rockmore, D.N. (2012).
Quantitative patterns of stylistic influence in the evolution of literature.
Proceedings of the PNAS, 109(20), pp. 7682-7686.
Juola, P. (2006). Authorship attribution. Foundations and Trends in Information,
1(3):233-334.
Juola P. (2016). The Rowling case: A proposed standard analytic protocol for
authorship questions. Digital Scholarship in the Humanities, 30(1), i100-i113.
Kocher, M., & Savoy, J. (2017). Distributed language representation for
authorship attribution. Digital Scholarship in the Humanities, 2017, to
appear.
Labbé, D. (2007). Experiments on authorship attribution by intertextual
distance in English. Journal of Quantitative Linguistics, 14(1):33-80.
Mosteller, F., & Wallace, D.L. (1964). Applied Bayesian and Classical Inference:
The Case of the Federalist Papers. Addison-Wesley, Reading.
Rexha, A., Klampfl, S., Kröll, M., & Kern, R. (2016). Towards a more fine
grained analysis of scientific authorship. Proceedings ECIR 2016, pp. 26–31.

JADT’ 18

715

Savoy, J. (2015). Comparative evaluation of term selection functions for
authorship attribution. Digital Scholarship in the Humanities, 30(2):246-261.
Stamatatos, E. (2009). A survey of modern authorship attribution methods.
Journal of the American Society for Information Science and Technology,
60(3):433-214.
Tuzzi, A., & Cortelazzo, M. (2018). What is Elena Ferrante? A Comparative
Analysis of a Secretive Bestselling Italian Writer. Digital Scholarship in the
Humanities, to appear.

716

JADT’ 18

Regroupement d’auteurs dans
la littérature du XIXe siècle
Jacques Savoy
Université de Neuchâtel (Suisse) – Jacques.Savoy@unine.ch

Abstract
This paper presents the author clustering problem in which a set of n texts
written by several distinct authors must be regrouped into k clusters, each of
them corresponding to a single author. The proposed model can use different
distance measures and feature sets (composed of the most frequent word
types). The evaluation is based on a French corpus composed of 200 excerpts
of novels written during the 19th century. By varying different parameter
settings, the evaluation indicates a better performance achieved with words
instead of n-grams of letters. The Cosine distance achieves lower
performance levels compared to the Tanimoto (L1) or Matusita (L2) functions.
The text size plays an important role in the effectiveness of the solution,
showing that size of 10,000 tokens produces significantly better results than
text size of 5,000 to 500 tokens. A more detailed analysis provides reasons
explaining stylistic aspects of some authors.
Résumé
Cette communication présente le problème du regroupement d’auteurs dans
lequel un ensemble de n textes écrits doit être regroupé dans k grappes
distinctes, une pour chaque auteur. Le modèle proposé permet l’emploi de
différentes mesures de distance et divers ensembles d’attributs (vocables les
plus fréquents). L’évaluation s’appuie sur un corpus composé de 200 extraits
de romans français du XIXe siècle. En variant différents paramètres, notre
étude indique que les vocables s’avèrent meilleur que les n-grammes de
lettres. La fonction cosinus génère un taux de réussite plus faible que le
fonction Tanimoto (L1) ou Matusita (L2). La taille des textes joue un rôle
important dans la qualité de réponse et une longueur de 10 000 mots permet
une performance significativement supérieure à des valeurs variant de 5 000
à 500 mots. Une analyse apporte quelques explications sur le style de
différents auteurs.
Keywords : Automatic classification, unsupervised machine learning,
authorship attribution.
Mots-clés : Classification automatique, apprentissage non-supervisé,
attribution d’auteur.

JADT’ 18

717

1. Introduction
Le problème d’attribution d’auteur (Juola, 2006) rencontre un intérêt
grandissant avec la multiplication des canaux électroniques. La présence de
messages anonymes ou pseudo-anonymes soulève de nombreux défis en
criminalité (Olsson, 2008), (Chaski, 2013) à l’exemple des chats calomnieux
ou des courriels menaçants. Pourtant des questions plus classiques méritent
notre attention comme, par exemple, déterminer la véritable identité de la
romancière Elena Ferrante (Gatti, 2016) ou sur les relations de Shakespeare et
de ses co-auteurs (Michell, 1996), (Craig & Kinney, 2009).
Dans ce cadre, notre communication présente les problèmes liés à la question
du regroupement d’auteurs avec une application en littérature française du
XIXe siècle. Ce problème se résume ainsi. Disposant d’un ensemble de n
extraits de romans, on doit regrouper en k classes disjointes, chacune
contenant tous les écrits du même auteur. Ce problème a été posé lors de la
campagne d’évaluation CLEF-PAN 2016 et 2017 (Stamatatos et al., 2016) mais
les collections tests n’ont pas été rendues publiques. Ce problème présente
une difficulté majeure par l’absence de données d’entrainement.
2. Travaux reliés
Afin d’identifier l'auteur d’un écrit, trois familles d’approches ont été
proposées (Juola, 2006). En premier lieu, des mesures stylométriques
supposées invariantes ont été évoquées comme la longueur moyenne des
phrases, la taille du vocabulaire par rapport à la longueur du document
(rapport TTR) (Rexha et al., 2016). Toutes ces mesures possèdent
l’inconvénient d’être instables face à des textes de tailles différentes (Baayen,
2008). Une deuxième famille d’approches se fonde sur le vocabulaire.
Mosteller & Wallace (1964) proposent de sélectionner de manière semiautomatique les vocables les plus pertinents. Burrows (2002) sélectionne les
mots les plus fréquents et, en particulier, les mots fonctionnels (déterminants,
prépositions, conjonctions, pronoms et verbes auxiliaires). Ces derniers
possèdent l’avantage d’être plus fortement reliés au style de l’auteur qu’à la
sémantique. Cette liste comprendra entre 50 à 1 000 vocables les plus
fréquents (Hoover, 2007). D’autres auteurs proposent de définir a priori une
telle liste (Hughes et al., 2012). Ainsi, chaque texte peut être représenté par les
fréquences d’occurrence de ces vocables. Ensuite, une mesure de distance (ou
de similarité) permet d’estimer la proximité de deux textes. L’attribution
s’établit habituellement selon la règle du plus proche voisin.
Troisièmement, des modèles d’apprentissage automatique (Stamatatos, 2009)
permettent de sélectionner les attributs (mots, bigrammes de mots ou de
lettres, POS, émoticons, etc.) possédant le meilleur pouvoir discriminant.
Ensuite un classifier est entraîné sur un ensemble d’apprentissage (SVM,

718

JADT’ 18

régression logistique, etc.). Cependant, dans le cadre du regroupement
d’auteurs, aucune donnée d’entraînement n’est disponible rendant caduc de
telles approches. Dès lors, pour résoudre ce problème, des approches
proposent de déterminer en premier lieu le nombre k d’auteurs sur
l’ensemble n d’écrits (Stamatatos et al., 2016). Cette valeur fixée, on applique
un algorithme de classification k-means afin d’identifier les différents groupes
de textes. Par itération, le nombre k d’auteurs peut être affiné. Comme second
paradigme, la distance entre chaque écrit est calculée, puis on applique un
algorithme de classification hiérarchique (Lebart et al., 1998) pour former les
grappes de documents. Dans cette étude, nous suivrons cette seconde
stratégie de résolution, choix qui nous a permis d’obtenir le deuxième rang
lors de la dernière campagne d’évaluation PAN-CLEF 2016.
3. Corpus de test et méthodologie d’évaluation
L’évaluation empirique tient une place importante en attribution d’auteur.
Comme les corpus des campagnes PAN-CLEF 2016 et 2017 n’ont pas été
rendus publics, nos évaluations seront basées sur une collection extraite de la
littérature française du XIXe siècle. Ce corpus nommé St-Jean1 contient 200
extraits de romans écrit par 30 auteurs (entre 1801 (Châteaubriant, Attala) et
1901 (Régnier, Les Rencontres de Monsieur de Bréot)). Ce nombre d’écrivains et
de textes étant élevé, la tâche demeure ardue. Chaque auteur est représenté
par au moins trois extraits (avec un maximum de treize pour Balzac)
provenant d’un à six romans et aucun écrivain ne représente plus de 5 % du
corpus. Chaque extrait contient en moyenne 10 073 formes (min : 10 026 ;
max : 10 230 ; standard déviation : 25). Au total, ce corpus contient 2 014 641
formes pour 51 661 vocables extraits de 67 romans. Disposant de n textes,
notre approche produira une liste ordonnée de liens entre textes avec une
indication de la distance entre eux. Un exemple est présenté dans le
tableau 1. Avec ce corpus, la solution se compose de 30 groupes requérant la
présence de 670 liens intra-auteurs. Comme mesure d’évaluation, nous
reprenons la précision moyenne (AP) (la moyenne des précisions obtenues
pour chaque lien pertinent), mesure usitée lors des campagnes PAN-CLEF
2016 et 2017. Ainsi, une valeur unique de performance reflète la qualité de
chaque modèle de classification. Comme seconde mesure, la valeur HP
(haute précision) indique le nombre de liens correctement établis depuis le
début jusqu’à la présence du premier lien erroné. Dans notre tableau 1, la
valeur HP = 168 signalant que les 168 premiers liens sont justes.

1 Ce corpus a été créé par D. Labbé et est disponible
(www.unine.ch/clc/home/corpus.html) soit sous la forme de textes, soit lemmatisé.
Les encodages UTF-8 et Windows sont disponibles.

JADT’ 18

719

Tableau 1 : Exemple d’un extrait d’une liste ordonnée selon la distance (Tanimoto)
Rang
Distance
Texte 1
Texte 2
1
0,239
51 Flaubert
62 Flaubert
2
0,242
3 Flaubert
20 Flaubert
3
0,248
29 Sand
115 Sand
4
0,248
122 Staël
140 Staël
5
0,253
125 Fromentin
159 Fromentin
6
0,255
37 Flaubert
62 Flaubert
7
0,256
132 Régnier
162 Régnier
...
…
…
…
169
0,324
42 Maupassant
51 Flaubert

4. Sélection des attributs et mesure de distance
Afin de regrouper les documents selon leur auteur, nous devons les
représenter en fonction de leur style et non en fonction des thèmes qu’ils
abordent. Comme mentionné précédemment, plusieurs études ont démontré
que les vocables les plus fréquents constituent des attributs pertinents pour
détecter le style d’un auteur. Dans le cadre de l’attribution d’auteur, le thème
pourrait perturber des affectations correctes lorsque, par exemple, deux
auteurs abordent des sujets similaires. Pour cerner les aspects stylistiques,
une étude récente a démontré que tenir compte des 200 à 300 mots les plus
fréquents (Savoy, 2015) apporte de bonnes performances comparées à
d’autres fonctions de sélection (rapport des cotes, gain d’information, chicarré, etc.). Sur la base du corpus St-Jean, les mots les plus fréquents de notre
corpus sont : de (4,11 % des occurrences), et (2,44 %), la (2,36 %), le (1,94 %),
et à (1,9 %). Comme alternative, plusieurs études proposent de recourir aux
fréquences des lettres et des bigrammes de lettres et, plus généralement, des
n-grammes afin de distinguer les différents styles (Kjell, 1994), (Juola, 2006).
On remarquera toutefois que les composantes stylistiques et thématiques
seront toutes les deux présentes dans la génération de tels n-grammes. Dans
cette étude, la distinction entre majuscules et minuscules est ignorée et les
signes de ponctuation sont éliminés. Par contre, on tiendra compte du fait
qu’une lettre débute ou termine un mot. Le nombre maximal d’attributs
s’élève à (27 x 27) + 27 = 756. Pour la langue française, on retrouve 594 (ou
78,6 %) combinaisons possibles dans notre corpus. Les lettres françaises les
plus fréquentes sont : e (15.6 % des lettres), s (8,3 %), a (8,3 %), i (7,5 %), et t
(7,2 %). En indiquant par _ l’espace, les bigrammes de lettres les plus usuels
sont : e_ (5,1 % des bigrammes), s_ (3,5 %), t_ (2,7 %), _d (2,4 %), et _l (1,8 %).
Dès que chaque document est représenté par m de mots (ou de n-grammes de
lettres), on peut calculer sa distance avec les autres entités du corpus. Le
choix de cette fonction de distance (ou de similarité) peut s’opérer selon des
critères théoriques (par exemple, symétrie, inégalité triangulaire) ou

720

JADT’ 18

empiriques (efficacité). Basée sur le profilage d’auteur, une étude récente
(Kocher & Savoy, 2017) indique qu’aucune mesure de distance s’avère
toujours la meilleure. Par contre un groupe restreint permet d’obtenir de
bonnes performances comme la distance de Manhattan ou de Tanimoto basée
sur la norme L1, ou celle de Matusita (norme L2). Nous avons repris ces
mesures en y ajoutant la distance du cosinus. Ces quatre mesures respectent
la symétrique et respectent l’inégalité triangulaire (Kocher & Savoy, 2017).
Dans la définition de ces mesures de distance, les lettres majuscules
indiquent les vecteurs représentants les documents. Les minuscules (ai, bi)
correspondent aux fréquences relatives des termes sélectionnés.

5. Évaluation
Notre première évaluation concerne l’efficience des différentes mesures de
distance ainsi que la performance du nombre de vocables les plus fréquents
retenus comme attributs. Le tableau 2 indique les valeurs de précision
moyenne (AP) et de haute précision (HP) en représentant les textes par les
100 à 1 000 vocables les plus fréquents, ou tout le vocabulaire. La dernière
ligne et colonne nous renseigne sur la moyenne des APs.
Tableau 2 : Précision moyenne (AP) et haute précision (HP) selon diverses mesures de
distance avec des représentations construites entre 100 vocables et tout le vocabulaire
Manhattan
Tanimoto
Matusita
Cosinus
Moyenne
Attributs
AP
HP
AP
HP
AP
HP
AP
HP
AP
100
0,674
185
0,695
192
0,655
181 0,626 152
0,663
200
0,692
186
0,708
193
0,687
222 0,628 145
0,679
190
196
244 0,629 148
300
0,705
0,720
0,727
0,695
0,750
500
0,720
186
0,735
189
212 0,627 149
0,708
0,730
0,743
0,709
1 000
183
186
0,745
204 0,617 142
Tout
0,713
166
0,672
168
0,568
135 0,599 142
0,691
Moyenne
0,706
183
0,712
187
0,689
200 0,621 146
0,681

JADT’ 18

721

Ces résultats indiquent que les différences de précision moyenne restent
faibles entre les mesures de Manhattan, Tanimoto et Matusita. Toutes les
trois s’avèrent supérieures au cosinus. En considérant la haute précision
(HP), Matusita tend à apporter une meilleure efficacité. Reste à déterminer a
priori cette valeur maximale, sans connaître les attributions correctes. Enfin,
une représentation par 300 à 500 voire 1 000 vocables les plus fréquents
fournit les meilleurs taux de succès. En remplaçant les vocables par des ngrammes de lettres (performances indiquées dans le tableau 3), les valeurs de
performance s’avèrent inférieures aux vocables. La variation des taux de
succès entre une combinaison uni- et bigrammes de lettres (deuxième ligne
du tableau 3) ou des séquences plus longues s’avère peu élevée. Par contre
les temps de traitement s’accroissent rapidement (8,2 minutes pour les uni- et
bigrammes à plus de 4 heures pour les 5-grammes comparé à 3 minutes avec
les 500 mots les plus fréquents). Enfin, la fonction cosinus retourne les
performances les moins bonnes. Nos premières évaluations se fondaient sur
l’ensemble du texte disponible, soit environ 10 000 mots. Si l’on réduit cette
taille à 5 000 voire à 500, les taux de réussite obtenus sont indiqués dans le
tableau 4. La première ligne est reprise du tableau 2 puis les tailles
décroissent comme le signale la première colonne. La réduction moyenne des
performances est reprise dans la dernière colonne. Ainsi, en réduisant les
textes à 5 000 mots, la baisse moyenne s’élève à 25,8 %. Si l’on doit œuvrer
avec des longueurs de 1 000 à 500 mots, les taux de réussite s’avèrent faibles
générant une réduction de 80 à 90 %. Est-il vraiment raisonnable d’effectuer
des attributions d’auteur avec de telles tailles ?
Tableau 3 : AP et HP selon diverses mesures de distance avec des n-grammes de lettres
Matusita
Cosinus
Moyenne
Manhattan
Tanimoto
HP
HP
AP
HP
AP
HP
AP
n-grams
AP
AP
uni & bi
0,559
139 0,559
139
0,503
128 0,538
94
0,540
3-gram
0,527
108 0,527
108
0,471
130 0,476
108
0,500
112
0,532
4-gram
0,570
153 0,570
153
0,507
147 0,481
5-gram
0,587
177 0,587
177
0,541
181 0,543
73
0,565
0,588
200 0,588
200
0,557
188 0,415
6-gram
36
0,588
Moyenne
0,566
155 0,566
155
0,506
147 0,510
97
0,545
Tableau 4 : AP et HP selon diverses mesures de distance avec des textes de tailles différentes
(représentation sur la base de 300 vocables)
Matusita
Cosinus
Moyenne Différence
Manhattan
Tanimoto
HP
HP
AP
HP
AP
HP
Taille
AP
AP
10 000 0,705 190 0,720
196 0,727 244 0,629 148 0,695
5 000
0,526 55
0,545
58
0,526 85
0,466 74
0,516
-25,8%
2 500
0,326 31
0,342
39
0,306 35
0,284 11
0,315
-54,8%
1 000
0,152 4
0,152
2
0,116 1
0,141 3
0,140
-79,8%
500
0,093 2
0,089
2
0,079 3
0,086 2
0,087
-87,5%

722

JADT’ 18

En analysant la liste triée obtenue avec la fonction Matusita et en
représentant les textes par les 300 vocables les plus fréquents, les distances
les plus faibles se retrouvent entre des extraits de la même œuvre. La
distance la plus faible se trouve avec le roman Les Rencontres de Mr de Bréot
(1901) de Régnier, puis on trouve Bouvard et Pécuchet (1881) de G. Flaubert,
Delphine de Mme de Staël (1803), Mme Bovary (1857) de G. Flaubert et La Petite
Fadette (1832) de G. Sand. Si l’on analyse les appariements les plus difficiles
entre deux œuvres du même auteur, les romans Graziella (1852) et Geneviève
(1863) de A. de Lamartine constitue le lien le plus distant. Ensuite, on
rencontre La double Maîtresse (1900) de H. de Régnier, Aurélia (1855) et Les
Illuminés (1852) de G. de Nerval et Le père Goriot (1833) et La Maison Nucingen
(1838) de H. de Balzac. Ces auteurs peuvent adopter des styles assez
dissemblables, rendant une attribution plus ardue. Parmi les œuvres dont le
style est perçu comme proche par la machine mais qui sont écrites par deux
auteurs distincts, on trouve en tête Bel-Ami (Maupassant, 1885) et Mme Bovary
(Flaubert, 1857), puis Volupté (Sainte-Beuve, 1834) et Dominique (Fromentin,
1862), Notre Cœur (Maupassant, 1890) et Mme Bovary (Flaubert, 1857), et enfin
L’Assommoir (Zola, 1879) et Mme Bovary (Flaubert, 1857).
6. Conclusion
Parmi les fonctions de distance, notre étude indique que le cosinus n’apporte
pas de bons résultats. Par contre, les différences de performance entre les
fonctions Manhattan, Tanimoto ou Matusita demeurent faibles. Afin de
cerner une partie importante du style des auteurs, le recours à une
représentation sur la base de vocables s’avère plus efficiente que le recours
aux n-grammes de lettres (pour n variant de 1 à 6). Représenter le style avec
les 300 à 500 vocables les plus fréquents s’avère pertinent.
Lorsque l’on compare la précision moyenne (AP) et la haute précision (HP),
le choix des paramètres optimaux diffère quelque peu d’une mesure de
performance à l’autre. Notons que l’AP ne punit pas sévèrement les erreurs
d’affectation, erreurs qui entraînent immédiatement une baisse de la valeur
HP. Enfin, la taille des textes joue un rôle essentiel dans une attribution
d’auteur et des valeurs inférieures à 1 000 mots ne permettent que des
affectations souvent douteuses. Parmi les auteurs retenus, le style du roman
Mme Bovary se rapproche de celui de Maupassant (Bel-Ami) ou de Zola
(L’Assommoir).
Remerciements
L’auteur remercie D. Labbé pour avoir mis à sa disposition le corpus St-Jean.

JADT’ 18

723

Références
Baayen, H.R. (2008). Analysis Linguistic Data: A Practical Introduction to Statistics
using R. Cambridge University Press, Cambridge.
Burrows, J.F. (2002). Delta: A measure of stylistic difference and a guide to likely
authorship. Literary and Linguistic Computing, 17(3):267-287.
Chaski, C. (2013). Best practices and admissibility of forensic author
identification. Journal of Law and Policy, 21(2):333-376.
Craig, H., & Kinney, A.F. (2009). Shakespeare, Computers, and the Mystery of
Authorship. Cambridge University Press, Cambridge.
Gatti, C. 2016. La véritable identité d’Elena Ferrante révélée. BublioObs, 2 octobre
2016.
Hoover, D.L. (2007). Corpus stylistics, and the styles of Henry James. Style,
41(2):160-189.
Hughes, J.M., Foti, N.J., Krakauer, D.C., & Rockmore, D.N. (2012). Quantitative
patterns of stylistic influence in the evolution of literature. Proceedings of the
PNAS, 109(20), pp. 7682-7686.
Juola, P. (2006). Authorship attribution. Foundations and Trends in Information,
1(3):233-334.
Kjell, B. (1994). Authorship determination using letter pair frequency features
with neural network classifier. Literary and Linguistics Computing, 9(2):119-124.
Kocher, M., & Savoy, J. (2017). Distance measures in author profiling. Information
Processing & Management, 53(5):1103-1119.
Labbé, D. (2007). Experiments on authorship attribution by intertextual distance
in English. Journal of Quantitative Linguistics, 14(1):33-80.
Lebart, L., Salem, A. and Berry, L. (1998). Exploring Textual Data. Dordrecht,
Kluwer.
Michell, J. (1996). Who Wrote Shakespeare? Thames and Hudson: New York (NY).
Mosteller, F., & Wallace, D.L. (1964). Applied Bayesian and Classical Inference: The
Case of the Federalist Papers. Addison-Wesley, Reading.
Muller, C. (1992). Principes et méthodes de statistique lexicale. Honoré Champion,
Paris.
Olsson, J. (2008). Forensic Linguistics. Continuum, London.
Rexha, A., Klampfl, S., Kröll, M., & Kern, R. (2016). Towards a more fine grained
analysis of scientific authorship. Proceedings ECIR 2016, pp. 26–31.
Savoy, J. (2015). Comparative evaluation of term selection functions for
authorship attribution. Digital Scholarship in the Humanities, 30(2):246-261.
Stamatatos, E. (2009). A survey of modern authorship attribution methods.
Journal of the American Society for Information Science and Technology, 60(3):433214.
Stamatatos, E., Tschuggnall, M., Verhoeven, B., Daelemans, W., Specht, G., Stein,
B., & Potthast, M. (2016). Clustering by authorship within and across
documents. Working Papers, CLEF-2016.

724

JADT’ 18

What’s Old and New? Discovering Topics in the
American Journal of Sociology1
Stefano Sbalchiero, Arjuna Tuzzi
University of Padova – stefano.sbalchiero@unipd.it; arjuna.tuzzi@unipd.it

Abstract
Nowadays the field of text mining techniques seems to be very active in
dealing with the increasing mass of available digital texts and several
algorithms have been proposed to analyze and synthesize the vast amount of
data that today represents a challenging source of information overload.
Topic modeling is a collection of algorithms which are useful for discovering
themes, i.e. topics, in unstructured text. The Latent Dirichlet Allocation
(LDA) by Blei (et al., 2003) was one of the first topic modeling algorithms and
since then the field seems to be active and many variants and other
algorithms have been suggested. The present study considers a topic as an
indicator of the relevance of a research area in a specific time-span and its
temporal evolution pattern as a way to identify the paradigm changes in
terms of theories, ideas, forgotten topics, evergreen subjects and new
emerging research interests. The study aims to contribute to a substantive
reflection in Sociology by exploring the temporal evolution of topics in the
abstracts of articles published by the American Journal of Sociology in the
last century (1921-2016). Within the classical LDA perspective, the study also
focus on topics with a significant increasing or decreasing trend (Griffiths et
Steyvers, 2004). The results show different shifts that involved relevant
reflections on various issues, from the early debate on the
“institutionalization” process of Sociology as a scientific discipline to recent
developments of sociological topics that clearly indicate how sociologists
have reacted to new social problem.
Keywords: Chronological corpus, History of Sociology, Academic Journals,
Text Mining, Latent Dirichelet Allocation

1 This study was supported by the University of Padova, fund CPDA145940
(2014) “Tracing the History of Words. A Portrait of a Discipline Through Analyses of
Keyword Counts in Large Corpora of Scientific Literature" (P.I. Arjuna Tuzzi,
University of Padova).

JADT’ 18

725

1. Introduction: topic modeling
As evidenced by the literature on topic modelling (Blei et al., 2003;
Ponweiser, 2012; Grimmer et Stewart, 2013; Griffiths et Steyvers, 2004), text
mining approaches can mitigate the problem of analysing huge collections of
textual data when they increase in number and size and complicate all
information processing. From a methodological point of view, since the
topics emerge directly from data, text mining approaches can tone down
some problems about the role of analysts in coding and interpreting the
content hidden in corpora, e.g. research bias or errors that notoriously affect
most approaches in comparative and quanti-qualitative researche (Strauss et
Corbin, 1990; Corbetta, 2003). A popular approache to extract information by
summarizing the main contents embedded in relevant collection of texts in
digital form is known as topic modeling (Blei et Lafferty, 2009), which is
essentially a collection of algorithms that are exploited to discover themes,
i.e. topics, in unstructured and complex texts. The Latent Dirichlet Allocation
(LDA) is one of the first topic modeling algorithms, namely a “generative
probabilistic model of a corpus. The basic idea is that documents are
represented as random mixtures over latent topics, where each topic is
characterized by a distribution over words” (Blei et al., 2003, p. 996). LDA is a
technique that facilitates the automatic discovery of themes in a collection of
documents. Since a text document can deal with different topics and the
words that occur in that document reflect a set of possible topics, in
“statistical natural language processing, one common way of modeling the
contributions of different topics to a document is to treat each topic as a
probability distribution over words, viewing a document as a probabilistic
mixture of these topics” (Griffiths et Steyvers, 2004, p. 5228). Actually we
cannot directly observe topics but only documents and words, as topics are
part of the latent and hidden text structure. The model infers the latent topic
structure given by observed words and documents: this is the LDA's
generative processes which recreate (generate) the documents of the corpus
by assigning the probability of topics (the relevance) to documents and the
probability of words to topics. The result is a probabilistic distribution of
topics over documents that is characterized and described by a cluster of cooccurring words (Blei et al., 2003). This list o words enable the researcher to
interpret the meaning of all the generated topics. For the purposes of the
present study, the temporal variable is crucial to analyse the direction and
evolution of topics, and particularly to the extent that they have a direct
relationship with the most significant shifts in the development of Sociology
as a discipline over time. For these reasons, we propose a LDA-based topic
detection procedure as this “method discovers a set of topics expressed by
documents, providing quantitative measures that can be used to identify the

726

JADT’ 18

content of those documents, track changes in content over time” (Griffiths et
Steyvers, 2004, p. 5228). An additional estimation procedure exploits a
metavariable (year) to explore the topics trends: LDA offers the opportunity
to estimate the slope of a linear model that represents the distribution of
topics by year. The model permits to identify “hot and cold topics” (Griffiths
et Steyvers, 2004), i.e. topics with significant increasing (hot) and decreasing
(cold) trends through time.
2. Corpus and data
The American Journal of Sociology (AJS), established in 1895 as the first U.S.
scholarly journal in its field, can be considered one of the world’s preeminent
journals and a leading voice for research in social sciences. The journal
fosters pathbreaking work from all areas of sociology, with an emphasis on
theory building and innovative methods. AJS is a multi-disciplinary journal
that strives to speak to a "general sociological reader" and is open to
sociologically informed contributions from anthropologists, statisticians,
economists, educators, historians, and political scientists. Manuscripts are
subjected to a double-blind review process and published articles are
considered representative of the best current theoretical and methodological
debates. Our corpus includes all the abstracts of the papers published by AJS
that have been retrieved from popular archives (Scopus and Web of science)
and the journal webpages. We decided to work on the abstracts since they
provide concise information about the main contents of all articles. With
regard to selection criteria, they were based on the following consideration:
when abstracts did not provide any information about the content or did not
refer to relevant scientific contributes (e.g. editorials, master heads, errata,
acknowledgements, rejoinders, notes, announcements, corrections, list of
consultants, obituary, etc.) we decided to disregard them in further analyses.
The corpus is composed of 3,992 abstracts, collected for a period of almost a
century (mean: 41 per year), from the Volume No. 27, Issue No. 1 (1921) to
the latest, No. 121, Issue No. 4 (2016). The collected texts had relevant
contents for the purpose of the present analysis based on the following
consideration and hypothesis: If we consider a topic as an indicator of the
relevance of a research area in a specific time-span, then the temporal
evolution pattern of subject matters can portray main paradigm changes in
terms of theories, ideas, forgotten topics, evergreen subjects and new
emerging research interests in Sociology. The corpus has been pre-processed
by means of TaLTaC2 software package. After the tokenization (the
identification of words given character sequences chopping it up into pieces),
the corpus has been normalized replacing uppercase with lowercase letters.
An automatic search procedure identified relevant multi-words (MWs), i.e.

JADT’ 18

727

informative sequences of words (Pavone, 2010) repeated at least five times in
the corpus (849 MWs in total). This procedure retrieved most interesting
MWs in the abstract (e.g. united states, fr. 395; social structure, fr. 115; social
science, fr. 101; labor market, fr. 89; social change, fr. 78) and contributed to
increase the amount of information conveyed by sequences of words2. Then,
the corpus has been processed by means of R software packages3:
punctuation marks and numbers have been removed, as well as some
grammatical words (articles, conjunction, prepositions, pronouns). The
corpus is composed of 24,418 word-types and 512,410 word-tokens (tab. 1),
and the measures show that there is a sufficient level of redundancy to
proceed with statistical analyses of textual data (Lebart et al., 1998; Trevisani
et Tuzzi, 2015; Bolasco, 2013).
Table 1. Basic lexical measures of the corpus of AJS abstracts

(V) WORD-TYPES
(N) WORD-TOKENS
(V/N)*100 = TYPE/TOKEN RATIO
(V1/V)*100 = PERCENTAGE OF HAPAX

24,418
512,410
4.76
47.08

3. Topic detection
As the LDA algorithm “fits” the terms in the document into a number of
topics that must be specified apriori, this represents an important and
sensitive decision that affects results and findings: few topics will produce
broad subjects and mixed-up contents, while too many topics will produce
minimal subjects and results too detailed to be readable and interpretable. To
set the number of topics in a data driven manner we have the opportunity to
calculate different metrics (Arun et al., 2010) and estimate the optimal
number of topics (Griffiths et Steyvers, 2004) by means of the maximum loglikelihood of LDA for a number of topics ranging from 2 to 50 (Fig. 1).

2 If MWs did not appear at least 5 times in the corpus, that is about once every 20
years, it was not considered important; however, the MWs that appeared with a
frequency equal to or greater than 10 are 417.
3 The analysis were implemented by R pakages: Tm, Lda, Topic model.

728

JADT’ 18

Fig. 1: Fitting the model: log-likelihood calculated for increasing number of topics

The best number of topics is the one with the highest value of log-likelihood
that is around 30 and can be established as the optimal number of topics.
Figure 2 shows the general trend of all the 30 topics as depicted by the fitted
model A clue of how these topics change over time is shown by 30 panels
with a topic trend line each, that lists the number of topics with positive or
negative trends. All of the topics are ordered by slope: decreasing topics
appear in the first panels (top left), and increasing ones in the last panels
(bottom right)., Since the main aim of this study is to detect the temporal
evolution of old, new and emerging topics in Sociology, we can resort to a
limited
number
of
topics
that
show
prototipycal
temporal
patterns(Ponweiser, 2012; Griffiths et Steyvers, 2004).

Fig. 2: Temporal patterns of the 30 topics in Sociology sorted by slope of linear models

JADT’ 18

729

Consistent with the idea that topics show different trends and embrace
theoretical, conceptual, and methodological shifts, the analysis of timedependent phenomena identifies three specific temporal patterns of topics:
topics whose trajectory has grown in time and it is increasing over time (28,
4, 2, 27, 15, 11); topics whose trajectory decreased (7, 3, 21, 9, 13, 18); and
topics whose peak-like journey (meteor) was high only in a specific interval
of time (14, 17, 28, 15) or shows more irregular temporal trajectories.
4. What’s old and new in Sociology?
To focus on major increasing or decreasing topics from 1921 to 2016, we
explored the contents of five coldest and hottest topics. Figure 3 provides the
top term for these topics.
The groups of coldest topics correspond on one hand to the methodological
development of sociological perspectives, and on the other hand to some
specific objects of research. These topics were very popular in about 20s and
50s. First of all, the debate on the “institutionalization” process of Sociology
as a scientific discipline characterized the early debate (topic 7). The main
need was to create a strong scientific and knowledge base from the
development of ideas advanced by the "founding fathers", e.g. Durkheim. At
the same time, the debate on the “measurement” of social phenomena arose.
The issue of migration between cities and farms (topic 3) by economic and
social groups gives the net law of rural-urban social selection. The emerging
of a scientific social reflections about health and illness (topic 21) by using
empirical data to evaluate how social life affects morbidity and mortality
rate, and vice versa, increased in efforts for better educated public and to
improve health legislation. The development of psychological sociology
(topic 9) and the general progress of psychological interpretations of social
processes and institutions have decreased over time; researches in this
tradition have been criticized because they mainly exemplified the biological
background of social interpretations, also supplied by the impulse from the
Darwinian doctrine. Class culture, conflict and leisure (topic 13) were
popular issues in the 30s and 50s: the industrialization had raised many
questions, from the class conflict to the growth of leisure hours of after work
hours, providing new insights for social thought. The group of hottest topics
(Fig. 4) is related to articles that have a focus of interest in a wide range of
empirical case studies that underline most significant changes that have
occurred since the mid-1960s.

730

JADT’ 18

Fig. 3: Decreasing topics: the five coldest (significant neg., p level 0.005)

Fig. 4: Increasing topic attention: the five hottest (significant pos., p level 0.005)

In those years, gender revolution (topic 11), ethnic discrimination (topic 2),
mobilization, power and élite (topic 15), protests and social movements (topic
27), and the “measurement” of social phenomena in a post-positivist fashion,
especially until the 70s (topic 4), offered to sociologists the opportunity to
deal with a social effervescence of a particular historical moment. These hot
topics indicates the ‘birth’ and recent developments of some sociological
topics that clearly indicate how Sociology (as a discipline) and sociologists
have reacted to new social problems.
In conclusion, through the topic detection analysis of the abstracts of articles,
different shifts that involved reflections on various issues have been
identified. During the twentieth century, Sociology expanded its scope and
influence, and motivated much research studies as well as a diversification of
the field. Other studies have offered a remarkable theoretical contribution to
the historical ‘shape’ of Sociology as a discipline (Kalekin-Fishman et Denis,
2012), even in a critical perspective (Turner, 1998), either emphasizing the
content of the various domains of sociology (Scott et Desfor Edles, 2011; Blau,
2004), or specifically within the intellectual ground of American Sociology
since the mid-nineteenth century (Calhoun, 2007). Even if they show an

JADT’ 18

731

interesting round of paradigmatic reflection in Sociology, there has been a
lack of research studies on the history of Sociology through empirical data
and evidence to fast-moving sociological topics over time. To the extent that
the history of Sociology is a continuous approach to the Sociology of the
present, a new way of reading the history of a discipline is rely on topic
detection of articles published in mainstream journals which mirror the
sociological scientific debate of a specific historical moment. We analysed
these trends exploiting topics as emerged from a text corpus and highlighted
two distinct directions of topics, characterized by different theoretical and
methodological implications that coexist within the same period considered:
the hot-increasing and cold-decreasing topics. Results show how Sociology
has become one of the main social science to provide fresh thinking about a
whole range of topics affecting the public sphere and, as a consequence, the
discipline developed shifting priorities in universities and social research
agenda towards specialization and fostered the birth of a wide range of subdisciplines over time. This is just the tip of the iceberg: further analyses will
shed light on many more aspects that need a deeper reflection.
References
Arun R., Suresh V., Veni Madhavan C. E. and Narasimha Murthy M. N.,
(2010). On finding the natural number of topics with latent dirichlet
allocation: Some observations. In Mohammed J. Zaki, Jeffrey Xu Yu,
Balaraman Ravindran and Vikram Pudi (eds.), Advances in knowledge
discovery and data mining, Springer Berlin Heidelberg, pp. 391-402.
Blau J. R. (2004). The Blackwell Companion to Sociology, Malden, MA: Blackwell.
Blei D. M., Ng A. and Jordan M. I., (2003). Latent Dirichlet allocation. Journal
of Machine Learning Research, 3: 993-1022.
Blei D. M and Lafferty J.D., (2009). Topic Models. In A. Srivastava, M. Sahami
(eds.), Text Mining: Classification, Clustering, and Applications. Chapman &
Hall/CRC Press.
Bolasco, S. (2013). L’analisi automatica dei testi. Fare ricerca con il text mining.
Carocci, Rome.
Calhoun G. (2007). Sociology in America: A History. Chicago: University of
Chicago Press
Corbetta P. (2003). Social Research: Theory, Methods and Techniques, SAGE
Publications Ltd., London.
Griffiths T. and Steyvers M., (2004). Finding scientific topics. Proceedings of the
National Academy of Sciences of the United States of America (PNAS),
101(Supplement 1):5228-5235.
Grimmer G. and Stewart B. M., (2013). Text as Data: The Promise and Pitfalls
of Automatic Content Analysis Methods for Political Texts, in Political

732

JADT’ 18

Analysis, 21 (3): 267-297.
Kalekin-Fishman D. and Denis A. (2012). The Shape of Sociology for the 21st
Century: Tradition and Renewal, London, SAGE.
Lebart, L., Salem, A. and Berry, L. (1998). Exploring textual data. Kluwer
Academic Publishers: Dordrecht
Pavone, P. (2010). Sintagmazione del testo: una scelta per disambiguare la
terminologia e ridurre le variabili di un’analisi del contenuto di un
corpus. In S. Bolasco, I. Chiari and L. Giuliano (Eds.) Statistical Analysis of
Textual Data: Proceedings of 10th International Conference Journées d’Analyse
statistique des Données Textuelles, 9-11 June 2010, Sapienza University of
Rome, pp. 131-140. LED.
Ponweiser M., (2012). Latent Dirichlet Allocation in R, Vienna University of
Business and Economics.
Scott A. and Desfor Edles A. (2011). Sociological Theory in the Contemporary
Era: Text and Readings, Thousand Oaks, Pine Forge Press.
Strauss, A.L. and Corbin, J. (1990). Basics for Qualitative Research: Grounded
Theory Procedures and Techniques, Newbury Park, Sage.
Trevisani, M. and Tuzzi, A. (2015). A portrait of JASA: the History of
Statistics through analysis of keyword counts in an early scientific journal.
Quality & Quantity, 49(3): 1287-1304.
Trevisani, M. and Tuzzi, A. (in press). Learning the evolution of disciplines
from scientific literature. A functional clustering approach to normalized
keyword count trajectories. Knowledge-Based System.
Turner S. (1998). Who’s Afraid of the History of Sociology? Swiss Journal of
Sociology, 24: 3-10.

JADT’ 18

733

Comparison of Neural Models for Gender Profiling
Nils Schaetti, Jacques Savoy
Université de Neuchâtel - Rue Emile-Argand 11 - CH2000 Neuchâtel - Switzerland

Abstract
This paper describes and evaluates two neural models for gender profiling
on the PAN@CLEF 2017 tweet collection. The first model is a character-based
Convolutional Neural Network (CNN) and the second an Echo State
Network-based (ESN) recurrent neural network with various features. We
applied these models to the gender profiling task of the PAN17 challenge
and have demonstrated that it can be applied to gender profiling. As
features, we propose using pre-trained word vectors, part-of-speech (POS)
and function words (FW) for the ESN model, and character 2-grams matrix
with punctuation marks, smilies, beginning and ending 2-grams for the deep
learning model. We finally compared these strategies to a baseline and found
that an ESN model based on Glove pre-trained word vectors achieves the
highest success rate and outperforms the baseline and the character-based
CNN model.
Keywords:
Author
Profiling,
Gender
Profiling,
Deep-Learning,
Convolutional Neural Network, Reservoir Computing, Echo State Network,
Natural Language Processing
1. Introduction
At the age of big data, a large number of applications are based on an
exponential amount of various data such as pictures, videos, articles, links
and blogs shared directly from computers, websites, smartphones and
sensors. Social networks and blogs are the new platforms of communication
based on fast interactions, generating a large varieties of content with their
own characteristics. These contents are difficult to compare with traditional
texts, such as novels and articles.
This issue raises new questions : Can we determine if the author of a textual
content is a man or a woman ? Can we identify the author’s place of origin,
his age group or his (or part of) psychological profile ? Answering these
questions can help solve current issues of the social network era, such as fake
news, plagiarism and identity theft. Author profiling is, therefore, a
particular and pertinent subject interest.
In addition, author profiling is central to applications involving marketing,
security and forensics. For example, forensic linguistics and police

734

JADT’ 18

investigation forces would like to know specific defining characteristics, such
as the gender, the age group and the socio-cultural background of an author
of harassing messages. When we apply this to marketing, companies and
resellers could make use of these profile characteristics while targeting their
consumers’ preferences, based on the analysis of individual consumer social
network posts and online product consulting. In order, to extract this
information, the classic statistical methods are employed as they have proven
to be effective for text classification.
Deep learning has gained increasing popularity just over the last decade,
becoming a "breakthrough" technology in image recognition and computer
vision. Yet, it faces difficulties in natural language processing (NLP) tasks.
But recurrent neural networks (RNN), as well as long short-term memory
(LSTM) obtained better results in such tasks. In this view, we therefore
decided to test such an approach on the gender profiling tasks with two
neural models, one based on Convolutional Neural Networks (CNN) and 2grams of characters, and the other on the Reservoir Computing Paradigm.
Finally, we compare them to a baseline composed of both a random and a
naive Bayes classifier
This paper is organized as follows. Section 2 introduces the data set used to
train and test both of the models and the methodology used for evaluation.
Section 3 describes and evaluates our deep-learning model. Section 4
introduces the proposed echo state network-based reservoir computing
model. Section 6 compares the results with the baseline. In the last section,
we draw conclusions on our findings and possible future improvements.
2. Methodology
To compare our two models on the gender profiling task, we needed a
common ground composed of the same dataset and evaluation measures. To
create this common ground, the PAN CLEF evaluation campaign was
launched [1] and allowed multiple research groups to propose and compare
profiling algorithms with the same methodology.
For the PAN CLEF 2017 evaluation campaign, four test collections of tweets
were generated written in several languages including English. Based on
these collections, the challenge was to classify Twitter profiles per language
variety (e.g., UK vs. US English) and gender. We were then able to use this
common ground for our two models and compare their capacities on the
gender profiling task.
The dataset was collected on Twitter and is composed of tweets from
different authors with 100 per author. For each author, a label indicates the
correct gender (male, female). The collection included 3,600 authors, residing
in the United-States, Great Britain, Ireland, New Zealand, Australia and

JADT’ 18

735

Canada, 600 per country, and 1,800 for each group, for a total of 360’000
tweets. The table velow resumes dataset properties.
Authors

Tweets

Genders

3600

360k

(male) 1800 ; (female) 1800

The overall performance of a model is based on the accuracy on the gender
profiling task. The accuracy is the number of correctly classified author
gender divided by the number of authors. Based on data depicted in the table
above, a random baseline will produce an accuracy rate of 0.5 (or 50%).
3. Character N-grams Matrix-based Convolutional Neural Networks
A Convolutional Neural Network (or CNN) is a variety of feed-forward
artificial neural networks inspired by the visual cortex [2]. In our first model,
we applied a CNN to a character bigram representation matrix for an author
in a collection. The first shows the structure of the representation matrix.

For each letter, one can find one row. In the first position, the relative
frequency of this letter is provided. Then, from left to right, the matrix is
composed of the relative frequencies of each character bigram (e.g., at row "t"
and column "h", the relative frequency of the bigram "th" is given). The hird
part is optional and composed of relative frequencies of ending character
bigrams, and finally, the last part is the same optional matrix representing
the starting character bigrams of each word. This matrix representing an
author is the input for the CNN
The first two layers are composed of 20 and 10 kernels respectively, with a
size of 5 × 5. These layers are followed by a drop-out layer. The last two are
linear layers based on ReLU. The outputs are finally obtained by a Softmax
function and give the author’s predicted class. The predicted class is
therefore the class with the highest corresponding output from this function.
The training set is composed of 90% of the dataset and the remaining 10% is
use to estimate the performance. This procedure is repeated 10 times with

736

JADT’ 18

non-overlapping test sets to obtain the 10-fold cross validation estimator.
Matrix / Alphabet

English

+ Punctuation

+ Punctuation &
Smilies

Bigrams

75.26%

76.16%

76.51%

+ starting bigrams

76.02%∗

77.63%∗†

77.50%∗

+ ending bigrams

75.94%

77.22%†

77.25%

+
starting
&
ending bigrams

76.12%

77.83%†

78.33%∗†

4. Echo State Network-based Reservoir Computing models
4.1. Echo State Networks
An Echo State Network was introduced in [3] and corresponds to the first
equation. In this model, the highly non-linear dimensional vector xt, denoting
the activation vector at time t, is defined by
xt+1 = (1 − a) * xt + a * f(Win * ut+1 + W * xt + W)
where xt ∈ R Nx with Nx the number of neurons in the reservoir. The scalar a
represents the leaky rate allowing to adapt the network’s dynamic to the the
task to be learned. The input signal is represented by the vector ut with
dimension Nu, multiplied by the weight matrix in W∈RNx×Nu. In addition, the
matrix W∈RNx×Nx stores the internal weights. Finally, Wbiais is the bias, and
usually the initial vector is fixed to x0 = 0, corresponding to a null state.
The network’s output ŷ is defined by ŷt = g(W * xt) and the learning phase
consists of finding the values of the matrix Wout∈RNy×Nx , e.g., by applying the
ridge regression method. This matrix is defined by
Wout = Y * XT(X * XT + λ * I)−1
Where Y∈RNy×T is the matrix containing each target output ŷ for t = 1, 2, . . ., T
where T denotes the training size, and Ny the number of outputs (categories).
Similarly, the matrix X∈RNx×T stores the reservoir states xt obtained during the
training phase. Finally, the parameter λ is a regularization factor.
4.2. From texts to temporal signals
In order to apply ESN for text classification, we must first transform input
texts as a temporal signal. In this study, we have evaluated three signal
converter methods. First, each word sequence in a text (e.g. "to the citizens

JADT’ 18

737

of") can be viewed as a word vector (WV) (e.g., vec(to), vec(the), vec(citizens),
vec(of), each vector extracted from word embeddings pre-trained with
Glove), Part-Of-Speech (POS) vector (size : number of POS tags), and as a
function word (FW) (size : number of FW).
As output, the ESN generated the vector yt,g with g∈{male, female} denoting
the probability that the tokens in the ESN’s memory at time t as been written
by a man or a woman. We then end up with an output temporal signal of
gender probabilities (over t = 1, 2, . . ., T), and the final predicted class of a
document is the one with the highest average across time.

4.3. State-Gram
In addition, the output layer can take account of more than one state to
estimate the class probabilities. A state-gram value of 2 means that the
training is performed, not only on a single xt , but on xt−1 ∪ xt . Such a model
was effective for handwritten digits recognition [4].
5. Results
In the second table, one can see the results of the deep-learning CNN model
with different vocabulary and starting and ending bigrams. The statistical
tests indicate that the starting bigrams can significantly improve the
performance with respect to the base model (first row). The combination of
starting and ending bigrams (last row) shows a significant improvement only
for the vocabulary composed of punctuation marks and smilies. The best
result (78.33%) is achieved by a CNN model with punctuation and smilies,
with starting and ending character bigrams.

738

JADT’ 18

The left plot in the second figure shows the three features (WV, POS, FW)
with a leak rate between 0.01 and 1.0. Using the same three feature sets, the
right-side plot indicates the accuracy rate obtained by the state-gram model
with value between 1 and 5. With a solid line, the best leak-rate parameter
value is used, and with the dotted curves, a leak-rate value of 1 was used.
Overall, Figure 2 indicates that the pre-trained word vector (WV) is the best
feature set with a maximum value of 80.81% with a leak rate of 0.01. As the
best accuracy rates is obtained with a leak rate between 0.01 and 0.05 (left
plot in Figure 2), we can conclude that the author profiling task has a very
slow temporal dynamics. The right-side plot signals that no significant
improvement is achieved by increasing the value of the stage-gram
parameter for the best leak-rate parameter value. Moreover, a high value of
Ns decreases the performance for POS feature. The performance slightly
increases for a leak-rate parameter value of 1, but these results show that the
leak-rate parameter is a better lever to increase the accuracy rates.
The following table compares the accuracy rates that can be achieved by a
random classifier, the naive Bayes model together with the CNN and ESN
models (with Nx = 1,000).
Classifier

10-CV success rate

Random baseline

50.0 %

Naive Bayes classifier baseline

75.5 %

CNN 2-grams + starting-grams + ending-grams

78.3 %

ESN on Glove with Nx = 1000

80.6 %

6. Conclusion
This paper presents a comparison of two neural models composed of a
character-based CNN model and an echo-state network (ESN) model with
POS, function words (FW) or pre-trained word vectors (WV) as possible
feature sets. Based on the CLEF-PAN 2017 dataset, the best CNN model
achieves a success rate of 78.3% with a feature set composed of the
vocabulary, the punctuation marks, and smilies. The best ESN model obtains
a success rate of 80.6% with 1,000 neurons and a leak-rate of 0.01. Based on
our experiment setting, this model achieves the best performance. In
comparison, the naive Bayes classifier obtains a success rate of 75.5% and the
average and best performance for the gender profiling task in PAN 2017 was
respectively 75.88% and 82.5%.
Our results indicate that the two models can significantly improve the
accuracy rate on the gender profiling task. Moreover, they demonstrated that

JADT’ 18

739

a simple model, thanks to its simple linear regression algorithm, such as the
echo state network can achieve a higher success rate than a more complex
model such as a character-based CNNs. This higher result level can be
explain by the recurrent architecture of the ESN model, allowing it to take
into account word order. In the future, we want to explore more features for
the ESN and word vectors pre-trained for Twitter applications to achieve
hopefully a better performance. We will also apply classical and deep ESN
architectures to other natural language processing tasks such as authorship
identification and author diarization.
References
Francisco Rangel, Paolo Rosso, Ben Verhoeven, Walter Daelemans, Martin
Potthast, and Benno Stein. Overview of the 4th author profiling task at
pan 2016 : cross-genre evaluations. Working Notes Papers of the CLEF,
2016.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradientbased learning applied to document recognition. Proceedings of the IEEE,
86(11) :2278–2324, 1998.
Herbert Jaeger. The “echo state” approach to analysing and training
recurrent neural networks-with an erratum note. Bonn, Germany :
German National Research Center for Information Technology GMD
Technical Report, 148(34) :13, 2001.
Nils Schaetti, Michel Salomon, and Raphaël Couturier. Echo state networksbased reservoir computing for mnist handwritten digits recognition. In
Computational Science and Engineering (CSE) and IEEE Intl Conference
on Embedded and Ubiquitous Computing (EUC), 2016 IEEE Intl
Conference on, pages 484–491. IEEE, 2016.

740

JADT’ 18

Segments répétés appliqués à l'extraction de
connaissances trilingues
Lionel SHEN
Université Sorbonne Nouvelle - Paris 3 – lionel.shen@sorbonne-nouvelle.fr

Abstract
In a context of globalized societies, multilingualism is becoming an economic
and social phenomenon. Translation constitutes a crucial element for
communication. A good translation guarantees the quality of the
transmission of all information. However, face to the challenge of
multilingual information monitoring, can we simply use translation? With
the advent of the digital age and the integration of all new technologies,
corporate governance is undergoing a complete metamorphosis. One of the
priorities remains the efficient exploitation of accumulated big data. The
objectives of this paper hope to highlight the specificity and efficiency of the
Repeated Segments tool through information discovering of trilingual
thematic corpora (French, English and Chinese).
Résumé
Dans un contexte de sociétés mondialisées, on peut parler de multilinguisme
ou encore de plurilinguisme. Aujourd’hui, la frénésie autour du phénomène
des mégadonnées et le multilinguisme sont en train de métamorphoser tous
les services et les comportements de notre époque. La traduction devient
alors un élément capital pour la communication entre les peuples. Une bonne
traduction garantit la qualité de la transmission de toutes les informations.
Cependant, devant la gageure que constitue le projet de réaliser une veille
multilingue, peut-on utiliser simplement la traduction ? Cet article s’articule
autour d’explorations de corpus thématiques trilingues appliquées à
l'extraction de connaissances et tente de mettre en lumière la spécificité et
l’efficacité des cooccurrences en trois langues, français, anglais et chinois.
Keywords: segments répétés, textométrie, veille multilingue, multilinguisme,
fouille d’informations, text-mining, cooccurrences, poly-cooccurrences
1. Introduction
Le monde, qui utilise des centaines de langages depuis des millénaires, a
formalisé les mots et les grammaires pour transcrire, enseigner et transmettre
sur des supports, les savoirs, les faits et les pensées. Des hiéroglyphes aux

JADT’ 18

741

idéogrammes, en passant par les alphabets, ces représentations diffusent
ainsi l'image du monde à travers les époques, les évolutions, les moeurs et les
courants de pensée. Cela représente aujourd’hui des centaines de milliards
de mots dans des corpus différents, avec des occurrences variables. Il n'est
pas possible à un être humain d'aborder par lui-même la masse des
publications archivées ou en circulation. Seul l'usage de l'informatique peut,
à présent, dans le cadre de la mondialisation, permettre un balayage massif
des séquences des corpus nécessaire à l'étude des occurrences et des usages
des mots, au moins dans les langues essentielles diffusant le savoir,
l'information et la communication entre les humains. L'utilité de ces
recherches est étendue, allant des besoins sociaux, humains, scientifiques aux
guerres économiques, en passant par les médias et les enjeux stratégiques des
politiques. C'est la capacité à détecter, enregistrer, analyser et comprendre
dans les meilleurs délais, qui va permettre aux différentes forces de pouvoirs
d'anticiper les décisions et d'agir efficacement. Cette force de veille,
implantée de manière continue et basée sur des outils performants, élaborés
et mis en œuvre par des chercheurs, des informaticiens, des stratèges, des
économistes, sous l'autorité des décideurs... va donc construire les forces de
demain, parfois à l'échelle de la planète. Dans un contexte de sociétés
mondialisées, on peut parler de multilinguisme ou encore de plurilinguisme.
Aujourd’hui, la frénésie autour du phénomène Big-data et le multilinguisme
sont en train de métamorphoser tous les services et les comportements de
notre époque. La traduction devient alors un élément capital pour la
communication entre les peuples. Une bonne traduction garantit la qualité de
la transmission de toutes les informations. Cependant, devant la gageure que
constitue le projet de réaliser une veille multilingue, peut-on utiliser
simplement la traduction ?
Cet article s’articule autour d’explorations de corpus thématiques trilingues
appliquées à l'extraction de connaissances et tente de mettre en lumière la
spécificité et l’efficacité de l’outil Segments répétés en trois langues.
2. Corpus
Pour constituer ce travail, deux types de corpus sont mobilisés : un corpus
comparable (nommé ENRG) et un corpus parallèle (nommé CLRG),
composés de données textuelles extraites des discours de presse, ainsi que
ceux des ONG. La construction de ces deux corpus s’effectue autour de trois
thèmes d’actualité ayant pour objet, l’environnement, l’énergie et le
changement climatique.
La construction de ces deux corpus s’opère à partir d’articles de journaux
issus de nos trois sphères de communication, à savoir, le Monde pour la
France (4 817 articles), le NYT pour les États-Unis (3 993 articles) et 1200

742

JADT’ 18

médias pour la Chine (14 509 articles) comme le présente les deux figures
(figure 1 et figure 2) ci-dessous.
Les données textuelles extraites du corpus comparable proviennent des
discours de la presse, tandis que celles du corpus parallèle sont issues de
ceux des ONG.

Figure 1 : volumétrie du corpus comparable ENRG

Figure 2 : volumétrie du corpus parallèle CLRG

Quant à l’aspect temporel des données du corpus comparable, il diffère selon
les sources et couvre des périodes plus ou moins étendues : de 1999 à 2012
pour le Monde, de 2005 à 2012 pour le NYT, de 2008 à 2013 pour les médias
chinois. Concernant le corpus parallèle, les articles datent de 2006 à 2014. La
figure 3, ci-dessous montre les différentes périodes couvertes par les médias
retenus.

JADT’ 18

743

Figure 3 : périodes couvertes par les corpus ENRG et CLRG

Les dépouillements sont réalisés à l’aide des outils de la textométrie,
notamment grâce aux analyses factorielles des correspondances (AFC), aux
spécificités du modèle hypergéométrique, aux segments répétés, aux réseaux
cooccurrentiels et poly-cooccurrentiels ou encore à la carte des sections. Les
caractéristiques locales et globales, les convergences, les divergences et les
particularités de ces différents corpus ont été mises successivement en
évidence. Après avoir présenté rapidement les deux corpus utilisés, nous
allons nous polariser sur l’outil Segments répétés appliqué d’abord au corpus
parallèle puis ensuite au corpus comparable. Nous nous intéresserons, plus
particulièrement dans cet article à la spécificité des segments répétés
appliqués à l’extraction de connaissances multilingues. Comme le souligne
André Salem, « L’outil prend toute sa valeur lorsque l’unité linguistique traitée
n’est pas le mot, mais le segment répété (suite de mots d’une longueur 2, 3, 4, 5) »
(Salem 1987).
Nous rappelons que «Un segment répété est une suite de formes dont la fréquence
est supérieure ou égale à 2 dans le corpus».
Nous émettons l’hypothèse suivante : l’outil Segments répétés serait plus
performant en chinois, qu’en anglais et qu’en français.
Corpus parallèle : segments répétés anglais-chinois
Nous examinons maintenant les segments les plus répétés obtenus à partir
des deux volets (anglais-chinois) du corpus parallèle CLRG.
Tableau 1 : segments les plus répétés du corpus parallèle CLRG

Le tableau 1 ci-dessus illustre les 14 segments les plus répétés de CLRG.
Nous constatons que la fréquence de segments répétés du volet anglais est

744

JADT’ 18

beaucoup plus élevée que celle du chinois.
Par exemple, la fréquence du segment climate change est de 2 468 dans le volet
anglais, tandis que dans le volet chinois, la fréquence est de 830.
La signification des segments répétés du volet anglais relève peu
d’informations intéressantes. Les mots-outils ou les mots syntaxiques sont les
plus répétés, un seul thème relatif à notre recherche est présent, climate
change. En revanche, les segments répétés en chinois nous révèlent les
véritables thèmes de notre recherche, gaz à effet de serre, changement climatique,
énergies renouvelables, nouveau/nouvelle.
Nous pouvons dire que deux types de répétitions se manifestent : d’une part
de mots grammaticaux pour l’anglais, et d’autre part, de mots de contenu
pour le chinois. Rappelons que la forte répétition de mots grammaticaux est
la cause du grand nombre d’occurrences en anglais. Plus l’emploi des mots
grammaticaux est intensif, plus le nombre d’occurrences est important. Ce
phénomène dissymétrique des segments répétés dans les deux volets est
absolument normal, car la structure syntaxique des deux langues est
complètement différente. Le fait d’avoir des traductions de l’un à l’autre ne
prouve nullement l’emploi symétrique des segments qui se répètent de la
même manière dans les deux langues. Cependant, un prétraitement de
l’anglais pour éliminer les mots outils donnerait plus de sens à l’étude des
segments répétés (Shen, 2016). Les remarques formulées par André Salem
viennent étayer notre hypothèse, renforcée également par celles de Damon
Mayaffre. « L'analyse des voisinages récurrents permet d'utiliser les segments
répétés pour documenter les analyses statistiques faites à partir des formes simples.
On trouvera enfin une analyse typologique effectuée à partir des segments répétés. »
(Salem, 1986). «Moins encore que la fréquence d’un mot, la récurrence de segments
ne peut être naïvement attribuée au hasard : soit elle pointe une contrainte
syntaxique, soit elle indique une détermination ou option sémantique. Dit
rapidement, le mot est une unité graphique, le plus souvent ambiguë, sans sens
explicite, pas même doté de signification. Le segment, lui, devient une unité
linguistique porteuse de sens» (Mayaffre, 2007).
Ces résultats de l’étude bilingue (anglais-chinois) des segments répétés
parallèles ainsi que leurs analyses montrent que, pour une même information
énoncée et décrite en deux langues, la répétition événementielle et
thématique est plus saillante en chinois en raison de la faible pratique des
anaphores (Shen, 2016). De plus le contenu est plus diversifié, puisque nous
retrouvons nos principaux thèmes de recherche.
Nous abordons l’étude des segments répétés dans le corpus comparable
ENRG, composé de trois sous-corpus : sous-corpus français ENRG-FR, souscorpus américain ENRG-US, sous-corpus chinois ENRG-CN.

JADT’ 18

745

Corpus comparable : segments répétés trilingues (français, anglais/US,
chinois)
Tableau 2 : segments les plus répétés du corpus comparable ENRG

Le tableau 2 ci-dessus présente les 16 segments les plus répétés d’ENRG.
Comme pour le corpus parallèle, notre premier constat est une répétition
thématique particulièrement saillante pour le sous-corpus chinois (ENRGCN). Par exemple, la fréquence du segment réduire les émissions est de 12
554, placé comme le segment le plus répété dans ENRG-CN, formes absentes
dans le haut du tableau des deux autres sous-corpus. Cependant, ces formes
existent, mais sont classées bien plus bas dans les résultats des segments
répétés. Les autres sous-thèmes représentés par les séquences répétées
comme faible teneur en carbone, énergie éolienne, photovoltaïque, etc., directement
liés aux énergies et au changement climatique sont également mis en valeur
dans le tableau 2. Pour les sous-corpus français et américain, seuls des mots
grammaticaux ou mots-outils apparaissent dans les segments les plus
répétés. Ce phénomène est dû essentiellement au mécanisme des anaphores
ou au mécanisme déictique qui n'est pas le même en français et en anglais
américain (Shen, 2016). Toutefois, nous remarquons qu’en chinois, ce sont
des termes clés qui se répètent, tandis qu’en anglais et en français, il s’agit
souvent d’entités nommées (noms propres, toponymes, etc.).
3. Conclusion
Dans le processus d’extraction de connaissances trilingues, nous pouvons
conclure que les segments répétés mettent en lumière très efficacement les
caractéristiques les plus saillantes en chinois que dans les deux autres
langues occidentales. Deux types de répétitions se manifestent : d’une part
des mots grammaticaux pour le français et l’anglais, et d’autre part, des mots
de contenu pour le chinois.
De plus, nous soulignons que les cooccurrences ou poly-cooccurrences
permettent également d’extraire des connaissances du corpus grâce à la

746

JADT’ 18

coprésence de formes éloignées. Selon Mayaffre, «L’étude des segments répétés
offre une alternative à la lemmatisation. Elle permet de désambiguïser les termes de
manière formelle et surtout de manière endogène, en corpus et non en référence
(arbitraire) au dictionnaire ou à la langue» (Mayaffre, 2007).
A juste titre, en raison de la forte présence des mots-outils, les cooccurrences
ou poly-cooccurrences par rapport aux segments répétés permettent de
récupérer les séquences répétées non contigües au travers des phrases ou des
paragraphes.
A partir des résultats des segments répétés des deux corpus, nous pouvons
affirmer que l’outil Segments répétés présente l’avantage d’extraire
rapidement des informations clés en chinois, alors qu’en français et en
anglais, le mécanisme des cooccurrences et poly-cooccurrences met en valeur
des informations non détectables par des moyens traditionnels (par exemple,
les concordances).
Aussi, l’outil Segments répétés constitue un atout fondamental pour la fouille
d’informations multilingues.
Bibliographie
Bonnafous S. and Tournier M. (1995). Analyse de discours, lexicométrie,
communication et politique. In : Langages, 29e année, n°117, Paris,
Larousse, pp. 67-81.
Habert B., Nazarenko, A., Salem A. (1997). Les linguistiques de corpus. Paris,
Armand Colin/Masson, 254 p.
Habert B. and Zweigenbaum P. (2002). Problèmes épistémologiques : Régler
les règles. TAL. Paris, Association pour le traitement automatique des
langues, vol. 43, no3, pp. 83-105.
Lafon P. (1981). Analyse lexicométrique et recherche des cooccurrences. In:
Mots, n°3, octobre 1981. Butor-Rousseau, Péguy, Presse du Zaïre, "la
nouvelle droite", vocabulaires, communiste et socialiste, cooccurrences?
pp. 95-148.
Lafon P. and Salem A. (1983). L'inventaire des segments répétés d'un texte.
In: Mots, n°6, mars 1983. L'oeuvre de Robert-Léon Wagner. Vocabulaire
et idéologie. Analyses automatiques. pp. 161-177.
Lamalle C. and Salem A. (2002). « Types généralisés et topographie textuelle
dans l’analyse quantitative des corpus textuels » in A. Morin et P.
Sébillot (éds), JADT 2002. Saint-Malo : IRISA-INRIA, vol. 1, 403-411.
Lebart L. and Salem A. (1995). Statistique textuelle
Longrée D., Luong X., Mellet S. (2004). « Temps verbaux, axe syntagmatique,
topologie textuelle : analyses d’un corpus lemmatisé » in G. Purnelle, C.
Fairon, A. Dister (éds), JADT04. Louvain : Presses universitaires de
Louvain, vol. 2, 743-752.

JADT’ 18

747

Longrée D., Luong X., Mellet S. (2006). « Distance intertextuelle et classement
des textes d’après leur structure : méthodes de découpage et analyses
arborées » in J.-M. Viprey (textes réunis par), JADT’ 06. Besançon :
Presses universitaires de Franche-Comté, vol. 2, 643-654.
Mayaffre D. (2004). Paroles de président. Jacques Chirac (1995-2003) et le
discours présidentiel sous la Vème République. Paris :
Champion.Mayaffre, Damon (2004). Paroles de président. Jacques
Chirac (1995-2003) et le discours présidentiel sous la Vème République.
Paris : Champion.
Mayaffre D. (2007). L'analyse de données textuelles aujourd'hui : du corpus
comme une urne au corpus comme un plan : Retour sur les travaux
actuels de topographie/topologie textuelle. Lexicometrica, Andrée
Salem, Serge Fleury, 2007, pp.1-12.
Rastier F. (2001). Arts et sciences du texte. Paris : Puf.
Salem A. (1986). Segments répétés et analyse statistique des données
textuelles. In: Histoire & Mesure, 1986 volume 1 - n°2. Varia. pp. 5-28.
Salem A. (1987). Pratique des segments répétés. Essai de statistique textuelle,
1987
Shen L. (2016). Méthodes de veille textométrique multilingue appliquées à
des corpus de l’environnement et de l’énergie : « Restitution, prévision
et anticipation d’événements par poly-résonances croisées », Thèse :
Sciences du langage, Université Sorbonne Nouvelle – Paris 3, octobre
2016, 474 p.
Viprey J. (2005-a). « Philologie numérique et herméneutique intégrative », in
Adam J.-M. et Heidmann U. (éds.), Sciences du texte et analyse de
discours. Genève : Slatkine, 51-68.
Viprey J. (2005-b). « Corpus et sémantique discursive : éléments de méthode
pour la lecture d corpus », in A. Condamines (dir.), Sémantique et
corpus. Paris : Lavoisier, pp. 245-276.
Viprey J. (2006). « Structure non-séquentielle des textes », Langages, 163, 7185.

748

JADT’ 18

Misurare, Monitorare e Governare
le città con i Big Data
Sandro Stancampiano
Istat – stancamp@istat.it

Abstract
Several new data sources are investigated in the production process of
official statistics. This paper describes the results of the analysis of online
reviews about four points of interest in Rome, Italy. The reviews, collected
from the web using web scraping and data wrangling techniques, was
written by tourists and visitors during the 2017. The general aim of this
research is to extract useful information to help civil servants and citizens in
decision-making processes. Within the activities related to this study were
automatically collected and stored in a Data Base 9227 documents (each
document is a review) used to build the corpora. The paper intends to
classify the reviews and qualify the sentiment of the texts using tools and
techniques of text mining.
Abstract
Numerose nuove fonti di dati vengono analizzate nel processo di produzione
delle statistiche ufficiali. Questo documento descrive i risultati dell'analisi
delle recensioni online su quattro punti di interesse della città di Roma, in
Italia. Le recensioni, raccolte con tecniche di web scraping e data wrangling,
sono state scritte da turisti e visitatori nel corso del 2017. Lo scopo generale di
questa ricerca è di estrarre informazioni a supporto dei processi decisionali
sia dei dipendenti pubblici sia dei cittadini. Tra le attività correlate a questo
studio sono stati raccolti e archiviati automaticamente in una base di dati
9227 commenti utilizzati per creare un corpora analizzato utilizzando
strumenti e tecniche di text mining. Il documento intende classificare le
recensioni e qualificare il sentimento dei testi.
Keywords: big data, Internet as data source, text mining, cluster analysis,
web scraping.
1. Introduzione
Questo progetto si propone di indagare soluzioni relative all’uso dei Big Data
per produrre statistiche ufficiali a supporto della pubblica amministrazione.
L’Istat ha incluso questo tema, condiviso a livello europeo, nel Piano

JADT’ 18

749

triennale della ricerca tematica e metodologica1. L’Istat sta considerando la
possibilità di utilizzare i Big Data nel processo di produzione dei dati, in
modo da attenuare il trade-off tra tempestività e accuratezza (Alleva, 2016).
2. Background della ricerca
Questo lavoro si focalizza sul tema della gestione dei beni culturali,
indagando mediante tecniche esplorative multivariate (Bolasco, 2014) fonti
dati non convenzionali. Si vogliono mostrare le enormi potenzialità dei dati
presenti sul web per produrre statistiche al fine di ottimizzare i processi
decisionali. Il risultato della ricerca potrà essere di ausilio agli amministratori
nella gestione dei servizi dedicati ai fruitori dei beni culturali presenti sul
territorio. L’esperimento, che si concretizza in un progetto pilota replicabile
ed estendibile su ampia scala, utilizza l’analisi testuale (text mining) per
estrarre informazioni da dati scaricati dal web mediante tecniche di web
scraping. Si vogliono scoprire regolarità nei testi esaminati utilizzando la
cluster analysis (analisi dei gruppi). Questa tecnica, applicata attraverso il
software IRaMuTeQ, consente di definire la distanza tra gli oggetti che si
vogliono classificare (Ceron et al., 2013).
3. Obiettivo e ipotesi di ricerca
Tra i molti siti web utilizzati dagli utenti per produrre contenuti, è stato
scelto Tripadvisor. Gli utenti registrati utilizzano il sito per scrivere le loro
recensioni sui luoghi in cui si sono recati condividendo le loro esperienze
(Iezzi e Mastrangelo, 2012). Sono state scelte quattro tra le più celebri
attrazioni della città di Roma frequentate quotidianamente da numerosi
turisti (Colosseo, Pantheon, Fontana di Trevi e Piazza Navona). Il Colosseo
con oltre sei milioni di visitatori ha determinato, anche per il 2016,
l'incremento degli incassi garantiti dai musei italiani2 e la supremazia della
regione Lazio in questa graduatoria. Molti visitatori lasciano valutazioni
relative ai luoghi aggiungendo considerazioni sullo stato di conservazione
dei beni, sui servizi e i disservizi che hanno notato. Si ritiene che analizzando
questi commenti, sia possibile dedurre preziose informazioni.
L’analisi ha permesso di ottenere una classificazione gerarchica delle
recensioni basata sui termini caratterizzati da un utilizzo superiore alla
media con riferimento alla variabile monumento.

https://www.istat.it/it/files/2011/07/Piano-strategico-2017-2019.pdf (pp.27-28)
http://www.beniculturali.it/mibac/export/MiBAC/sitoMiBAC/Contenuti/Mibac
Unif/Comunicati/visualizza_asset.html_892096923.html
1
2

750

JADT’ 18

4. Corpus e metodo
I commenti sono stati raccolti in una base dati mediante l’applicativo
Diogene3: progettato con il paradigma OOA/D e realizzato con metodologia
agile (Larman, 2005). Utilizzando lo stesso software è stato creato il corpus
delle recensioni.Le 9227 recensioni raccolte, pubblicate dal 1 gennaio al 31
dicembre 2017, sono così suddivise: Colosseo 3483 (37.8%), Piazza Navona
1020 (11%), Fontana di Trevi 2829 (30.6%) e Pantheon 1895 (20.5%).
Si è proceduto in prima istanza con l’analisi lessicale ricavando informazioni
utili alla successiva analisi testuale volta a localizzare unità di testo di rilevo
per gli obiettivi del presente studio (Bolasco, 2013). L’analisi ha permesso di
individuare gruppi di parole omogenei al loro interno ed eterogenei tra loro
riguardo ai “concetti” espressi nelle recensioni. Il corpus analizzato è
composto da 9227 testi, 1788819 occorrenze, 11891 forme, 366 hapax di cui il
3.08% relativi alle forme e lo 0.02% relativi alle occorrenze e media 193.87.
La ricchezza lessicale del corpus è molto bassa4 (V/N*100 = 0.66%), difatti a
fronte di un testo ampio si riscontra un vocabolario ridotto. Osservando le 30
forme attive con la frequenza assoluta maggiore, notiamo come il linguaggio
utilizzato privilegi i sostantivi e gli aggettivi rispetto ai verbi. Gli aggettivi
esprimono positività (bello, bellissima, grande) e i sostantivi sono legati alla
fruizione dei beni oggetto di studio (monumento, piazza, visita, luogo, consiglio,
interno) così come i verbi (visitare, fare, vedere, dire, entrare, trovare).
5. Gli scriventi e le recensioni
I dati relativi ai giorni della settimana in cui è stata scritta la recensione,
evidenziano la tendenza degli utenti a mettere nero su bianco i dettagli delle
loro esperienze nei giorni centrali della settimana, con una predilezione per i
mercoledì (vedi Figura 5.1).
Le persone durante i fine settimana si dedicano alle visite dei beni culturali e
preferiscono descrivere quanto visto e vissuto martedì, mercoledì e giovedì.
Nel periodo oggetto di studio le recensioni relative alle quattro piazze sono
state in media 741 al mese con un minimo di 572 a giugno e un massimo di
1129 a gennaio.
Dalla Figura 5.2 risulta che i primi mesi dell’anno, da gennaio ad aprile, sono
quelli in cui si concentra il maggior numero di recensioni (oltre il 42% del
totale).

3 Diogene è un software sviluppato in java per effettuare processi di data
wrangling.
4 Il calcolo è stato effettuato applicando la formula RL=V/N dove V = ampiezza
del vocabolario e N = numero totale di parole nel testo.

JADT’ 18

751

Figura 5.1: Numero di recensioni per giorno
della settimana (gennaio – dicembre 2017)

Figura 5.2: Numero di recensioni per mese
(gennaio – dicembre 2017)

6. Cluster Analysis
La cluster analysis ci consente di raggruppare le unità statistiche
massimizzando coesione e omogeneità delle parole incluse in ciascun gruppo
e minimizzando al tempo stesso il legame logico tra quelle assegnate a
gruppi/classi differenti.

Figura 6.1: Dendrogramma delle classi secondo similarità

752

JADT’ 18

Il dendrogramma (Figura 6.1) mostra la divisione del corpus in 4 classi. Le
parole contenute in ciascuna classe permettono di individuare le tipologie di
argomenti trattati nel corpus, applicando la metodologia Alceste proposta da
Max Reinert e implementata nel software IRaMuTeQ5.
In Figura 6.2 osserviamo le parole appartenenti ai quattro gruppi e come si
dispongono sul piano fattoriale. Questa visualizzazione chiarisce meglio il
significato delle classi individuate.
Il gruppo di parole in rosso (65.4%), che si concentrano intorno all’origine, è
composto dai termini più utilizzati: trasversali a tutto il corpus e di
conseguenza a tutti e quattro i beni esaminati. Si tratta di parole tema come
roma, simbolo, monumento, città, storia, dei verbi visitare, vedere, tornare, dire e di
sostantivi e aggettivi come bello, emozione, luce, bellezza che esprimono
positività e azioni legate alla visita.
La classe 2, in verde (10.9%), rappresenta i commenti pubblicati da persone
che sono attente a quello che accade nei luoghi e considerano prioritaria la
sicurezza, la legalità e la qualità dei servizi che trovano.
Si distinguono parole come venditore, abusivo, presenza, peccato, fastidioso,
ordine, municipale, polizia, fischietto. Ci sono inoltre parecchi riferimenti alle
attività commerciali (bar, bancarella, locale, ristorante, gelateria, trattoria) con
particolare riguardo a cosa si può mangiare (aperitivo, pizza, granita, gelato,
vino) e alle modalità di fruizione (tavolino, tavolo, panchina). Questo gruppo di
parole evidenzia considerazioni che non sono strettamente correlate alla
visita culturale ma piuttosto a tutto quello che ruota intorno a una escursione
turistica.
La classe 3, in celeste (12.7%), rappresenta tematiche connesse ad aspetti
economici e pratici che in alcuni casi possono causare disagio durante la
visita. Emergono parole come acquistare, prenotare, saltare, fila, coda,
interminabile, biglietto, pagare, guida, audioguida, gratis, costo, euro, ticket.
Gli argomenti sottesi sono relativi al costo del biglietto, all'attesa per
l’ingresso e alla modalità della visita con connotazione sia positiva sia
negativa a seconda della situazione particolare descritta dall’utente.
La classe 4, in viola (11%), rappresenta coloro che descrivono e raccontano
l’esperienza dal punto di vista culturale citando eventi, luoghi e personaggi
storici. Le parole più utilizzate sono tomba, re, raffaello, sanzio, chiesa, colonna,
fiume, barocco, agone, agnese, borromini, savoia, papa, pagano, cristiano. Si tratta di
riferimenti a luoghi di culto e opere (Sant’Agnese in Agone, la fontana dei
Quattro Fiumi, le tombe dei re custodite nel Pantheon, ecc.), agli artisti che le

5 IRaMuTeQ è un software realizzato per effettuare analisi multidimensionali di
testi che fornisce una interfaccia grafica a R, altro software di elaborazione dati
particolarmente efficiente per l’analisi di grandi dataset.

JADT’ 18

753

Figura 6.2: La disposizione delle parole sul piano fattoriale

hanno realizzate (Raffaello Sanzio e Borromini su tutti), alla storia e al
contesto sociale e culturale di pertinenza dei siti visitati.
La disposizione dei termini sul piano fattoriale, a prescindere dai gruppi,
evidenzia il continuum della visita, che inizia con la prenotazione, la biglietteria
e il successivo acquisto seguito dalla fila per entrare e dalla constatazione della
bellezza del monumento per poi visitare e immergersi negli aspetti artistici e
nella storia del luogo in cui ci si trova.
7. Conclusioni e sviluppi futuri
Le tematiche palesate sono di sicuro interesse per gli amministratori pubblici,
che possono ascoltare direttamente dalla voce dei cittadini quali sono i
principali problemi dal punto di vista degli utenti. Sulla base di questo
genere di analisi il decisore può valutare se e come intervenire per migliorare

754

JADT’ 18

la gestione dei luoghi e dei beni culturali.
Il flusso informativo parte dal cittadino che alla fine del processo può
ottenere dei benefici tangibili grazie ai dati che lui stesso ha immesso in rete.
Il processo descritto in questo lavoro mostra un uso classico di Big Data: dati
prodotti con una finalità specifica vengono utilizzati successivamente per
raggiungere altri obiettivi apportando un innegabile valore aggiunto
(Rudder, 2015).
Le tecniche di text mining applicate hanno permesso di valorizzare
informazioni che altrimenti sarebbero rimaste inutilizzate.
Ulteriori e più approfondite analisi potranno essere condotte con la stessa
metodologia e i medesimi software adoperati in questo lavoro. Si potrà
continuare il monitoraggio, incrementando il corpus per condurre un’analisi
longitudinale su questi stessi monumenti o studiare altre città e altri beni
culturali al fine di migliorare le politiche di gestione e ottimizzare i processi
decisionali.
References
Alleva G. (2016). Più forza ai dati: un valore per il Paese. Relazione di apertura
della 12° conferenza nazionale di statistica.
Bolasco S. (2014). Analisi Multidimensionale dei dati. Metodi, strategie e criteri
d’interpretazione. Carocci editore.
Bolasco S. (2013). L’analisi automatica dei testi. Fare ricerca con il text mining.
Carocci editore.
Ceron A., Curini L., Iacus S. M. (2014). Social Media e Sentiment Analysis.
L’evoluzione dei fenomeni sociali attraverso la Rete. Springer Italia.
Iezzi Domenica F., and Mastrangelo M. (2012). Il passaparola digitale nei
forum di viaggio: mappe esplorative per l’analisi dei contenuti. Rivista
Italiana di Economia, Demografia e Statistica, 66 (3-4), pp. 143-150.
Larman C. (2005). Applicare UML e i Pattern. Analisi e progettazione orientata
agli oggetti. Luca Cabibbo (a cura di), Pearson Education Italia.
Rudder C. (2015). Dataclisma. Chi siamo quando pensiamo che nessuno ci stia
guardando. Mondadori.

JADT’ 18

755

Exploration textométrique d’un corpus de motifs
juridiques dans le droit international des transports
Fadila Taleb1, Maryvonne Holzem2
Université Rouen Normandie – fadila.taleb@etu.univ-rouen.fr
2Université Rouen Normandie– maryvonne.holzem@univ-rouen.fr
1

Abstract
Within the framework of a research whose objective consists of responding to
a need formulated by the IDIT, which helps to interpret the jurisprudential
texts contained in its database, we are looking to highlight the interpretive
paths considered as modal scenarios. We propose here a preliminary
textometric analysis in order to define the linguistic profile of the corpus and
to detect certain repeated segments that may represent a relevant constraint
to complete and enrich the interpretive paths identified in the case law.
Résumé
Dans le cadre d’une recherche dont l’objectif consiste à répondre à un besoin
formulé par l’IDIT1, celui d’aider à l’interprétation des textes jurisprudentiels
contenus dans sa base de données, nous cherchons à mettre au jour des
parcours interprétatifs envisagés comme des scénarios modaux. Nous
proposons ici une analyse textométrique préalable afin de cerner le profil
linguistique du corpus et de détecter certains segments répétés pouvant
représenter une contrainte pertinente pour compléter et enrichir les parcours
interprétatifs identifiés dans les textes jurisprudentiels.
Keywords: textométrie, parcours interprétatif, scénario modal, segments
répétés, motifs juridiques, droit des transports.
1. Introduction
1.1. Contexte
Dans le cadre d’un projet pluridisciplinaire « PlaIR »2, des chercheurs
informaticiens, linguistes, juristes posent la question de l’aide à
l’interprétation3 du fond jurisprudentiel de la base de données de l’IDIT. Du
point de vue linguistique, notre tâche préalable à une implémentation
consiste en l’étude de décisions de justice dans le but de comprendre leur

Institut du Droit International des Transports.
Plateforme d’Indexation Régionale
3 Notre objectif est celui d’une aide instrumentée centrée sur l’agir de
l’utilisateur cf. travaux du groupe ʋ (Holzem et Labiche, 2017).
1
2

756

JADT’ 18

structure, le mécanisme argumentatif mis en œuvre et les mouvements de
transformations textuelles susceptibles de déclencher des parcours
interprétatifs pouvant aider à la lecture de ces décisions. Notre recherche
s’écarte des modèles prédictifs, justice prédictive ou legaltech, qui, sous
l’influence des big data et du Machine Learning, produisent des résultats de
contentieux sur des bases algorithmiques. De ce point de vue, nous
partageons les craintes de bon nombre de juristes de voir ces legaltech
« devenir eux mêmes une nouvelle forme de justice » (Garapon, 2017). Il s’agit
d’une pratique textuelle (et intertextuelle) comprise comme régime de
transformation et d’interprétation. Dans cette perspective, notre recherche se
place donc du côté de la jurilinguistique et son objectif est d’essayer de
comprendre dans une approche linguistique et à travers l’étude du matériel
textuel les décisions de justice et les stratégies argumentatives mises en
œuvre pour ainsi aider à leur interprétation.
1.2. Questionnement et hypothèse
Pour aider à l’interprétation nous cherchons à cerner les stratégies
argumentatives mises en œuvre par le juge, notamment dans sa manière
d’intégrer et de prendre en charge les discours des autres (celui des parties
du procès, celui des experts, celui du législateur etc.). Notre hypothèse est
fondée sur des recherches antérieures (Holzem 2014 et Taleb 20144) qui ont
montré l’intérêt de la prise en compte des modalités linguistiques suivant le
modèle développé dans (Gosselin 2010) pour la constitution d’un parcours
interprétatif (Rastier 2001) envisagé ici comme scénario modal susceptible
d’aider à l’interprétation. Mais avant de procéder à une telle analyse textuelle
menée directement sur des textes pleins, nous avons eu besoin de cerner dans
sa globalité et ses spécificités le profil linguistique de notre corpus d’étude.
Pour cela, nous avons eu recours à une analyse textométrique approfondie,
menée avec le logiciel TXM. Au fil de nos investigations textométriques, nous
nous sommes rendu compte de l’importance de certaines fonctions offertes
par ces outils pour la détection, par exemple, de segments répétés5, qui peuvent
représenter une contrainte pertinente pour compléter les parcours
interprétatifs identifiés grâce à l’étude des modalités. L’objectif de cet article
est de présenter, dans ses grandes lignes, en raison de la place, l’analyse
textométrique menée sur notre corpus.

4 Un mémoire de master 2 recherche en science du langage soutenu en juin 2014 :
« Étude du scénario modal et du syllogisme juridique pour la compréhension du processus de
production du texte. Cas des textes du droit. »
5 Suite de formes graphiques identiques attestées plusieurs fois dans le texte.

JADT’ 18

2.

757

Corpus et méthodologie

2.1. Description globale
Nous avons, à la suite de Rastier (2011), retenu le critère du genre comme
critère définitoire du corpus de référence. Il regroupe des textes (décisions de
justice) relevant du discours judiciaire6 et appartenant au genre
jurisprudentiel7. En reprenant la typologie du corpus proposée par B.
Pincemin (1999) et reprise par (Rastier 2011), nous avons distingué quatre
niveaux de corpus : (i) un corpus existant/latent (archives pour Rastier) qui
correspond dans notre recherche à la base de données de l’IDIT ; (ii) un
corpus de référence qui renvoie à l’ensemble des documents numérisés dans
le fond jurisprudentiel de l’IDIT ; (iii) un corpus d’étude qui contient un
nombre délimité de ces décisions sélectionnées pour les besoins de notre
recherche et enfin (iv) un corpus distingué (corpus d’élection ou sous-corpus
pour Rastier) correspondant à des passages précis des textes étudiés nommés
« les motifs ». Ces derniers constituent le cœur du jugement, le juge exposant «
(…) les raisons de faits et de droit qui justifient la décision (…).» (Cohen et
Pasquino, 2013). Notre intérêt pour cette zone textuelle est doublement
motivé. Premièrement notre objectif consiste à repérer les moments clés de
transformations du jugement pour cerner les stratégies argumentatives mises
en œuvre et partant aider à leur interprétation. Deuxièmement la motivation
est une composante commune8 à toutes les décisions de toutes les
juridictions. Elle doit faire face à une double exigence : logique et persuasive.
L’une est due à la forme syllogistique du raisonnement juridique imposée et
l’autre à la nécessité de persuader l’auditoire de la décision9 de sorte à éviter
les recours et faire accepter la solution juridique apportée comme étant la
seule possible.

Il renvoie aux discours produits par (ou au sein) des juridictions. Il est à
distinguer du discours juridique qui désigne, entre autre, les domaines du droit ou ses
sources (lois, réglementation etc.). L’un concerne la création du droit, l’autre rend
compte de son aspect applicatif.
7Le terme de jurisprudence renvoie ici à l’ « ensemble des décisions rendues par les
tribunaux d’un pays, pendant une certaine période dans une certaine manière. » (Dictionnaire
du vocabulaire juridique 2017,éd. LexisNexis) P.322)
8 Ce qui n’est pas le cas pour les autres composantes. Ainsi, l’exposé du litige ne
figure pas dans les arrêts de la cour de cassation, car celle-ci étant une juridiction
d’ordre suprême, son rôle est de veiller à la bonne application des normes juridiques,
elle considère l’appréciation des faits par les juges de fond comme étant souveraine.
9 Composée certes des parties du litige directement concernées par la décision,
mais aussi les autres juges des autres juridictions et un public encore plus large, le
destinataire universel.
6

758

JADT’ 18

2.2. Caractéristiques quantitatives
Le volume textuel du corpus d’étude est de 878848 occurrences dont 22456
formes. Le sous-corpus des motifs représente à lui seul près de la moitié des
occurrences du corpus d’étude. Il contient 393092 occurrences pour 14742
formes. La dysmétrie de la distribution des formes dans les différentes zones
délimitées montre l’importance et le rôle des motifs dans les décisions de
justice, ils sont leur raison d’être, et tout juge est dans l’obligation de motiver
son jugement.
2.3. Encodage et prétraitement
Notre corpus présente l’avantage d’être accessible en ligne. Cependant,
l’ensemble des textes au format PDF n’est pas homogène : certains
documents proviennent d’un format image (non océrisé10). Le format PDF
n’étant pas pris en charge par TXM, nous avons tout d’abord procédé à une
conversion (avec la technique d’océrisation pour les fichiers annotés et
numérisés) au format TXT, puis dans un second temps à un codage XML en
s’inspirant des recommandations de la TEI11 pour l’encodage des données
textuelles. Ce dernier nous permet une navigation plus fine dans le corpus
grâce à des métadonnées péritextuelles, comme celles relatives au type de la
juridiction : tribunal de commerce (TC), cour d’appel (CA), cour de cassation (CC),
à la date et au lieu, et des métadonnées intratextuelles, telles que celles
relatives à des parties spécifiques dans les textes. Nous avons relevé quatre
parties principales : faits, moyens, motifs, conclusions. Les motifs et les conclusions
sont présents dans toutes décisions étudiées. Les faits sont absents des CC, et
les moyens ne sont pas toujours indiqués comme tels dans les arrêts CA, ils
sont souvent rappelés dans la zone des faits sous forme de discours indirect.
La figure suivante représente les différentes phases de préparation du corpus
avant son traitement textométrique :

Figure 1 Les étapes de préparation du corpus

10 OCR (Optical Character Recognition) Reconnaissance Optique de Caractères,
étape nécessaire pour déchiffrer les formes et les traduire ici en lettres.
11 Text Encoding Initiative : recommandations standard pour l’encodage des
documents numériques.

JADT’ 18

759

Pour le passage du format TXT au format XML-TEI nous avons créé les
balises spécifiques au genre du corpus étudié : <TypeJuridiction>,
<DateProcès>, <LieuProcès> etc. Nous avons eu recours à un encodage semiautomatique au moyen d’un tagger conçu spécialement pour notre étude par
Eric Trupin, MCF en informatique au laboratoire LITIS12. Cette étape
indispensable de préparation du corpus pour le traitement textométrique a
été à la fois chronophage et délicate : traitement des annotations manuscrites
et nettoyage de documents plus anciens.
3.

Exploration textométrique du corpus distingué : la zone des motifs

3.1. Etude occurrentielle : les spécificités lexicales
Une première étude contrastive au moyen d’un traitement textométrique
phare, le calcul de spécificités13, permet d’avoir une vue globale sur les
caractéristiques lexicales du corpus distingué « les motifs ». Le tableau cidessus dresse la liste des 20 premières formes les plus spécifiques à cette
zone. Il est trié par ordre décroissant sur l’indice de spécificité de celle-ci :

Figure 2 : spécificité lexicales de la zone des motifs

Nous portons ici attention à un usage excessif d’occurrences caractéristiques
du discours judiciaire et constitutif de la zone des motifs : Attendu, que,
Considérant14, attendu, de même pour les connecteurs : Mais et donc.
Laboratoire d'Informatique, du Traitement de l'Information et des Systèmes,
Université Rouen Normandie
13 Le calcul de spécificités implémenté dans TXM repose sur la loi
hypergéométrique développée par Lafon (1984). Le seuil de pertinence d’une
distribution est fixé à 2 : +2 l’indice de spécificité est positivement significatif, -2 il est
négativement significatif. L’indice se situant entre les deux est banal.
14 Dans notre corpus la forme Considérant n’apparaît que dans les CA. Son
absence dans les CC serait donc significative.
12

760

JADT’ 18

L’ensemble de ces marqueurs jouent un rôle spécifique ici, celui de ponctuer
l’argumentation du juge en assurant sa progression syllogistique. L’usage
excessif du futur, représenté avec les verbes être (sera : 22,9) et condamner
(condamnera : 14,6) n’est pas surprenant, car avant de prononcer le verdict
final dans un acte exclusivement directif (énoncé réservé à la zone des
dispositifs), les juges avancent au préalable dans la zone des motifs les
résultats (comme le montre d’ailleurs le suremploi du verbe résulte (19,6)) de
leurs argumentations : « Le jugement entrepris sera confirmé en ses autres
dispositions qui ne sont pas critiquées».15; « Le tribunal condamnera Monsieur le
capitaine du […] ;».16 L’emploi significatif d’autres mots, comme équité,
marchandises, inéquitable renvoie à la thématique des textes étudiés : le droit des
transports. L’emploi significatif des adverbes de négation : ne (+50,3), pas
(+38,5) révèle une caractéristique particulière de l’argumentation juridique
car, fidèle au principe spinoziste Determinatio negatio es, la négation manifeste
une valeur réplicative et résultative (i.e. portée référentielle en réponse à ce
qui a été énoncé précédemment et qui n'a plus lieu d'être) préparatoire à la
transformation juridique de l’énoncé.
3.2. Etude contextuelle
Au-delà des investigations menées sur des unités lexicales minimales, les
outils que propose la communauté ADT problématisent la notion de contexte
selon des paliers différents pour privilégier un retour au texte. Nous allons
ici donner l’exemple de la contextualisation des « attendu » dans la zone des
motifs dont le suremploi a été relevé dans le tableau ci-dessus.
Suite à
une étude cooccurrentielle autour du mot-pôle « attendu », nous avons
repéré une très forte attractivité avec le connecteur « Mais » (l’indice de
spécificité17= +95).

CA Rouen, 03/10/2013
TC de Rouen, 15/12/2003
17 Le calcul des cooccurrences qui repère les affinités et répulsions lexicales selon
un indicateur de probabilité de rencontre repose sur le même modèle que celui du
calcul des spécificités (Lafon, 1984).
15
16

JADT’ 18

761

Figure 3: Concordancier "Mais attendu" dan la zone des motifs

Nous avons remarqué une systématicité dans l’usage des « Mais attendu » qui
vient clore un enchaînement de propositions subordonnées introduites par
des « Attendu que », repris parfois par la conjonction que. L’étude approfondie
des contextes de ce « Mais attendu » révèle une incidence particulière de celuici sur ses contextes droits :
« Attendu que les marchandises ont été totalement perdues du fait de leur
décongélation. Attendu que la première évaluation des marchandises a été établie à
18. 498, 85 € départ usine, Mais attendu qu'en application de la loi française du 18
juin 1966, le montant de la marchandise s'évalue en valeur CIF (coût + assurance +
fret). Attendu qu'en l'espèce la valeur CIF des marchandises se monte à 21. 163, 96
€, c'est bien ce montant que le tribunal retiendra en préjudice principal. ».
Dans l’extrait ci-dessus Mais attendu introduit non seulement un mécanisme
de renforcement argumentatif18, mais il joue également le rôle de déclencheur
de transformation modale entre deux modalités de type axiologiques19. Dans
l’exemple cité ici, le Mais attendu accompagné d’une référence juridique
« application de la loi française […] assure cette transformation entre une norme
liée au domaine du transport (marchandises totalement perdues du fait de leur
décongélation : modalité axiologique négative) et les modes d’édiction d’une
norme juridique cette fois. La marchandise dépréciée se trouve alors
revalorisée (axiologique positif du point de vue juridique) par le changement
des co-occurrents à droite (valeur CIF (coût + assurance + fret)).

18 Voire les travaux pionniers de A. Ducrot (1984) sur les valeurs argumentatives
de Mais.
19 Les modalités axiologiques sont propres aux jugements de valeur de nature morale,
idéologique et/ou légale. (Gosselin, 2010).

762

JADT’ 18

4. Conclusion
À travers cette contribution nous avons voulu montrer l’intérêt que
représente une étude textométrique pour l’appréhension de son corpus
d’étude. Si notre objectif principal, celui de mettre au jour des parcours
interprétatifs nommés scénarios modaux (Taleb 2015), est difficilement
envisageable en se limitant à une stricte étude textométrique (car elle repose
sur l’étude modale propre à chaque texte). L’approche textométrique s’est
avérée néanmoins pertinente pour décrire et cerner le profil linguistique du
corpus. Son principe différentiel essentiel du point de vue sémantique, nous
a incitées à adopter cette démarche d’analyse contrastive indispensable.
L’analyse contextuelle à plusieurs paliers nous a permis le repérage de
constructions lexicales répétitives, comme l’exemple des « Mais attendu » exposé
ici, qui se révèlent être des moments clés du jugement et donc parcours
interprétatifs corrélatifs à une transformation modale.
Références
Cohen M. et Pasquino P. (2013). La motivation des décisions de justice, entre
épistémologie sociale et théorie du droit. Le cas des Cours souveraines et des cours
constitutionnelles. CNRS, New York University, University of Connecticut.
Ducrot A. (1982). Le dire et le dit. Les Éditions de minuit, Paris.
Garapon A. (2017). Les enjeux de la justice prédictive. La semaine juridique
LexisNexis, N°12: 47-52.
Gosselin L. (2010). Les modalités en français. La validation des représentations.
Amsterdam-New-York : Rodopi B.V.
Holzem M. (2014). Le Parcours interprétatif sous l’angle d’une
transformation d’états modaux, dans Numes Correia C. et Coutinho M. A.
(eds), Estudos Linguisticos : Linguistic studies , n° 10, p. 283-295.
Holzem M. Labiche J (2017) Dessillement numérique : énaction, interprétation,
connaissances. Bruxelles, Bern, Berlin : PIE Peter Lang.
Lafon P. (1984). Dépouillements et Statistiques en Lexicométrie. SlatkineChampion.
Pincemin B. (1999). Diffusion ciblée automatique d’informations : conception et
mise en œuvre d’une linguistique textuelle pour la caractérisation des
destinataires et des documents, Thése de Doctorat en Linguistique, Universit.
Paris IV Sorbonne, chapitre VII.
Rastier F. (2001). Art et science du texte. Puf. Rastier 2011
Rastier F. (2011). La mesure et le grain. Paris, Éditions Honoré Champion.
Taleb F. (2015). Les modalités linguistiques pour aider à l’interprétation de
textes juridiques. Actes Interface TAL IHM (ITI'2015), 22ème Congrès TALn,
Caen.

JADT’ 18

763

The Framing of the Migrant:
Re-imagining a Fractured Methodology
in the Context of the British Media.
James M. Teasdale
Sapienza University of Rome - teasdale.1650019@studenti.uniroma1.it

Abstract 1
This study analyses the portrayal of migrants and migration in the British
press over two periods, using frame analysis as a foundation methodology,
while attempting to improve upon the methodology used in similar studies.
The study holds the ‘frame’ to be the key organising feature in the portrayal
of migrants and these frames can be located through a cluster analysis of
textual data. The first aim of the work is to ascertain how far location and
time affect the deployment of one frame or another, what these frames
consist of and, therefore, provide a detailed analysis of how migration is
portrayed in the British press: a focus sorely lacking in previous frame
analysis studies to date. The study demonstrates that six frames can be
identified over two periods; four being thematic and two being episodic. The
‘negative’ and ‘positive’ migrant frames were present in the first period, as
the ‘local’ focus provided an ideal ground for the former’s deployment as the
subject was located closer to home and was depicted as a threat. While the
second period saw the dominance of the ‘positive’ migrant frame with the
death of Alan Kurdi and the corresponding conceptual shift to the ‘global’
removing the subject from the immediate border and placing them in a wider
context. This was coupled with the overlap of the domestic responsibility
frame with the ‘positive’ migrant frame as the two became intimately linked
in the second period, while the European responsibility frame also arose.
This demonstrated that the hegemony of one frame can be challenged but
only if the corresponding situation is ‘drastic’ enough to allow.
Abstract 2
Questo studio analizza la raffigurazione dei migranti e della migrazione nella
stampa britannica durante il corso di due periodi di tempo, utilizzando la
teoria del frame analysis come metodologia di base e cercando di migliorare
il procedimento di analisi utilizzato in studi analoghi. La ricerca pone il
“frame” come principio organizzatore di base nella rappresentazione dei
migranti. Questi frames possono essere rintracciati attraverso l'analisi
clustering di dati testuali. Il primo scopo dello studio è quello di accertare

764

JADT’ 18

quanto posizione e tempistiche possano influenzare l’impiego di un frame
rispetto ad un altro, in che cosa consistano questi frames e dunque fornire
un’analisi dettagliata di come il processo migratorio venga descritto nella
stampa britannica. Si tratta di un focus fortemente mancante negli studi
basati sulla teoria del frame sino ad oggi. L’osservazione dimostra che,
durante i sopra citati due periodi di tempo, sono sei i frame che possono
essere identificati: si tratta di quattro di tipo tematico e due di tipo episodico.
I frame “negativo” e “positivo” riguardo i migranti si possono rintracciare
nel primo periodo, dal momento che il focus “locale” ha fornito un terreno
ideale per l'impiego degli stessi. I soggetti erano infatti situati in prossimità
del territorio ed erano dunque raffigurati come una minaccia. Al contrario, il
secondo periodo di tempo vede il prevalere del frame “positivo” riguardo ai
migranti, innescato dalla morte di Alan Kurdi e dal corrispondente
slittamento concettuale che ha portato alla rimozione “globale” del soggetto
dai confini immediatamente prossimi per ricollocarlo in un contesto più
ampio. Questo si è appaiato al sovrapporsi del frame della responsabilità
nazionale con il frame “positivo” riguardo ai migranti. Si può notare come i
due frames siano diventati profondamente interconnessi durante il secondo
periodo, proprio mentre si registrava l'insorgere del frame della
responsabilità europea. Ciò dimostra come l'egemonia di un singolo frame
possa essere sfidata, ma solo nel caso in cui la situazione corrispondente sia
“drastica” al punto da permetterlo.
Keywords: migration, frame analysis, cluster analysis, British media, text
mining
1. Introduction
1.1 Frame analysis and the migration crisis
Over the last two decades frame analysis has become an increasingly popular
tool for analysing the portrayal of a subject in the media, due to its ability to
demonstrate the latent and manifest meaning of the news and the recurring
themes and elements that exist in common between individual texts
(Zhongdang and Kosicki, 1993). According to Entman, ‘framing essentially
involves selection and salience. To frame is to select some aspects of a
perceived reality and make them more salient in a communicating text, in
such a way as to promote a particular problem definition, causal
interpretation, moral evaluation, and/or treatment recommendation for the
item described.’ (Entman 1993). A reality is presented to the audience, a
reality that can be considered a package of information of which the
constituent parts together form the frame being deployed (Gamson et al.
1983). One frame is distinguishable from another precisely because this
collective package is the sum of its parts. These parts are defined as framing

JADT’ 18

765

devices and reasoning devices, which can be discovered alongside one
another thereby indicating the presence of one frame or another. These
framing devices can consist of metaphors, visual images, lexical choices,
stereotypes, idioms etc. (Tankard et al. 2001) which in turn support reasoning
devices within the same frame which define the problem, assign
responsibility, pass judgement and present possible solutions (Entman 1993).
As a relatively new approach, and apart from the shared inheritance from
cognitive psychology (Bartlett 1932), anthropology (Bateson 1972) and the
seminal work of Erving Goffman (Goffman 1974), frame analysis remains a
fluid approach with a lack of empirical and methodological consistency
across studies. Some authors have even contended if the school in of itself
can even be considered a paradigm due to this diversity (D’angelo 2002:871;
Entman 1993:51). This paper is not concerned with this contention, but does
strive to arrive at a methodology which incorporates various elements of
previous techniques in order to arrive at a complimentary approach which in
turn minimises the criticism normally fired at more extreme approaches
deployed in the past due to their perceived rigidity and shortcomings.
To date very little frame analysis has been directed towards migration,
especially in the British context. Despite the migration crisis showing no
signs of abating, the response of Europe has generally been categorised by
two approaches; (i) strengthening internal and external borders to restrict
movement throughout Europe (ii) disrupting attempted crossings by means
of the Mediterranean. Britain is particularly interesting within this context,
not only as a state which has consistently tried to curb entry at an official
level, but also because of the media’s and public’s keen obsession with
migration which was ultimately exemplified in the Brexit referendum. The
media can be considered as central to this response. Whether one considers it
to be the embodiment of public opinion or of elite opinion, it is nonetheless
an incarnation of a country’s position and can be seen as acting as an arbiter
of said country’s opinion. The current migration crisis is as complex as it is
pressing, and the ‘reality’ presented by the media should not be seen as
natural, ready to be recorded and transmitted from one human being to
another, but rather as something that is constructed and then transmitted
according to constructivist theory (Goffman 1974). The media is therefore
able to set the agenda and frame the debate on the migration crisis, in turn
affecting the reality in the mind of the population and government.
This paper has two aims in mind. The first is to develop a methodology
which combines previous qualitative and quantitative approaches in order to
improve validity and reliability while the second is to use said methodology
to ascertain how migration is portrayed by the British media and how far this
portrayal is affected by factors such as time and geographical focus.

766

JADT’ 18

2. Methodology
The study’s methodology was constructed with historical criticisms directed
at frame analysis in mind; either that the process is too qualitative and
therefore lacks reliability, or that it is conducted too quantitively, and
therefore lacks reliability. The first step was to collect the data, which was
obtained manually from four daily British newspapers’ online archives (the
Daily Express, the Guardian, the Telegraph and the Daily Mail), and
included all newspaper articles which included ‘migration’, ‘migrant’,
‘refugee’ etc. in the title, or whose content largely dealt with such topics. The
two periods of investigation are 28th to 31st July 2015 and 2nd to 6th September
2015, these dates were chosen in order to ascertain whether frames could be
consistently identified across two periods, even in the short term, but also to
investigate whether dominant frames can be challenged if events are deemed
drastic enough (the tragic death of Alan Kurdi became the dominant news
story in the second period, whereas the first was primarily concerned with
the Calais crisis). In total 505 were gathered, 160 for the first period and 345
for the second.
The quantitative aspect of the study consists of a computer assisted
approach, by using cluster analysis to process the data and indicate the
presence of ‘frames’. Because, as mentioned above, framing is considered to
be the grouping and salience of certain elements to the neglect of others, one
can consider the cluster generated by a computer to precisely be a direct
indication of the presence of one frame or another, as words are the primary
form framing elements assume. The software used was the R program in
conjunction with the Iramuteq interface. The clustering method used is that
of Reinert (Reinert 1983), whose conception of clusters as a ‘cognitiveperceptive framework’ lends itself perfectly to frame analysis, concerned as it
is with discerning different representations of a perceived reality. The
second, more qualitative step of the study, was to conduct a deep read of all
the texts, where the researcher intuitively coded texts and created a frame
matrix which allowed an awareness of the context of the text as well as those
framing and reasoning devices which seemed re-occurring and therefore
significant. Combined, this allowed the reliability of the initial cluster
analysis generated by the computer to be complemented by the in depth
familiarity of the researcher, which provided a validity to the interpretation
of results.

JADT’ 18

767

3. Results

Figure 1. Cluster analysis for first period

Figure 2. Cluster analysis for the second period

The two cluster analyses seem to identify three distinct clusters, yet those
identified in the second period varying dramatically in respect to the first.

768

JADT’ 18

The first period under investigation generated three clusters, which have
been labled The Refugee Cluster (Red), The Migrant Cluster (Green) and the
Calais Crisis Cluster (Blue). However, the second period produced three
different clusters: Migration as a Domestic Issue Cluster (Red), Migration as a
European Issue Cluster (Green) and the Migrant Crisis Cluster (Blue). At first
glance these results seem to refute the basis of framing theory; that frames
are not produced by the journalist, but are deployed from the cultural
repertoire they cognitively hold in common with the rest of society (Goffman
1974). This is because, if framing theory is correct, then in the space of one
month it would be impossible for frames to mutate completely, and one
would expect the clusters identified in the first period to be identical to those
found in the second. However, if one makes a distinction between issuespecific and generic frames and episodic and thematic frames (de Vreese
2005) the two cluster groups are far more similar than first meets the eye.
For instance, the first period produced two frames which are predominantly
concerned with the figure of the migrant and two differing portrayals of the
migrant; the migrant as a helpless victim and the migrant as an opportunistic
individual. These are both clusters which one can consider thematic frames
as the clusters do not refer to one story but rather represent a thematic
perspective. The third frame, however, can be categorised as being an issue
specific frame, concerned as it is only with the Calais crisis, the ‘Jungle’ camp
and the stories of migrants attempting to enter the channel tunnel. The
second period, similarly, consists of two thematic frames (that which
considers migration as an issue for the British government and that which
considers it to belong to the realm of European governance) and one episodic
frame (those stories relating specifically to the death of Alan Kurdi and those
migrants attempting to move through Hungary and Austria in the early days
of September 2015). If the two episodic frames are laid aside, one is left with
four remaining; the ‘negative’ migrant frame, the ‘positive’ migrant frame, the
domestic responsibility frame and the European responsibility frame. What is
interesting to note in the second period, is that ‘positive’ migrant frame from
the first period does not disappear, but overlaps with and bolsters/is
bolstered by the the arising domestic responsibility frame. For example,
many of the key terms of the ‘positive’ migrant frame (vulnerable, refugee,
conflict, persecution, support, receive, community etc.) are emblematic of
those found in the so-called domestic responsibility frame (vulnerable,
refugee, sanctuary, hazardous, save, help etc.) This means that rather than
‘disappearing’, the frame which represents migrants as individuals in need
has been combined with arising domestic responsibility frame.
However, this does not account for the disappearance of the ‘negative’
migrant frame. The reason for this lack of presence, and likewise the merging

JADT’ 18

769

of the ‘positive’ migrant frame and the domestic responsibility frame in the
second period, is due to the shock events linked to the tragic death of Alan
Kurdi on September 2nd 2015. The event seems to have made the deployment
of the ‘negative’ migrant frame untenable in the second period, while at the
same time the ‘positive’ migrant frame persists as the period proved more
fertile for this perspective. This is one reason why the two frames overlapped
in the second period; the outrage and shock at the death of a toddler
ultimately led to the locating of the solution to the ‘positive’ migrant frame in
the domestic responsibility frame. Interestingly, this overlap did not occur
with the European responsibility frame, which may be due to British political
actors (the majority of those interviewed across the articles) actively
positioning themselves as ready to help migrants in order to show
themselves in a positive light.
Another interesting finding is how location affected or at least was linked to
the change in hegemony between the ‘positive’ and ‘negative’ migrant
frames. In the first period, the obsession with the Calais crisis (demonstrated
by the presence of the corresponding episodic frame) seemingly provided
conceptual ground in which the ‘negative’ migrant frame could flourish,
whereas in the second period, dominated as it was by news of the death of
Alan Kurdi (and the presence of a more international episodic frame)
ensured the continued presence of the ‘positive’ migrant frame. One reason
for this could be that as the migrant is located nearer to the British boarder,
the ‘negative’ migrant frame (characterised by terms such as arrest, siege,
repel, overwhelm) was more easily deployed due to the greater unease of
foreign migrants entering the country, whereas when the focus was
positioned more globally this unease was overcome by the moral shock of
Alan Kurdi’s death, lessening the unease and therefore the appropriateness
of the previous frame.
Despite demonstrating some continuity of frames across the two periods, that
geographical focus affects the deployment of one frame or another and that
shock events can seemingly shift the frames in play to a great extent, the
study is not without shortcomings. Firstly, the two time periods, and the
limitation of four days to each, has greatly reduced the data available. This in
turn makes it impossible to understand how far and how robust the
identified frames are across an extended period of time and whether other
frames come into play depending on the specific moment or the dominating
news story. One solution could be to extend the time frame, but this might in
turn lead to a drop in validity and insight due to the limitations of the
researcher to deal with the data to the same extent as a computer. The second
issue, as has already been mentioned, is determining precisely the
characteristics of one frame in relation to another. One possible solution

770

JADT’ 18

would be to predetermine those terms which are identified as framing
elements or reasoning devices as variables in the cluster analysis, which
would in turn limit the identification of episodic frames in favour of thematic
frames and over a longer period more clearly define the continuation, and
the fluctuation in presence, of identified frames. The drawback of this,
however, is that arguably the subjectivity of the researcher enters at too early
a stage and harms the validity of the methodology. A third point is that,
although the cluster analysis did capture many of the framing devices (as
they are commonly exhibited as words), it was unable to capture all (for
instance accompanying images) and was largely unable to identify the
presence of reasoning devices (as the unit of analysis needs to be bigger than
single word choice).
References
Bartlett, F. (1932). Remembering: A Study in Experimental and Social Psychology.
Cambridge University Press.
Bateson, G. (1972). Steps to an Ecology of Mind: Collected Essays in Anthropology,
Psychiatry, Evolution, and Epistemology. University of Chicago press.
D’Angelo, P. (2002). News Framing as a Multiparadigmatic Research
Program: A Response to Entman. Journal of Communication, 52(4): 870888.
Entman, R.M. (1993). Framing: Toward Clarification of a Fractured Paradigm.
Journal of Communication, 43(4): 51-58.
Gamson, William A. and Kathryn E. Lash. (1983). The Political Culture of
Social Welfare Policy. In S.E. Spiro and E. Yuchtman-Yaar, Evaluating the
Welfare State: Social and Political Perspectives. Academic Press.
Goffman, E. (1974). Frame analysis: An essay on the organization of experience.
Harper and Row.
Reinert, M. (1983). Une méthode de classification descendante hiérarchique:
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, 8(2): 187-198.
De Vreese, C.H. (2005). News Framing: Theory and Typology. Information
Design Journal and Document Design, 13(1): 51-62.
Zhongdang, P. and Kosicki G.M.. (1993). Framing Analysis: An approach to
news discourse, Political Communication, 10(1): 55-75.
Tankard, J.W. and Severin W.J. (2001). Communication Theories: Origins,
Methods and Uses in the Mass Media, 5th Edition. Pearson.

JADT’ 18

771

Results from two complementary textual analysis
software (Iramuteq and Tropes) to analyze social
representation of contaminated brownfields
Marjorie Tendero1, Cécile Bazart2
1

University of Rouen – CREAM and Agrocampus Ouest - marjorie.tendero@agrocampusouest.fr
2University of Montpellier, Montpellier – CEE-M - cecile.bazart@umontpellier.fr

Abstract
The aim of this paper is to demonstrate the complementarity of two types of
textual analysis software, Iramuteq and Tropes, to analyze a corpus of data
extracted from an open-ended question from a national cross-sectional
survey. Descendant hierarchical classification made with Iramuteq lead to
more homogeneous and less groups of discourse than the references fields
made with Tropes. References fields allow to reveal how the corpus’ thematic
are articulated made with Iramuteq.
Résumé
Cette communication présente l’apport complémentaire de deux logiciels
d’analyse de contenu, Iramuteq et Tropes, pour analyser les représentations
sociales à partir de réponses données à une question ouverte dans un
questionnaire d’enquête. Il montre que les classifications hiérarchiques
descendantes opérées à l’aide du logiciel Iramuteq peuvent être approfondies
de façon complémentaire à l’aide des classifications sémantiques par univers
de références et l’outil scénario du logiciel Tropes. Les classes de discours
sont moins nombreuses et plus homogènes que les univers de références mis
en évidence par logiciel Tropes. Ces derniers montrent l’articulation des
thématiques du corpus.
Keywords: Brownfield; Classifications; Iramuteq; textual data analysis;
Tropes.
1. Introduction
L’analyse de contenu regroupe les techniques permettant une analyse
systématique et objective des communications écrites et orales. Il s’agit d’une
approche multidisciplinaire croisant des méthodes quantitatives et
qualitatives, et dont les domaines d’application sont très nombreux : sciences
de la communication, sociologie, psychologie, informatique, et économie par
exemple. Ces techniques étudient la structure d’un texte, ou d’un discours,

772

JADT’ 18

ainsi que sa logique afin de mettre en évidence le contexte dans lequel il est
produit, et sa signification réelle à partir de données objectives. Ces méthodes
permettent de traiter les réponses à des questions ouvertes en soutenant
l’interprétation du phénomène étudié sur des critères quantitatifs et objectifs
(Garnier and Guérin-Pace 2010). Pour analyser les réponses données à des
questions ouvertes, un des avantages de ces méthodes et d’éviter les biais liés
à la codification thématique a posteriori. Toutefois, cette méthode fait l’objet
de critiques. Ces dernières sont relatives aux étapes à mettre en place pour
préparer le corpus, pour effectuer les analyses, et interpréter les résultats.
Ainsi, lors de la phase de préparation du corpus, une lemmatisation peut être
effectuée. Or, celle-ci regroupe parfois des formes dont l’emploi, dans un
contexte donné, mène à des contresens (Lemaire 2008). C’est le cas lorsqu’une
forme au pluriel est lemmatisée au singulier. De plus, les dictionnaires des
expressions utilisés par les logiciels peuvent ne pas rendre compte des
marqueurs de modalités comme la négation (Fallery and Rodhain 2007). Par
ailleurs, des différences interprétées en termes d’analyse de contenu peuvent
en réalité provenir de différences sociales dans la façon dont un individu
s’exprime à l’oral ou à l’écrit. Les problèmes d’homonymies, de polysémies,
de synonymies peuvent donc amener à construire des classes lexicales
différentes alors qu’elles relèvent de modes d’expression hétérogènes sur la
forme mais en réalité très similaire sur le fond ; ce qui est le cas des opinions
exprimées par des périphrases, des paraphrases ou des ellipses. Une
attention particulière doit donc être portée au traitement des ambiguïtés afin
d’éviter toute erreur d’interprétation. Pour cette raison, il est intéressant de
combiner deux approches complémentaires, et donc différents logiciels,
d’analyse de contenu ; ce qui permet d’assurer la validité des résultats
(Vander Putten and Nolen 2010; Lejeune 2017). C’est par exemple ce qui a été
fait sur un corpus d’entretien pour comparer les logiciels Nvivo et
Wordmapper (Peyrat-Guillard 2006). Dans cette communication nous
soulignons l’apport complémentaire des logiciels Iramuteq et Tropes pour
l’analyse des représentations sociales associées aux friches polluées à partir
des réponses données à une question ouverte dans le cadre d’une enquête
administrée au niveau national auprès de 803 individus résidant sur une
commune impactée par ce type de foncier. Nous présentons dans la section
qui suit la méthodologie adoptée, les données récoltées et les analyses
effectuées. Dans une troisième section, nous présentons les résultats obtenus
à l’aide du logiciel Iramuteq ; puis ceux obtenus à partir du logiciel Tropes
dans une quatrième section. Nous discutons des apports complémentaires de
ces deux logiciels pour l’étude des représentations sociales à partir de
l’analyse des réponses données à une question ouverte dans une dernière
section.

JADT’ 18

773

2. Méthodologie
Nous avons élaboré un questionnaire afin d’étudier la perception
individuelle vis-à-vis du risque de pollution du sol, et les représentations, et
perceptions relatives aux friches urbaines et à leur reconversion. Le
questionnaire a été administré aux riverains résidant sur les communes
impactées par une friche polluée1. Au total, 803 réponses complètes ont été
collectées sur 503 communes impactées par la présence d'une friche polluée.
Pour analyser les représentations sociales, associées aux friches polluées,
nous avons utilisé la question ouverte suivante : « à quoi associez-vous
l’expression de friches urbaines ? ». Nous avons procédé à une analyse de
données textuelles car cette technique d’analyse des données se prête
particulièrement bien à l’étude des représentations, individuelles ou sociales,
en rendant compte de la dynamique représentationnelle et cognitive d’un
phénomène (Abric 2003; Beaudouin and Lahlou 1993; Kalampalikis 2005;
Negura 2006).
Toutes les questions étaient obligatoires. Cependant, tous les participants
n’ont pas réussi à y répondre : certaines réponses n’étaient qu’une suite de
caractères permettant de passer à la question suivante. De plus, cette
question ouverte se situait dans la seconde partie du questionnaire. Ce
dernier était relativement long ; il en a résulté une perte d’attrition. Nous
avons donc écarté ces réponses de notre analyse. Au total, 539 réponses ont
pu être conservées ; soit 67,12 % des réponses collectées.
Les données ont été formatées pour pouvoir être analysées à partir du
logiciel IRaMuteQ (Interface de R pour les analyses multidimensionnelles de
textes et de questionnaires) version 0.7 alpha 2 dans un premier temps. C’est
un logiciel libre développé par Pierre Ratinaud au sein du LERASS
(Laboratoire d’Études et de Recherche Appliquées en Sciences Sociales)
distribué sous les termes de la licence GNU GPL (v2) (Baril and Garnier 2015;
Ratinaud and Déjean 2009). Le tableau 1 ci-dessous montre un extrait des
réponses analysées.
Tableau 1 : Extrait du corpus analysé
0001

percept_eleve
danger_oui
confiance_non
Abandonnée, sale, nuisible
0002
percept_eleve
danger_oui

affecte_non
prevent_non
exist_non
gestfri_non
sexe_h age_4059 reg_centre

gestion_non
intention_oui

affecte_non
prevent_oui
gestion_non
exist_oui gestfri_non
intention_oui
confiance_oui

1 Ces communes ont été identifiées à partir d’une extraction de la base de
données BASOL sur les sites et sols pollués (ou potentiellement pollués) appelant une
action des pouvoirs publics, à titre préventif ou curatif.

774

JADT’ 18

sexe_f
age_1924 reg_als
Zones non_habité
0003
percept_eleve
affecte_non
prevent_non
gestion_non
danger_non
exist_non
gestfri_non
intention_non
confiance_non
sexe_f
age_4059 reg_als
Un jardin en ville, laissé à l'abandon.
0004
percept_moyen affecte_non prevent_non
gestion_non
danger_non
exist_non
gestfri_non
intention_oui
confiance_non
sexe_f
age_4059 reg_rha
zone abandonnée, zone polluée ville

Le corpus de texte analysé a les caractéristiques décrites dans le tableau cidessous.
Tableau 2 : Statistiques descriptives associées au corpus analysé

Nombre de réponses
Nombre de mots (occurrences)
Nombre moyen de mots utilisés
Nombre de formes actives (total)
Nombre de formes supplémentaires (total)
Nombre d’hapax
Nombre de formes
Nombre de formes actives (différentes)
Nombre de formes supplémentaires (différentes)

Corpus « friche »
539
2 177
4,04
1 537
640
275
482
402
80

Nous comparons les analyses suivantes : statistiques descriptives et
classification hiérarchique descendante effectuée à l’aide du logiciel Iramuteq
et univers de références et scénario à l’aide du logiciel Tropes. Il s’agit d’un
logiciel d’analyse sémantique de textes créé en 1994 par Pierre Molette et
Agnès Landré à partir des travaux de Rodolphe Ghiglione sur l’analyse
propositionnelle de discours (Molette, Landré, and Ghiglione 2013).
3. Résultats de l’analyse avec Iramuteq
3.1. Statistiques descriptives
Le tableau ci-dessous décrit les termes les plus fréquemment employés par
les individus (effectif ≥ 20) lorsqu’ils évoquent les friches polluées. Ces
dernières sont des « terrains » (99 occurrences), des « zones » (36) laissées à
« l’abandon » (106). Il s’agit de terrains sur lesquels étaient implantées
d’anciennes « usines » (29) aujourd’hui « désaffectées » (17).

JADT’ 18

775

Tableau 3 : Termes les plus fréquemment employés (statistiques descriptives à partir du
logiciel Iramuteq)

Formes
actives
Abandon
Terrain
Laisser
Abandonner
Ville
Zone
Terrain
vague

Effectif

Type

106
99
63
49
46
36
34

Nom
Nom
Verbe
Verbe
Nom
Nom
Nom

Forme
active
Usine
Pollution
Ancien
Espace
Bâtiment
Sol

Effectif

Type

29
28
28
25
25
20

Nom
Nom
Adjectif
Nom
Nom
Nom

3.2. Classification hiérarchique descendante
65.49 % des réponses données sont classifiées au sein de quatre catégories. Le
tableau 4 ci-après indique la significativité des termes associés à chaque
classe. La première classe regroupe les termes faisant référence aux anciennes
activités industrielles. La deuxième classe renvoie aux problème de la gestion
de déchets en milieu urbain en évoquant les « décharges », les « saletés », et
la « pollution ». La troisième classe correspond aux termes caractérisant ce
type d’espace. La quatrième classe, quant à elle, fait référence aux espaces de
nature auxquels les friches correspondent, en particulier dans le cas de
parcelles agricoles laissées en jachère.
4. Résultats complémentaires apportés par Tropes
Nous avons formaté le corpus pour l’analyser avec le logiciel Tropes.
L’analyse des univers de références nous permet de mettre en évidence les
principaux thèmes utilisés dans le texte en regroupant les termes dans des
classes d’équivalent sémantiques. Le tableau 4 ci-après présente les résultats
obtenus par les univers de références à l’aide du logiciel Tropes.
Les classifications sont données par ordre décroissant et indiquent le nombre
de termes qui s’y rapportent. Ces classifications ne permettent pas toujours
de couvrir l’ensemble des termes utilisés dans le corpus : seuls les substantifs
les plus significatifs du texte y apparaissent. Il est toutefois possible de
paramétrer ces classifications à partir du mode scénario du logiciel ; la figure
1 en montre un extrait.
5. Discussion et conclusion
Le tableau 6 précise les avantages et contraintes respectifs liés à l’utilisation
de ces deux logiciels pour analyser les représentations sociales des friches
polluées. En particulier, la classification sémantique par univers de références

776

JADT’ 18

et l’outil scénario font apparaître des classes plus nombreuses et moins
homogènes que dans le cas de la classification hiérarchique descendante
effectuée sous Iramuteq.
Tableau 4 Résultats de la classification hiérarchique descendante à partir du logiciel Iramuteq
Classe 1 (39,7 %)

Classe 2 (15 %)

Anciennes activités industrielles

Problèmes de gestion
déchets en milieu urbain

Forme active

²

p

Abandonner

58,95

Usine

42,73

<
0,0001
<
0,0001

Ancien

Bâtiment

Industriel

Polluer

Désaffecté

Site

29,1

28,82

22,13

17,24

15,66

15,66

Immeuble

14,05

Forme
active
Pollution

²

p

151,38

Sol

59,79

<
0,0001
<
0,0001
<
0,0001

Laisser

<
0,0001

Friche

<
0,0001

Milieu_urbain

0,00073

Sauvage

<
0,0001

Ville

<
0,0001

Repos

<
0,0001

Désert

<
0,0001

Saleté

<
0,0001

Décharge

<
0,0001

Terre

0,00017

Culture

Zone

13,72

0,00021

Industrie

10,9

0,00096

Lieu

10,87

0,0023

Non_construit

9,29

0,00513

Endroit

7,83

0,00547

Vieux

7,72

0,00547

des

Classe 3 (33,7
%)
Zone
abandonnée et
inutilisée
Forme active

32,66

17,13

17,13

11,41

Classe 4 (11,6 %)
Espace
jachère

agricole

en

²

p

Terrain

107,94

< 0,0001

Forme
active
Espace

Abandon

84,27

< 0,0001

Nature

²

p

114
.47

<
0,0001
<
0,0001

62.
29
82,57

< 0,0001

Vert
46.
45

16,1

< 0,0001

Libre
41.
31

12,58

10,6

0,00038

0,00113

Non_exp
loité

38.
60

Champ
30.
79

Non_cultivé

8,25

7,85

7,85

4,06

0,00408

0,00507

Aller

5,95

Non_utilisé

2 ,97

0,01471

NS
(0,08500)

0,00507

Non_ent
retenu
Rntreten
ir
Non_cul
tivé

24.
89
5.8
1
2.7
1

<
0,0001
<
0,0001
<
0,0001
<
0,0001
<
0,0001
0,0159
6
NS
(0,099
62)

0,04402

Tableau 5 : Principaux univers de références associés au corpus
Univers de références 1
Référence
Eff.
Exemple de termes
associés
Ville
74
Ville, taudis, zone
urbaine
Lieu
59
Zone
Habitat
55
Bâtiments, immeubles,
logement, appartements

Référence
Ville
Lieu
Industrie

Univers de références 2
Eff.
Exemple de termes
associés
73
Ville, taudis, milieu urbain,
zone urbaine
59
Site, zone, lieu
50
Industrie, zone industrielle,
usines

JADT’ 18

777

Industrie

50

Immeuble

36

Bâtiments, immeuble

39

Zone industrielle,
industrie, usine
Pollution, dépotoir

Écologie

Pollution

33

33
22
22
20

Végétation, herbe, ronce
Déchet, détritus
Jachère, cultures
Terre

Déchet
Agriculture
Terre

22
21
20

Polluant, pollution,
dépotoir
Déchet, détritus
Jachère, cultures
Sols, terre

Plantes
Déchet
Agriculture
Terre

Figure 1 : Extrait des scénarios sous Tropes (ordre croissant)

Cet outil permet d’approfondir et de valider l’interprétation effectuée à partir
de la classification hiérarchique descendante à l’aide du logiciel Iramuteq.
Ces deux logiciels apparaissent donc comme complémentaires. Ces
complémentarités restent toutefois à vérifier à l’aide d’autres type de corpus
(entretiens par exemple). Enfin, pour étudier les représentations sociales de
friches polluées auprès de populations impactées par ce type de site, il serait
intéressant d’identifier le lexique émotionnel et affectif utilisée à l’aide
d’EMOTAIX par exemple (Piolat and Bannour 2009). En effet, cela
permettrait de mieux identifier la dimension affective dans les intentions
comportementales à l’égard de ce type de site.

778

JADT’ 18

Tableau 6 : comparaison des fonctionnalités d'Iramuteq et de Tropes pour l'analyse
des représentations sociales
Logiciels
Procédures
Découpage du
texte
Style du texte
Mise en scène
Épisodes et rafales
Classifications
Scénario
Statistiques
descriptives
Analyse de
similitude
Analyse de
spécificité et
analyse factorielle
des
correspondances
Analyse
prototypique
Principaux atout
pour l’étude des
représentation
sociales
Principaux
inconvénients pour
l’étude des
représentations
sociales

Iramuteq

Tropes

Segments de texte

Propositions canoniques

Classification hiérarchique
descendante






Univers de références


Indirectement par mots avec des
graphes en aire ou étoilé




Richesse des analyses et des
résultats

Formatage des corpus moins
contraignant

Formatage des corpus longs

Lemmatisation et classification
automatisées aboutissent à des
résultats peu lisible

References
Abric, Jean-Claude. 2003. Méthodes D’étude Des Représentations Sociales. ERES.
Baril, Élodie, and Bénédicte Garnier. 2015. ‘Utilisation d’un outil de
statistiques textuelles : IRaMuteQ 0.7 alpha 2. Interface de R pour les
analyses multidimensionnelles de textes et de questionnaires’. Institut
National d’Études Démographique.
Beaudouin, V, and S Lahlou. 1993. ‘L’analyse Lexicale : Outil D’exploration
Des Représentations’. Cahier de Recherche C (48): 25–92.
Fallery, Bernard, and Florence Rodhain. 2007. ‘Quatre approches pour
l’analyse de données textuelles :lexicale, linguistique, cognitive,
thématique’. In XVIème Conférence de l’Association Internationale de
Management Stratégique. Montréal, Canada.
Garnier, Bénédicte, and France Guérin-Pace. 2010. Appliquer les méthodes de la
statistique textuelle. Les collections du CEPED (Centre Population et

JADT’ 18

779

Développement). Paris: CEPED.
Kalampalikis, Nikos. 2005. ‘L’apport de la méthode Alceste dans l’analyse
des représentations sociales’. In Méthodes d’étude des représentations
sociales, edited by Jean-Claude Abric, 147–63. Hors collection. ERES.
Lejeune, Christophe. 2017. ‘Analyser Les Contenus, Les Discours, Ou Les
Vécus ? À Chaque Méthode Ses Logiciels !’ In Les Méthodes Qualitatives
En Psychologie et Sciences Humaines de La Santé, Dunod, 203–24. Psycho
Sup.
Lemaire, Benoît. 2008. ‘Limites de La Lemmatisation Pour L’extraction de
Significations’. In 9ème Journées Internationales d’Analyse Statistique Des
Données Textuelles, 725–32. Lyon, France.
Molette, Pierre, Agnès Landré, and Rodolphe Ghiglione. 2013. Tropes. Version
8.4. Manuel de référence. http://tropes.fr/doc.htm.
Negura, Lilian. 2006. ‘L’analyse de Contenu Dans L’étude Des
Représentations Sociales’. SociologieS Théories et recherches (October).
Peyrat-Guillard, Dominique. 2006. ‘Alceste et WordMapper : L’apport
Complémentaire de Deux Logiciels Pour Analyser Un Même Corpus
D’entretien’. In Journées d’Analyse Statistique Des Données Textuelles, 725–
36. Besançon, France.
Piolat, Annie, and Rachid Bannour. 2009. ‘EMOTAIX : Un Scénario de Tropes
Pour L’identification Automatisée Du Lexique Émotionnel et Affectif’.
L’Année Psychologique 109 (04): 655. https://doi.org/10.4074/S00035033
09004047.
Ratinaud, Pierre, and Sébastien Déjean. 2009. ‘IRaMuTeQ: Implémentation de
La Méthode ALCESTE D’analyse de Texte Dans Un Logiciel Libre’.
Modélisation Appliquée Aux Sciences Humaines et Sociales MASHS, 8–9.
Vander Putten, Jim, and Amanda L Nolen. 2010. ‘Comparing Results from
Constant Comparative and Computer Software Methods: A Reflection
About Qualitative Data Analysis’. Journal of Ethnographic and Qualitative
Research 5: 99–112.
Remerciements
Nous remercions Jean-Marc Rousselle pour avoir administré en ligne ce
questionnaire sous Limesurvey. Cette enquête a bénéficié du soutien
financier du SRUM 2015, de l’université de Montpellier, du CEE-M
(LAMETA), de l’ADEME, de la Région Pays-de-la-Loire, et du CREAM
(Université de Rouen).

780

JADT’ 18

Multilingual Sentiment Analysis
Matteo Testi1, Andrea Mercuri1,2, Francesco Pugliese1,3
Deep Learning Italia – m.testi@deeplearningitalia.com
2Tozzi Institute – a.mercuri@deeplearningitalia.com
3Italian National Institute of Statistics – francesco.pugliese@istat.it
1

Abstract
In recent years, Sentiment Analysis (SA) has attracted significant attention in
different areas of Research and Business. This is because “sentiments” can
influence opinions of product vendors, politicians and the public opinion.
The sentiments of users are generally categorised into three classes: negative,
positive or neutral. Lately, more and more Deep Learning (DL) models have
been employed to SA thanks to their automatic high-dimensional feature
extraction capability. However, DL supervised models are greedy of data
and the shortage of sentiment’s data sets in specific languages (other than
English) is a big issue. In order to address this multilingual issue of training
sets we propose a very deep Recurrent Convolutional Neural Network
model (RCNN) which achieves “state-of-art” accuracy in sentiment
classification. Extracting keywords from the final max-pooling layer we are
able to create a corpus of domain-specific keywords. By exploiting these
“discriminative” extracted words we scrape a long sequence of sentences (in
two different languages) in order to feed a Neural Machine Translation
model. A sequence-to-sequence model with attention and beam-search has
been implemented to translate one language sentences (i.e. English) into
another language sentences (i.e. Italian). As example, we train our RCNN on
an English twitter sentiment training-set and extract keywords to generate
the machine translation model. During the test stage, we translate our test
sentences (i.e tweets) into another language for which we have poor training
set (i.e. Italian). Results highlight a significant accuracy gain of this technique
with regard to a model exclusively trained on a poor training set expressed in
a language different from English.
Keywords: sentiment, analysis, multilingual, deep, learning, recurrent,
convolutional, neural, machine, translation
1. Introduction
In recent years, Sentiment Analysis (SA) has attracted significant attention in
different areas of Research and Business. This is mainly due to the fact that
“sentiments” (which are exhibited on the web by users) can affect opinions of
product vendors, politicians and readers in general, namely the public

JADT’ 18

781

opinion. According to one of the most accredited definitions: Sentiment
Analysis is the field of study that analyses people’s opinions, sentiments,
evaluations, appraisals, attitudes, and emotions towards entities such as
products, services, organisations, individuals, issues, events, topics, and their
attributes (Qurat Tul Ain et al, 2017; Liu, 2012). This user point of view may
usually be expressed under the unstructured form of an opinion, review,
news, disapproval, etc. The rising demand of SA comes from the need of
summarising a general direction of user opinions from social media
(Haenlein et Kaplan, 2010). In fact, the aggregate data from Sentiment
Analysis can represent a valuable information in order to orient decisions in
politics, digital marketing or finance. Therefore, SA arises as a
multidisciplinary field joining computational linguistics, information
retrieval, semantics, natural language processing and artificial intelligence in
general (Aydogan et Akcayol, 2016). Ultimately, SA can be seen as the
process of automatically categorise utterances into three different classes:
negative, positive or neutral. Generally these sequences of text or sentences
come from social networks, opinion web-sites, e-commerce feedbacks, etc.
Twitter is one of the most useful microblogging platforms for Sentiment
Analysis and Opinion Mining since it offers very good API to download
tweets and it is very popular amongst different categories of people (Pak et
Paroubek, 2010). Traditionally, SA is a text classification problem and relies
on two kinds of approaches: a) “lexicon-based” which is usually applied to
problems without a training set. This technique generally makes use of a
fixed number of keywords to orient the classification process by means of
decision trees such as k-Nearest Neighbours (k-NN) or Hidden Markov
Model (HMM); b) “machine learning-based” where extracted features typically
consist of Parts of Speech (POS) tags, n-grams, bi-grams, uni-grams and bagof-words. Classification can be performed by Naïve Bayes or Support Vector
Machines (SVMs) (Singh et al., 2016). Traditional lexicon-based approaches
are not effective anymore in combination with the modern textual Big Data
corpuses, especially as far as sentiment concerns. On the other hand, Machine
learning approach can be supervised and unsupervised (less common) and it
is a methodology able of automation over enormous corpus of data, this is a
critical requirement for a reliable Sentiment Analysis. Deep Learning is a
branch of Machine Learning proposed by G.E. Hinton in 2006 and adopts
Deep Neural Network for text classification (Hinton et Salakhutdinov, 2006).
Deep Learning enhance traditional neural networks introducing more than
thousands of neurons, millions of connections, new regularisation techniques
(dropout, data augmentation, batch normalisation), new pre-processing
(skip-gram, word embeddings, etc) and different new models both
supervised and unsupervised: Convolutional Neural Networks (CNN)

782

JADT’ 18

(Krizhevsky et al., 2012), Deep Belief Networks (DBN) (Hinton et al., 2006).
and many more. Lately, more and more Deep Learning (DL) models have
been employed to SA thanks to their automatic high-dimensional feature
extraction capability (Vateekul and Koomsubha, 2016). For instance, in
Financial Sentiment Analysis (FinTech), Deep Learning has contributed to
investigate how to harness different media and financial resources in order to
improve the accuracy of stock price forecasting (Day et Lee, 2016). The
experimental results show how news sentiment categorisation, by means of
Deep Neural Networks, has different effects to investors and their
investments. However, SA is a challenging field due to the lack of supervised
data and to the nature inherently subjective of sentiments. In this work we
tackle one of the biggest problems for modern machine learning-based
Sentiment Analysis: the shortage of data sets in specific less common
languages (Italian, German, etc.). In order to address the classification of
sentiments we examined some of “state-of-art” text classifiers: many deep
learning models have been employed in Sentiment Analysis previously, such
as those invented by Stanford University: Recursive Neural Networks
(RNNs) (Socher et al., 2011b) and Recursive Neural Tensor Networks
(RNTNs) (Socher et al., 2013). Furthermore, Stanford released the Sentiment
Treebank that is the first corpus with fully labeled parse trees to train RNTSs.
RNTNs reach an accuracy ranging from 80% up to 85.4% on a Sentiment
Treebank’s test set. Although Recursive Models are very efficient in terms of
constructing sentences’ sentiment representations, their performance heavily
depends on the performance of the textual tree construction. Constructing
such a textual tree exhibits a time complexity of at least O(n^2), where n is
the length of the text. For this reason, we decided to make use of a Recurrent
Convolutional Neural Network model (RCNN) (Lai et al., 2015) achieving a
rather competitive accuracy in sentiment classification with regard to
Recursive Models. RCNNs exploit a recurrent structure to capture contextual
information as much as possible when learning word representations, which
may introduce considerably less noise compared to traditional windowbased neural networks. Moreover, the benefit of exhibiting a time complexity
of O(n) is a big added-value of RCNNs. To provide the support to a
Multilingual Sentiment Analysis, a Neural Machine Translation (NMT)
model has been employed in order to translate one language sentence (i.e.
English) into another language sentence (i.e. Italian). Basically, a NMT model
is a Neural Network structured in an encoder-decoder pattern which turned
out as a competitive alternative to the traditional Statistical Machine
Translation (SMT). The encoder consists of two independent recurrent
networks: “forward” which reads the the sentence in the natural order and
“backward“ which reads the sentence in reverse order. Instead, the decoder is

JADT’ 18

783

an RNN capable to compose the sentence to be translated. This sequence-tosequence model can be trained on a training set made of pairs of sentences:
the first is expressed into the source language and the second into the target
language (Cho et al., 2014).
2. Materials and Methods
The novelty of our Recurrent Convolutional Neural Network, with respect to
the original paper, is that we introduced two new recurrent models called
Long Short Term Memories (LSTM) instead of simple RNNs. These two
LSTM bi-directionally scan the text. The topology of the RCNN (see Fig. 1) is
intentionally designed to capture the context of each word (see the original
paper for further details). The RCNN has been trained on a corpus of 1.6
million tweets composed from various Semeval training-sets (Strapparava et
Mihalcea, 2007) and divided into positives (800k) and negatives (800k). To
input textual sequences into the neural network we insert a pre-trained
embedding layer on top (Mikolov et al, 2013). The embedding layer, which
has been pre-trained on an English Wikipedia Corpus, transforms indexed
words into numerical vector. Embedding vectors are characterised by a
semantical relationship amongst them according a chosen metrics, a cosine
distance in this case. Size of embedding vecotrs is 300.

Figure 1. The structure of the RCNN scanning the sentence “A sunset stroll along the South
Bank affords an array of stunning vantage points” (Lai et al., 2015).

During the training stage, the RCNN achieves 84% of accuracy on a
validation set (selected at the 20% of the original dataset). On a test set of 380
tweets (provided by Semeval), the model returns around 82% of accuracy on
positive tweets and 78% of accuracy on negatives, with an approximative

784

JADT’ 18

80% overall on a mixed tweets set. We followed recomended settings within
the original paper for the hyper-parameters selection.
Finally, we have modified the RCNN in order to extract the most significant
keywords that are specific for the model to drive the sentiment classification.
Basically, the third layer, that is the max-pooling layer, relies on an elementwise “max” function as follows :

The most "discriminative” words for the sentiment classification are those
most frequently selected in the max-pooling layer. Hence, we extracted the
indices of words corresponding to the max values of activation identified
within the third layer. During the training we determined 3.2 millions of
keywords, namely 2 for each tweet, the most important and the second in
order of signinifancy. Many of the resulting keywords come duplicated or
altered for multiple reasons : they might belong to a common slang or
undergo typing errors. Then, we removed doubles and we matched the rest
with the embedding corpus cointaining 2.5 mln words of the English
Language. This process turned out with 85,000 correct english keywords. By
exploiting these keywords as seed, we scraped a long sequence of sentences
in English from a website of Contextual Translations such as “Reverso
Context” (context.reverso.net) and its Italian translation in many different
form of expression. This stage led to a training set of 800,000 pairs of
sentences English-Italian and a Validation set of 50,000 pairs. A multi-level
sequence-to-sequence model with attention and beam-search has been
implemented to be trained on the training set of pairs (see Fig. 2) (Bahdanau
et al., 2014; Luong et Manning, 2016).

Figure 2. Multiple levels encoder-decoder (Luong et Manning, 2016).

JADT’ 18

785

“Attention-based” models enable the decoder to “focus” specifically on some
words rather than others, selectively orienting towards a more efficient
combination of words within the destination language sentences (Chorowski,
et al., 2015). “Beam search” is a greedy algorithm maximising the probability
of the ouput words (Britz et al., 2017). The NMT model was trained with an
embedding matrix randomly initialised and trained within the same process.
Embedding vectors size was 512. Both encoder and decoder are made of two
LSTM cells with an hidden state size equal to 512. Training algorithm was the
Stochastic Gradient Descent (SGD) with 32 sized batches; initial learning rate
of 1 and a decay factor of 0.5 starting from the 5-th epoch, plus early stopping
to reduce the overfitting. Beam search amplitude has been set to 5. In Fig.3
they are reported some resulting translations from Italian to English, on a test
example.

Figure 3. Some translations from Italian to English by means of the neural model trained by
us.

In the same time, we have trained the RCNN model on the most popular
Italian Sentiment Polarity Training set of tweets called SentiPolc 2016
(Barbieri et al., 2016). which is made of 7,000 annotated tweets and 300 test
tweets. In this case (Italian language) our model reaches 45% of validation set
accuracy and 43% on test set. For the embedding layer we have adpoted a
pre-trained language model on an Italian Wikipedia Embedding Corpus.
3. Results
We have tested the English RCNN model on the same italian SENTIPOLC
2016 test-set translated into English by our neural machine translation model.
Results highlight a boost of performance : 78% of accuracy on the test set
versus the 43% of the Italian trained RCNN model proving our strategy of
stacking NMT and RCNN models is successful.
4. Conclusion
Despite of the imperfections of the Neural Machine Translation producing
translations with some errors, the RCNN is tolerant to minimal errors and is

786

JADT’ 18

able to hold the accuracy to high levels on a test set. This is because RCNN
was previously trained on a solid and huge English corpus of tweets. This
entire process of keywords extraction, specifically to the task of sentiment
classification from the training set, is a fully novel approach to tackle the
problem of the lack of Sentiment training sets in other languages. Keywords
allow generating a domain-specific training set for the Neural Machine
Translation. Arguably, we believe this way of stacking NMT and RCNN lead
to a cutting-edge Multilingual Sentiment Classifier that can benefit other
fields of Text Classification in future. Future directions might be towards a
closer integration of NMT and Text Classifier and a reduction of translation
errors.
References
Qurat Tul Ain, Mubashir Ali, Amna Riaz, Amna Noureen, Muhammad
Kamran, Babar Hayat and A. Rehman (2017). Sentiment Analysis Using
Deep Learning Techniques: A Review. International Journal of Advanced
Computer Science and Applications (ijacsa).
Haenlein, M., and Kaplan, A. M. (2010). An empirical analysis of attitudinal and
behavioral reactions toward the abandonment of unprofitable customer
relationships. J. Relatsh. Mark.
Aydogan, E. and Akcayol, M. A. (2016). A comprehensive survey for sentiment
analysis tasks using machine learning techniques. Int. Symp. Innov.
Liu, B. (2012). Sentiment analysis and opinion mining (synthesis lectures on human
language technologies). Morgan & Claypool Publishers.
Pak, A., and Paroubek, P. (2010, May). Twitter as a corpus for sentiment analysis
and opinion mining. In LREc (Vol. 10, No. 2010).
Singh, J., Singh, G., and Singh, R. (2016) A review of sentiment analysis
techniques for opinionated web text. CSI Trans. ICT.
Hinton, G. E., and Salakhutdinov, R. R. (2006). Reducing the dimensionality of
data with neural networks. science, 313(5786), 504-507.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification
with deep convolutional neural networks. In Advances in neural
information processing systems (pp. 1097-1105).
Hinton, G. E., Osindero, S., and Teh, Y. W. (2006). A fast learning algorithm for
deep belief nets. Neural computation. 18(7), 1527-1554.
Vateekul, P., and Koomsubha, T. (2016, July). A study of sentiment analysis
using deep learning techniques on Thai Twitter data. In Computer Science
and Software Engineering (JCSSE), 2016 13th International Joint
Conference on (pp. 1-6). IEEE.
Day. M., and Lee C. (2016) Deep Learning for Financial Sentiment Analysis on
Finance News Providers. no. 1, pp. 11271134.

JADT’ 18

787

Socher, R., Pennington, J., Huang, E. H., Ng, A. Y., and Manning, C. D. 2011b.
Semi-supervised recursive autoencoders for predicting sentiment distributions.
In EMNLP, 151–161.
Socher, R., Perelygin, A., Wu, J. Y., Chuang, J., Manning, C. D., Ng, A. Y., and
Potts, C. 2013. Recursive deep models for semantic compositionality
over a sentiment treebank. In EMNLP, 1631–1642.
Lai, S., Xu, L., Liu, K., and Zhao, J. (2015). Recurrent Convolutional Neural
Networks for Text Classification. In AAAI.
Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F.,
Schwenk, H., and Bengio, Y. (2014). Learning phrase representations using
RNN encoder-decoder for statistical machine translation. arXiv preprint
arXiv:1406.1078.
Strapparava, C., and Mihalcea, R. (2007, June). Semeval-2007 task 14: Affective
text. In Proceedings of the 4th International Workshop on Semantic
Evaluations (pp. 70-74). Association for Computational Linguistics.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Efficient estimation of
word representations in vector space. arXiv preprint arXiv:1301.3781.
Bahdanau, D., Cho, K., and Bengio, Y. (2014). Neural machine translation by
jointly learning to align and translate. arXiv preprint arXiv:1409.0473.
Luong, M. T., and Manning, C. D. (2016). Achieving open vocabulary neural
machine translation with hybrid word-character models. arXiv preprint
arXiv:1604.00788.
Chorowski, J. K., Bahdanau, D., Serdyuk, D., Cho, K., and Bengio, Y. (2015).
Attention-based models for speech recognition. In Advances in Neural
Information Processing Systems (pp. 577-585).
Britz, D., Goldie, A., Luong, T., and Le, Q. (2017). Massive exploration of neural
machine translation architectures. arXiv preprint arXiv:1703.03906.
Barbieri, F., Basile, V., Croce, D., Nissim, M., Novielli, N., and Patti, V. (2016,
December). Overview of the EVALITA 2016 SENTiment POLarity
Classification Task. In Proceedings of Third Italian Conference on
Computational Linguistics (CLiC-it 2016) & Fifth Evaluation Campaign
of Natural Language Processing and Speech Tools for Italian. Final
Workshop (EVALITA 2016).

788

JADT’ 18

A linguistic analysis of the image of immigrants’
gender in Spanish newspapers
Juan Martínez Torvisco
Universidad de la Laguna – jtorvisc@ull.edu.es

Abstract 1 (in English)
The phenomenon of immigration has been studied from diverse perspectives
is important to understand that immigration is a fact associated with times of
crisis. The reason for the avalanche of immigrants to the Canary Islands
(Spain) is because it is the gateway to Europe, and therefore, immigrants
want to enter from this point. This research arises from the need to
linguistically determine the treatment of the phenomenon of immigration in
the Spanish press as a result of the arrival of thousands of foreign citizens to
the coast of the Canary Islands in 2006 and in 2015. It attempts to analyse
four Spanish newspapers using Iramuteq qualitative analysis software, two
from the Canary Islands (El Día and Canarias 7) and two Spanish national
newspapers (El País and ABC). Also, we wanted to know how it is the
informative treatment of gender. Our hypothesis is that the word male
(immigrant) appear more than woman and on the contrary woman (refugee)
has a higher frequency than male. Results are presented on a dendogram
figures.
Abstract 2 (in Spanish)
El fenómeno de la inmigración se ha estudiado desde diversas perspectivas,
y es un hecho asociado a tiempos de crisis. El motivo de la avalancha de
inmigrantes en las Islas Canarias (España) se debe a que es la puerta de
entrada a Europa y, por lo tanto, los inmigrantes quieren entrar desde esta
parte de Europa, buscando una major vida. Esta investigación surge de la
necesidad de determinar lingüísticamente el tratamiento del fenómeno de la
inmigración en la prensa española como resultado de la llegada de miles de
ciudadanos extranjeros a la costa de las Islas Canarias en 2006 y 2015. Se
analizan cuatro periódicos españoles utilizando el software Iramuteq de
análisis cualitativo, dos de ámbito regional de Canarias (El Día y Canarias 7)
y dos periódicos de ámbito nacional (El País y ABC). También queríamos
saber cómo aparece el género en las noticias de estos diarios. Nuestra
hipótesis es que los inmigrantes son mayoritariamente hombres por tanto
debe aparece más que la mujer y al contrario, la palabra mujer (refugiada)
tiene una frecuencia mayor que la del hombre. Los resultados se presentan

JADT’ 18

789

dos figuras de dendograma con el Análisis Jerárquico Descendiente (DHC) y
reflejan que la mujer aparece en 2015 pero no está presente en las noticias de
los diarios en 2006 y a la inversa ocurre con el hombre.Keywords: a set of
keywords describing the content of the paper.
1. Introduction
The media have become a powerful tool to make visible conflicts, or show
realities that sometimes remain hidden from the world. Such a fact seems
unquestionable. One of the most-recent cases are the so-called “immigration
crisis” or the “refugees’ crisis,” it began before the dates analyzed in the
current research, however, achieve an uncertain projection until these
citizens reached the coasts of Europe, in the case of the Canary Archipelago.
The concept “immigrant” as Shier, Engstrom & Graham (2011) suggest that
they define an “immigrant” is a person arriving (immigrating) who has come
to live in a country from some other country with the purpose to settle there.
The journalistic enterprises face the challenge of attracting new audiences,
being aware of the transformation of the sector and the emergence of a new
ecosystem. These companies require narrative treatments contrasting from
those already known, since these information units synthesize the content
and preponderance of the published news; these elements are deciding to
capture the attention of the readers (Jarvis, 2014).
Through the selection of the headlines, it is possible to highlight the role of
new professionals in the newsrooms that are responsible for defining what
kind of news be published. As Ramonet (1998) makes evident, the variety of
sources guarantee objectivity. However, information is a social good that
concerns and understand the whole society. This society must establish
moral norms that govern the responsibility of the media (Fraerman, 1998).
The phenomenon of immigration has been analyzed from diverse
perspectives is important to understand that immigration is a fact associated
with times of crisis. But the gender issues are not treated deeply. Thus, one
important aim is to know whether journalists take account that fact.
The Canary Islands (Spain) is a point of gateway to Europe and this is the
reason for the avalanche of immigrants, males and females. The evidence
suggests immigrant’s networks wanting to enter by this point to reach
European land. Most migration researchers understand these networks as
consisting of a set of “strong ties” based on kinship, friendship, or a shared
community of origin that connects migrants and non-migrants (Massey et al.
1998). Migration network approach is that a multidirectional flow of
information and resources forms the basis of every migratory process
(Dekker & Engbersen, 2014).
The migration phenomenon in Europe has had two phases of maximum

790

JADT’ 18

activity in the years 2006 and 2015 where, despite being displaced people
from the place of origin to another destination, including a change of
residence. In the first case, the citizens who enter Europe through the Canary
Islands are the so-called undocumented immigrants. These people left their
countries as a free choice and for a “personal interest,” in line with the
definition of International Conference on Migration (IOM). In the second
case, refugees have carried out the displacement (also present in 2006, but in
a very small percentage) to save their lives or preserve their freedom, as
United Nations High Commissioner for Refugees (UNHCR) states.
The data analyzed in this paper focuses on international migration and the
movement across national borders, consequently this work takes care of the
time-span analysis that separates two massive arrivals and the evolution that
originates in the field of communication in that period. The search terms
“immigrant,” in 2006 and “refugee” in 2015 and also the words “man” and
“woman” were used as keywords to search the headlines and full news of
database and locate information about immigration, and refugees (MUGAK,
2016). The study analyses the year 2006 matching with 2015 and aims to
probe the narrative production generated by two Spanish newspapers (ABC
and El País) and two Spanish regional newspapers (Canarias 7 and El Día), in
relation to the immigration phenomenon that took place in the Canary
Islands in those years.
2. Method
In the present study carried out in the years 2006 and 2015, statistical
methods are mainly concerned with the non-linguistic information from a
text; e.g. term frequencies, inverse frequency and the position of a keyword
in a text. For data analysis, for the study we apply Iramuteq software
(Interface de R pour les Analyses Multidimensionnelles de Textes et de
Questionnaires; Ratinaud, 2009; Ratinaud & Marchand, 2012, 2015).
In our study, for the data processing, apply the Descending Hierarchical
Classification (DHC) by Reinert method (1983, 1986, 1990) defined by lexical
classes, where each of them represents a subject matter, and they can be
described according to the vocabulary that defines them.
From the most frequent words given in the text segments, lexical analysis
was performed. This analysis overcomes the dichotomy between quantitative
and qualitative research, as it allows employing statistical calculations on
qualitative data, the texts. The vocabulary related to “immigration,
immigrant/s, refugee/s, man and woman etc.” are identified and quantified
in the frequency and, in some cases in relation to its position within the text.

JADT’ 18

791

3. Results
Below, the author illustrate the data of the text corpus of the years 2006 and
2015 period of study. The corpus used in this analysis is ad-hoc constructed.
It contains 4.703 newspaper headlines and news published throughout 2006
and 2015 in Spanish. We used four newspapers two of nationwide (El Pais
and ABC) and two of regional scope (Canarias 7 and El DIA), of which 169
news corresponds to El Pais, 291 news for ABC, whereas Canarias 7
published 512. The information of three newspapers was obtained through
MUGAK (Centre of Studies and Documentation on Immigration, Racism and
Xenophobia, Basque Country, Spain, 2016) database, in case of the
newspaper El DIA, 3.731 news; the information was taken directly from the
newspaper database.
Table 1 - Statistical data from the text corpus of study
Corpus 2006

Corpus 2015

Subcorpus 2006

Subcorpus 2015 (text
in web editions)

Occurrences

426.135

30.531

147.468

6.148

Forms

11.993

4.792

9.747

1.487

Hapax

5.093

2.440

4.525

827

Texts

7

11

7

4

In addition, the characteristics of each text, the number of occurrences
detected in the online version of the newspapers is broad and reflects 20% of
the occurrences of the entire corpus, observed the lexicometry while the
remaining 60% belongs to the activity developed in the profiles enabled in
the social networks of each newspaper. It can be observed the following
cloud of words by collecting in generic terms the forms that characterize the
selected texts.
As it can observe some of the words, with bulkier characters and therefore
most relevant, are related to the area of study that concerns us: period 2006
the word immigrant is the most used in the newspapers analyzed, followed
by Canarias, patera and cayuco (two types of small boats) as a form to arrive to
the Canary Islands. However, in 2015 appears the term refugee (refugiado),
immigrant (immigrant), welcome (bienvenida), government (gobierno), rescue
(rescate) or the Canary Islands (Canarias).
In addition, some forms of refugeing, offering, asking or rescue appear, as
Crespo (2008) points out, a certain ideological position that undoubtedly
helps to construct a certain image about the migratory phenomenon and its

792

JADT’ 18

consequences for the receiving countries. The graphs generated by the
Iramuteq software of this corpus of text can be inferred that some specific
forms give positive or negative value. Depending on the verbs used for this
purpose and the profile of the migrant to which reference is made, in our
case display the data of the two analyzed periods. These appear related to the
terminology of the topic that occupies us and previously used in the
construction of the press holders.
3.1. Data from Descending Hierarchical Classification Analysis 2006
Iramuteq 0.7 alpha 2 software (Ratinaud, 2014) provides multivariate
analysis through DHC and calculates descriptive results of clusters according
to its main vocabulary (Camargo & Justo, 2013). Likewise, its location in the
dendrogram, the resulting forms’ clusters reflect the different work scenarios
beside how some social realities cross: class 1 (social, immigrant aid), class 2
(immigrants and their local rescue), Class 3 (social and family), class 4
(institutional). As well, a concept that appears common to two conglomerates
in “immigrant” and “immigration” as can be seen in the figure below. (Fig.1).
The word “male” appears 184 times, X2 =521,9.

Figure 1 - DHC Dendrogram 2006

JADT’ 18

793

3.2 Data from 2015 DHC
The data shown in the graphs below (Fig. 2) of this text offer an estimated
viewing on the figure of the “refugee” and the “immigrant” and their
evolution in the context of the knowledge acquired by the media as the
phenomenon is going forward. In such a way, we find two words, “refugee”
and “immigrant”, that appear in the journalistic headlines.

Figure 2 - DHC Dendrogram 2015

The result of the above dendrogram reflects the different work scenarios and
how some social realities are mixed: class 4 (local), class 2 (institutional), class
3 (social) and class 1 (European). The word “woman” appears 20 times with
X2=28,9. It is worth mentioning the founding of the term "to receive", an
element that is similar to the rest of the verbs that accompany it in the
constellation of words in which it is lodged (to propose, to find, to celebrate
or to dispose among many others). However, it becomes more relevant due

794

JADT’ 18

to its preponderance and strategic situation in an environment in which it
appears with vocabulary with which it keeps linguistic similarities.
4. Conclusion
This object of study that evolves in parallel to the population movement, as
well as certain informative personalization through the introduction of
adjectives that indicate narrative subjectivity. Our findings suggest a vast of
knowledge that covers countless issues related immigrants and refugees and
woman and man. It can be said that the word “man” does not appear during
the 2006 and it does “male”, however in 2015 appears “woman” instead
“female and it does not “male” like in 2006. The mechanization of publishing
systems marks a clear dividing line between some texts and others and the
shortage of human and technical resources used for this activity, causes local
media to be less interventionist in drafting their texts than national ones.
Finally, it should be notice for the future researches the role of journalists and
the usage they do of the gender topic as a way to know how the immigration
phenomenon man/woman behaves.
References
Crespo, E (2008). El léxico de la inmigración: atenuación y ofensa verbal en la
prensa alicantina. En M. Martínez (Ed.) Inmigración, discurso y medios de
comunicación (pp.45-62). Alicante: Instituto Alicantino de Cultura Juan Gil
Albert, Diputación Provincial de Alicante.
Dekker, R & Engbersen, G. (2014). How social media transform migrant
networks and facilitate migration. Global Networks 14, 4, 401–418.
Jarvis, J. (2014). Geeks Bearing Gifts. CUNY Journalism Press, New York.
Spanish El fin de los medios de comunicación de masas. ¿Cómo serán las noticias
del futuro? Barcelona: Ediciones Gestión 2000.
Massey, D. S., J. Arango, G. Hugo, A. Kouaouci, A. Pellegrino and J. E. Taylor
(1998) Worlds in motion: understanding international migration at the end of the
millennium, New York: Oxford University Press.
Mugak (2016) Centre of Studies and Documentation on Immigration, Racism and
Xenophobia, Basque Country, Spain. Available in www.mugak.eu
Ramonet, I (2011). La tiranía de la comunicación. Madrid: Debate.
Ratinaud, P. (2009). IRAMUTEQ: Interface de R pour les Analyses
Multidimensionnelles de Textes et de Questionnaires [Computer software]
Retrieved 5th march 2013 in http://www.iramuteq.org.
Ratinaud, P. (2014). Visualisation chronologique des analyses ALCESTE :
application à Twitter avec l’exemple du hashtag #mariagepourtous. In
Actes des 12eme Journées internationales d’Analyse statistique des Données
Textuelles. JADT 2014 (p. 553- 565). Paris, France. Disponible

JADT’ 18

795

Ratinaud, P. & Marchand, P. (2012). Application de la méthode ALCESTE à
de “gros” corpus et stabilité des “mondes lexicaux”: analyse du
“CableGate” avec IraMuTeQ. Em: Actes des 11eme Journées
Internationales d’Analyse statistique des Données Textuelles. JADT 2012.
Liège.
Ratinaud, P., & Marchand, P. (2015). Des mondes lexicaux aux
représentations sociales. Une première approche des thématiques dans les
débats à l’Assemblée nationale (1998-2014). Mots. Les langages du politique,
108, 57- 77
Reinert, M. (1983). Une méthode de classification descendante hiérarchique:
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, 8, 2, 187- 198.
Reinert, M. (1986). Un logiciel d’analyse lexicale: ALCESTE. Les cashiers de
l’Analyse des Données, 4, 471-484.
Reinert, M. (1990). ALCESTE. Une méthologie d’analyse des données
textuales et une application: Aurelia de G. de Neval. Bulletin de méthologie
sociologique, 28, 24-54
Shier ML, Engstrom S & Graham JR (2011) International migration and social
work: A review of the literature, Journal of Immigrant and Refugee Studies, 9,
1, pp. 38-56. http://dx.doi.org/10.1080/15562948.2011.547825.

796

JADT’ 18

Lo strano caso delle frequenze zero nei testi
legislativi euroistituzionali
Francesco Urzì
combinazioni.lessicali@gmail.com

Abstract
In this paper we intend to verify the actual impact of the so-called universals
of translation – i.e. those linguistic features which typically occur in translated
rather than original texts - on the legislative texts produced by the European
Union. To this aim, a number of text segments have been heuristically
selected in order to ascertain if their statistical absence, or quasi-absence,
from European legislation should be traced back to the effects of the
abovementioned universals and to identify possible EU-internal factors that
might explain such conspicuous statistical absences.
Keywords: universals of translation. European Union, Eur-lex, euroitaliano,
terminology.
1. Introduzione
Negli ultimi tempi si sono moltiplicati gli studi su corpora comparabili volti a
verificare l’effettiva incidenza dei cosiddetti universali della traduzione, ossia
dei tratti linguistici comuni ai testi tradotti e non riconducibili a un’influenza
sistemica della lingua sorgente (Baker 1993 e 1996 e Laviosa 2002). Per
l’italiano disponiamo delle analisi di Garzone 2005 e di Ondelli-Viale 2010.
Ondelli-Viale, che si avvalgono esclusivamente di un corpus di estrazione
giornalistica, rilevano ad esempio la minore ricchezza lessicale e la frequenza
lievemente maggiore del Vocabolario di base nelle traduzioni, per effetto
dell’universale traduttivo della semplificazione.
Meno numerosi sono gli studi sui tratti specifici dell’euroitaliano, ossia di
quella varietà della nostra lingua rappresentata dall’italiano delle traduzioni
dell’UE. In tale ambito Cortelazzo 2013 ha operato un confronto quantitativo
di due corpora di una certa ampiezza costituiti rispettivamente da direttive
europee e leggi italiane di recepimento, utilizzando tra l’altro misure
lessicometriche (ad es. type/token ratio e hapax) e prendendo anche in
considerazione i “segmenti ricorrenti” (che secondo l’autore confermano per
il corpus UE scelte lessicali “leggermente più povere e omogenee di quelle
nazionali”).
Con il presente contributo ci proponiamo di stabilire sulla scorta di segmenti
scelti euristicamente, casi eclatanti di frequenze zero o prossime allo zero sul

JADT’ 18

797

dominio di secondo livello europa.eu, e più specificamente su Eur-lex, che ne
costituisce un sottoinsieme. Lo scopo di tale esercizio è di verificare
• se l’irrilevanza statistica di determinate lessie in questi corpora,
praticamente costituiti solo da testi tradotti - ricordiamo la pluricitata
affermazione di Umberto Eco secondo cui “la lingua dell’Europa è la
traduzione” - non forniscano una prova incontrovertibile degli effetti
degli universali traduttivi, in particolare quelli della semplificazione e
della normalizzazione (o conservatorismo linguistico);
• se non sia pure ravvisabile un processo di “autoinibizione” da parte
dei traduttori UE all’utilizzo di tali lessie. Non opererebbero in altre
parole solo le tendenze generali ascrivibili al processo traduttivo in
sé (gli universali della traduzione appunto), ma anche e soprattutto la
specifica cultura traduttiva euroistituzionale e lo specifico contesto
tecnico-operativo che contraddistingue i servizi di traduzione delle
Istituzioni europee.
Essendo tale analisi di tipo eminentemente qualitativo, l’utilizzo di un corpus
“rumoroso” come Google non inficia la rilevanza dei risultati quantitativi,
che tendono unicamente a individuare solo grandi scarti di frequenza, per
cui è vero in questo caso che “more data is better data”.
2. La cultura traduttiva delle Istituzioni europee
2.1 Confusione fra ‘termine’e ‘parola’
Un tratto soggiacente della cultura di categoria dei traduttori
euroistituzionali è la non percezione della differenza teorica fondamentale fra
‘termine’ e ‘parola’. E’ diversa infatti nel termine e nella parola la natura del
riferimento,
“che nel termine è specializzata all’interno di una particolare disciplina,
mentre nella parola è generale in una varietà di argomenti (Cfr. Scarpa 2008:
52, che cita Sager 1994: 43).
Cabré (1999, 33-34), sulle orme di Wüster (1981), menziona due specificità
della terminologia. La prima è che
“words in dictionaries are described with respect to their use in context; they
are considered as elements of discourse. For terminology, on the other hand,
terms are of interest on their own account”;
la seconda che
“lexicology and terminology present their inventories of words or terms (…)
in different ways because they start from different viewpoints: terminology
starts with the concept and lexicology, with the word”.
Cabré (ibidem, 36) nota inoltre che
“whereas a terminological inventory usually contains only nouns, in a
general language dictionary all grammatical categories are represented”.

798

JADT’ 18

2.2 Referenzialità intertestuale
La natura “ciclica” degli atti legislativi dell’Unione - che molto spesso
modificano e aggiornano testi legislativi precedenti – che fa sì che le soluzioni
traduttive già consacrate dall’ufficialità finiscano per essere trasferite di peso
sui nuovi atti, con un fenomeno che si potrebbe definire di common law
linguistica, in cui il precedente esercita forza vincolante sul giudizio
linguistico autonomo del traduttore. E' in questa fase che il traduttore UE
spesso assegna status di ‘termini’ a sintagmi che pur non rispondendo
teoricamente a tale definizione (v. 2.1) hanno comunque acquisito il crisma
dell'ufficialità per essere stati "validati" in testi legislativi precedentemente
pubblicati o anche solo verificati sul piano qualitativo e ritenuti idonei a a
essere immessi nel successivo iter legislativo. E’ così che determinate
soluzioni traduttive tendono a perpetuarsi all’interno delle “filiera testuale”
della materia trattata. Al riguardo va citato anche l’effetto di
condizionamento subito dai traduttori più giovani, i quali trovano arduo
sostenere scelte linguistiche innovative in contrasto con la "tradizione" dei
testi dell'acquis communautaire e, soprattutto, tendono a non discostarsi
dall'approccio traduttivo dei colleghi più anziani.
3. Il contesto tecnico-operativo dei servizi di traduzione delle Istituzioni
europee
3.1 House Rules
I servizi di traduzione delle Istituzioni europee hanno a disposizione un
“Manuale di convenzioni redazionali” (OPOCE 2011), nella cui pagina di
benvenuto si legge che "la sua applicazione [del Manuale] è obbligatoria
[grassetto originale] per chiunque intervenga nella preparazione di ogni
documento (su carta o elettronico) nelle istituzioni, organi o servizi
dell’Unione europea". Non viene fatta nel Manuale alcuna distinzione fra le
varie tipologie di testi e le differenti funzioni comunicative che competono a
ciascuna di esse. Inoltre molte regole di redazione sono presentate sotto
forma di prescrizione assoluta Ad esempio, si prescrive "direttiva" (atto
legislativo) con la minuscola (il che non sorprende visto il numero di volte in
cui il termine viene utilizzato nei testi UE), nonostante la regola secondo cui
(Lesina 2009) "nei casi in cui un nome generalmente usato in senso comune
viene utilizzato in senso proprio, con un significato restrittivo o particolare
(…) l'iniziale maiuscola può [corsivo mio] essere utile per ragioni di
chiarezza, al fine di segnalare al lettore la particolare accezione del nome".
Conoscendo la scarsa frequentazione degli italiani (anche di buona cultura)
con la terminologia degli atti legislativi comunitari, sorprende che il Manuale
di convenzioni redazionali prescriva che "direttiva", anche quando non
seguita dagli estremi completi dell'atto legislativo (ad es. direttiva

JADT’ 18

799

2049/39/CE), debba essere sempre scritta con la minuscola (dunque anche nei
testi a carattere divulgativo destinati alle pagine web).
3.2 Effetto standardizzante delle tecnologie CAT e MT
Attualmente i traduttori delle Istituzioni europee beneficiano di una
memoria di traduzione comune a tutti i servizi denominata “Euramis” e che
provvede alla pretraduzione dei testi sia quando la traduzione è curata dai
servizi interni sia quando è esternalizzata ad agenzie di traduzione. Da
qualche anno è entrata in servizio anche la traduzione automatica che, su
richiesta del traduttore, integra l’output della traduzione assistita. Poiché ad
alimentare la memoria Euramis sono esclusivamente segmenti di testo
“validati” (ossia già sottoposti al processo interno di controllo di qualità e
dunque ritenuti idonei al successivo dibattito politico o alla pubblicazione) i
traduttori preferiscono non discostarsi da soluzioni ritenute “sicure” (e la cui
adozione, va pure sottolineato, si traduce in un notevole risparmio di tempo).
4. Esempi paradigmatici di "grandi assenti"
Ad esemplificazione di quanto sopra passiamo di seguito in rassegna una
serie di sintagmi, che presentano casi clamorosi di frequenze zero o prossime
allo zero. Nelle relative tabelle il numero di occorrenze preceduto da
asterisco indica dei “falsi positivi”. L’asterisco fra parentesi segnala che sono
dei falsi positivi almeno una parte delle occorrenze. Le forme prese in
considerazione sono una forma aggettivale gerundiva (costruendi), alcuni
sintagmi nominali con aggettivo relazionale (indagini poliziesche, attività
manutentive, servizi consulenziali), un composto aggettivale determinativo
formato da due aggettivi relazionali (politico-programmatico) e due costrutti,
rispettivamente con fattorizzazione (dati quali- quantitativi) e zeugma
preposizionale (valutare e tener conto [di]). Laddove utile sono state proposte,
a titolo comparativo, le statistiche relative alla forme più in uso nel corpus
legislativo europeo.
4.1 Gerundivo
Token
Costruendi

Google
11.800

Europa.eu
*2

Eur-lex
*1

I due unici esempi di europa.eu - ‘i costruendi locali’ e ‘sepolcri esistenti e
costruendi’, entrambi provenienti dalla banca elettronica TED1, sono riferiti
ad aree territoriali italiane. In questo caso sembra aver operato il
1 TED - Tenders Electronic Daily, ossia il supplemento alla Gazzetta ufficiale
dell'Unione europea dedicato agli appalti pubblici europei

800

JADT’ 18

conservatorismo linguistico, che ha indotto ad evitare una forma non
registrata dai dizionari2 e probabilmente ritenuta dai traduttori troppo ardita.
4.2 Aggettivi relazionali semplici e composti
Un analogo comportamento linguistico convenzionale e semplificatorio da
parte dei traduttori si osserva nel caso degli aggettivi relazionali. Non tutti i
suffissi che formano aggettivi relazionali sono infatti suffissi "dedicati", ossia
deputati a codificare esclusivamente il rapporto di relazione; alcuni formano
anche aggettivi qualificativi. Tale è ad esempio il suffisso -ivo3 come in attività
produttive vs. prefisso produttivo. Spesso basta questa ambivalenza semantica a
dissuadere il traduttore dall'utilizzare tali aggettivi in funzione relazionale e
a indurlo a fargli preferire soluzioni alternative (ad es. con l'impiego della
preposizione ‘di’ o con locuzioni preposizionali del tipo ‘relativo/riguardo
a/in materia di’. Nel caso di ‘indagini poliziesche’, potrebbe forse aver agito
anche il proposito di evitare una indesiderata connotazione.
Token
Indagini di polizia
Indagini
poliziesche

Google
164.000
14.700

Europa.eu
793
(*)2

Eur-lex
85
0

Da notare che una delle 2 occorrenze di ‘indagini poliziesche’ in europa.eu è
un comunicato stampa, dunque scritto con ogni probabilità da un giornalista
e non da un traduttore.
Token
Attività
manutenzione
Attività
manutentive

di

Google
1.230.000

Europa.eu
6.730

Eur-lex
354

89.400

(*)139

*1

Da osservare che l’unico risultato di Eur-lex per ‘attività manutentive’ lo si
ritrova in un testo italiano, che riportiamo (grassetto mio)
“Regolamento del sottosegretario di Stato per l'Edilizia abitativa, la
Pianificazione territoriale e l'Ambiente recante definizione di nuove
Tale forma non registrata ad esempio nel Sabatini Coletti 2008 che però riporta
‘istituendo’ e ‘costituendo’, mentre il Grande dizionario Garzanti riporta solo
‘costituendo’.
3 Suffisso usato prevalentemente per la formazione di aggettivi qualificativi
(Wandruska 2004: 391)
2

JADT’ 18

801

prescrizioni relative alla prevenzione di perdite accidentali di fluidi
frigorigeni nell'ambito dell'utilizzo di o dell'esecuzione di attività
manutentive su impianti di refrigerazione e, in relazione alle stesse, recante
modifica del regolamento prescrizioni impermeabilità impianti di
refrigerazione 1997”
Dei 139 risultati in europa.eu 114 provengono dalla banca TED e, come
conferma un controllo a campione eseguito da chi scrive, si riferiscono ad
avvisi di appalto riguardanti il territorio italiano.
Token
Servizi di consulenza
Servizi consulenziali

Google
6.870.000
96.600

Europa.eu
29.300
(*)21

Eur-lex
16
0

Anche in questo caso, dei 21 risultati di europa.eu 3 provengono da TED,
altri (anche se non tutti) da regioni italiane.
Per quanto riguarda gli aggettivi relazionali composti, del tipo: libero
professionale (relativo alla libera professione) oppure marittimo-portuale
(relativo ai porti marittimi), si è scelto come caso eclatante di assenza il
composto ‘politico-programmatico’. L’assenza è tanto più significativa in
quanto non mancano certo nell’Unione europea i documenti funzionalmente
analoghi al Documento politico-programmatico italiano, ma è solo a
quest’ultimo documento che fanno riferimento le pochissime occorrenze di
questo termine riscontrate su europa.eu e Eur-lex. Ancor più che nel caso
degli aggettivi relazionali semplici, l’assenza si spiega con il senso di
incertezza semantica che le formazioni aggettivali costituite da due aggettivi
relazionali possono ingenerare, visto che spesso la loro disambiguazione
(stabilire cioè se si tratta di composto coordinativo o determinativo) può
avvenire solo in relazione a un dato cotesto.
Token
Politico-programmatico

Google
34.900

Europa.eu
8

Eur-lex
*1

Delle 8 occorrenze di europa.eu, almeno 2 provengono da documenti redatti
da curatori italiani. L’unica occorrenza in Eurlex (dove la versione inglese è
policy and planning platform), fa pensare a un brano di testo originariamente
redatto in italiano e a una lettura coordinativa, anziché determinativa, del
composto in sede di traduzione.

802

JADT’ 18

4.3 Fattorizzazioni e costruzioni zeugmatiche
Questi due costrutti, i cui meccanismi sono di difficile reperimento nelle
grammatiche, sono ampiamente utilizzati nel linguaggio giuridico e
amministrativo italiano per evidenti ragioni di economia linguistica. Si è
scelta a tal fine la sequenza 'dati qualitativi e quantitativi', che è
un’espressione che ricorre sovente in testi che riportano dati statistici e che
viene pertanto utilizzata in una pluralità di settori. Per lo zeugma
grammaticale si sono ricercate le occorrenze della sequenza ‘valutare e tener
conto’4, che è risultata non ben accetta dai traduttori in quanto probabilmente
troppo “audace”. Oltretutto costrutti di questo tipo vengono sovente
attribuiti a un’influenza della lingua inglese5, motivo questo di ulteriori
spinte puristiche da parte dei traduttori.
Token
Dati qualitativi e
quantitativi
Dati
qualiquantitativi

Google
23.100

Europa.eu
370

Eur-lex
1

10.400

*9

0

I 9 risultati europa.eu si riferiscono tutti a progetti italiano nati in ambito
regionale
Token
Valutare e tener
conto

Google
1930

Europa.eu
(*)5

Eur-lex
0

Dei 5 esempi in europa.eu 2 si devono all’eurodeputata Pasqualina
Napolitano (doc. A6-0502/2008) mentre 3 sono di provenienza esterna all’UE.

Come nel seguente esempio (grassetto mio):
Art. 5. (Coordinamento per la sicurezza e salute ex decreto legislativo n. 81 del
4

2008)
1. Ai sensi dell’articolo 90, comma 1-bis, del decreto legislativo n. 81 del 2008, il
Tecnico incaricato è obbligato a considerare, valutare e tener conto, al momento delle
scelte tecniche per la fase progettuale oggetto dell'incarico, dei principi e delle misure
generali di tutela di cui all’articolo 15 del citato decreto legislativo n. 81 del 2008.
(http://bandieconcorsi.comune.trieste.it/contenuti/allegati/schema_contratto_incarico.
pdf).
5 Fanfani 2010

JADT’ 18

803

Riferimenti bibliografici
Baker M. (1993), “Corpus Linguistics and Translation Studies – Implications
and Applications”, in: M. Baker/G. Francis/Tognini Bonelli (a cura di), Text
and Technology: In Honour of John Sinclair, Amsterdam-Philadelphia:
Benjamins, 233-250.
Baker M. (1996), “Corpus-based Translation Studies: The challenges that Lie
Ahead”, in: H. Somers (a cura di), Terminology, LSP and Translation: Studies
in Language Engineering in Honour of Juan C. Sager, AmsterdamPhiladelphia: Benjamins, 175-186.
Cabré, M. T. (1999), Terminology – Theory, methods and applications,
Amsterdam-Philadelphia: John Benjamins.
Cortelazzo M. A (2013), "Leggi italiane e direttive europee a confronto", in:
Stefano Ondelli (a cura di), "Realizzazioni testuali ibride in contesto
europeo. Lingue dell’UE e lingue nazionali a confronto", Trieste, EUT
Edizioni Università di Trieste, 2013, pp. 57-66.
Fanfani M. (2010) Anglicismi, in Simone R., Berruto G. D’Achille P. (a cura di)
“Enciclopedia dell’italiano”. Istituto della Enciclopedia italiana, Roma
Garzone G. (2005), “Osservazioni sull’assetto del testo italiano tradotto
dall’inglese”, in: A. Cardinaletti/G. Garzone (a cura di), L’italiano delle
traduzioni, Milano: Franco Angeli, 35-58.
Grande Dizionario Garzanti di italiano (2017), De Agostini Scuola s.p.a. –
Garzanti linguistica (versione elettronica)
Laviosa S. (2002), Corpus-based Translation Studies. Theory, Findings,
Applications, Amsterdam-New York: Rodopi.
Laviosa S. (2002), Corpus-based Translation Studies. Theory, Findings,
Applications, Amsterdam-New York: Rodopi.
Lesina R. (2009), Il Nuovo Manuale di stile¸Bologna: Zanichelli
Manuale interistituzionale di convenzioni redazionali, Ufficio delle pubblicazioni
dell’Unione europea (OPOCE), 2011, ISBN 978-92-78-40704-9
Ondelli S. e Viale M. (2010), L’assetto dell’italiano delle traduzioni in un corpus
giornalistico. Aspetti qualitativi e quantitativi. In Rivista internazionale di
tecnica della traduzione, n.12/2010, pp. 1-62. ISSN 1722-5906.
Sabatini F e Coletti V. (2008), Il Sabatini Coletti. Dizionario della lingua italiana,
Milano, Rizzoli-Larousse.
Sager J. (1994), Language Engineering and Translation Consequences of
Automation, Amsterdam-Philadelphia: John Benjamins.
Scarpa F. (2008), La traduzione specializzata, seconda edizione, Milano:
Hoepli.
Urzì F. (2016), “Il paradosso degli aggettivi di relazione composti derivati da
sintagmi N+A. Una risorsa non utilizzata in traduzione”, in: R. Bombi/V.
Orioles (a cura di), Lingue in contatto-Contact Linguistics, Roma: Bulzoni,

804

JADT’ 18

163-178.
Wandruszka U. (2004), “Aggettivi di relazione”, In M.Grossmann/F. Rainer
(a cura di), La formazione delle parole in italiano, Tübingen, Niemeyer, 382394.
Wüster E. (1976), "La théorie générale de la terminologie - un domaine
interdisciplinaire impliquant la linguistique, la logique, l'ontologie,
l'informatique et les sciences des objets", in H. Dupuis (a cura di), Essai de
définition de la terminologie. Actes du colloque international de terminologie
(Québec, Manoir du lac Delage, 5-8 octobre 1975), Québec, Régie de la langue
française, pp. 49-57.
Wüster E. (1981), “L’étude scientifique générale de la terminologie, zone
frontalière entre la linguistique, la logique, l’ontologie, l’informatique e les
sciences des choses”, in Rondeau, Guy/Felber, Helmut (a cura di), Textes
choisis de terminologie – I Fondements théorique de la terminologie, Québec,
GIRSTERM, 55-114.

JADT’ 18

805

Les traductions françaises de The Origin of Species :
pistes lexicométriques
Sylvie Vandaele
Université de Montréal – sylvie.vandaele@umontreal.ca

Abstract
In order to develop a sound methodology that would guide the analysis of
the translations of important writings, we used Hyperbase to perform a
lexicometic analysis of specificities on two corpora based on the various
English and translated editions of Charles Darwin’s The Origin of Species. We
show that the translated corpus is characterized by a notable lexical
dispersion. compared to the source corpus. By combining the use of
Hyperbase with Logiterm. a text alignment software. we were able to target
and analyse contexts of interest. This approach allows for the rapid
identification of contexts that are significant both statistically and in terms of
the analysis of the translation strategies themselves.
Résumé
Afin de mettre au point une méthode raisonnée d’analyse des traductions
d’œuvres conséquentes, nous avons soumis les versions originales de The
Origin of Species, de Charles Darwin ainsi que leurs traductions en français à
une analyse lexicométrique des spécificités à l’aide du logiciel Hyperbase.
Nous montrons que le corpus de traductions se caractérise par une
dispersion lexicale notable, contrairement au corpus anglais source. Les
spécificités ont permis, à l’aide du logiciel d’alignement bilingue Logiterm,
de cibler l’analyse de contextes bilingues montrant les différences de choix de
traduction, Cette approche permet de repérer rapidement des contextes
significatifs tant sur le plan statistique que sur le plan de l’analyse des
stratégies de traduction.
Keywords: The Origin of Species;
specificities; Hyperbase; Logiterm,

retranslation;

translation

choices;

1, Introduction
La retraduction, fréquente en littérature (voir Monti et Schnyder, 2011), est
rare en science, The Origin of Species [désormais OS], l’œuvre célèbre de
Charles Darwin, fait exception : six éditions de langue anglaise (de 1859 à
1872), six traductions en français dont deux modernes (voir Vandaele et

806

JADT’ 18

Gendron-Pontbriand [2014] pour les détails). Cependant, l’ampleur de
l’œuvre rend l’analyse des traductions difficile. Nous proposons une
méthode consistant à isoler les spécificités lexicales des originaux et des
traductions, puis à repérer les contextes bilingues alignés correspondants,
soumis ensuite à une analyse qualitative. Nous accédons ainsi rapidement
aux éléments saillants de l’évolution de l’œuvre et de ses traductions.
2. Corpus et méthodologie
Les deux corpus1 sont constitués par les chapitres intégraux des six éditions
originales anglaises de l’OS (1859-1872) et les six traductions en français, à
l’exclusion du paratexte et des notes de bas de page. Les césures en fin de
ligne ont été éliminées, les numéros de page, placés entre deux phrases, les
appels de notes, enlevés. Nous avons eu recours au logiciel Hyperbase v. 102
réalisé par Étienne Brunet (Brunet 2011). L’annotation syntaxique et la
lemmatisation ont été réalisées au préalable avec Cordial v. 14 (Synapse)
pour le français, et à la volée, pour l’anglais, avec la version de TreeTagger
incluse dans Hyperbase. L’alignement des versions originales et traduites a
été réalisé avec Logiterm v, 5.7.1. (Terminotix).
3. Les versions originales anglaises de l’OS
Le corpus anglais compte un peu plus d’un million d’occurrences, Darwin a
procédé à des ajouts, mais aussi à des retraits.3 La 6e édition (18724) est 28 %
plus longue que la 1re (1859), soit 48 000 occurrences de plus. L’analyse de la
richesse du vocabulaire montre la proximité lexicale des six éditions
originales : on compte 8559 lemmes pour tout le corpus, 6082, pour la 1re
édition et 7431, pour la 6e (tableau 1).
Les lemmes communs forment la majorité du corpus : pour les textes 2 à 2,
leur nombre varie de 5597 à 6600, tandis que le nombre des lemmes privatifs
fluctue de 136 à 1795. L’examen des formes donne des résultats du même
ordre. L’accroissement chronologique des lemmes montre un léger
appauvrissement pour la 2e et la 3e édition, mais un enrichissement notable

Les textes anglais viennent du site Darwin Online (John van Wyhe, dir. 2002. The Complete Work of Charles Darwin Online - http://darwin-online.org.uk/). Les textes
français ont été obtenus par Gallica ou Google livres, ou ont été numérisés par nous.
2 Téléchargeable à <http://ancilla.unice.fr/>.
3
Voir le variorum en ligne (van Wyhe, 2002-; < http://darwinonline.org.uk/Variorum/1859/1859-1-dns.html>).
4 Celle de 1876, dite 6b, est quasiment identique à celle de 1872. C’est l’édition de
1872 qui a été traduite par Edmond Barbier (1876), raison pour laquelle nous l’avons
choisie sans notre analyse.
1

JADT’ 18

807

du vocabulaire dans la 6e édition (tableau 1), essentiellement redevable à un
grand nombre d’hapax, souvent des noms d’espèces.5 Ce résultat reflète le
fait que Darwin apporte de plus en plus de données à l’appui de sa théorie.

Année de
publication et
édition
1859, 1re éd,
1860, 2e éd,
1861, 3e éd,
1866, 4e éd,
1869, 5e éd,
1872, 6e éd,
Total

Tableau 1 – Corpus des éditions originales de l’OS
Richesse du
vocabulaire
Nombre
Effectif des
Code
d’occurrences6
lemmes
N (écarts
réduits)
OS01
170 634
6082 (2,67)
OS02
171 665
6210 (4,21)
OS03
181 974
6019 (0,34)
OS04
200 608
6914 (9,59)
OS05
199 963
7072 (11,67)
OS06
218 870
7431 (14,06)
1 143 714
8559

Accroissement
chronologique
Écarts réduits
(calculés sur les
lemmes
4,5
-6,5
-4,9
1,8
0,3
16,5

L’analyse arborée (selon Luong, 1994; cité dans Brunet 2011) met en évidence
la faible distance séparant les textes, ce qui est attendu (figure 1), mais
permet de situer les différentes éditions entre elles : qu’il s’agisse des
fréquences (1A) ou des présences (1B)7, on note une grande proximité entre
les 1re et 2e éditions, ce qui est corroboré dans les préfaces. La 5e et la 6e sont
proches, cette dernière se distinguant par les nombreux hapax. La 3e et la 4e
sont intermédiaires. Nombre de lemmes privatifs passent sous la barre des
5 %, les spécificités sont peu nombreuses, ce qui est attendu, mais révélateur.
Les spécificités positives ne repèrent aucun mot plein pour les quatre
premières éditions, mais font apparaître le pronom I et le déterminant my.
C’est à la 5e édition que l’on note l’apparition de deux spécificités de mots
pleins statistiquement significatives : survival et fittest, avec un écart réduit de
4,6 et de 4, respectivement, pour les formes, ou survival (substantif, 4,6) et fit
(adjectif, 4) pour les lemmes. Dans la 6e édition, apparaissent Mr (7,1), through
(6,1) cambrian (5,8) orchids (4,3), developed (4,9) et development (4,2), lower (4,2),

Le nombre d’hapax augmente considérablement dans la 6e édition :
respectivement, 45, 40, 61, 133, 134, et 622 occurrences (lemmes) de la 1re à la 6e édition
(écart réduit de 33,5 pour la 6e édition).
6 Les valeurs reportées dans les tableaux sont fournies par Hyperbase. Il y a de
légères différences avec des valeurs publiées antérieurement, dues à la préparation
des textes et aux logiciels utilisés pour le décompte.
7 Respectivement selon Labbé et Jaccard, cités dans Brunet 2011.
5

808

JADT’ 18

beneficial (4,1) et spontaneous (4,1). L’analyse des lemmes fait, en plus des
précédents, remonter survival (substantif, 4,6), spine (substantif, 5,3), increased
(adjectif, 4,2), movement (substantif, 4,1), fit (adjectif, 4,1), beneficial (adjectif,
4,1) et spontaneous (adjectif, 4,1).

A

B

Figure 1 – Analyse arborée sur les lemmes : A – sur les fréquences; B – sur les présences

Le regroupement des spécificités en catégories reflétant le contenu
sémantique (établi à partir des contextes) est instructif : concepts théoriques
(fittest, fit, survival, through [expression de la causation]), données et citations
(cambrian, orchids, spine, Mr), vision dynamique du vivant de Darwin (develop,
development, increased, movement, spontaneous), jugements de valeur (beneficial,
lower [certaines occurrences]). Ainsi , les spécificités, même rares, se
démarquent par leur saillance : elles captent l’introduction du fameux
concept de Spencer (1864), survival of the fittest et permettent de présumer une
affirmation de la pensée de Darwin – à savoir sa vision profondément
dynamique de la nature. Enfin, les spécificités négatives signalent que les
fréquences relatives du déterminant possessif my et du pronom I diminuent
avec le temps, ce qui traduit l’ajout de passages non argumentatifs contenant
des données, et ce qui corrobore l’augmentation des hapax, constitués par
majoritairement par des noms d’espèces.
4. Analyse du corpus français
Le corpus français comprend un peu plus de deux millions d’occurrences
(tableau 2) : trois traductions d’époque (Clémence Royer [1862, 3e éd.], JeanJacques Moulinié [1873, 5e éd.], Edmond Barbier [1876, 6e éd.]); celle de
Daniel Becquemont (2008), qui part de la traduction de Barbier et la modifie
pour remonter à la 1re édition; deux modernes, par Augustin Berra (2009, 6e
éd.) et Thierry Hoquet (2013, 1re éd.) (voir Vandaele et Gendron-Pontbriand
[2014] pour les références bibliographiques). Les textes comptent de 181 785 à

JADT’ 18

809

248 863 occurrences, soit un écart de 67 078 occurrences. Les différences de
coefficients de foisonnement8 révèlent déjà que les traducteurs ont travaillé
avec des stratégies de traduction distinctes. L’homogénéité lexicale diminue
par rapport aux originaux. La contribution de chacun des textes à la richesse
lexicale est beaucoup plus importante en français qu’en anglais : les lemmes
partagés dans les textes pris deux à deux se situent entre 4498 (13Ho et 62Ro)
et 5649 (73Mo et 76Ba) pour un total de 11712 lemmes (soit 3153 lemmes de
plus que dans le corpus anglais). Chacun des textes français contribue pour
un pourcentage moindre au vocabulaire commun (figure 2A). Les effectifs
des lemmes privatifs sont plus importants (de 772 à 3000) et fluctuent d’un
traducteur à l’autre (figure 2B). Sont mises en évidence les différences entre
Becquemont (08Bq) et Hoquet (13Ho) pour la 1re édition, et entre Barbier
(76Ba) et Berra (09Be) pour la 6e édition, mais aussi la proximité (attendue)
entre Barbier et Becquemont.
Tableau 2 – Traductions françaises de l’OS – * d’après la traduction de Barbier de la 6e édition,
Année
de
publication
1862
1873
1876
2008
2009

Richesse du
vocabulaire
Effectif des lemmes
N (écart réduit)
6357 (-6,7)
7036 (0,8)
6971 (-3,8)

08Bq

186 440

9%

6260 (-4,8)

09Be

248 863

14 %

7804 (5,0)

13Ho

181 785
1 277 582

7%

6579 (-0,2)
11 712

Traduit par

Code

Nombre
d’occurrences

1861 (3e)
1869 (5e)
1872
(6ea)
1859 (1e)*

C. Royer
J.-J,.Moulinié
E. Barbier

62Ro
73Mo
76Ba

D.
Becquemont
A. Berra
T. Hoquet

1876
(6eb)
1859 (1e)

2013
Total

207 633
211 691
241 170

Coefficient
de
foisonnement
14 %
6%
10 %

Édition
originale
anglaise

Les distances lexicales intertextuelles (figure 3) confirment la proximité de
Becquemont et de Barbier, mais révèlent deux faits inattendus : 1) Royer
(62Ro) se situe sur la même branche que Berra et Hoquet; 2) Moulinié (73Mo)
se place entre Becquemont et Barbier lorsque l’on passe des fréquences aux
présences.

Le coefficient de foisonnement est l’accroissement du nombre d’occurrences
observé lorsque l’on traduit de l’anglais au français. Il est généralement admis, en
traduction dite « pragmatique » (par opposition à la traduction littéraire) que le taux
de foisonnement se situe généralement entre 10 % et 15 %, une des causes étant que le
français recourt à plus de mots grammaticaux que l’anglais. Une forte concision peut
diminuer ce taux.
8

810

JADT’ 18

Figure 2 – A – Contributions respectives de chacun des textes aux parties communes des
corpus anglais et français (lemmes)9 – B – Richesse lexicale (lemmes), Le pointillé indique le
seuil de 5 %,

Diverses hypothèses explicatives doivent être explorées, mais il n’est en tout
cas plus permis de douter que les manières de traduire sont décisives au
point de brouiller, sur le plan lexical, la chronologie des versions originales,
et que cette approche permet de mettre ces particularités en évidence.

A

B

Figure 3 – Analyse arborée (méthode Luong) sur les lemmes
A – calculée sur les fréquences (Labbé); B - calculée sur les présences (Jacquard)

Nous nous sommes ensuite concentrée sur les spécificités positives des lemmes des
mots pleins et, parmi elles, avons sélectionné les unités dont la signification
paraissait la plus caractéristique du propos central de l’OS : ainsi, sélection,
préservation, pouvoir… ont été retenus, mais pas aujourd’hui, grandement,

Le schéma a été obtenu à partir des effectifs des lemmes pour chacun des
textes, ramenés en pourcentage du nombre total de lemmes par corpus
(représentation « radar » fournie par Excel v.16). Les effectifs des lemmes des textes
traduits ont été disposés en regard des textes anglais (ceux de OS1 et OS6 ont donc été
dupliqués); de plus, la forme assymétrique du tracé pour le français rend compte de
l’absence de traduction d’OS2 et d’OS4. À cause de ces particularités, l’aire délimitée
par les traits n’est pas représentative des valeurs totales pour chacun des corpus, mais
le schéma reste visuellement parlant.
9

JADT’ 18

811

inclure… Nous nous sommes ensuite concentrée sur les spécificités positives
des lemmes des mots pleins et, parmi elles, avons sélectionné les unités dont
la signification paraissait la plus caractéristique du propos central de l’OS :
ainsi, sélection, préservation, pouvoir… ont été retenus, mais pas
aujourd’hui, grandement, inclure…

Figure 4 – Analyse factorielle de correspondances : sélection de lemmes parmi les spécificités

La quarantaine de lemmes ainsi obtenus a permis de générer un graphe
(figure 4) représentant le résultat d’une analyse de correspondances (menée
selon le programme de Lebart, inclus dans Hyperbase, sur les données
pondérées). Le graphe montre que les modernes (Berra, Hoquet) s’opposent
aux anciens (Barbier, Moulinié) ou quasi-ancien (Becquemont), Royer se
situant à part. La consultation des contextes ciblés par cette méthode dans les
corpus alignés par Logiterm permet d’analyser qualitativement les choix de
traduction. L’exemple le plus frappant est le choix de élection et de électif par
Royer, qui s’oppose au choix de sélection par les autres traducteurs (tab. 3).

812

JADT’ 18

Tableau 3 – Traductions alignées d’une phrase commune à toutes les éditions anglaises
(Introduction)
Darwin and we shall then see how Natural Selection almost inevitably causes much
Extinction of the less improved forms of life…
62Ro
Nous verrons comment cette élection naturelle cause presque inévitablement de
fréquentes extinctions d’espèces parmi les formes de vie moins parfaites…
73Mo
Nous y verrons comment la sélection naturelle détermine presque inévitablement
l'extinction des formes moins perfectionnées…
76Ba
Nous verrons alors que la sélection naturelle cause, presque inévitablement, une
extinction considérable des formes moins bien organisées…
08Bq
Nous verrons alors que la sélection naturelle cause presque inévitablement une
extinction considérable des formes moins bien organisées
09Be
nous verrons alors de quelle façon la sélection naturelle cause presque
inévitablement une forte extinction des formes de vie moins améliorées…
13Ho
Et nous verrons comment la Sélection Naturelle cause presque inévitablement
une grande Extinction des formes de vie moins améliorées…

5. Conclusion
Le ciblage de contextes, repérés au moyen d’une analyse lexicométrique
préalable, dans des corpus alignés conséquents est une stratégie de choix.
Elle permet d’arriver assez vite à des observations statistiquement
significatives et de pointer d’emblée sur des éléments majeurs sans
hypothèse préalable. Comme le souligne Brunet (2002), l’intérêt de travailler
sur des traductions est que certains paramètres sont fixés. L’inconvénient
actuel de l’entreprise tient à la faible ergonomie du processus, c’est-à-dire
aux nombres de clics liés au passage d’un logiciel à l’autre. Restent les
nombreuses modifications sous le seuil de 5 %, qui peuvent recéler, malgré
l’absence de signification statistique, des éléments cruciaux en matière de
choix de traduction. D’autres stratégies de filtrage sont alors nécessaires pour
leur étude.
Remerciements
Nous remercions vivement Étienne Brunet, Damon Mayaffre et Laurent
Vanni pour leurs conseils sur l’utilisation d’Hyperbase. Il va de soi que les
éventuelles erreurs sont nôtres. Merci aussi à Marie-Joëlle StratfordDesjardins, étudiante auxiliaire de recherche, pour son aide à la préparation
du corpus. La présente recherche a bénéficié d’une subvention de recherche
du Conseil de recherche en sciences humaines du Canada (2015-2018).

JADT’ 18

813

Références
Brunet É. (2002). Un texte sacré peut-il changer ? Variations sur l’Evangile. In
Cook J., dir. Bible and Computer, Leiden / Boston : Brill, pp. 79-98.
Brunet É. (2011). Hyperbase – Manuel de référence. Hyperbase pour Windows,
version 8.0 et 9.0.
Luong X. (1994). L’analyse arborée des données textuelles : mode d’emploi.
Travaux du cercle linguistique de Nice, 16 : 27-42.
Monti E. et Schnyder, P., dir. (2011). Autour de la retraduction : Perspectives
littéraires européennes. Coll. Universités, Paris : Orizons,
Spencer H. (1864). The Principles of biology. Vol. 1, New York: Appleton.
Vandaele S. et Gendron-Pontbriand E.-M. (2014). Des « vilaines infidèles » aux
grands classiques : traduction et retraduction de l’œuvre de Charles Darwin. In:
Pinilla J. et Lépinette B., dir, Traducción y difusión de la ciencia y de la
técnica en España en los siglos XVIII y XIX,Valence : Universitat de
València, pp. 249-276.

814

JADT’ 18

Circuits courts en agriculture :
utilisation de la textométrie dans le traitement d’une
enquête sur 2 marchés
Pierre Wavresky1, Matthieu Duboys de Labarre2,
Jean-Loup Lecoeur3
2

1Umr Cesaer Inra-Agrosup Dijon – pierre.wavresky@inra.fr
Umr Cesaer Inra-Agrosup Dijon – matthieu.duboys-de-labarre@inra.fr
3Umr Cesaer Inra-Agrosup Dijon – yajintei@hotmail.fr

Abstract
Semi-structured interviews about short food supply chains have been done
with producers and consumers on two different markets. Our work gives an
insight to the themes common to producers and consumers that are not
attributable to the interviews guides. It also underlines the advantages of a
textometric approach and the precautions necessary to interpret such a
corpus.
Résumé
Des entretiens semi-directifs sur le thème des circuits courts alimentaires ont
été menés sur deux marchés, auprès de producteurs et des consommateurs.
Notre travail s'intéresse notamment aux thématiques communes aux
producteurs et consommateurs et qui ne soient pas imputables aux grilles
d’entretiens. Il souligne par ailleurs les apports d'une approche
textométrique, ainsi que les précautions d'interprétation sur un tel corpus.
Keywords: short food supply chain, semi-structured interviews, textometry
1. Introduction et méthodologie
Les circuits courts alimentaires interviennent de plus en plus dans le débat
social. Ils sont devenus l’emblème d’une opposition au « modèle
conventionnel ». Ils s’inscrivent également dans des enjeux de politique
publique (définition légale en 2009 avec le plan Barnier1), et scientifique. Ils
comprennent des formes innovantes comme les AMAP, mais aussi des
formes plus anciennes comme les marchés ou la vente à la ferme.
La sociologie a abordé les circuits courts sous des angles variés : la
consommation engagée (Dubuisson-Quellier, 2009), la sociologie de

1 Circuit de commercialisation comprenant au plus un intermédiaire entre le
producteur et le consommateur.

JADT’ 18

815

l’innovation (Chiffoleau et Prévost, 2012), d’autres ont approché la question
en décalant le point de vue vers le développement local (Traversac, 2010) ou
au travers de la notion de proximité (Mundler et Rouchier, 2016). Les travaux
de sociologie insistent sur l’intérêt économique des circuits courts, mais aussi
sur leur capacité à recréer du lien social (Prigent-Simonin et HéraultFournier, 2014). De nombreux dispositifs s’appuyant sur les circuits courts de
commercialisation se caractérisent par un rapport direct entre
consommateurs et producteurs. Ce lien a été l’objet de différentes analyses et
interprétations dans la littérature. Il est perçu comme un déplacement de
l’espace de référence des agriculteurs vers celui des consommateurs (Dufour
et Lanciano, 2012). Il a aussi été analysé comme le lieu de rencontre autour
d’attentes plurielles (Chiffoleau et Prévost, 2012). Plus généralement, il
s’ancrerait dans des logiques communes de re-localisation des pratiques
agricoles et alimentaires (Duboys de Labarre, 2005). C’est ce lien que nous
allons analyser au travers d’un dispositif textométrique. Nous mettrons en
lumière les intérêts et les éventuelles limites interprétatives liés au type de
corpus (faible nombre d’entretiens semi-directifs). Cela nous éclairera
également sur les thématiques abordées et leur spécificité. Dans le cadre du
projet européen H2020 « Strength2food » 2 , pour la France, nous avons
interrogé 23 personnes3 (12 vendeurs-producteurs et 11 consommateurs) sur
deux marchés (en milieu rural et en milieu urbain) par entretien semidirectifs. Nos deux sous-populations relèvent d’initiatives différentes dans
leur structuration et leur ancienneté4. Dans les deux cas, les parties-prenantes
restent attachées à la consommation/production bio et sont assez engagés. Ce
corpus n’est donc pas représentatif (ni des consommateurs ni des
producteurs) et nous considérons ce travail comme exploratoire.
Le corpus est analysé grâce au logiciel de textométrie Iramuteq5, les thèmes
communs ou spécifiques des producteurs et consommateurs seront
recherchés essentiellement par classification descendante hiérarchique
(Reinert, 1983) et par analyse de spécificité. Parmi les variables caractérisant
les textes, a été incluse une variable à 4 modalités : consommateur-rural,

https://www.strength2food.eu/. Ce projet a été financé par le programme de
recherche et d'innovation Horizon 2020 de l'Union européenne dans le cadre de la
convention de subvention n° 678024
3 Ces entretiens, structurées autour de 6 thèmes, sont semi-directifs et visent à
favoriser l’expression des acteurs. Ils sont retranscrits mot à mot et incluent des
annotations de l’intervieweur.
4 Celle en milieu urbain est un marché de plein vent traditionnel, celle en milieu
rural est un marché de producteurs innovant.
5 http://www.iramuteq.org/ (Pierre Ratinaud)
2

816

JADT’ 18

consommateur-urbain, producteur-rural, producteur-urbain6. Comme la
longueur des interviews est très variable (de 102 à 560 segments de texte) et
le nombre d’interviewés assez faible (23), les statistiques relatives à cette
variable peuvent être essentiellement imputables à une interview, il est donc
d’autant plus nécessaire de revenir à l’interview. De plus il peut arriver que
le lien, en termes de Khi², entre une des quatre catégories (ou une interview)
et une thématique (classe de la classification) soit faible. Or quelques
segments de textes énoncés par cette catégorie sous-représentée sont parfois
très liés à cette thématique, et dire que le lien est faible serait erroné. D’où
l’analyse, aidée par une représentation graphique, des segments de textes les
plus caractéristiques d’une classe, pour chaque catégorie étudiée.
Deux annotations de l’intervieweur, caractérisant la parole de l’interviewé,
ont été conservées au sein du corpus, et seront donc analysées comme les
autres mots : « rire » (codé « _rire ») et « blanc », signifiant un délai avant la
réponse ou en son sein (codé « _blanc). Le but étant de voir si des hésitations
(« _blanc ») sont cooccurrentes d’autres lemmes.
2. Analyse statistique du corpus réponse
Les 5 lemmes les plus courants sont : aller, voir, bio, gens, marché. Ce qui
ressemble à un programme : aller au marché, donc favoriser un mode de
circuit court, pour acheter ou vendre des produits bio et pour voir des gens,
donc avec un aspect relationnel important. Il est probable que les lemmes bio,
aller et marché soient liés au contexte d’enquête (nature des enquêtés pour bio
et nature des dispositifs pour aller et marché). Enfin, le caractère assez
homogène de l’importance quantitative de ces 5 lemmes peut être interprété
comme le reflet d’un horizon commun partagé par nos informateurs et ce en
dépit de de leur groupe d’appartenance (producteur ou consommateur) ou
du dispositif étudié.
2.1. Classification descendante hiérarchique : 12 types de discours
Une classification descendante hiérarchique7 (Reinert 1983) a permis de
dégager 12 types de discours. Nous nous focaliserons sur 2 ensembles de
classes8, selon qu'elles sont plutôt spécifiques ou peu spécifiques
d'une catégorie (producteur ou consommateur).

Producteur-urbain signifiant producteur vendant sur le marché de la ville
moyenne, en opposition avec producteur-rural qui vend sur le marché du village.
7 5264 segments de texte sur les 6231, soit 84%, ont été retenus par la
classification.
8 Nous écartons la classe 3 (12,5%) car elle est peu interprétable (lemmes
polysémiques : chose, gens, monde...).
6

JADT’ 18

817

Graphique 1 : les 12 classes de discours

Le premier ensemble regroupe les classes 1, 2, 6, 9 et 11 qui sont
caractéristiques d’un sous-groupe. Les classes 1 et 11 concernent surtout les
producteurs, par contre les classes 2, 6 et 9 émanent principalement de
consommateurs. Dans la classe 1 (14.4%) il est question des aides, de projet,
d’installation, de reprise (d’exploitation), d’investissement. Il y a des critiques sur
la PAC (notamment sur le fait que ce soit compliqué), mais pas seulement :
« Bah comme on a de la surface un peu ouais ça commence c’est super compliqué la
PAC je sais pas si tu veux qu’on en parle _rire même nous on a du mal » (Lydie,
productrice rurale). La classe 11 (11,7%) est orientée autour des produits
laitiers (lait, chèvre, fromage, yaourt, vache, faisselle, litre, cabri…), avec un aspect
monétaire (euro, prix). Dans la classe 6 (8.1%) c’est de nourriture dont il est
question, notamment le fait de manger des fruits et légumes de saison
(manger, tomate, fraise, saison, pas en hiver). C’est un discours de
consommateurs, surtout urbains. Melissa et Jennifer parlent surtout des
courses qu’elles font, où elles les font (sur le marché de la ville moyenne
essentiellement, où elles ont été interrogées). Toutefois l’autre thème (manger
des fruits de saison) est celui qui est le plus typique de cette classe.
Dans la classe 9 (3.3%) il est question de ville (vivre en ville/à la campagne) et
de distance, aussi bien en termes de proximité que de nombre
d’intermédiaires (distance, kilomètre, circuit_court, intermédiaire). C’est plutôt
une classe de consommateurs. Enfin dans la classe 2 (12.4%) les 4 premiers
lemmes forment une phrase : acheter produit bio producteur. Revendeur et local
sont présents aussi. Il est donc question du comportement d’achat, mais pas
des produits qu’on achète, comme dans la classe 6, plutôt de certaines de
leurs propriétés (bio) et de la qualité du vendeur (producteur). Les classes 1, 2
et 6 renvoient directement à des thèmes abordés dans les guides d’entretiens
respectifs des groupes et la classe 11 à une catégorie de produit agricole

818

JADT’ 18

spécifique qui était surreprésentée dans l’échantillon des producteurs
transformateurs (5 informateurs sur 12). Ces classes parlent des pratiques
liées aux groupes (professionnelles, d’achat et de consommation alimentaire)
et permettent de les caractériser. Nous noterons que les classes 1, 2 et 6
renvoient à la notion de maîtrise ou de contrôle. Pour la classe 1 parce que les
aides PAC sont parfois perçues comme extérieures et complexes. Pour les
classes 2 et 6 au contraire parce qu’elles traduisent l’idée que le
consommateur maîtrise sa pratique (choix de se fournir directement auprès
d’un producteur et en aliments bio, locaux et de saison).
Le second ensemble regroupe les classes 4, 5, 7, 8, 10 et 12. Elles sont peu
spécifiques d’une catégorie. La Classe 10 (7.3%) est celle du respect des
animaux et plus généralement du respect du vivant. On peut remarquer que
le lemme _rire y est particulièrement rare : dans cette classe, le respect des
animaux est abordé comme une question sérieuse. « C’est un animal pour
l’élevage donc je le mange s’il a été élevé dans le respect des lois de la nature et de
l’univers s’il a été élevé d’une manière respectueuse par rapport à
l’environnement » (Théophile, producteur urbain) [Les mots en gras sont
spécifiques de la classe]. Il n’y a pas de différence marquée rural/urbain ou
producteur/consommateur.

Graphique 2 : Score des segments de texte (classe 10)

Mais si on considère le nombre de
segments de texte caractéristiques
(graphique 2), on voit que Jacques n’en
parle pas beaucoup mais il en a énoncé
certains très caractéristiques. Autrement
dit, il parle peu mais intensément du
bien-être animal : « Et nous nos animaux
on est en bio on fait attention au bien-être
animal on fait le choix de garder tous les
petits pour pas qu’ils partent dans des
élevages industriels intensifs et la suite
logique» (Jacques, producteur rural
[score=925]9).

La classe 7 (4,5%) renvoie à deux univers de sens différents autour du lemme
vie : d’une part la notion de trajectoire de vie en relation avec la parentèle
(famille, parent [d’origine agricole], grand_parent, enfant), et d’autre part à une
forme de souci de soi (mode de vie sain, santé reliée à nourriture et alimentation).
« En amont dans un mode de vie qui devrait te permettre d’avoir une vie plus

9 La somme des Khi² (mesurant le lien entre chaque lemme et la classe) donne le
score du segment de texte.

JADT’ 18

819

harmonieuse plus saine plus en meilleure santé physique psychique mentale
sociale parce_que tu crées du lien aussi enfin y a une… ça va dans une même
mouvance » (Claire, consommatrice rurale).
La classe 8 (5.5%) concerne les céréales (farine, pain, gluten, variété, vieux,
boulanger), notamment les vieilles variétés.
La classe 5 (6.3%) est celle du doute (on se pose des questions, il y a des _blanc :
ces 3 lemmes sont entre 8 et 9 fois plus nombreux qu’attendu). « Se poser des
questions » et penser évoque aussi une prise de conscience de problèmes.
Mais c’est également « poser des questions » aux vendeurs sur leur
production.
La classe 4 (5.2%) est celle des relations et de leur importance. « Eh ben les
relations humaines on côtoie une diversité de population quoi des gens et en fait
on se parle c’est agréable _rire » (Christine, consommatrice rurale).
Enfin la classe 12 (8.8%) est celle du temps (temps passé [heure], horaire
précis [h]). Les jours de la semaine sont cités, les moments de la journée aussi,
avec matinée, nuit, café, boire… Les 2 individus les plus impliqués dans cette
classe sont François et Thérèse (éleveurs urbains). Il n’y a pas de spécificité
forte d’une des 4 catégories car s’il y a surreprésentation de certains
producteurs dans cette classe, d’autres parlent très peu de cet aspect (David
et Théophile). Or les deux producteurs qui sont principalement impliqués
dans cette classe se sont installés dans un cadre familial (ils ont repris
l’exploitation de leurs parents). Alors que ceux qui en parlent le moins sont
des hors cadres familiaux. La littérature (Dufour et Lanciano, 2012) souligne
que les contraintes temporelles sont plus importantes dans le cadre d’une
production en circuits courts. Cette dernière serait vécue différemment en
fonction de la trajectoire des agriculteurs (cadres ou hors cadres familiaux).
Le caractère commun de ces classes nous permet de proposer quelques pistes
de réflexions concernant les liens qui se nouent entre producteurs et
consommateurs. La classe 5 (celle du doute) renvoie partiellement à une
forme de réflexivité partagée par ces deux groupes. Le respect des animaux
et de la nature (classe 10)10 et l’aspiration à un mode de vie, un souci de soi
(classe 7) dessinent un lien entre préoccupations personnelles et engagements
globaux (respect des animaux et cause environnementale) (Pleyers, 2011).
Enfin, la classe 4 souligne l’horizon commun que constitue l’importance du
lien social attaché aux circuits courts.

10 Cette classe commune émerge dans le discours alors qu’elle n’est pas un thème
des deux guides d’entretiens.

820

JADT’ 18

2.2. Pronoms personnels et spécificités
L’analyse des spécificités des 4 catégories d’interviewés, toutes classes
confondues, a mis notamment en évidence un emploi très différencié des
pronoms personnels. Les consommateurs ruraux citent souvent deux des
producteurs par leur prénom. Le lemme discuter est également présent. Donc
ils parlent de gens avec lesquels ils sont en lien fort.
Les consommateurs urbains citent beaucoup je et j, ainsi que vous : « Oui et
puis […] si vous voulez vos salades au bout de 3 ou 4 jours en grande_surface elles
ont pas été vendues elles ont quand même pas la même tête que celles que j’achète qui
ont été cueillies la veille hein » (Mélissa, consommatrice urbaine). Il est donc
question de ce que l’interviewé fait (je, j) et de ce qu’il ne fait pas (vous). Donc
de son comportement d’achat : ce qu’il achète, du lieu où il achète ou
pas (marché, supermarché, …), de la façon dont c’est produit ou vendu (bio,
label, équitable, local, transport). Il y a également le lemme rencontre : le lien est
présent, mais de façon plus conceptuelle, moins proche que dans le groupe
des consommateurs ruraux.
Chez les producteurs ruraux les pronoms tu et nous sont très employés. Le
nous peut renvoyer à un couple de producteurs (Georges et Gina) ou à une
communauté à laquelle on appartient : (les producteurs diversifiés, les
producteurs du marché du village rural) : « Nous ce qui fait la caractéristique du
secteur c’est que c’est des exploitations qui sont tournées vers beaucoup d’espèces on
n’a pas de spécialisation enfin pas de très très grosse spécialisation » (David,
producteur rural). Il nous semble que cette spécificité dans l’utilisation des
pronoms peut-être rattachée à la nature différente des dispositifs (et non à
leur caractère rural ou urbain). Dans un cas, le marché de plein vent
traditionnel, nous avons affaire à une structure de taille importante qui
préexiste aux acteurs. S’il est bien un lieu de rencontre, il est plus fortement
marqué par une dimension individuelle tant pour les producteurs que pour
les consommateurs (d’où la présence du je). Dans l’autre, le petit marché de
producteurs engagés, nous avons affaire à un projet de taille plus réduite
construit par une partie des acteurs. Les relations interpersonnelles,
l’identification à un ou des collectifs mais également la dimension
participative y sont donc plus marquées.
3. Conclusion et perspectives
De nombreux thèmes sont apparus fortement dans le discours des
interviewés : l’importance des relations, l’importance d’acheter au
producteur des produits bio, de manger des produits de saison, d’utiliser des
variétés de blé ancienne, de respecter l’environnement et les animaux.
D’autre part, l’emploi de pronoms personnels différents et l’usage ou non de
prénoms, révèlent une proximité avec les producteurs locaux (discours des

JADT’ 18

821

consommateurs ruraux), l’appartenance à un groupe (discours des
producteurs ruraux), une norme dans le comportement d’achat (discours des
consommateurs urbains). Il est important de ne pas tenir compte uniquement
de la spécificité globale d’une catégorie (ou d’un interviewé) pour juger de sa
plus ou moins grande implication dans une thématique (cas de Jacques). De
ce fait, les thèmes révélés par la classification ne sont pas toujours très
spécifiques d’une catégorie. Malgré un corpus restreint et spécifique, la
textométrie permet de mettre au jour des éléments factuels identifiés dans la
littérature et d’esquisser des liens analytiques avec des approches théoriques
plus générales. Ces résultats nous amèneront à poursuivre ce travail, dans le
cadre du projet Strenght2Food, en y intégrant une comparaison
internationale (avec tout ou partie du corpus des 6 pays partenaires sur cette
thématique).
Références
Chiffoleau Y., Prévost B. (2012). Les circuits courts, des innovations sociales
pour une alimentation durable dans les territoires, Norois, 224.
Duboys de Labarre M. (2005). Le mangeur contemporain, une sociologie de
l’alimentation. Thèse de sociologie, soutenue à Bordeaux, 426p.
Dubuisson-Quellier S. (2009). La consommation engagée. Paris, Presses de la
Fondation nationale des sciences politiques (Contester).
Dufour A., Lanciano E. (2012). Les circuits courts de commercialisation: un
retour de l'acteur paysan ? Revue Française de Socio-Économie (n° 9), pp.
153-169.
Mundler P., Rouchier J. (2016). Alimentation et proximités: Jeux d’acteurs et
territoires. Educagri.
Pleyers G. (dir.) (2011) La consommation critique, mouvements pour une
alimentation responsable et solidaire. Desclée de Brouwer.
Prigent-Simonin A-H., Hérault-Fournier C. (2014). Au plus près de l’assiette.
Editions Quæ.
Reinert M. (1983). Une méthode de classification descendante hiérarchique :
application à l’analyse lexicale par contexte. Les cahiers de l’analyse des
données, VIII(2) :187-198.
Traversac J.B. (2010). Circuits courts : contribution au développement
régional. Educagri.

822

JADT’ 18

On the phraseology of spoken French: initial
salience, prominence and lexicogrammatical
recurrence in a prosodic-syntactic treebank
Rhapsodie
Maria Zimina, Nicolas Ballier
Université Paris Diderot
mzimina@eila.univ-paris-diderot.fr; nicolas.ballier@univ-paris-diderot.fr

Abstract
This paper focuses on specific quantitative characteristics of spoken language
phraseology in the Rhapsodie speech database (ANR Rhapsodie 07 Corp-03001). A recent study (Zimina & Ballier, 2017) has shown that prosodic
segmentation into IPE: Intonational PEriods (segments of speech with
distinctive pitch and rhythm contours) available within the Rhapsodie
database offers new insights for the observation of the functions of formulaic
expressions in speech. Recurrent lexicogrammatical patterns at the beginning
of Intonational PEriods (IPE) are strongly related to spoken formulaic
language. These variations of initial salience depend upon several factors
(interactional needs, social context, genres, etc.). Further experiments have
shown that initially salient patterns also have specific prosodic characteristics
in terms of prominence (prosodic stress) across major speech genres of the
Rhapsodie dataset (oratory, narrative, description, argumentation, procedural)
and corresponding speaking tasks. These specific prosodic characteristics are
likely to reflect communicative needs of speakers and listeners (interactions,
uptakes, speaking turns, etc.).
Keywords: phraseology, prosodic constituents, prominence, salience,
textometrics
1. Introduction
Our research examines the notions of phraseology and formulaic language in
speech production on the basis of prosodic transcriptions indicating specific
events in speech: boundary tones, pitch accents, disfluent segments, etc. (Yoo
et Delais-Roussarie, 2009). We believe that such speech events coded in
spoken corpora are relevant for identifying the prosodic characteristics of
formulaic language.
Corpus-based studies of phraseology often exploit recurrent patterns
detected using repeated segments, co-occurrences and pattern-matching

JADT’ 18

823

techniques to explore formulaic strings of written texts (Granger, 2005; Sitri et
Tutin, 2016). This approach seems equally applicable to oral discourse.
Following this approach, our initial objects of study are predictable and
productive sequences of signs called lexicogrammatical patterns (lexical signs,
grammatical constructions). Made of permanent ‘pivotal’ signs and a more
productive ‘paradigm’, these patterns may be discontinuous and may or may
not be syntactic constituents (Gledhill, 2011; Gledhill et al., 2017). For
example:
§ et donc euh c'est pour ça qu'aujourd'hui je suis en italien en XXX …
§ c'est-à-dire § ouais § un mois c'est pour ça que ça s'appelle radio Timsit …
§ mais bien sûr donc <c'est pour ça bien sûr bien sûr que je parlais oui XXX …
§ c'est pour cela que je tenais à vous rencontrer la veille de notre fête …
We then explore the ways in which prosodic features may correlate with
extended lexical patterns, as well as the extent to which prosody corresponds
to patterns which have a particular register or discourse function. These
lexicogrammatical patterns combined with prosodic features extracted from
speech databases are possible methodological tools for identifying
phraseological characteristics of oral discourse.

Figure 1: Characteristic elements of initial prominence across speech genres: Clit + Verb in
initial position of Intonational PEriods (IPE)

824

JADT’ 18

2. The Rhapsodie speech database
Large spoken corpora are rarely distributed with a fine-grained prosodic
annotation. Fortunately, for French, a reference corpus, the Rhapsodie speech
database (ANR Rhapsodie 07 Corp-030-01), is freely available online
(http://www.projet-rhapsodie.fr/). This syntactic and prosodic treebank is
composed of 57 short samples of spoken French (approximately 5 minutes
long), orthographically and phonetically transcribed (approximately 33,000
words).
2.1. Database structure
The Rhapsodie data file is available online in tabular form. The corpus data is
structured as follows: all Rhapsodie texts are first identified by codes. Each
text is further divided into separate units and segmented into tokens. The
remaining columns display more than 60 linguistic annotations of the tokens,
including microsyntax (rection, dependency, constituency), macrosyntax and
prosody. This data set was transformed from the spreadsheet format into a
Trameur base file using regular expressions (Fleury, 2013). The data structure
is composed of two parts: (1) a Thread, which is a list of items with position
identifiers; (2) a Frame, which is a list of corpus partitions defined on the
Thread. Each partition has a name and a list of named constituents identified
through their first and last token positions on the Thread. Thus, each
annotated token from the Rhapsodie corpus becomes an item identified by its
position on the Thread (Fleury et Zimina, 2014).
2.2. Prosodic annotation
The corpus covers several discourse types and speaking styles: oratory,
narrative, description, argumentation, procedural; interactive, semiinteractive and non-interactive; public and private, planned, spontaneous
and semi-spontaneous, etc. (Lacheret et al., 2017). The transcriptions and the
annotations are aligned on the speech signal (Lacheret et al., 2014). A
combination of manual and automated annotations allowed a segmentation
of speech into prosodic periods (Lacheret et Victorri, 2002), which relies on
the initial characterization of two types of speech events retained from the
manual annotation: prosodic prominence and disfluencies. Organized
around rhythmic and melodic components, the hierarchy of prosodic
constituents includes: Intonational PEriods (IPE); Intonational PAckages
(IPA): sub-constituents internal to periods; Rhythmic Groups (RG): subconstituents internal to intonational packages; Metrical Feet (MF): subconstituents inside rhythmic groups; Syllables, with Prominence levels,
including: 0 (non-prominent), W (weak) and S (strong).

JADT’ 18

825

3. Quantitative analysis of the prosodic dimension of phraseology
As the link between the “marked status” as a +phrase/expression/formulaic
expression etc. and prosodic constituents is still to be revealed, some of our
research questions are of an exploratory nature and more than 60 layers of
morpho-syntactic, syntactic, macro-syntactic and prosodic annotation in
Rhapsodie necessarily open new perspectives for the exploration of the
prosodic dimension of phraseology.
3.1. Preliminary research
Previous research on spoken phraseological units (Lin, 2013) did not take
into account the prosodic hierarchy, in other words, the various sizes of the
prosodic constituents (Nespor et Vogel, 2007). We first replicated this
methodology, which consisted in describing the prosodic characterisations of
phraseological units attested in speech-to-text transcription. For example, in
the categorization of the stress hierarchy, we can distinguish between
stressed (strong) and less prominent (weak) syllables.
Preliminary analyses of repeated segments from the Rhapsodie corpus, such
as jeune fille (F=20) or je veux dire (F=21) led us to observe the non-congruency
of the recurrence of prosodic features, such as prominence, and traditional
phraseological units recovered from transcribed speech data:
… une|dis-weak jeune|strong fille|weak euh|weak habillée|weak tout|strong
en|strong
…
… c|tail'|tail est|tail une|tail jeune|tail fille|tail # pauvre|strong et|strong
affamée|strong
…
… dans|strong la|strong rue|strong avec|weak une|weak jeune|weak
fille|filled-dis
#
§ vous|weak voyez|weak ce|weak que|strong je|strong veux|strong
dire|strong #
§ euh|tail vous|weak voyez|weak ce|weak que|weak je|weak veux|weak
dire|weak §
Our first analyses of these examples made us simultaneously consider
multiple prosodic properties of these collocations, as well as the necessity of
taking into account the hierarchy of prosodic units corresponding to these
speech contexts.
3.2. Prosodic constituents: a genre-based analysis of lexicogrammatical
recurrence and initial salience
A recent study (Zimina et Ballier, 2017) has shown that prosodic
segmentation into IPE: Intonational PEriods (segments of speech with

826

JADT’ 18

distinctive pitch and rhythm contours) available within the Rhapsodie
database offers new insights for the observation of the functions of formulaic
expressions in speech. Lexicogrammatical patterns at the beginning of IPE
are strongly related to spoken formulaic language. Recurrent prominences
observable after speech breaks can be revealed by textometric analysis of
repeated Part-Of-Speech (POS) segments (Salem, 1987) at the beginning of
the IPEs. This method can be used to isolate a set of pivotal elements
associated with what is commonly perceived as a strong prosodic boundary.
Computation of characteristic elements (Lebart et al., 1998), applied to
repeated segments (Salem, 1987), describes these regularities with respect to
genre and speaking styles, categorized in Rhapsodie as ‘subgenres’ (Lacheret
et al., 2014).
Discovered variations of initial salience depend upon several factors
(interactional needs, social context, genres, etc.). They reflect specific
communicative needs of the speakers. For example, Cl + V is a positive
characteristic element at the beginning of the IPE in the speech contexts of
oratory genre (specificity index: +10). The following examples reveal some
lexicogrammatical realizations of this productive pattern in Rhapsodie
(categories corresponding to Cl + V appear in bold):
# il faut les faire grandir #
# § ce sera un coup franc #
# je souhaite que l'Europe #
These pivotal elements reflect the structure of regular rhetorical units with a
predictable/definable discourse function (performative utterances).
3.4. Prosodic salience and prominence: systemic combination
To fine-tune our study of regular prosodic features of phraseology, we have
added another layer of analysis, namely the prominence of final syllables. For
these purposes, we have combined three annotation levels: (1) the positional
properties of units within the IPE structure with BILOU tags (Beginning,
Inside, Last, Unit-length and Outside); (2) POS tags; (3) final stress
prominence: strong, weak, pause_ and % (inaudible or non-transcribed due to
overlap). The base file has been automatically re-annotated with Le Trameur
to add this new combined annotation layer to the Rhapsodie data. Repeated
segments have then revealed that at the beginning (B) of IPEs, the final
prominence of the pivot Cl + V is either weak weak (F=114) or strong strong
(F=95). On Figure 1, the results of characteristic elements analysis in different
speech genres show that these weak realisations of final syllables are positive
characteristic elements (specificity index: +06) of procedural utterances, such

JADT’ 18

827

as instructions in travel planning, while both strong (+06) and weak (+05)
realisations of Cl + V are characteristic elements of initial prosodic salience in
oratory genre. A finer-grained analysis according to speaking task, presented
on Figure 2, shows that this prosodic richness can be attributed to the
subgenre of political discourse, a well-known subtle and complex discourse
phenomenon (Dorna, 1995; Mayaffre, 2002). For political speech, pragmatic
strategies influence the choice of a weak weak (+06) or of a strong strong (+03)
sequence to realize specific discourse functions, for example:
... reculer la pauvreté # § ce|strong sera|strong tout le sens du combat de la
France… (focus) … ces valeurs # § en les faisant vivre # nous|weak
serons|weak plus forts pour aborder les temps qui viennent… (fonction
performative)
The strong prominence of Cl + V also corresponds to emphatic realisations at
the beginning of IPE in sermons (+03), as evidenced in Figure 2. Similarly,
advertising favours recurrent overuse of stressed syllables in this position
(+04).

Figure 2: Characteristic elements of initial prominence in different speaking tasks: Clit + Verb
in initial position of Intonational PEriods (IPE)

4. Conclusions
The prosodic hierarchy (Nespor et Vogel, 2007) acknowledges several layers

828

JADT’ 18

of granularity, from the prosodic utterance to the phoneme. For our
investigation, the IPE was the best initial candidate for the proper level of
granularity in the prosodic hierarchy. There were structural reasons for this,
such as the fact that speaking turns were likely to be signalled in initial
position of IPE. The textometric analysis of this prosodic constituent of the
prosodic hierarchy has shown revealing features such as the limited
distribution of POS categories in the initial position, as well as the role of
prosodic prominence (stressed syllables) and its relevance for the distinction
of speech genres. It appears that the recurrent patterns reflected by such
sequences as “je salue”, “elle souhaite”, “il faut”, “on continue” are not unlike
the stable lexicogrammatical patterns that can be observed in written data. In
all likelihood, the initial characteristic distributions with specific prosodic
characteristics correspond to communicative needs (interactions, uptakes,
speaking turns, etc.).
Because of the complexity of the Rhapsodie speech database, we regard these
explorations as preliminary: we have only based our analysis on few layers
of annotations. Besides, other layers of granularity within the prosodic
hierarchy might also be relevant: Intonational PAckage (IPA), Rhythmic
Group (RG), etc. We also surmise that other variables (channel, planning
type, event structure: monologal vs. dialogal tasks) are likely to reveal related
features.
5. Future research
In future work, various lines of investigation can be pursued, such as the
examination of the various layers of the prosodic hierarchy. Looking for
collocational structures may lead us to question the recurrence of patterns
within prosodic units, in other words, the embedding of prosodic
constituents or the complexity of the boundaries of the constituents across
the layers of the prosodic hierarchy. Other correlates might be considered
such as duration and prosodic contours (the whole corpus also includes the
sound files). It may well be the case that annotators were influenced by the
genre of the recordings, and auditory analysis across genres based on our
characteristic elements analysis may nuance the levels of prominence
assigned by the annotators. If the identification of these collocational
prosodic patterns is robust, it should also remain transparent and decisive for
subjects when resynthesized. The acoustic signal can be modified so as to
erase lexical contents, only keeping the melody (humming). Resorting to
humming should enable us to test the relevance of prosodic sequences which
should robustly remain identifiable as characteristic signals of collocations in
perception tests.

JADT’ 18

829

References
Dorna, A. (1995). Les effets langagiers du discours politique. Hermès, La Revue
1995/2 16, 131–146.
Fleury, S. (2013). Le Trameur. Propositions de description et d’implémentation des
objets textométriques. Sorbonne nouvelle – Paris 3, http://www.tal.univparis3.fr/trameur/trameur-propositions-definitions-objetstextometriques.pdf
Fleury, S., Zimina, M. (2014). Trameur: A Framework for Annotated Text
Corpora Exploration. In: Proceedings of 25th International Conference on
Computational Linguistics (COLING 2014), Dublin, Ireland, pp.57–61,
http://www.aclweb.org/anthology/C14-2013.pdf
Gledhill, C. (2011). The ‘lexicogrammar’ approach to analysing phraseology
and collocation in ESP texts. ASp (Anglais de Spécialité) 59, 05–23.
Gledhill C., Patin S., Zimina M. (2017). Identification et visualisation de
schémas lexico-grammaticaux caractéristiques dans deux corpus
juridiques comparables en français. CORPUS 17, 113–143.
Granger, S. (2005). Pushing back the limits of phraseology. How far can we
go? In: Cosme, C., Gouverneur, C., Meunier, F., Paquot, M. (eds.):
Proceedings of PHRASEOLOGY 2005. An Interdisciplinary Conference,
Université Catholique de Louvain, Louvain-la-Neuve, pp. 165–168.
Lacheret, A., Kahane, S., Beliao, J., Dister, A., Gerdes, K., Goldman, J-P., Obin,
N., Pietrandrea, P., Tchobanov, A. (2014). Rhapsodie: a Prosodic-Syntactic
Treebank for Spoken French. In: Proceedings of the Ninth International
Conference on Language Resources and Evaluation (LREC'14), Reykjavik,
Iceland.
Lacheret, A., Kahane, S., Pietrandrea, P. (eds.) (2017). Rhapsodie: a prosodic and
syntactic treebank of spoken French, John Benjamins, AmsterdamPhiladelphia.
Lacheret, A., Victorri, B. (2002). La période intonative comme unité d’analyse
pour l’étude du français parlé : modélisation prosodique et enjeux
linguistiques. Verbum 24 (1-2), 55–73.
Lebart, L., Salem, A., Berry, L. (1998). Exploring Textual Data. Kluwer
Academic Publishers, Dordrecht, Boston.
Lin, Ph. M.S. (2013). The prosody of formulaic expression in the
IBM/Lancaster Spoken English Corpus International Journal of Corpus
Linguistics. International Journal of Corpus Linguistics 18(4), 561–588.
Mayaffre, D. (2002). Discours politique, genres et individuation sociolinguistique. In: Morin, A., Sébilot, P. (eds.). Actes des JADT 2002, SaintMalo, France, IRISA-INRIA, pp.517–529.
Nespor, M., Vogel, I. (2007). Prosodic Phonology. Berlin. Mouton De Gruyter.
RHAPSODIE Homepage, http://www.projet-rhapsodie.fr

830

JADT’ 18

Salem, A. (1987). Pratique des segments répétés. Essai de statistique textuelle.
Klincksieck, Paris.
Sitri, F., Tutin, A. (dir.) (2016). Phraséologie et genres de discours. Patrons,
motifs, routines. LIDIL 53.
Yoo, H-Y, Delais-Roussarie, E. (eds.) (2009). Actes de la conférence Interface
Discours
&
Prosodie
(IDP
2009),
Paris,
France,
http://makino.linguist.jussieu.fr/idp09/actes_fr.html
Zimina, M., Ballier, N. (2017). Intonational PEriods (IPE) and Formulaic
Language: A Genre-based Analysis of a French Speech Database.
Proceedings of Europhras 2017 Conference: Computational and Corpus-based
Phraseology: Recent Advances and Interdisciplinary Approaches, London,
http://www.tradulex.com/varia/Europhras2017-II.pdf

Abstracts

JADT’ 18

833

What kind of contributions does research provides?
Mapping issue based statements in research abstracts
Filippo Chiarello1, Gualtiero Fantoni1,
Andrea Bonaccorsi1, Silvia Fareri2
1

School of Enginner, University of Pisa – filippo.chiarello@destec.unipi.it
2Marco Biagi Foundation, University of Modena and Reggio Emilia

Abstract
Sentiment analysis is the study of the polarity (positive/negative) of
documents (Pang and Lee 2008). Lexicon based techniques are one of the
most used methods to compute the polarity of a document. Lexicons are
collections of words, each annotated with its own positive or negative
orientation. The overall sentiment of a document is therefore computed upon
the prior polarity of the contained words.
Given that one of the most used approach to build sentiment lexicons is
terms extraction from polarized documents (e.g. restaurant or movie reviews,
social network post), we may hesitate to apply it to styles of text dramatically
different from what they were validated on. Lexicons need to be redesigned
every time we want to shift from one context to another.
Another well-known problem of lexicon approaches is that writers make use
of valence shifters. These are words that affects the polarized words: a
negation flips the sign of a polarized word (e.g., “it is not good”), an
amplifier increases the impact of a polarized word (e.g., “I totally enjoy
that.”), or a de-amplifier reduces the impact of a polarized word (e.g., “it is
almost perfect”). So if valence shifters occur frequently a simple dictionary
lookup may not be measuring the sentiment appropriately.
In this paper, we try to address both these problems with a specific focus on
extracting negative sentences from abstracts of scientific articles. We thus
propose a negative sentence extractor for appropriately dealing with the
research paper domain. We will start with the collection of a set of abstracts
belonging to the same field of knowledge. These documents are then preprocessed using state of the art natural language processing tools (sentence
splitter, tokenizer, lemmatization). Then for each sentence we will compute a
negative sentiment polarity and take in to consideration only the sentences
having a negative polarity score below a thresh-hold level. Finally we will
apply topic modelling algorithm on the negative sentences, with the aim of
give a graphical synthesis of the main problems of a research field.
For the polarity computation, we will rely on a lexicon developed by the
authors that extracts advantages and disadvantages of inventions from

834

JADT’ 18

patents (Chiarello, 2017) and a novel dictionary lookup approach that tries to
incorporate weighting for valence shifters (Rinker 2018). Following a bottom
up-approach we will redesign these lexicon for an optimal application on
paper documents.
References
Pang B. and Lee L. (2008). Opinion Mining and Sentiment Analysis,
Foundation and Trends in Information Retrieval, vol.(2)
Chiarello, F., Fantoni, G., Bonaccorsi, A., (2017). Product description in terms
of advantages and drawbacks. Exploiting patent information in novel
ways. ICED
Rinker, T. W. (2018). sentimentr: Calculate Text Polarity Sentiment version
2.3.2. http://github.com/trinker/sentimentr

JADT’ 18

835

Technical sentiment analysis: predicting the success
of new products using social media
Filippo Chiarello1, Giacomo Ossola1, Gualtiero Fantoni1,
Andrea Bonaccorsi1, Andrea Cimino2, Felice Dell’Orletta2
2

1School of Engineer, University of Pisa – filippo.chiarello@destec.unipi.it
Institute for Computational Linguistics of the Italian National Research Council

Abstract
Nowadays, social media has become an inseparable part of modern life,
providing a vast record of mankind’s everyday thoughts, feelings, and
actions. For this reason, there has been an increasing interest in research of
exploiting social media as information source of knowledge although
extracting a valuable signal is not a trivial task. In fact, social media data is
noisy and must be filtered before proceeding with the analysis. In this
domain, sentiment analysis, which aims to determine the sentiment content
of a text unit, is considered one of the best data mining method. It relies on
different approaches (Collomb, 2013): machine learning, lexicon-based,
statistical and rule-based.
In this work, we try to understand if sentiment analysis is really the best
available method to analyze consumer’s perception of products. In
particular, we compare state of the art sentiment analysis based on machine
learning methods with a lexicon approach based on a dictionary of
advantages and drawbacks related to products, important aspects evaluated
by consumers during the buying decision process. The lexicon has been
developed by researchers of the authors (Chiarello, 2017) to extract
advantages and drawbacks of inventions from patents.
Our work started with the selection of an event able to polarize Twitter users’
attention and a set products to analyze. In particular, we chose a premiere
trade-show for the video game industry, and two video game consoles
disclosed during the event. We collected tweets about products published
before, during and after the trade-show. Since social media data is noisy (for
example it may contains spam and advertising), before proceeding with the
analyses, we filtered our dataset. In particular, after removing too short and
non-English tweets, we manually classified a randomly extracted subset of
posts to train the automatic classifier which provide us the cleansed dataset.
Finally, we built product-related clusters of tweets.
Once obtained the final dataset, we conducted a sentiment analysis of the
posts using state of the art machine learning techniques. We classified each
tweet as positive, negative or neutral. Then, we applied our lexicon

836

JADT’ 18

identifying advantages tweets and drawbacks tweets. In order to compare
the outputs of the two analyses, we considered advantages tweets as
positive, drawbacks as negative, and tweets without words from our lexicon
as neutral. We found consistent and interesting differences between the two
methodologies. In particular we found that when a product has a certain
technological complexity and fuels a more technical social media discourse,
sentiment analysis seems to be less performing while advantages and
drawbacks analysis is abler to produce technical-functional judgements
about the products.
References
Collomb C. , Costea ,C. and Brunie L. (2013). A Study and Comparison of
Sentiment Analysis Methods for Reputation Evaluation.
Chiarello, F., Fantoni, G., Bonaccorsi, A., (2017). Product description in terms
of advantages and drawbacks. Exploiting patent information in novel ways.
ICED

JADT’ 18

837

Citizens and neighbourhood life: mapping
population sentiment in Italian cities
Fiorenza Deriu, Domenica Fioredistella Iezzi2
1

Sapienza University of Rome – email@provider.be
2Tor Vergata University – email@provider.be

Abstract 1
In recent years, citizens are increasingly taking an active part in neighbourhood life
by writing comments on social media. Citynews is an online platform, that is active in
45 Italian provinces, and offers to readers facts and news about Italy. It records every
month 85 million visits and counts over 565 thousand registered users. It has the
characteristic of being a citizen journalism: an increasing part of these contents are in
fact write to the platform by readers. We scraped the comments of users a sample of
6,418 posts published on the Citynews site from January 2017 to January 2018, on 9
metropolitan cities (Bari, Bologna, Florence, Genoa, Milan, Naples, Rome, Turin and
Venice). We applied two analysis procedure: 1. Exploratory analysis of the contents of
citynews posts with the aim of classifying the opinions of active citizens; 2) Sentiment
analysis of the sampled cities, with assignment of scores by postcode. In the first
phase, we mapped the contents of the reviews and identify the areas in which citizens
report major events (positive or negative) occurred in the zone. In the second phase,
we built a sentiment index, normalized by postcode and designed the maps of
citizenship mood.
Abstract 2
Negli ultimi anni, i cittadini partecipano sempre più attivamente alla vita di quartiere
scrivendo commenti sui social media. Citynews è una piattaforma online, attiva in 45
province italiane, che offre ai lettori notizie e commenti sull'Italia. Registra ogni mese
85 milioni di visite e conta oltre 565 mila utenti registrati. Ha la caratteristica di essere
un citizen journalism: una parte crescente di questi contenuti è infatti scritta sulla
piattaforma dai lettori. Abbiamo raccolto i commenti degli utenti su un campione di
6.418 post pubblicati sul sito Citynews da gennaio 2017 a gennaio 2018, su 9 città
metropolitane (Bari, Bologna, Firenze, Genova, Milano, Napoli, Roma, Torino e
Venezia). Abbiamo applicato due procedure di analisi: 1. Analisi esplorativa dei
contenuti dei post delle città con l'obiettivo di classificare le opinioni dei cittadini
attivi; 2) Analisi del sentimento delle città campionate, con assegnazione dei punteggi
per codice postale. Nella prima fase dell'analisi, abbiamo mappato i contenuti delle
recensioni e identificato le aree in cui i cittadini segnalano i principali eventi (positivi
o negativi) accaduti nella zona. Nella seconda fase, abbiamo costruito un indice di
sentiment normalizzato, e disegnato le mappe dell'umore della cittadinanza.
Keywords: active citizenship, map, sentiment analysis.

838

JADT’ 18

Vax network: profiling influential nodes with social
network analysis on twitter
Francesca Di Carlo, Rosy Innarella, Brizio Leonardo Tommasi
Tor Vergata University – francesca.dicarlo90@gmail.com;
rosy.innarella@gmail.com; brizio.tommasi@gmail.com

Abstract
We live in a society increasingly conditioned by opinions of third parties. If our
objective is to spread a new concept or idea effectively into the society and the
public opinion, we must consider two basic factors: the presence of influential
individuals that accelerate the spread of the process and the susceptibility of
people to be “infected” by the new idea. In order to keep under control the
spread of new behaviors, it is very important to collect as much as information
about the profiles of the people belonging to the two categories (Sawyer, 2011).
This paper analyses a social network based on two clusters of people. The first
cluster identifies people with opinion “pro-vax” the second one identified by
“anti-vax” opinions emerged in the last three years. This analysis consists of a
description of a data collection of public tweets available online; every collection
is based on about one month of tweets per year. Tweets are extracted with
specific “key-words” in the context of vaccine and anti-vaccine factors. In
particular the data collection is based on a “key-words search model” consisting
of a combination of context words, for example: (i) “core cluster” with:
vaccination, vaccine, vaxxer, antivaccine, antivaxxer, anti-vaccine, anti-vaxxer; (ii)
“effects cluster” with vaccine dependent variable: Autism, MMR, Pharma and
vaccine; (iii) “community cluster” with vaccine dependent variable: Cdc, Aaps,
Fda, Hrc, Wakefield and vaccine. We used twitter fetcher included in the
semantic and social network analysis software Condor (Gloor, 2009); in this way
we could fetch 120 days of tweets distributed over three years: apr-2015, nov2016, jun-2017. The total volume of collected tweets is about 300,000. All the
collected tweets were written in English. In this elaboration we were able to
distinguish behavior emerged in tweets in each cluster analysis by carrying out a
preliminary sentiment analysis of the collected tweets. This network of Twitter,
named “vax network”, consists of about 800,000 links with an average degree of 3
relations. The analysis of “vax network” is based on the mainly centrality
measures (e.g. degree, betweenneess, closeness, clustering coefficient,
eigenvector) for the identification of relevant node in terms of potential influence
of “vax networks”. Instead of the sentiment analysis is based on main methods
(e.g. word frequencies, activity, emotionality, sentiment, complexity) of natural
language used in the tweets of “vax network”, for profiling the relevant node in
terms of its main characteristic correlated with own centrality in the network.

JADT’ 18

839

Consequently we can identify the profiled nodes that have prevalent behavior for
influencing or clustering the trend of the “vax network”. In the last three years we
have identified a specific trend of the sentiment analysis, calculated on the twitter
discussions about vaccine network. The sentiment analysis of the tweets has a
negative meaning, regardless of using positive or negative words in the vaccine
context. In general the 80/20 law emerges between “pro” and “anti” nodes. The
sentiment together with emotionality has a growing positive trend, due to an
awareness communication on vaccine utility. Keywords: network analysis,
sentiment analysis, text mining, twitter, profiling, influence, vaccine, complex
network.

840

JADT’ 18

Alteryx
Davide Donna
Managing Partner, The Information Lab

Alteryx is an analytics platform that allows users to build complex
workflows for an end to end data management and analytics without
scripting. This is possible through an intuitive interface where users drag and
drop on the canvas the different tools that he needs to connect, blend,
transform, analyze and export the data.
Alteryx is a Data Science platform that includes three macro areas: ETL with
capability to connect to most of the databases and services, clean, prepare
and blend the data; Advance Analytics, integrating several predictive and
statistical models scripted in R and Pyton; Geospatial analytics.
In the textual data analytics contest Alteryx finds several applications,
presented during the speech:
- web scraping: download html pages and parsing data through regular
expressions (Regex)
- big data management
- Connection to external services for advance analytics. I.e. with the direct
interaction with Azure Cognitive Services Alteryx can implement key
phrases extraction, sentiment analysis and language dectection
- Connection to external services for data input: i.e. with the connectors to
Twitter and other social media the information can be directly imported into
the analytical platform
- Integration of R and Python libraries: the models are written into Alteryx
and integrated and run within the entire flow
- Cleaning, blending and transformation of the information downloaded
from different sources
- Export the results in the favorite format. I.e. Visualanalytcs or reporting tool
like Tableau for a clear and smart representation.

JADT’ 18

841

Complexity of US President Speeches
Valerio Ficcadenti1, Roy Cerqueti2, Marcel Ausloos3,4
Marche Polytechnic University - v.ficcadenti@univpm.it
2University of Macerata - roy.cerqueti@unimc.it
3University of Leicester - ma683@le.ac.uk
4GRAPES - Angleur, Liege, Belgium - marcel.ausloos@ulg.ac.be
1

Abstract
This work is devoted to the exploration of the rhetoric dynamics of a large
collection of US Presidents’ speeches. In particular, speeches are viewed as
complex systems and are first analysed through rank-size laws, being the
words of each speech ranked in terms of their frequencies. The building of
the dataset itself represents a relevant step of the study. In this respect, by
using a web scraping routine on the Miller Center website, a large span of
978 speeches have been downloaded. After a pre-processing phase the set is
reduced to 951; for each one, the words’ frequencies are stored. A best fit
procedure with Zipf-Mandelbrot laws is performed over the 951 talks
individually. Thanks to these estimations, it is possible to reach interesting
conclusions on how 45 United States Presidents, from April 30, 1789 till
February 28, 2017 delivered political messages. Our analysis shows some
remarkable regularities, not only inside a given speech, but also between
different speeches. Results are discussed under a political and linguistics
points of view.
Keywords: US Presidents’ speeches, speeches’ framework, US Presidents’
rhetoric, rank-size analysis Zipf-Mandelbrot law

842

JADT’ 18

Measuring the Dynamics
of Social Networks with Condor
Peter A. Gloor
MIT Center for Collective Intelligence, Cambridge MA – pgloor@mit.edu

Condor is an easy-to-use dynamic semantic social network analysis
tool, which automatically imports many types of communication data such
as e-mail, Twitter, Blogs, Wikipedia, and Facebook, converting it into social
networks for longitudinal network, content and sentiment analysis using
machine learning and deep learning. It has been developed over the last 15
year at the MIT Center for Collective Intelligence together with other
universities (Colgone, Rome Tor Vergata, etc.).
Condor consists of four parts:
1. A series of fetchers to directly load data from e-mail, for example from
Gmail or Exchange, from Twitter, from Google, Wikipedia, and
Facebook.
2. Interactive preprocessing functions to modify and reduce the graph, filter
by content and by structure, annotate by geocode, merge multiple e-mail
addresses, and create modified graphs.
3. Visualization functions, to show the static network, a dynamic movie of
the network over time, geographical word maps, and different views
for structure, content, and sentiment of actors.
4. Export functions to export time series of all variables for later
longitudinal analysis in statistics packages such as R or SPSS (or Excel).

JADT’ 18

843

Condor can….
…Analyze structure, dynamics, and content of E-Mail, constructing the
communication matrix between different influencers.
…Coolhunt the Internet using Twitter, Blogs, and Facebook for trends
and trendsetter to analyze interactions among influencers in different
online information spheres (Twitter, Blogs, Forums, Wikipedia,…)
…Measure collaboration and communication efficiency on seven proprietary
metrics (honest signal of collaboration) by structure, time, and contents:
central
leadership,
rotating
leadership,
balanced
contribution,
responsiveness, honest sentiment, shared context, social capital.
…Find digital tribes on social media using deep learning, and mapping tribal
affiliations to brands, products, and people.

Requirements
Condor runs on Windows, Mac, and Linux, and needs Java (MySQL
optional) installed.
Download Condor and the Condor manual
from http://guardian.galaxyadvisors.com/guardian
Condor is described in Gloor, P. Sociometrics and Human Relationships:
Analyzing Social Networks to Manage Brands, Predict Trends, and Improve
Organizational Performance , Emerald Publishing, London 2017
Contact: Peter A. Gloor, MIT Center for Collective Intelligence and
Galaxyadvisors AG
www.galaxyadvisors.com, www.transparencyengine.org, pgloor@mit.edu

844

JADT’ 18

“BIG DATA” Words Trend Analysis using the
multidimensional analysis of texts
Iolanda Maggio1, Domenica Fioredistella Iezzi2, Matteo Fatighenti3
2

1Rhea Group – iolanda.maggio@esa.int
Tor Vergata University – stella.iezzi@uniroma2.it
3CapGemini – matteo.fatighenti@gmail.com

Abstract
Big data is a term that describes large volumes of high velocity, complex and
variable data that require advanced techniques and technologies to enable
the capture, storage, distribution, management, and analysis of the
information. Big Data includes so many specialized terms that it’s hard to
know where to begin. An evolution in languages was experienced in the last
few years and this paper shows how the relevant terms are changing. This
analysis started as a complement of a thesis activity for the Master
Programme in Data Science of University of Tor Vergata. To perform the text
analysis the IRaMuTeQ software has been used that will be described in the
next paragraphs. The exercise aims at analyse both contents with a
tool/software for multidimensional analysis of texts to give evidence of BIG
DATA words trends. This approach provides the users with different text
analyses, either simple ones, such as the basic lexicography related to
lemmatization and word frequency; or more complex ones such as
descending hierarchical classification, post- hoc correspondence factor
analysis and similarity analysis. The vocabulary distribution is presented in a
comprehensive and clear way with graphical representations derived from
the lexicographic analysis. These analyses can be performed using texts
referring to a certain thematic (text corpus) grouped in one text archive; or
data from spreadsheets (matrices with individuals in a row and words in a
column), like the dataset derived from free evocation tests.
Keywords: big data, text mining, Zipf diagram, Clustering, Dendograms,
Corpus.

JADT’ 18

845

Itinerari turistici, network analysis e text mining
Mario Mastrangelo
Tor Vergata University – mario.mastrangelo@uniroma2.it

Abstract
Appare ormai evidente come gli strumenti propri del web 2.0 abbiano
modificato in maniera sostanziale il mondo in cui viviamo, e che le ricadute
di tali strumenti siano di grande entità soprattutto in alcuni settori economici.
Tra questi, sicuramente il turismo. Le opportunità fornite dai cosiddetti user
generated Contents -e dall’avvento dei Big Data- hanno trasformato
sensibimente questo settore, sia dal lato della domanda, in termini di scelta
delle destinazioni, di pianificazione del viaggio, di soddisfazione in merito ai
servizi e ai fornitori, sia dal lato dell’offerta, in termini di analisi dei flussi, di
marketing, e di customer satisfaction. L’ipotesi, sempre più suffragata dai
fatti, è pertanto quella che per una quota crescente dei turisti contemporanei
il parere espresso da propri pari tramite i social media, sia di tipo specifico,
come Tripadvisor, che di tipo generico, come ad esempio Facebook e Twitter,
sia ormai da porre sullo stesso piano e in certi casi possa risultare addirittura
più incisivo di gran parte dei messaggi veicolati dai canali informativi
tradizionali. In questa direzione, il presente contributo intende illustrare una
applicazione pratica di alcune tecniche di analisi dei dati testuali (analisi
delle corrispondenze lessicali, text clustering, sentiment analysis, etc) che,
impiegate in serie o in parallelo, possono contribuire a fornire un valido
supporto per l’individuazione di itinerari turistici sempre più individuali e
"sociali" in un determinato contesto territoriale.
In particolare, in questa simulazione sono stati presi in considerazione 38 siti
di Roma, tutti molto noti, piuttosto diversificati - da monumenti, a palazzi
istituzionali, a sedi museali – e tutti vicini tra loro, così da poter creare piccoli
itinerari raggiungibili a piedi. Questi siti divengono i nodi di un grafo
orientato, i cui archi costituiscono le distanze tra ciascun sito, caratterizzati
da una serie di attributi, alcuni dei quali di natura “tradizionale” come la
classificazione iniziale in base a parametri storico-artistico-funzionali
(periodo romano, periodo barocco, periodo rinascimentale, periodo
contemporaneo, siti museali e siti “misti”, ovvero quelli che non rientrano
con esattezza in nessuna delle precedenti categorie) e altri derivanti dai social
media, nel caso specifico mediante opportune operazioni di text mining su

846

JADT’ 18

un Corpus assemblato a tal fine da un insieme di recensioni in lingua inglese
dei 38 siti su Tripadvisor.
La rappresentazione dei siti sotto forma di grafo, insieme ai due insiemi di
attributi associati ai nodi, permette di eseguire con facilità alcune operazioni
a partire dai desiderata del turista; ad esempio suddividere i siti in base a
come i turisti parlano di loro, e individuare insiemi di percorsi tra i siti
caratterizzati da un sentiment più elevato, e dunque dal fatto che altri turisti,
in posizione simmetrica, hanno espresso con le loro parole un alto
gradimento. In altre parole, dati un punto di partenza e un punto di arrivo, a
percorsi più o meno tradizionali, definiti in modo “oggettivo” (ad esempio,
tour barocco, tour religioso, e, in prospettiva, tour gastronomico, etc) si
potrebbero aggiungere percorsi definiti in modo soggettivo sulla base delle
valutazioni espresse dagli altri turisti (ad esempio, tour delle tappe più
ammirate, più emozionanti, più commentate, etc). In tal modo si unirrebbero
le due anime attuali dell’informazione turistica, quella tradizionale
indispensabile ma più statica e quella basata sui contenuti generati da altri
utenti, effimera e soggettiva ma più dinamica; dal punto di vista scientifico,
questo approccio permette di applicare tecniche ben consolidate ma
raramente utilizzate in maniera congiunta, allo scopo di estrapolare,
interpretare ed impiegare in maniera efficace ed originale la grande mole dei
contenuti generati dai vari utenti in un settore particolarmente fecondo in
questo senso.

JADT’ 18

847

Text Mining per l’analisi qualitativa e quantitativa
dei dati amministrativi utilizzati dalla Pubblica
Amministrazione
Maria Francesca Romano1, Guido Rey2,
Antonella Baldassarini2 Pasquale Pavone4
1Istituto di Economia, Scuola Superiore Sant’Anna, 2Istituto di Management, Scuola Superiore
Sant’Anna, 3ISTAT, 4Centro di Analisi delle Politiche Pubbliche – Università di Modena e Reggio
Emilia

Abstract
Nella Pubblica Amministrazione la maggior parte delle informazioni (sia
qualitative che quantitative) è contenuta in testi in linguaggio naturale e
spesso i testi “nascondono” al loro interno anche molti dati numerici. Articoli
di giornale e interventi in dibattiti televisivi richiamano spesso l’attenzione
dell’opinione pubblica sulla mancata o incompleta informatizzazione della
giustizia. Un esempio positivo sono le sentenze emesse dalla Corte di
Cassazione già visualizzabili ed interrogabili sul sito www.italgiure.it. Si
tratta di un totale di oltre 475.000 documenti a partire dall’anno 2012. E’ già
stata effettuata l’estrazione dal sito Italgiure di un sottoinsieme di sentenze di
merito (circa 4.700), e le analisi testuali condotte con il software TalTac2.0
mostrano (Romano 2017) come sia possibile giungere a buoni risultati nella
verifica della :
 completezza rispetto alla individuazione delle parti offese, degli
imputati, delle eventuali parti civili, e di enti, /aziende pubbliche e
private: I nominativi sono tutti in chiaro ed esistono con pochissime
eccezioni, coperte da omissis.
 attendibilità / certezza rispetto ad eventi criminosi e successiva
elaborabilità con strumenti informatici
 identificazione dei soggetti (Enti Pubblici) e del ruolo di imputati e di
enti / aziende private;
 dell’ammontare economico degli atti criminosi collegati ad attività
economiche (corruzione, aste pubbliche, gare di appalto, ecc.);
 luoghi in cui si sono svolti gli atti criminosi.
 presenza di appartenenti ad organizzazioni criminali e loro insediamenti
Il paper discute i problemi metodologici derivanti dall’uso di testi in
linguaggio naturale di carattere giuridico, propone una metodologia di
estrazione delle informazioni rilevanti ed i vantaggi derivanti da una

848

JADT’ 18

integrazione tra dati e informazioni estratti dalle sentenze di merito e archivi
di natura amministrativa.
Sarà discussa la possibilità di integrare più basi di dati, mettendo in relazione
i dati estratti dalle sentenze con i dati presenti in altre basi di dati anche di
natura statistica; saranno presentati i risultati di un esercizio su di un insieme
di sentenze della Corte di Cassazione da collegare con archivi statistici come,
ad esempio, il registro delle imprese attive tracciando possibili scenari
conoscitivi per un’analisi economica di eventi criminosi.
Componente centrale dell'analisi è il confronto con altre basi di dati,
statistiche e testuali, al fine di acquisire elementi qualitativi e quantitativi su
particolare fenomeni criminosi, nonché convalidare la correttezza dei dati
estratti dalle sentenze non sempre esenti da errori.
Riferimenti bibliografici
Romano M.F. (2017), Dalle parole ai numeri: estrarre dati dalle sentenze della
magistratura, in Rey G.M. (a cura di), La mafia come impresa. Analisi del
sistema economico criminale e delle politiche di contrasto, FrancoAngeli,
pp.121-154.

JADT’ 18

849

Taglio cesareo e Vbac in Italia al tempo dei Big Data:
una proposta di ulteriore contributo informativo.
Alessandro Cesare Rosa
Dipartimento di Epidemiologia del S.S.R. del Lazio – a.rosa@deplazio.it

Abstract 1
The project starts with the curiosity to investigate, for a topic of
epidemiological (but also popular) interest such as the caesarean section in
Italy, which is the level of information and awareness acquired by a sample
of citizens who frequents and writes an opinion on the web, comparing
freely. Conceptually linked to the caesarean section, the theme of VBAC,
acronym of "Vaginal Birth After Cesarean", will be analyzed in parallel with
the same methodology. The objective is to highlight connections, if they exist,
between aspects related to the correct practice of the caesarean section that
the woman and / or the single health care provider should adopt (ministerial
guidelines created for the public) with the opinion expressed by a sample of
web-users. To do so, it was decided to scrape, from the web, the textual
content of the opinions present in some websites and online forums of
frequentation, precisely in order to extract, where possible, semantic macroareas comparable to the concepts of departure
Abstract 2
Il progetto nasce dalla curiosità di indagare, per una tematica di interesse
epidemiologico ( ma anche divulgativo) quale il taglio cesareo in Italia, quale
sia il livello di informazione e consapevolezza acquisita da un campione di
cittadini che frequenta e scrive la propria opinione sul web, confrontandosi
liberamente. Concettualmente legato al taglio cesareo, anche il tema del
“Parto Vaginale dopo Cesareo”,acronimo inglese VBAC, verrà analizzato in
parallelo con la medesima metodologia. Ci si pone l’obiettivo di evidenziare
eventuali connessioni tra gli aspetti legati alla corretta pratica del taglio
cesareo che la donna e/o il singolo operatore sanitario dovrebbero
adottare,delineati nelle linee guida ministeriali rivolte al pubblico, con la
percezione in merito espressa da un campione di web-users. Per fare ciò, si è
deciso di attingere, dal web, al contenuto testuale delle opinioni presenti in
alcuni siti e forum online di stampo e frequentazione volutamente
generalista, proprio al fine di elicitare,dove possibile,macro-aree semantiche
confrontabili con i concetti di partenza.
Keywords: Exploratory Textual Data Analysis, Text Mining, Discourse
analysis.

850

JADT’ 18

Finito di stampare in proprio
nel mese di giugno 2018
UniversItalia di Onorati s. r. l.
Via di Passolombardo 421, 00133 Roma Tel: 06/2026342
email: editoria@universitaliasrl. it – www. universitaliaeditrice.it