Quantifying Approaches to Discourse for Social Scientists

Postdisciplinary Studies in Discourse

Series Editor Johannes Angermuller Centre for Applied Linguistics University of Warwick Coventry, UK

Postdisciplinary Studies in Discourse engages in the exchange between discourse theory and analysis while putting emphasis on the intellectual challenges in discourse research. Moving beyond disciplinary divisions in today’s social sciences, the contributions deal with critical issues at the intersections between language and society. Edited by Johannes Angermuller together with members of DiscourseNet, the series welcomes high-quality manuscripts in discourse research from all disciplinary and geographical backgrounds. DiscourseNet is an international and interdisciplinary network of researchers which is open to discourse analysts and theorists from all backgrounds. Editorial board: Cristina Arancibia, Aurora Fragonara, Péter Furkó, Tian Hailong, Jens Maesse, Eduardo Chávez Herrera, Michael Kranert, Jan Krasni, María Laura Pardo, Yannik Porsché, Kaushalya Perera, Luciana Radut-Gaghi, Marco Antonio Ruiz, Jan Zienkowski More information about this series at http://www.palgrave.com/gp/series/14534

Ronny Scholz Editor

Quantifying Approaches to Discourse for Social Scientists

Editor Ronny Scholz Centre for Applied Linguistics University of Warwick Coventry, UK

Postdisciplinary Studies in Discourse ISBN 978-3-319-97369-2    ISBN 978-3-319-97370-8 (eBook) https://doi.org/10.1007/978-3-319-97370-8 Library of Congress Control Number: 2018958470 © The Editor(s) (if applicable) and The Author(s) 2019 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Cover design by Tjasa Krivec This Palgrave Macmillan imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Acknowledgements

The idea for this volume was born during the First International DiscourseNet Congress in Bremen in late summer 2015. Together with Marcus Müller, Tony McEnery, and André Salem, we had organized the panel ‘Quantifying methods in discourse studies. Possibilities and limits for the analysis of discursive practices’. With scholars of international renown coming from linguistics, statistics, computer sciences, sociology and political sciences in countries such as Canada, France, Germany, Switzerland and the United Kingdom as well as guests from many other countries, the panel was a real success. Papers presented seminal work using a great variety of quantifying methods mostly in combination with qualitative methods. This edited volume is driven by the interdisciplinary and international attitude of the inspiring discussions that we led in the panel. Completing such a fascinating project would not have been possible without a network of international supporters, to name only a few of them: ERC DISCONEX Group and the Professional and Academic Discourses Group, both hosted in Applied Linguistics at the University of Warwick; the Centre d’Étude des Discours, Images, Textes Écrits, Communication (CEDITEC) at the University Paris Est-Créteil; DiscourseLab at the TU Darmstadt; and last but not least, DiscourseNet, which unites discourse researchers across national and disciplinary borders all over the world. v

vi Acknowledgements

This volume would not have seen the light without the support of many colleagues and friends. I am grateful to the anonymous reviewers who provided me with detailed and encouraging feedback. I also thank the series editor Johannes Angermuller and the editorial assistant Beth Farrow from Palgrave for supporting this publication project throughout with great enthusiasm. Finally, I am thankful to my wife Joy Malala and to our new-born son Gabriel Amani for tolerating the extra hours that I had to put into editing this volume.

Praise for Quantifying Approaches to Discourse for Social Scientists

“In today’s complex world of communication there is an urgent need to stand back, to analyse, and to ask what is going on. This multi-national collection of papers by communication scientists does just that. The papers in this book not only provide technical tools of both a quantitative and a qualitative kind, they make possible a perspective that gives objectivity to our understanding of the disturbing world of words in which we flounder.” —Paul Chilton, Emeritus Professor, Department of Linguistics and English Language, Lancaster University, UK “This is a very welcome addition to the literature on quantitative methods for the analysis of language and meaning-making processes. Taking into account textual and social contexts of language use including extra-­textual contexts the volume convincingly demonstrates that quantifying approaches to discourse should not – and cannot – be reduced to the mere counting of words. This book will be of use to students and researchers interested in particular in the challenges posed by big data and technology-­driven language practices.” —Alexandra Homolar, Associate Professor of International Security, Department of Politics and International Studies, The University of Warwick, UK “Bringing together a wide range of quantitative approaches to discourse analysis which stretch far beyond the relatively well established methods of corpus lin-

vii

viii 

Praise for Quantifying Approaches to Discourse for Social Scientists

guistics, this volume provides a great overview for both beginners and experts in the field of quantitative discourse analysis.” —Professor Annika Mattissek, Professor for Economic Geography and Sustainable Development, Department of Geography, University of Freiburg, Germany “In the fast-moving field of text processing, keeping up with methodological innovation can be a real challenge. This volume provides social scientists with a rich and varied palette of approaches and tools, combining sound theoretical foundations with practical advice on methods.” —Professor Gerlinde Mautner, Institute for English Business Communication, Vienna University of Economics and Business, Austria “With the discourse turn in the social sciences, the need for a state of the art guide to practice and theory of meaning construction is evident. In this volume, leading British and continental scholars present quantitative and qualitative methods of exploring discourse and the wider context into which texts are embedded, while discussing and bringing together the approaches of Critical Discourse Analysis and the Foucauldian dispositif. Long overdue!” —Wolfgang Teubert, Emeritus Professor, Department of English Language and Linguistics, University of Birmingham, UK

Contents

Part I Introductory Remarks

   1

1 Understanding Twenty-First-Century Societies Using Quantifying Text-­Processing Methods  3 Ronny Scholz 2 Beyond the Quantitative and Qualitative Cleavage: Confluence of Research Operations in Discourse Analysis 23 Jules Duchastel and Danielle Laberge

Part II Analysing Institutional Contexts of Discourses

  49

3 The Academic Dispositif: Towards a Context-Centred Discourse Analysis 51 Julian Hamann, Jens Maesse, Ronny Scholz, and Johannes Angermuller

ix

x Contents

4 On the Social Uses of Scientometrics: The Quantification of Academic Evaluation and the Rise of Numerocracy in Higher Education 89 Johannes Angermuller and Thed van Leeuwen

Part III Exploring Corpora: Heuristics, Topic Modelling and Text Mining  121 5 Lexicometry: A Quantifying Heuristic for Social Scientists in Discourse Studies123 Ronny Scholz 6 Words and Facts: Textual Analysis—Topic-Centred Methods for Social Scientists155 Karl M. van Meter 7 Text Mining for Discourse Analysis: An Exemplary Study of the Debate on Minimum Wages in Germany183 Gregor Wiedemann

Part IV New Developments in Corpus-Assisted Discourse Studies

 213

8 The Value of Revisiting and Extending Previous Studies: The Case of Islam in the UK Press215 Paul Baker and Tony McEnery 9 The Linguistic Construction of World: An Example of Visual Analysis and Methodological Challenges251 Noah Bubenhofer, Klaus Rothenhäusler, Katrin Affolter, and Danica Pajovic

 Contents 

xi

10 Multi-method Discourse Analysis of Twitter Communication: A Comparison of Two Global Political Issues285 Jörn Stegmeier, Wolf J. Schünemann, Marcus Müller, Maria Becker, Stefan Steiger, and Sebastian Stier Index315

Notes on Contributors

Katrin  Affolter  is a PhD student in a joined program between the Zurich University of Applied Science (ZHAW) and the University of Zurich (UZH), Switzerland. The research topic of her PhD thesis is natural language interfaces for databases. From 2012 to 2016, she studied computer science at the UZH, focusing on computational linguistics and databases. In her master’s thesis ‘Visualization of Narrative Structures’, she developed a web application to explore narrative structures based on an interactive graph visualization. Johannes Angermuller  is Professor of Discourse and the director of the ERC DISCONEX research group at the Centre for Applied Linguistics at Warwick, UK, and EHESS in Paris, France. He is a discourse researcher in linguistics and sociology. His recent publications deal with academic and political discourses and include books such as Poststructuralist Discourse Analysis (2014) and Why There Is No Poststructuralism in France (2015), which have come out in English, French, German, Turkish, Portuguese, and Spanish versions. Paul Baker  is Professor of English Language at the Department of Linguistics and English Language, Lancaster University, UK, where he is a member of the Corpus Approaches to Social Sciences ESRC Research Centre. He specializes in corpus linguistics, using and developing corpus methods to carry out discourse analysis, as well as being involved in research in media language, variation and change and social identities. He has written 16 books, including Using Corpora to Analyse Discourse (2006), Sexed Texts: Language, Gender and Sexuality (2008), and Discourse Analysis and Media Attitudes (2013). He is also commissioning editor of the journal Corpora (EUP). xiii

xiv 

Notes on Contributors

Maria  Becker  is a doctoral researcher at the Department of Computational Linguistics at Heidelberg University. She is also a researcher at Discourse Lab, a research environment for digital discourse analysis at TU Darmstadt. She studied German philology, linguistics, philosophy, communication science, and psychology at the universities of Heidelberg and Mannheim. For her PhD thesis, she works on the automated reconstruction of implicit knowledge in argumentative texts. Her research interests further include deep learning, corpus linguistics, discourse analysis, media linguistics, spoken language and medical communication. Noah  Bubenhofer  is head of the Digital Linguistics group at the ZHAW University of Applied Sciences in Winterthur, Switzerland. His research topics are linguistics, corpus linguistics, and language and digitality. From 2015 to 2018, he was a senior researcher at the Institute of Computational Linguistics at the University of Zurich. For his PhD thesis ‘Muster an der sprachlichen Oberfläche’ (patterns at the linguistic surface), he developed corpus linguistic methods for discourse and cultural analyses. Jules Duchastel  is Emeritus Professor of the Department of Sociology at the Université du Québec à Montréal. His main areas of research are transformation of the welfare state in Canada and Québec starting from the 1940s to present time and computer-assisted analysis of political discourse. He has held, from 2001 to 2008, a Canadian Research Chair on Globalization, Citizenship and Democracy and has founded the Centre for Computer Assisted Text Analysis (ATO) in 1983. Julian Hamann  is a postdoctoral researcher at the Leibniz Center for Science and Society at the Leibniz University Hannover, Germany. His research draws on the sociology of social sciences and humanities, the sociologies of knowledge and culture, and higher education studies. His present work touches on topics like evaluation and boundaries, subjectivity and performativity, academic knowledge and academic careers, and power and social inequality. His current research has recently appeared in Poetics, Higher Education, Zeitschrift für Soziologie, and History of Humanities. Danielle  Laberge  is Emeritus Professor of the Department of Management and Technology at the Université du Québec à Montréal. Throughout her career, she has taught methodology and epistemology and published extensively with Jules Duchastel on these questions. Her recent research deals with project management and governance.

  Notes on Contributors 

xv

Thed  van Leeuwen is a senior researcher at the Centre for Science and Technology Studies (CWTS) of Leiden University in the Netherlands. He is co-­ leading the research theme on Open Science, and the project leader of the Open Science Monitor. As a member of the SES research group, other research topics Thed is involved in relate to the evaluation of research, in particular in the social sciences and humanities, as well in the ways research quality is perceived. The overarching science policy context under which research assessments are organized and the role of bibliometric indicators therein are of major concern for this research agenda. Thed is co-editor of the OUP journal Research Evaluation, as well as associate editor of the Frontiers journal Research Metrics & Analytics. Jens Maesse  is Assistant Professor in the Department of Sociology, University of Giessen. His research focus is on discourse analysis, sociology of science and education, economic sociology and political economy. His publications include ‘Austerity discourses in Europe: How economic experts create identity projects’, Innovation: The European Journal of Social Science Research 31 (1): 8–24 (2018). ‘The elitism dispositif. Hierarchization, discourses of excellence and organisational change in European economics’, Higher Education 73: 909–927 (2017). Tony McEnery  is Distinguished Professor of English Language and Linguistics at Lancaster University. He is currently a Group Director (Sector Strategy) at Trinity College London, on secondment from Lancaster University. Tony was previously Director of Research and Interim Chief Executive at the UK’s Economic and Social Research Council (ESRC). He was also the Director of the ESRC Centre for Corpus Approaches to Social Science at Lancaster. He has published extensively on corpus linguistics. Karl M. van Meter  is a research sociologist at the Centre Maurice Halbwachs (ENS Paris) and an expert in sociological methods and methodologies. He is an American-French citizen with university degrees from the US, the UK and France. Although his PhD was in pure mathematics, he founded and directed for 34  years the bilingual Bulletin of Sociological Methodology/Bulletin de Méthodologie sociologique, which is now with Sage Publications. In his research he uses mainly quantitative text processing methods with which he traces major historical shifts in French, German and American sociologies, and the representation of politics in society. Marcus Müller  is full professor in German Studies—Digital Linguistics at the Department of Linguistics and Literature, Technische Universität Darmstadt. He studied German philology, romance studies and European art history at the

xvi 

Notes on Contributors

universities of Heidelberg and Granada. Müller leads the Discourse Lab, a research environment for digital discourse analysis (http://discourselab.de/). His research interests include digital linguistics, discourse analysis, language and art, and science communication. Danica  Pajovic  obtained her master’s degree in computational linguistics at the University of Zurich. She worked on the project ACQDIVIZ: Visualising Development in Longitudinal First Language Acquisition Data in the comparative linguistics department and was a collaborator in the project Visual Linguistics, led by Noah Bubenhofer at the University of Zurich. Klaus Rothenhäusler  is a junior researcher in the Digital Linguistics group at the ZHAW University of Applied Sciences in Winterthur, Switzerland. He received his master’s degree in computational linguistics from the University of Heidelberg and specialized in distributional semantic models during his time at the IMS Institute for Natural Language Processing in Stuttgart. Over the past years he has worked in numerous digital humanities projects. Ronny Scholz  coordinates the ERC-funded DISCONEX project on academic discourses at the University of Warwick, UK. He holds a master’s degree in discourse studies from the University Paris XII and a PhD in sociology and linguistics from Magdeburg and Paris Est. His work focuses on the question of legitimization of power in political discourses especially in the post-democratic era. He uses lexicometric tools as quantifying heuristic helping to explore new perspectives in various corpora of political discourse. Wolf J. Schünemann  is junior professor of political science with a focus on Internet and politics at Hildesheim University. His research and teaching cover the fields of Internet governance, international relations and European integration. After having studied political science, philosophy, German literature, and media at Kiel University and Sciences Po in Rennes, France, he worked as a research fellow and lecturer at the University of Koblenz-Landau. He received his doctoral degree with a comparative discourse study of referendum debates in France, the Netherlands and Ireland. Jörn Stegmeier  is a postdoctoral researcher at the Department of Linguistics and Literature, Technische Universität Darmstadt. Together with Marcus Müller, he heads the Discourse Lab, a research environment for digital discourse analysis (http://discourselab.de). His research interests include digital linguistics, corpus linguistics, and discourse analysis.

  Notes on Contributors 

xvii

Stefan Steiger  is a research associate at the University of Hildesheim and doctoral student at the Institute of Political Science at Heidelberg University. He studied political science, history and philosophy at Heidelberg University. His research interests include cybersecurity, Internet governance, political communication and foreign policy analysis. Sebastian Stier  is a postdoctoral researcher in the Department Computational Social Science at GESIS—Leibniz Institute for the Social Sciences in Cologne, Germany. He is an interim professor at the Institute of Political Science and the NRW School of Governance of the University of Duisburg-Essen. He received his PhD in political science at the University of Heidelberg. His main research interests include political communication, comparative politics, populism and computational social science methods. Gregor Wiedemann  is a post-doctoral researcher in the Language Technology group of the computer science department at Hamburg University, Germany. He studied political science and computer science in Leipzig and Miami. Due to his interdisciplinary background, he has worked in several projects in the fields of digital humanities and computational social science. In his research, he focuses on methods and workflows to analyse large text collections with technologies from natural language processing.

List of Figures

Fig. 2.1 Confluences in discourse analysis 25 Fig. 2.2 Transformation of the text. The figure is an adaption of a schema presented in Meunier (1993) 36 Fig. 3.1 Correspondence analysis of keywords and research interests of UK sociologists—rows only represented 75 Fig. 3.2 Correspondence analysis of keywords and research interests of UK sociologists—rows and columns represented 76 Fig. 3.3 Correspondence analysis of keywords and research interests represented on the websites of professors in sociology departments in the UK 77 Fig. 5.1 Correspondence analysis of the German press corpus on the financial crisis 2008 in the partition ‘month’ (Representation of column names only) 134 Fig. 5.2 Correspondence analysis of the sub-corpus German press interviews on the financial crisis 2008 134 Fig. 5.3 DHC in the German press corpus on the financial crisis 2008 (Analysed with Iramuteq)137 Fig. 5.4 The dominating semantic field in the German press corpus on the financial crisis 2008 139 Fig. 5.5 Over- and under-represented groups of words referring to discourse participants and discourse objects (partition ‘month’) 141 Fig. 5.6 Summary of the macrostructure of the financial crisis press corpus142 xix

xx 

List of Figures

Fig. 5.7 Map of text sections of interviewees displaying prototypical sentences of Angela Merkel’s interviews 143 Fig. 6.1 Factorial correspondence analysis of geographical terms in the official biographies of members of the last Central Committees of the USSR 159 Fig. 6.2 French Sociological Association (AFS) congress 2004 strategic diagram of all abstracts (femme)164 Fig. 6.3 French Sociological Association (AFS) congress 2004 ‘femme’ (woman) cluster with its keywords, including ‘travail’ (work) 165 Fig. 6.4 French Sociological Association (AFS) congress 2004 strategic diagram of all abstracts (without femme)166 Fig. 6.5 Strategic diagram of the first four months of the 2006 Association for the Right to Information corpus 171 Fig. 6.6 2006 keywords’ attractive power over the three four-month periods175 Fig. 6.7 Dominant 2007-1 terms over the four periods of 2007–2008 (Bush vs. UN) 176 Fig. 7.1 Document frequency of articles on minimum wages in two German newspapers 193 Fig. 7.2 Area plot of topic distribution over time 199 Fig. 7.3 Relative frequencies of documents containing stances on minimum wages 203 Fig. 8.1 Average number of articles about Islam per newspaper per month, 1998–2014 220 Fig. 8.2 Proportion of mentions of different branches of Islam for each newspaper237 Fig. 8.3 References to Sunni, Shia, Sufi, Salafi and Wahhabi over time. Dark grey denotes the proportion of mentions of references to branches of Islam (e.g. Sunni, Shia, Wahhabi); light grey bars denote references to Islam 238 Fig. 8.4 Summary of all data, comparing proportions of change over time241 Fig. 8.5 Claimed causes of radicalisation in the press in 1998–2009 244 Fig. 8.6 Claimed causes of radicalisation in the press in 2010–2014 245 Fig. 8.7 Claimed causes of radicalisation in the press in 2014 246 Fig. 9.1 Geocollocations control panel 262 Fig. 9.2 Geocollocations map view 264 Fig. 9.3 Dorling diagram view 265

  List of Figures 

Fig. 9.4 Fig. 9.5 Fig. 9.6 Fig. 9.7 Fig. 9.8 Fig. 10.1 Fig. 10.2 Fig. 10.3 Fig. 10.4

xxi

Reduced Dorling view, comparison of selected countries: collocate Beziehung (relationship) 267 Map view, selection of collocates Migration, Flüchtlinge (migration, refugees)—Spiegel/Zeit corpus 2010–2016 270 Map view, selection of collocates Migration, Flüchtlinge (migration, refugees)—Spiegel/Zeit corpus 1945–1960 271 Close view on the collocates in the migration discourse— Spiegel/Zeit corpus 2010–2016 272 Javascript library ‘D3.js’, ‘visual index’ of examples on the website277 Tweets on #ClimateChange 294 Tweets on #NetNeutrality 295 Network analysis of the tweets on #ClimateChange 297 Network analysis of the tweets on #NetNeutrality 298

List of Tables

Table 3.1 Table 3.2 Table 4.1 Table 7.1 Table 7.2 Table 8.1 Table 8.2 Table 8.3 Table 8.4 Table 8.5 Table 8.6 Table 8.7 Table 8.8 Table 10.1 Table 10.2 Table 10.3 Table 10.4 Table 10.5

Three levels of analysis 56 Four ideal typical dimensions of social context for the analysis of discourses 59 National evaluation schemes in some Western higher education systems 104 Topics terms and shares 196 Text classification of stances on minimum wages 202 The structure of the two newspaper corpora 219 Collocates of Muslim women 229 Patterns around veiling for Muslim women 232 Patterns around veiling—change over time (summary) 233 Arguments against veiling 234 Collocates of Muslim men 235 Levels of belief 239 Extremism keywords 243 Geolocated tweets and retweets of the ten most frequent countries293 The ten most represented countries in both samples 296 Categories derived from context analysis 302 Categorised keywords for tweets containing #ClimateChange 303 Categorised keywords for tweets containing #NetNeutrality 305

xxiii

Part I Introductory Remarks

1 Understanding Twenty-First-Century Societies Using Quantifying Text-­ Processing Methods Ronny Scholz

1

Analysing Knowledge-Based Post-­industrial Societies: Challenges and Chances

During the last 50 years, Western societies have experienced substantial changes. The phenomena of Europeanisation and globalisation as well as technical innovations such as the Internet and social media have revolutionised the way we use language when interacting, socialising with each other, or storing and recalling knowledge. In fact, the Internet has fostered access to globally produced information. In the abundance of sometimes contradicting information, the formation of knowledge in

I am thankful to Malcolm MacDonald, Joy Malala and Yannik Porsché for their helpful comments on earlier versions of this text.

R. Scholz (*) Centre for Applied Linguistics, University of Warwick, Coventry, UK e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_1

3

4 

R. Scholz

discourses with a particular inherent logic becomes evident. Moreover, recent debates in the public sphere about gender identities, national and cultural identities, and post-truth evidence an increased self-reflexivity in Western societies, emphasising the construction of knowledge as pivotal for the way we make sense of ourselves and our surrounding world. Drawing from Foucault, Althusser, Pêcheux, Laclau, and others, discourse studies has developed as a transdisciplinary research field that responds to a need for a better understanding of how contexts influence the making of meaning and knowledge on various levels of social interaction. This volume is a compilation of papers that present sophisticated quantifying methods used to analyse textual and social contexts of language use. The volume covers a large range of quantifying methods that could be used by social scientists who investigate the construction of knowledge in society. Before presenting the texts that have been compiled for this volume at the end of this introduction, the first part of the introduction will outline how trends in society have contributed to discourse becoming an object worthy of social sciences research. The second part will explain the importance of context for analysing the construction of meaning in different discourses. The third part of the introduction discusses the benefits of using quantifying methods to analyse meaning-­making process in the digital age, and the last part presents the purpose of this volume.

2

 ocietal Trends Influencing Knowledge S Production and Knowledge Exchange

This book showcases methods developed in natural language processing, quantitative linguistics, corpus linguistics and statistics for the analysis of discourses across the social sciences. Discourses have become a preeminent object of investigation across the social sciences and humanities because they have become palpable phenomena in everyday life. This is mainly due to an increased amount of coherent information contradicting previously established knowledge formations or narratives. The discursivity of knowledge is evidenced in a number of societal developments: (1) an increase in expert-based legitimation of decisions and competing, sometimes contradictory, expertise; (2) an increasingly rapid exchange of information on a global scale, sometimes contradicting mainstream

  Understanding Twenty-First-Century Societies Using… 

5

media narratives; and (3) a democratisation of information production through social media in which communities are formed beyond national borders, and in which historical events can be recontextualised by contradicting a narrative circulated, for instance, in the mass media. First, in post-industrial societies, work, as a primary object for sociological research, is often knowledge-based. Hence, the construction of knowledge and the conditions of its emergence become crucial for better understanding societies, and their dynamics and structuration. Moreover, its continuous production leads to a constant renewal of knowledge and thus, to a higher dynamic and fluidity of established bodies of knowledge. For example, knowledge about societal structures, including national, cultural, social and political boundaries seems to be constantly challenged by cross-boundary influences and experiences. One way of dealing with the emerging complexity of societies, is the instalment of knowledge authorities referred to as experts. They help to interpret, order, evaluate and interlink the abundance of produced information that becomes knowledge. Expertise, however, is the result of a knowledge industry that extends from classical universities to think tanks, which compete in the public space for recognition. Nowadays, political decisions are often legitimised by an army of experts and interpreted by political agents in line with their political argument (Maesse 2015). Some scholars have argued that the demonstration of scientific information has become part of a range of governance techniques contributing to the construction of Europe as a political and scientific community (Rosental 2015). Second, faster routes of communication have helped to raise awareness of political problems in the most remote areas on the planet. Nevertheless, the ever-larger news corporations and the deregulation of the media landscape have led to increasing competition in news production, favouring the coverage of dramatic events and information, which guarantee a high news value (Luhmann 2000), attracting high numbers of audience. Failing to provide balanced background information about a political issue enabling the citizens to engage fully with a political debate, these infrastructural changes have led to a new media style, a superficial ‘­infotainment’, which can easily be argued against with similar superficial but contradicting information. Third, important technological changes have taken place that impact on the way we socialise. Some scholars have stressed the importance of new media for economic development and political innovation (Barry 2001).

6 

R. Scholz

Indeed, new communication devices have enhanced communication on a global scale leading, among others, to similar cultural trends triggered by a globalised cultural industry. Moreover, social media have opened up a new social arena enhancing socialisation beyond geographical and social boundaries. Nowadays, personal and collective identities are often influenced to a substantial extent by experiences in the virtual space of social media. At the same time, knowledge emerges and is reproduced much more dynamically than it used to be. Large sets of information are available through the Internet and can be accessed and reproduced individually at any time and as often as desired. Smartphones and tablets have led to a popularisation of information production. Thereby the Internet  functions as a new social location, in which socialisation can take place often without control or influence of mainstream political stakeholders. This new social location has not only fostered democratic debates, it has also enhanced a particularisation of discourse communities, in which the flourishing of counter-discourses and identity-building that oppose official political positions is facilitated. This has given rise to populist and extremist discourses all over the world evidenced by the political success of movements like PEGIDA, SYRIZA, and PODEMOS or politicians like Trump, Orban, Duterte, or Le Pen. Social media, such as Facebook or Twitter, especially gives individuals easy access to communicating with the relatively anonymous masses. The twofold constellation of mainstream and social media has provided the grounds for a broader reproduction of knowledge, in which current and historical events can be recontextualised. As a result of the manifold recontextualisations available in the public space, various interpretations, narratives, and representations of historic and contemporary events seem to circulate in society that not only researchers but also the wider public refers to as ‘discourses’. In sum, we can identify a number of aspects that render discourse a relevant research object for social scientists.

3

Discourse, Context, and Meaning

The overall objective of discourse studies is to analyse language use in context in order to better understand the social construction of meaning and its influence on the knowledge circulating in any given society or

  Understanding Twenty-First-Century Societies Using… 

7

part of it. Thus, social scientists doing discourse analysis will, for instance, gain insights into: how knowledge about groups, communities, and social identities is constructed in discourses (inclusion and exclusion; gender, religion, race, class, nation); how this structuration is justified; how social spaces and positions are constructed, negotiated, and orchestrated; which narratives and ideologies drive different actors, groups, and communities in society; and how decisions in a given society or a part of it are being legitimised. Moreover, with its capacity to map prevailing argumentations and narratives, discourse studies can reveal how values are articulated in a particular way in order to justify the stance of a specific group, social strata, or class. In this sense, discourse analysts can show how society is governed and hence can feed into the formulation of a social critique that contributes to social progress (Herzog 2016). Foucault’s philosophy has helped to understand the formation of knowledge in terms of discourses that organise knowledge. Most importantly he has insisted on the fact that discourses are driven by power relations that are rooted in institutional, social, societal, and historical contexts in which language users have to operate (Foucault 1970, 1972, 1979). Foucault’s theoretical categories have informed discourse analytical approaches  not only in France but across the globe—to name only a few, the Discourse Linguistic Approach (Warnke and Spitzmüller 2008) and the Sociology of Knowledge Approach (Keller 2013) in Germany or the Discourse Historical Approach (Reisigl and Wodak 2016) and Critical Discourse Analysis (Dijk 1997; Fairclough 1995; Wodak and Meyer 2001). What is common to all approaches in discourse studies is their fundamental interest in meaning construction through natural language use in context. There are numerous definitions of discourse. I will touch upon two which best fit the purposes of this volume. First, there is Busse and Teubert’s definition which is common in German discourse linguistics. They define discourse as a ‘virtual text corpus’ containing all sorts of texts that have been produced on a particular topic. In order to analyse a discourse, a researcher has to compile a ‘concrete text corpus’ which is compiled from a representative selection of texts of the ‘virtual text corpus’ (Busse and Teubert 2014, 344). This definition might satisfy corpus linguists, but if we want to analyse discourse practices from a perspective that accommodates the broader spectrum of social sciences and humanities,

8 

R. Scholz

we have to add elements that refer to the social structure, the context of circulation, and the actions that allow an utterance to become effective. Achard’s (1993, 10) definition seems to suit this demand somewhat better. According to him, a discourse emerges from language practice in a situation in which a speech act is considered as effective as a result of its relation to a set of other linguistic and non-linguistic acts. Consequently, the analysis of discourse has to put emphasis on institutional and social context in which interlocutors relate to each other. At the same time, context cannot be understood as a stable, stereotypical, neutral, and self-­ contained entity, in which everything seems to happen (Blommaert 2005, 56). The conceptualisation and operationalisation of context is in fact a necessary analytical step in order to understand how meaning emerges in discourses. Discourse studies is a broad field of research integrating different disciplines and their foci on specific aspects of society, materiality, and context (Beetz and Schwab 2018). Hence, the understandings of contexts are quite diverse and multilayered (Porsché 2018, 81–129). This volume presents a collection of texts using quantitative methods which are applied to a range of inside and outside perspectives on language use. Whereas the former focuses on textual contexts (co-text) of discourse or the construction of context from within a particular situation of communication, the latter emphasises its institutional, social, societal, and historical contexts (Leimdorfer 2011, 74). Foucault’s concept of the dispositif articulates the inside and outside perspective on language use (Foucault 1972; Raffnsøe et al. 2016). This concept captures the nexus of power and knowledge reflecting institutional constraints in interpretive processes. Studies referring to this concept often distinguish, on the one hand, between a meso- and macro-level of society providing the institutional and societal structures in which discourses emerge, and, on the other hand, a micro-level on which discourses are performed (Bührmann et al. 2007). Therefore, the term ‘discourse’ often refers to institutionalised practices which follow certain sets of rules fixed over time. These practices are considered central for particular fields and areas of society and are evidenced in institutionalised ways of speaking (Bührmann and Schneider 2007, 5). The analysis focuses on the interplay between situational contexts and practices with discourses or the constitution of contexts through

  Understanding Twenty-First-Century Societies Using… 

9

discourses (Keller 2013). In this volume, Hamann et al. (Chap. 3) use quantitative  corpus methods to study the dispositif of academic discourse. Studies highlighting the analysis of dispositifs usually distinguish between discursive and non-discursive practices in discourse production (Bührmann and Schneider 2007). From a perspective of French discourse analysis or a post-structuralist discourse theory, the assumption of a pre-­ discursive world, in which non-discursive elements have their place and in which subjects can choose their position freely, would be considered as problematic.  Rather than asking how the knowledge is constructed, which a subject then uses to make a sense, post-structuralist approaches would ask how a discourse is constructed, in which subjects are placed and can emerge in a particular position. In this sense, pre-discursive elements can become relevant to discourse, and subsequently to society in general, only if they are articulated with a particular meaning in relation to other discursive elements. This idea goes back to Vološinov’s philosophy of language according to which the emergence of meaning is a result of a dialectic relation between the concrete utterance and the language system which, in opposition to Saussure, cannot be regarded as independent from communication. It is only in concrete social relation of people that a linguistic sign, together with the language system, can acquire meaning (Vološinov 1973, 157–158). Against this background, French discourse analysts have tried to go beyond a structuralist perspective to better account for discursive practices. Socio-pragmatic approaches to discourse tried to integrate contextual variables with the analysis of utterances. Aiming to capture social and cognitive dynamics, which are triggered by forms and implications on the textual level, they study the construction of meaning in its social context (e.g. Bronckart 1996). There are a couple of strands ­highlighting different aspects of context. Enunciative pragmatics has developed since the 1980s. It focuses on the reflexivity of the speech activity allowing the speakers to convert the system of language into discourse (Angermuller et  al. 2014, 135–139; Authier-Revuz 1984; Kerbrat-­Orecchioni 1980; Maingueneau 1997; Reboul and Moeschler 1998). Furthermore, a sociolinguistic tradition has developed that looked at how institutional structures influence the use of language, analysed through the perspective of a

10 

R. Scholz

sociology of language (Achard 1993; Bacot and Rémy-Giraud 2011; Boutet et  al. 1995). The operationalisation of an analysis of discourse practices has been taken most seriously in the praxematic approach to discourse cumulating in the analytical concept of the ‘praxem’ (Lafont 1978; Lafont et al. 1983; Siblot 1990, 1997). A similar idea has influenced sociologists in Québec when they analysed society as a result of the work on meaning production (Bourque and Duchastel 1984, 118). Here, the internal dynamics in discourses are considered to have a particular impact on this process (Bourque and Duchastel 1988, 51).

4

 hallenges and Chances for Discourse C Research with Quantifying Methods

Even though the societal changes aforementioned have rendered discourses important for social scientists, they also confront researchers with various challenges: Texts are produced and distributed at an enormous speed, in substantial numbers, by a large number of institutional and non-institutional actors; political discourses often emerge in an international institutional context and are reproduced in different languages (for a sociolinguistic account of this issue: Duchêne and Heller 2012). For example, the debates on the ‘Euro crisis’, the international trade agreement TTIP, or leaked document collections trigger political discourses in different countries. Moreover, computers influence more and more interactions between humans. Complex algorithms respond to our actions in virtual space and robots emulate individuals in social networks (Hegelich and Janetzko 2016) so that the concept of the author and the dialogue with such algorithmic agents (Antonijevic 2013, 95–97; Tufekci 2015) may need a new set of analytical categories and methods. The fact that large amounts of natural language data are stored in digitised form and can be accessed through the Internet is, in fact, to the advantage of computer-based methods as they are presented in this volume. There are now large databases for digitised political texts and press texts. In the ‘digital humanities’, enormous efforts are undertaken to create large digital archives of historical and literature texts from past centuries (e.g. the CLARIN centres). The abundance of available textual

  Understanding Twenty-First-Century Societies Using… 

11

data creates methodological challenges that are similar to those discussed around ‘big data’. Some argue that ‘big data’ will radically change the way in which we make sense of the world to the extent that theoretical and abstract models become obsolete. Others make a case for the opposite, contending for theories and methods of interpretation as necessary to maintain an analytical stance in the abundance of information (Boyd and Crawford 2012; Gonzalez-Baillón 2013; Schroeder 2014). Discourse researchers also have to face these points of debate. They have to discuss: To what extent can the analysis be based on text-processing software solutions? To what extent is there a need to understand the underlying algorithms in order to interpret the results? How important are discourse theories for the interpretation? How can results produced with quantitative methods complement qualitative approaches to discourse? The ‘digital age’ has indeed changed the way texts are produced, stored, made available, and read. This does not remain without consequences for discursive  practices. Social scientists that try to capture and study the production of knowledge under these circumstances need theories and methods that are able of accounting for the emergence of meaning in the different  social, historic, and textual contexts, in which phenomena become meaningful and therefore can become social. Thus the conceptualisation of context is as diverse as the disciplines that have subscribed to this transdisciplinary field. Consequently, discourse analysts have discussed on an epistemological (Maingueneau 2013; Paveau 2017) and a methodological level (Norris and Jones 2005; Jones et al. 2015) how to grasp social phenomena in the digital age. In this regard there has been a special emphasis on the analysis of discourses as perpetuated via social and new media (Barats 2013; KhosraviNik 2016). Complementary to these approaches, this volume emphasises the use of quantifying methods. Indeed, corpus linguistic methods have been used to analyse discourses for a long time, especially in France, where the lexicometric approach has been developed since the 1970s (see Chap. 5). Also, in Critical Discourse Analysis, the benefits of corpus methods, such as a better navigation in large text collections based on quantification, have been advocated for some time (Baker 2006; Mautner 2005, 2012) and are now being established in the field under the acronym CADS for Corpus-Assisted Discourse Studies (see Part IV in this volume and

12 

R. Scholz

Partington et al. 2013). Corpus methods applied in discourse studies can help to gain an overview over dominant discourse participants, discursive positions, topics, and arguments, and their interrelations relatively quickly. They can trace discursive dynamics over time within these discursive structures. And they permit the analyst to ‘zoom in’ into parts of the data which have proven to be of particular interest within the research process. In this sense corpus tools allow organising the research process in quantitative and qualitative stages—whereby the researcher can develop hypotheses and theories by alternating repeatedly in a reflexive loop between the analysis of the research material on different levels and the interpretation of the different results (Demazière et al. 2006; Leimdorfer and Salem 1995).

5

The Purpose of This Volume

Recent publications using corpus methods in discourse studies are numerous. Most of these publications focus on a particular approach such as corpus linguistics (Baker and McEnery 2015) or text mining (Ignatow and Mihalcea 2017; Wiedemann 2016). Some of them are very good introductions into technical and/or statistical aspects of the analysis (Biemann and Mehler 2014; Jockers 2014). The present volume is complementary to these works in that it takes a global perspective covering a broad range of quantifying methodologies that are used to analyse discourses. The volume aims to provide a useful introduction to researchers in social sciences facing the challenges of new technology-driven language practices and big data. It goes beyond the simple analysis of the textual level by taking into account the extra-textual context that is essential in order to understand how meaning is constructed in society. The aim is to give an overview of the broad range of quantifying corpus methods that can be deployed in order to address different levels of the co-text and the context. What all empirical contributions in this volume have in common is the use of some sort of quantification that reorganises the natural language data. The book is driven by a reflexive mindset regarding the possibilities and limits of quantifying methods, which are understood as complementary not as in opposition to qualitative methods.

  Understanding Twenty-First-Century Societies Using… 

13

The authors of this volume were asked to present their methods in an accessible way to corpus beginners and experts alike. For the first time, different national and disciplinary traditions of quantifying methods in discourse studies that have been developing to a great extent in parallel without taking much notice of the other approaches are presented in this collective volume. Moreover, this book draws on experiences from scholars in the Anglo-Saxon, German, and French-speaking academic world that work with similar methods, yet meet rarely, due to language, disciplinary, or institutional boundaries. With texts coming from authors in sociology (Chaps. 2 and 3), scientometrics (Chap. 4), quantitative linguistics and lexicometry (Chaps. 5 and 6), computational linguistics and text mining (Chap. 7) as well as corpus linguistics and political science (Chaps. 8, 9 and 10), the volume has a strong transdisciplinary outlook. The book is divided into Parts I–IV and comprises ten chapters. After the introductory remarks in the two chapters of Part I, the two following chapters of Part II look into how to integrate institutional contexts into discourse analysis. The three chapters in Part III set out complex algorithms developed in quantitative linguistics, lexicometry, and text mining. The benefit of such methods is clearly their heuristic strength. They take into account the complete vocabulary of all texts in the corpus at once and enable the development of new ideas not only on topics touched upon in a given corpus but also on the macro-structure of the discourse represented in it. Part IV covers new developments in Computer-Assisted Discourse Studies. All three chapters use classical corpus linguistic methods such as collocation analysis and keyword analysis and bring them to new splendour by taking innovative methodological steps. Chapter 1 has introduced the volume. By outlining societal developments that have rendered discourses a prominent research object across the social sciences, it has advocated the benefits of quantifying text-­ processing methods as a means of studying societies in the digital age. Second, the text has addressed some of the challenges discourse analysts have to face when studying digitised media communication. Third, it has further highlighted the importance of quantitative methods for an analysis of contextuality on different levels. Chapter 2 takes the reader on a methodological journey which will help him/her to stop thinking  of research in terms  of quantitative

14 

R. Scholz

and qualitative paradigms. After a short overview of theories that have influenced discourse studies, Jules Duchastel and Danielle Laberge present a general model of how social reality is being analysed with textprocessing methods. By highlighting strengths and weaknesses of qualitative and quantitative methods the text advocates an integrative mixed methods approach. The text shows that scientific interpretation is bound to specific explanatory operations that pave the way for a particular understanding of the world. Any scientific process, no matter whether qualitative or quantitative, is based on a common ground mobilising research operations for the identification of units, their description and their analysis. While the analytical paradigms differ in their epistemological and methodological assumptions, they are facing the same problem of reducing and restoring complexity. Chapter 3 outlines the dispositif approach, which combines a linguistic discourse analysis of texts with a sociological study of the social context (i.e. the dispositif understood as an institutional arrangement of practices and structures). Julian Hamann, Jens Maesse, Ronny Scholz, and Johannes Angermuller use the discourse of academic researchers to exemplify this approach. By articulating correspondence analysis of self-­ representations on researchers’ homepages with institutional data of sociology professors in the United Kingdom, they outline a research design that consists of three components: a linguistic analysis of texts, a sociological analysis of institutional contexts, and a theoretical account of how the two are related in the academic dispositif. The dispositif perspective on discourse aims to respond to a demand for systematic discourse research on the social and institutional contexts of discursive practices. Chapter 4 presents scientometrics as a type of corpus research which measures the scientific output of academic researchers and analyses underlying differences and inequalities among researchers based upon their scientific outputs. Johannes Angermuller and Thed van Leeuwen discuss the history of the field since Eugene Garfield launched the Science Citation Index in 1963 and investigate its practices and indicators, such as the Journal Impact Factor or the h-index. This contribution places the development of the field in the context of the rise of ‘numerocracy’—a regime of power knowledge which aims at governing large populations by numbers. By applying and extending Michel Foucault’s governmentality

  Understanding Twenty-First-Century Societies Using… 

15

thesis, the authors point out the non-scientific conditions of scientific practices in corpus research and make the case for a socially reflexive approach. In this contribution, Ronny Scholz draws a connection between Bachelard’s concept of ‘epistemic rupture’ and quantitative methods which allows the discovery of discursive phenomena prior to the interpretation of meaning in texts. Lexicometry is a corpus-driven approach that deploys, besides common corpus linguistic methods, complex algorithms that exhaustively analyse the lexis of a given corpus. It does so by contrasting different corpus parts organised in partitions. Taking examples from a corpus of 4000 press texts on the global financial crisis of 2008, the contribution illustrates how a large text corpus can be reduced systematically to a readable size. It also demonstrates different ways of exploring lexicosemantical macro-structures using correspondence analysis, descending hierarchical classification, and other methods. Chapter 6 explains how complex statistical methods such as factorial correspondence analysis, both descending hierarchical classification (Alceste, Topics) and ascending hierarchical classifications (Leximappe-­ Lexinet, Calliope) can be used to study which concepts and topics dominate a particular discourse in society at a certain period in time. Karl M. van Meter demonstrates how semantic and thematic shifts can be traced over time and which future developments might be more or less probable. The examples are taken from two projects: first, a synchronic and diachronic analysis of a corpus of conference abstracts submitted to the annual conferences of the American, French, and German national associations of sociology; second, drawing from Pete Dale Scott’s concept of ‘World Parapolitics’, a diachronic perspective on the representation of conflicts in the international press. Chapter 7 develops a discourse analysis  approach based  on the many opportunities provided by text mining. Gregor Wiedemann introduces unsupervised and supervised machine learning techniques to analyse a corpus covering twenty years of public discourse on statutory minimum wages in Germany. His contribution demonstrates how topic modelling can be used to reveal thematic clusters on the macro-level, and how text classification is able to trace utterances of political stance on the micro-­level of a discourse. In particular, the combination of data-driven clustering and theory-driven classification allows for complex analysis workflows on very

16 

R. Scholz

large text collections, thus making qualitative aspects of diachronic discourses quantifiable. In Chap. 8, Paul Baker and Tony McEnery introduce CADS, a means of using the methods of corpus linguistics to facilitate discourse analysis of large volumes of textual data. The chapter uses this framework not only to demonstrate the value of CADS but also to explore the importance of repeating studies over time to test the degree to which discourse is static, or changes, through time. By extending a study of the representation of Muslims and Islam in the UK press, the chapter shows the value of exploring the dynamic nature of discourse as a way of cautioning against the idea that discourse is necessarily stable across time. In Chap. 9, Noah Bubenhofer, Klaus Rothenhäusler, Katrin Affolter, and Danica Pajovic discuss common approaches using data visualisations within the field of digital humanities. They argue that by assigning equal importance to the development, as well as the usage of a visualisation framework, researchers can question dogmatic ‘best-practice’ norms for data visualisations which may prevent them from developing visualisations that can be used to find emergent phenomena within the data. They then focus on the question of how visualisations reconstitute language by using diagrammatic operations. Working with digital visualisations, the technological background is of great  importance for the interpretation and the development of new tools. As an example, they present a ­visualisation framework for ‘geocollocations’, which can be used as a tool to detect words that typically collocate with toponyms in text corpora. In Chap. 10, Jörn Stegmeier, Wolf J.  Schünemann, Marcus Müller, Maria Becker, Stefan Steiger, and Sebastian Stier present a multi-method discourse analytical approach to analyse Twitter communication on two political issues of global concern: environmental policy/climate change and Internet governance/net neutrality. Their corpus is compiled from Twitter messages containing #NetNeutrality or #ClimateChange, which the authors gathered between January and March 2015. First, they map and compare the geographical landscapes of the two policy fields by using geolocation information from the Twitter API and the Data Science Toolkit. Second, they carry out a comparative network analysis defining Twitter users as nodes, and Retweets (RT) and mentions (@) as links. Finally, the authors apply keyword analysis to identify discursive pat-

  Understanding Twenty-First-Century Societies Using… 

17

terns. Combining these three methods allows the authors to assess the degree of transnationalisation in the two fields.

References Achard, Pierre. 1993. La sociologie du langage. Paris: PUF. Angermuller, Johannes, Dominique Mainguenau, and Ruth Wodak, eds. 2014. The discourse studies reader. Main currents in theory and analysis. Amsterdam: John Benjamins. Antonijevic, Smiljana. 2013. The immersive hand: Non-verbal communication in virtual environments. In The immersive internet. Reflections on the entangling of the virtual with society, politics and the economy, ed. Dominic Power and Robin Teigland, 92–105. Basingstoke: Palgrave Macmillan. Authier-Revuz, Jacqueline. 1984. Hétérogénéité(s) énonciative(s). Langages 73: 98–111. Bacot, Paul, and Silvianne Rémy-Giraud. 2011. Mots de l’espace et conflictualité sociale. Paris: L’Harmattan. Baker, Paul. 2006. Using corpora in discourse analysis. London: Continuum. Baker, Paul, and Tony McEnery. 2015. Corpora and discourse studies. Integrating discourse and corpora, Palgrave advances in language and linguistics. Basingstoke: Palgrave Macmillan. Barats, Christine, ed. 2013. Manuel d’analyse du web en Sciences Humaines et Sociales. Paris: Armand Colin. Barry, Andrew, ed. 2001. Political machines. Governing a technological society. London: Athlone Press. Beetz, Johannes, and Veit Schwab, eds. 2018. Material discourse—Materialist analysis. Lanham, MD: Lexington Books. Biemann, Chris, and Alexander Mehler. 2014. Text mining. From ontology learning to automated text processing applications – Festschrift in honor of Gerhard Heyer. Translated by Gerhard Heyer. Theory and Applications of Natural Language Processing. Cham: Springer. Blommaert, Jan. 2005. Discourse, a critical introduction. Cambridge: Cambridge University Press. Bourque, Gilles, and Jules Duchastel. 1984. Analyser le discours politique duplessiste: méthode et illustration. Cahiers de recherche sociologique 2 (1, Le discours social et ses usages): 99–136.

18 

R. Scholz

———. 1988. Restons traditionnels et progressifs. Pour une nouvelle analyse du discours politique. Le cas du régime Duplessis au Québec. Montréal: Les Éditions du Boréal. Boutet, Josiane, Bernard Gardin, and Michèle Lacoste. 1995. Discours en situation de travail. Langages 29 (117): 12–31. Boyd, Danah, and Kate Crawford. 2012. Critical questions for big data. Information, Communication & Society 15 (5): 662–679. Bronckart, Jean-Paul. 1996. Activité langagière, textes et discours; pour un interactionnisme socio-discursif. Lausanne: Delachaux et Niestlé. Bührmann, Andrea D., Rainer Diaz-Bone, Encarnación Gutiérrez-Rodríguez, Werner Schneider, Gavin Kendall, and Francisco Tirado. 2007. Von Michel Foucaults Diskurstheorie zur empirischen Diskursforschung. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 8(2): Editorial FQS. Bührmann, Andrea D., and Werner Schneider. 2007. Mehr als nur diskursive Praxis?  – Konzeptionelle Grundlagen und methodische Aspekte der Dispositivanalyse. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 8 (2): Art. 28. Busse, Dietrich, and Wolfgang Teubert. 2014. Using corpora for historical semantics. In The discourse studies reader. Main currents in theory and analysis, edited by Johannes Angermuller, Dominique Mainguenau and Ruth Wodak, 340–349. Amsterdam: John Benjamins. Original edition, 1994. Demazière, Didier, Claire Brossaud, Patrick Trabal, and Karl Van Meter, eds. 2006. Analyses textuelles en sociologie. Logiciels, méthodes, usages. Rennes: Presses universitaires de Rennes. van Dijk, Teun A., ed. 1997. Discourse as social interaction. Discourse studies. London: Sage. Duchêne, Alexandre, and Monica Heller. 2012. Language in late capitalism. Pride and profit. New York: Routledge. Fairclough, Norman. 1995. Critical discourse analysis. The critical study of language. London: Longman. Foucault, Michel. 1970. The order of things. An archaeology of the human sciences (World of man). London: Tavistock. Original edition, 1966. ———. 1972. The archeology of knowledge. London: Tavistock. Original edition, 1969. ———. 1979. Discipline and punish. The birth of the prison. New York: Vintage Books. Original edition, 1975. Gonzalez-Baillón, Sandra. 2013. Social science in the era of big data. Policy & Internet 5 (2): 147–160.

  Understanding Twenty-First-Century Societies Using… 

19

Hegelich, Simon, and Dietmar Janetzko. 2016. Are social bots on Twitter political actors? Empirical evidence from a Ukrainian social botnet. Proceedings of the Tenth International AAAI Conference on Web and Social Media. Accessed July 1, 2018. https://www.aaai.org/ocs/index.php/ICWSM/ICWSM16/ paper/view/13015. Herzog, Benno. 2016. Discourse analysis as social critique—Discursive and non-­ discursive realities in critical social research. Basingstoke: Palgrave Macmillan. Ignatow, Gabe, and Rada Mihalcea. 2017. Text mining. A guidebook for the social sciences. Los Angeles: SAGE. Jockers, Matthew Lee. 2014. Text analysis with R for students of literature (Quantitative Methods in the Humanities and Social Sciences). Cham: Springer. Jones, Rodney H., Alice Chik, and Christoph A. Hafner, eds. 2015. Discourse and digital practices. Doing discourse analysis in the digital age. London: Routledge, Taylor & Francis Group. Keller, Reiner. 2013. Doing discourse research. An introduction for social scientists. London: Sage. Kerbrat-Orecchioni, Catherine. 1980. L’Énonciation. De la subjectivité dans le langage. Paris: Armand Colin. KhosraviNik, Majid. 2016. Social Media Critical Discourse Studies (SM-CDS): Towards a CDS understanding of discourse analysis on participatory web. In Handbook of critical discourse analysis, ed. John Flowerdew and John E. Richardson. London: Routledge. Lafont, Robert. 1978. Le travail et la langue. Paris: Flammarion. Lafont, Robert, Françoise Madray-Lesigne, and Paul Siblot. 1983. Pratiques praxématiques: introduction à une analyse matérialiste du sens. Numéro spécial de: Cahiers de linguistique sociale 6: 1–155. Leimdorfer, François. 2011. Les sociologues et le langage. Paris: Editions de la MSH. Leimdorfer, François, and André Salem. 1995. Usages de la lexicométrie en analyse de discours. Cahiers des Sciences humaines 31 (1): 131–143. Luhmann, Niklas. 2000. The reality of the mass media. Cambridge: Polity Press. Original edition, 1995. Maesse, Jens. 2015. Economic experts. A discursive political economy of economics. Journal of Multicultural Discourses 10 (3): 279–305. Maingueneau, Dominique. 1997. Pragmatique pour le discours littéraire. Paris: Dunod.

20 

R. Scholz

———. 2013. Genres de discours et web: existe-t-il des genres web? In Manuel d’analyse du web en Sciences Humaines et Sociales, ed. Christine Barats, 74–98. Paris: Armand Colin. Mautner, Gerlinde. 2005. Time to get wired: Using web-based corpora in critical discourse analysis. Discourse & Society 16 (6): 809–828. ———. 2012. Corpora and critical discourse analysis. In Contemporary corpus linguistics, ed. Paul Baker, 32–46. London: Continuum. Norris, Sigrid, and Rodney H.  Jones. 2005. Discourse in action. Introducing mediated discourse analysis. London: Routledge. Partington, Alan, Alison Duguid, and Charlotte Taylor. 2013. Patterns and meanings in discourse. Theory and practice in corpus-assisted discourse studies (CADS), Studies in corpus linguistics. Vol. 55. Amsterdam: Benjamins. Paveau, Marie-Anne. 2017. L’analyse du discours numérique. Dictionnaire des formes et des pratiques, Collection Cultures numériques. Paris: Hermann. Porsché, Yannik. 2018. Public representations of immigrants in museums— Exhibition and exposure in France and Germany. Basingstoke: Palgrave Macmillan. Raffnsøe, Sverre, Marius Gudmand-Høyer, and Morten S.  Thaning. 2016. Foucault’s dispositive: The perspicacity of dispositive analytics in organizational research. Organization 23 (2): 272–298. Reboul, Anne, and Jacques Moeschler, eds. 1998. Pragmatique du discours. De l’interprétation de l’énoncé à l’interprétation du discours. Paris: Colin. Reisigl, Martin, and Ruth Wodak. 2016. Discourse Historical Approach (DHA). In Methods of critical discourse studies, ed. Ruth Wodak and Michael Meyer, 23–61. Los Angeles, CA: Sage. Rosental, Claude. 2015. From numbers to demos: Assessing, managing and advertising European research. Histoire de la Recherche Contemporaine 4 (2): 163–170. Schroeder, Ralph. 2014. Big data and the brave new world of social media research. Big Data & Society 1 (2): 1–11. Siblot, Paul. 1990. Une linguistique qui n’a plus peur du réel. Cahiers de praxématique 15: 57–76. ———. 1997. Nomination et production de sens: le praxème. Langages 127: 38–55. Tufekci, Zeynep. 2015. Algorithmic harms beyond Facebook and Google: Emergent challenges of computational agency. Colorado Technology Law Journal 13 (2): 203–218. Vološinov, Valentin N. 1973. Marxism and the philosophy of language. New York: Seminar Press. Original edition, 1929.

  Understanding Twenty-First-Century Societies Using… 

21

Warnke, Ingo H., and Jürgen Spitzmüller, eds. 2008. Methoden der Diskurslinguistik. Sprachwissenschaftliche Zugänge zur transtextuellen Ebene. Berlin and New York: De Gruyter. Wiedemann, Gregor. 2016. Text mining for qualitative data analysis in the social sciences. A study on democratic discourse in Germany. Wiesbaden: Springer VS. Wodak, Ruth, and Michael Meyer. 2001. Methods of critical discourse analysis. Introducing qualitative methods. London: Sage.

2 Beyond the Quantitative and Qualitative Cleavage: Confluence of Research Operations in Discourse Analysis Jules Duchastel and Danielle Laberge

1

Introduction

The world of social and language sciences is characterised by many cleavages: between understanding and explaining, between structural and phenomenological analysis, between different fields and disciplines related to the study of language, between different national and continental traditions, and between qualitative and quantitative approaches. These oppositions often create new avenues of thought, but they become sterile when giving up important aspects of the analysis. We will ask ourselves how

J. Duchastel (*) Department of Sociology, UQAM – Université du Québec à Montréal, Montréal, QC, Canada e-mail: [email protected] D. Laberge Department of Management and Technology, UQAM – Université du Québec à Montréal, Montréal, QC, Canada e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_2

23

24 

J. Duchastel and D. Laberge

different approaches in discourse analysis deal with these oppositions, and eventually with their possible convergence. We will explore the capacity of mixed methods to overcome the opposition between qualitative and quantitative methods. We will see how interpretation and explanation are constitutive parts of the research process. First, we will show how discourse analysis stands at an intersection of disciplines, traditions, and approaches. We will then discuss the opposition between qualitative and quantitative methods and the mixed methods approach as a proposed solution. This will lead us to reconsider the distinction between explaining and understanding: we put forward the existence, in all sciences, of a hermeneutic arc that does not separate interpretation from explanation. Through the description of different states of the text in the process of discourse analysis, we will describe the necessary phases of reduction and restoration of complexity, whether the approach is quantitative or qualitative. We will illustrate the compatibility of these methods, showing that the concepts of causality and measurement can apply in either approach.

2

Oppositions and Convergences in the Field of Discourse Analysis

Discourse analysis stands at the confluence of various disciplines, traditions, and approaches. It arose from a dual need to overcome, in the humanities, the limited focus on content and, in the language sciences, the restricted structural approach to language. Discourse analysis introduced the need to consider language in its social context and apprehend content as it is materialised in linguistic forms and functions. Discourse analysis can be considered as a merger of two great traditions: the hermeneutical tradition of humanities and social sciences, based on the meaning of social practices and institutions, and the more functional and structural tradition of language sciences that focuses on the description of different aspects of language use. Within the context of this confluence, a third axis emerged, that of statistical and ­computer sciences, leading to the development of a tradition of computer-assisted discourse analysis. If

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

25

one can hardly speak of discourse analysis as a discipline, it is because of this profusion of influences. They are produced by as many analytical practices as there are many disciplines and intersections between them. Figure 2.1 represents the set of oppositions and similarities of the various traditions of discourse analysis as they emerged in the sixties. The diagram shows, at its centre, discourse analysis as the crossing point of all these traditions. Therefore, it is not to be regarded as a discipline but as a field of research practices sharing a number of designs from several disciplines. This confluence is also marked by numerous exchanges between national traditions. The diagram can be read as a set of oppositions, from top to bottom, left to right, and along the diagonals. The first major opposition from top to bottom distinguishes qualitative and quantitative approaches. It is possible to consider approaches at the top of the figure as belonging to ‘letters’, for example, quality, while the bottom part refers to ‘numbers’, for example, quantity (Pires 1982). The second major opposition can be read, from left to right, French versus Anglo-Saxon Hjelmslev (1931/1963) Linguistic Cercle of Copenhaguen

G. H Mead (1934) Symbolic interactionnism Austin (1962) Speech acts

Harris (1952) Discours Analysis

Berger & Luckman (1966) Social construction of reality

Barthes (1957) Semiology Dubois (1969), Benveniste (1966) Enonciation analysis

French school of Discourse analysis

Qualitative analysis

Garfinkel (1967) Ethnomethodology

Lacan (1966), Psychoanalysis Pêcheux (1969), Automatic discourse analysis

Searle (1970) Philosophy of langage

Foucault (1969), Discourse analysis

Sacks (1972) Conversation analysis

Althusser (1970), Ideology

Stone (1966) General Inquirer

analysis Muller (1968) Lexical statistics

Laswell (1952) Communication theory Lexicometry

Guiraud (1960) Linguistic statistics

Fig. 2.1  Confluences in discourse analysis

Content analysis

Berelson (1952) Content Analysis Holsti (1969) Content analysis

26 

J. Duchastel and D. Laberge

traditions,1 highlighting the relative preponderance of linguistic on the left and of social sciences on the right. Figure 2.1 illustrates a space where each term is opposed to the other, horizontally, vertically, or diagonally. At the top of the diagram, within the so-called qualitative perspective, the French school of discourse analysis and the Anglo-Saxon tradition of qualitative analysis form the first opposition. What distinguishes them most is that they belong to different disciplinary traditions. French discourse analysis is anchored in the distributional, functional, and pragmatic linguistics, aiming to study language as it is used in the real world. It owes much to the structuralist tradition: understanding symbolic phenomena in their systemic dimension. It has gradually given attention to speech as a theoretical and an empirical object (Foucault 1969), and evolved into a form of text linguistics (Adam 1999; Rastier 2001). On the other hand, the qualitative analysis has evolved from the bosom of symbolic interactionism and phenomenology, also under the influence of the philosophy of language and pragmatism. These traditions have a common interest in the intentional action manifested through speech acts. While the French tradition focuses on the linguistic aspects of situated speech, the American tradition is mostly interested in language as a vehicle for the social construction of reality. What particularly distinguishes the two traditions is the type of empirical speech that is favoured. From the beginning, the French tradition was interested in institutional discourse, that is, political or literary discourses. The American tradition was rather more inclined towards speech in everyday life, that is, localised interlocutions or conversation. On the bottom axis of the diagram, which represents the quantitative perspective, we can also contrast two different approaches. On one side, we have the French tradition of lexical analysis (lexicometry), and on the other, the American tradition of content analysis. Both approaches share a common interest for the quantification and measurement of linguistic phenomena, but they can be distinguished by their disciplinary origin.

 It has to be noted that both traditions are not hermetically closed. For instance, the French school of discourse analysis initially was inspired by Zellig Harris (1952) distributional approach to language. 1

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

27

While in France there is an interest in statistics applied to literary and political corpora, in America, it is the study of communication and propaganda that gave birth to a tradition of content analysis. While in both cases, there is a strong belief in the power of explanation with figures, the mathematical and statistical models greatly differ. On the one hand, complex statistical methods are applied to words in their ‘natural’ existence, that is to say, without coding, on the other hand, relatively simple counts of coded units are produced. But in both cases, the access to meaning is through the numbers. Observing the figure along the vertical axis, it is possible to distinguish on the left an opposition between the French tradition of discourse analysis at the top and the lexical approach at the bottom. This opposition has gradually evolved from a ‘dialogue of the deaf ’, during the sixties and seventies, to a mutual recognition in recent years, as computer-assisted discourse analysis systems began to impose their own legitimacy. Everything happens as if the requirements of formalisation of computing procedures made statistics less daunting in the eyes of those primarily interested in the description of language functions. On the right side, in the American tradition, the same opposition existed between qualitative and quantitative methods. In both cases, the interest lies primarily in the meaning of discourses, but the qualitative tradition emphasises the interpretive reading based on the coding of units, while content analysis is concerned, at least in its early stages, with the essentially quantitative count of units of speech. This opposition has also diminished over the years, and there aren’t hardly any purely orthodox researchers left. As proof of this, one has only to look at mixed qualitative and quantitative features in computer-­assisted qualitative data analysis systems. Finally, on the diagonal axes of the diagram, we oppose, two by two, each tradition. It is clear that the opposition between lexical and qualitative analyses follows the same logic as that between the two approaches in quantitative and qualitative content analyses in the American tradition. But this opposition is not really present in the literature. The opposition that puts face-to-face discourse analysis and content analysis took shape in the founding act of discourse analysis in France. We should remember that the French tradition of discourse analysis comes from the critique of the content analysis tradition (Haroche et  al. 1971). It criticises the

28 

J. Duchastel and D. Laberge

ignorance of the linguistic substratum of discourse in this tradition, although some authors, such as Osgood (1959), have justified its whole importance. Discourse analysis as a research practice has always had a syncretic character, each tradition drawing on several disciplinary and methodological sources. It follows that the oppositions described here have progressively moved towards a confluence of diverse perspectives. This is true of the reconciliation, in France, between the traditions of discourse analysis and of lexical analysis. A sign of this coming together is the growing place of the statistical analysis of textual dimensions, often referred to as ‘textometry’ or ‘logometry’ (Mayaffre 2007). This is also true of the combination of qualitative and quantitative methods in content analysis in the Anglo-Saxon tradition. Similarly, French and American traditions of discourse analysis have grown closer in recent decades. That which originally distinguished them—the nature of discourse analysed (in the first case, political and literary discourses and in the other, the everyday life discourses) and the disciplinary origin (for one, linguistic and for the other, pragmatic)—gradually converged. It is interesting to note that the authors of reference of the Anglo-Saxon tradition of critical discourse analysis (Fairclough 2007) or the school of social representations (Hall 2009) are the same as those of the French school: Barthes (1957), Althusser (1970), Foucault (1969), Derrida (1967), and Lévi-­Strauss (1949). It is equally interesting to note that the analysis of ordinary knowledge and conversation has crossed the Atlantic in the other direction. It is out of the question to define a fictional unity of discourse analysis domain, but it is certainly worth noting that the research practices in discourse analysis combine, rather than oppose, more and more disciplines, approaches, and methodologies.

3

Mixed Methods

The confluence of theoretical and methodological approaches in the current practices of discourse analysis involves the use of mixed methods. The idea of mixed methods fits into the broader project to overcome the opposition between qualitative and quantitative approaches, and to

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

29

somehow combine the two methodologies. While the quantitative methods are relatively easy to define, it is not the case for the qualitative ones. For example, the contrast between the upper left and the upper right of Fig. 2.1 indicates two different qualitative perspectives. Methods of discourse analysis aim to describe the forms and functions of language; in fact they take into account the qualitative aspects of speech. The latter refers more properly to the qualitative paradigm as such. But before going further in the characterisation of quantitative and qualitative paradigms, we must insist on the fundamental difference between the two approaches. While the definition of the quantitative approach is quite simple, for example, the use of mathematical and statistical tools in order to describe, explain, and predict phenomena through operationalised concepts as measurable variables, the qualitative approach refers to a large number of research practices, such as those listed by Denzin and Lincoln (1994): case study, ethnography, participant observation, phenomenology, ethnomethodology, grounded theory, biographical method, action research, and clinical research. More profoundly, quantitative and qualitative paradigms differ on three levels: epistemological, analytical, and operational. The paradigmatic configurations can vary in different ways according to the ontological positions adopted by researchers, but they generally indicate common positions regarding the task they are given. For the moment, we will not develop further the ontological questions regarding the existence of reality and truth that lies upstream of epistemological positions. These postures, positivist, post-positivist, critical, or constructivist give reality a more or less autonomous status. The same can be said about the regime of truth, the degree of relativity increasing, here, on the axis ranging from positivism to constructivism.2 These postures necessarily influence the various paradigmatic positions. We will instead concentrate on the analytical and operational plans characterising both qualitative and quantitative paradigms. These form a series of oppositions that should be thoroughly discussed. But the goal here  See also Table 6.2 ‘Paradigm positions on Selected Practical Issues’ in Guba and Lincoln (1994) and Table 1 ‘Trois positions ontologiques dans les sciences sociales contemporaines’ in Duchastel and Laberge (1999b). 2

30 

J. Duchastel and D. Laberge

is to give an overview of the main debates between the two viewpoints. At the epistemological level, three questions arise. The first question regards the viewpoint of the observer: while the quantitative approach adopts a positivist perspective, advocating a measure of distance between the observer and the data and procedural objectivity, the qualitative approach promotes empathy and subjectivity. The second question concerns the capacity for generalization. Quantitative scientists aim at formulating general and universal propositions, while the qualitative scientists insist on uniqueness and context. The third question is about the value of truth. Quantitative researchers put forward procedures’ validity and observers’ neutrality. The qualitative researchers prefer the ideas of transferability and credibility to those of validity and axiological commitment to neutrality. In analytical terms, quantitative methods proceed to the reduction of complexity, while qualitative methods favour its full apprehension. Quantitative scientists promote a deductive approach, at least in the confirmatory phase, while the qualitative researchers support induction or abduction. Moreover, the quantitative analysts encourage width (thin analysis) rather than depth (thick analysis) that characterizes the qualitative approach. Finally, in terms of operations, quantitative research works on variables while qualitative research is more interested in intentional actions. Quantitative research favours measurement rather than focus on qualitative processes. Consequently, quantitative researchers seek confirmatory statistical tests when qualitative researchers employ exploratory procedures. In summary, the purpose of quantitative m ­ ethods would be causal explanation and that of qualitative methods the understanding of meaning. The use of mixed methods can be explained by the relative weakening of the paradigmatic oppositions between quantitative and qualitative methods, and the adoption of a more pragmatic attitude. Aware of the variable nature of the data and of their actual availability, researchers have come to use materials or analytical approaches that have previously tended to be opposed. These changes are mostly based on pragmatic arguments: ‘It works!’ A review of practices in the area of mixed methods shows that there are essentially three arguments to justify these combinations. The first argument can be described as functional. It consists of simply juxtaposing the use of various types of methods according to the

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

31

needs of the research project and the nature of the data. The choice is up to the researcher to establish the sequence of qualitative and quantitative methods and their relative importance (QUAN > qual, QUAL > quan, QUAN = QUAL) as part of the research process. The second argument is more substantive. It justifies the hybridization of methods according to the nature of data. For example, discourse analysis and content analysis are applied to phenomena including aspects of both qualitative and quantitative nature. The third argument is epistemological. The use of mixed methods is legitimated by the idea of triangulation. Triangulation is seen as a way to increase confidence in the research results. However, we must recognize that the use of the term ‘triangulation’ is mostly metaphorical (Kelle 2001) and does not formally ensure a greater validity, except in the form of convergence or confirmation of findings. In sum, the use of mixed methods only proves that there should not be mutually exclusive types of methods. It seems, however, insufficient to reduce the issue of mixed methods to their sole effectiveness without trying to understand the implications of epistemological, analytical, and operational oppositions characterizing both qualitative and quantitative paradigms on these new forms of empirical approaches.

4

Explaining and Understanding

What can be drawn from the above? On the one hand, we have established that the practice of discourse analysis is at the confluence of several disciplines, themselves, relying on more or less quantitative or qualitative, phenomenological or structural, linguistic or sociological approaches. While each tradition has established itself on epistemological, theoretical, and methodological oppositions with other traditions, we can nevertheless observe a certain convergence in the use of methods and the mitigation of previous fractures. On the other hand, the fundamental opposition between qualitative and quantitative methods seems to dissolve in the pragmatic choice of mixed methods. This pragmatism often avoids examination of ontological and epistemological foundations of this practice. This is why we have to question the possible reconciliation of these two so strongly opposed paradigms.

32 

J. Duchastel and D. Laberge

To elucidate this question, it is useful to return to the starting point of the distinction between natural science and humanities as established by Dilthey in the late nineteenth century. This distinction was built on the contrast between explaining and understanding. According to this view, the natural sciences were entirely dedicated to the identification of causal relationships between phenomena, while the humanities sought to uncover the meaning of historically situated experiences. It is this design that better differentiates the paradigmatic opposition between quantitative and qualitative methods. But instead, we will rather rely on the assumption of Ricœur (1981, 161) that “it seems possible to situate explanation and interpretation along a unique hermeneutical arc and to integrate the opposed attitudes of explanation and understanding within an overall conception of reading as the recovery of meaning.” In fact, Ricœur defines a hermeneutical arc, from explanation to understanding, that is to say that the interpretation unfolds in a set of objective procedures for observation, description, and analysis resulting in the understanding of the research object. Hermeneutics cannot be reduced to the immediate interpretation of the observed reality, as might be the case for everyday knowledge. In scientific knowledge, the interpretation is necessarily supported by the mediation of operations that can be named explanatory procedures. This assumption allows us to reject two common conceptions of interpretation. The first comes from within the qualitative paradigm where interpretation is often seen as a hermeneutical comment. One textbook defines qualitative analysis as “a deliberate and rigorous representation and conscious transposition of the ‘self-other-world’ system, in order to make a new exploration in the particular perspective of the humanities and social sciences, which strive to bring out the sense rendering it understandable.” (Our translation, Paillé and Mucchielli 2008, 24) The researchers set out to reveal the meaning of speech in context. In fact, they are mostly interested in the referential function of discourse. But should we not consider that the essence of discourse analysis is to highlight the various linguistic and paralinguistic aspects of speech whose disclosure is necessary for an overall understanding? Interpretation cannot stand on its own and it requires the work of description and explanation.

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

33

The interpretative process’s second conception is restricted to the interpretation of results. In quantitative or qualitative frameworks, the work of interpretation is often limited in establishing the meaning of the results generated by research operations. It then maintains the illusion that these operations are absolutely objective until meaning is assigned to the results they produce. Such a point of view ignores the importance of interpretive acts that mark each stage of the research process. The projection of a theoretical framework, the identification of analytical dimensions, and the choice of values lent to research objects are all housed in the same interpretive acts within objectification procedures. What then is interpretation? In the broadest sense, there is a tendency to confuse this concept with that of understanding or appropriating, for ourselves, the meaning of an action, an intention, or a thought. The researcher would then be asked to develop his empathic abilities, which could give him access to the consciousness of the observed subject. It is true that, at the end of every project, the researcher arrives at a global interpretation of the observed phenomenon that is somehow detached from observation, description, and analytical procedures. This holistic interpretation can be seen as an appropriation for ourselves of the object, the global comprehension of the phenomenon (Duchastel and Laberge 1999a). But in the context of a scientific process, interpretation must be seen as the continuous confrontation of the researcher with discursive materiality (Conein et al. 1981) or language materiality (Paveau 2012). For several authors, we find this strong intuition that access to meaning cannot dodge the discursive materiality. Pêcheux (1975), and later on Paveau (2012) and Molino (1989), insisted that only the very materiality of speech could render analysis possible. Similarly, Ricœur (1981, 149) speaks of “the eclipse of the circumstantial world by the quasi-world of texts” as a condition for reading and interpreting. In sum, hermeneutics as the art of interpretation should be based on a set of procedures for the description, exploration, and analysis of material units of discourse. The intuition behind the project of discourse analysis was, from the outset, to go beyond content analysis and take into account the linguistic dimension of speech. Speech was not to be reduced to its purely linguistic dimensions—lexical or semantic. The hypothesis was to find various traces of discourse functions, such as those developed by Jakobson (1963),

34 

J. Duchastel and D. Laberge

in the material fabric of language. This is the case with enunciation analysis that seeks the inscription of speaker and audience in the thread of discourse. The same is true with the study of markers of argumentation. According to Gee (2011), discourse analysis is about the study of speech on three levels: the analysis of the information it conveys (saying), that of action it raises (doing) and of identity it formulates (being). Each of these dimensions is identifiable only through linguistic forms that make them intelligible. The interpretation must rely on certain classes of observation units and the description of their properties. This process is objectifying as well as interpretative. If this is true, a restrictive approach of interpretation cannot be sustained. Interpretation cannot be limited to the final act of the research process when making sense of results. Rather, interpretation should be present at the very beginning of the research process. Interpretation is part of every research procedure, and all procedures rely on interpretation. This means that explanatory procedures and interpretation go hand in hand and do not oppose each other, as the quarrel of paradigms would suggest. Rather than designing two general paradigms defined by their purpose, explaining, or understanding, it is more productive to integrate both actions within a single process. No science can do without a proper pre-comprehension of the object. There is always a knowledge frame, more or less theoretical, which predetermines the grasping of reality. What is sought is to increase this preliminary understanding. Explanation is most often thought of as establishing a relationship between two phenomena. But, it also has a semantic sense. Kaplan (1964) has defined interpretation as a semantical explanation, thus explaining the meaning of a statement. In both cases, the goal is to better understand. The various procedures for observation, description and analysis of objects are designed to enhance understanding by distancing the object from the subject and by linking the object with the cognitive frameworks at play. However, we must consider the asymmetry of both processes of explanation and interpretation. While explanatory procedures can be controlled to a certain point, the act of interpretation, even if it is well framed, remains difficult to define. The cognitive capacities of the researcher, semantic, emotional, or cultural, will result in some uncertainty of interpretation. However, it is easier to control the micro-level of the interpre-

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

35

tive process in various descriptive and analytical procedures than in the overall understanding of a phenomenon. That is why we distinguish ‘local interpretations’ that can be thought of, if not perfectly controlled, at all stages of the research process and ‘global interpretations’ that bring meaning to the complexity of the object at the expense of an assured mastery of the cognitive processes at work (Duchastel and Laberge 1999a).

5

The Problem of Complexity

One of the most fundamental criticisms addressed to the quantitative paradigm is its reductive approach to the problem of complexity. On the other hand, the comprehensive paradigm is based on the idea that the full complexity of any phenomena must be preserved. It shows strong resistance to any reduction that may dissipate meaning. Instead, an empathic approach is advocated. But is it possible to grasp an object without reducing its complexity and describing it? Qualitative methods are not exempt from this requirement as they shall, themselves, proceed to the identification of units of varying size (words, textual segments, sentences, and paragraphs) to which they affix referential or factual categories. Yet, proponents of the qualitative paradigm insist on the whole rather than the parts. The question may be ill defined. It is rather more appropriate to distinguish between systematic reduction of complexity and oversimplification. Admittedly, the distinction between in-depth analysis (thick) and wide analysis (thin) remains relevant and it is understandable that the first type embraces more complexity. But in all cases, reducing the phenomenon under study is unavoidable. It is not possible to grasp an object in its totality, if not intuitively. Thus we need to temporarily neglect some of its components to retain only a few. Ricœur (1986) explains that discourse analysis can be only done through the mediation of the text. This methodical act of concealing the complexity of the social conditions of discourse allows the proper identification of textual materiality and the observation of its properties and relationships. Such mixed interpretative and explanatory procedures will progressively lead to a more comprehensive understanding at the very end of the research process.

36 

J. Duchastel and D. Laberge

We see the process of understanding as a spiral formed by overlapping circles each having a point of origin based on a prior understanding of the object and an endpoint defined as the enriched understanding of the same object. Between these two points, there is a set of operations of construction, description, and analysis involving both explanation and interpretation procedures. These procedures are formed by the identification of dimensions and units, the description of units based on ­conceptual dimensions, and the exploration of their relationship. All these operations can be performed only on a well-defined materiality. This materiality is that of the text and the text is the main trace of the speech situation. The text is thus some reduction of the situated discourse. It is not possible to carry out the analysis without the use of a textual support, in contrast to mundane understanding in everyday life. The transformation of the text over the course of research will show how a dual process of reduction and recovery of complexity operates. Figure 2.2 shows the various stages in the transformation of the text with each specific methodical operations of discourse analysis. The initial form of the text is speech itself. It consists of the raw material on which we will perform various research operations. The ‘speech’ text is the starting point, a complex object produced within a socio-historical, cultural, cognitive and linguistic context, and a specific communication situation. The first transformation is to establish a ‘manuscript’ text. Initially, we may have a spoken or written speech, already in the form of a text. In the case of written speech, we then must select and authenticate a version of

Fig. 2.2  Transformation of the text. The figure is an adaption of a schema presented in Meunier (1993)

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

37

the speech that will become a text ‘outside of the world’, in the words of Ricœur. In the case of oral discourse, we first proceed to its transcription. Oral discourse includes a set of prosodic and contextual features that can be recorded in a more or less developed format using established conventions (e.g., Jefferson 2004). The ‘­manuscript’ text is an object both different and less complex than the original, in the sense that the conditions and context of its production and enunciation are no more present otherwise than within the text itself. The next transformation will produce an ‘edited’ text. Whatever the characterization of the manuscripts, transcripts of oral, in paper or computerized format, standardization and normalization work must be done in order to make the various elements of a corpus comparable. Information about the conditions of production of speech and of enunciation (speaker, support, place, time, etc.) must define each document of a corpus. We get a new ‘edited’ text which will be subsequently the object of description, exploration, and analysis. In summary, the ‘manuscript’ text is a derivation of the original discourse which  version has been established by authentication or transcription and the edited text is, in turn, the result of standardization and indexation according to a system of rules and descriptive categories. It is on the basis of this ‘edited’ text that the work of description, exploration, and analysis can be further performed. Which actions should then be performed on this textual material? We can define two universal research operations whatever the approach. The first task is to establish the observation units: What is to be observed? The second task consists of the description of these units based on one or more systems of categories: How is it to be observed? Observation units can be represented as a set of nested elements, from the global corpus to the sub-corpora, to the collection of texts that constitute each of them, to the various parts of each text, and finally to the middle and micro-level text units. Each nesting level of units may be described into a  system of categories. The corpus itself and its subsets are indexed with a metadata system. Every text component (section, paragraph, verbal exchanges, etc.) can be marked. Finally, speech units (textual segments, turns of speech, sentences, words) are coded depending on the research target (e.g., morpho-syntactic, semantic, pragmatic, enunciative, argumentative coding). Thus, the descriptive system unfolds at three levels: The corpus

38 

J. Duchastel and D. Laberge

is described by meta-categories, the parts of text are described by structural variables, and the speech units are described by a variety of properties associated with the research design. Arguably, the ‘edited’ text is actually transformed into a series of ‘representations’, in the sense that the text is now enriched with descriptions, and in some way, a form of complexity is partially restored. It represents, however, multiple images of the original text, but in no way corresponds fully to the context of its utterance. All text descriptions can be sorted and compiled. They may or may not be the subject of counts, crossovers, comparisons based on various segments established on the basis of metadata or structural variables. Each data mining operations described will result in the production of many new texts in the form of comments or numerical results. Each of these subtexts will only be a distant image of the original text. It is the accumulation of these images which will allow further exploration of the original speech and lead to the interpretation of the data, producing a new transformation of the text in the form of ‘interpretation’. The interpretation of the results can be partial or global, depending on whether we choose to interpret the empirical data produced by different sets of explorations or we attempt to give an overall sense of the whole data. Global interpretation will then mobilize much more than methodological devices. Theoretical and socio-­historical knowledge are needed to restore the full complexity of discourse in action. The final form of the text is a new text, the ‘interpretation’ text taking the form of an article or monograph aiming at the increased understanding of the phenomenon being studied. This more or less metaphorical representation of a succession of states of the text goes to show that speech can only be grasped in the form of its textual materiality which must be later subjected to methodical operations. From this point of view, it does not seem appropriate to distinguish between quantitative and qualitative methods. On the epistemological level, it is not productive to oppose complexity and simplicity. We have seen that understanding and explanation should form a hermeneutical arc. Any methodological approach necessarily implies a reduction of the object allowing some objectification of data. As we saw earlier, this process involves both operations of explanation and interpretation. These operations ultimately lead to the formulation of interpretative hypotheses

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

39

that allow for the appropriation of the object for ourselves, that is to say, its understanding.

6

 ausality and Measurement in Discourse C Analysis

We have tried so far to show how discourse analysis is, as its name suggests, a practice that focuses on the discursive materiality and implements systematic operations, both explanatory and interpretative. We have challenged the strict opposition between the qualitative and quantitative paradigms while recognizing the existence of distinctive practices concerned with quantitative or qualitative aspects of phenomena. The paradigmatic opposition between qualitative and quantitative approaches emphasizes two distinct criteria. As we have pointed out, the quantitative approach would favour measurement and causal explanation, and the qualitative approach would rather choose the global understanding of phenomena. To be convinced of the compatibility of the two approaches, it is useful to examine the presence of causal reasoning in the practice of discourse analysis and the relevance of measuring as an operation favouring at the same time reduction and restoration of complexity. We will attempt to illustrate how causal explanation and measurement have their place in the qualitative approach. With regard to causation, we refer to Tacq’s proposal (2011) that causal reasoning is present in both quantitative and qualitative researches. He gives an overview of different theories of causality in the social sciences to stress the idea of an experimental logic present in both approaches. He starts from the premise that in science, the causal relationship is rarely apprehended directly, but rather is considered in an indirect way, a sort of encirclement process. Thus, science most often uses probabilistic or statistical approaches to examine the necessary and sufficient conditions explaining a phenomenon, without being able to establish a direct causal link between phenomena. To support his conception of experimental logic, Tacq relies on the INUS model (Insufficient but Necessary part of a set, which is Unnecessary but Sufficient for the Result, Mackie 1974), which bases the nature of reasoning on all the conditions making possible the occurrence of an event.

40 

J. Duchastel and D. Laberge

According to the INUS model, an event may be the product of a necessary condition but insufficient in general, while being sufficient although not necessary under the circumstances. Tacq gives the following example: Experts may say that fire is the result of a short circuit. The cause cannot be declared necessary because other factors could cause fire. It cannot be declared sufficient since other conditions may contribute to the spread of fire. All we can say is that, combined with the short circuit, there is a set of positive or negative conditions that are sufficient without being necessary to trigger the fire. It is a counterfactual argument that questions the possibility of the occurrence of an event in the absence of an identified causal factor. The perspective is that of a causal field rather than a logical causation. According to the author, this type of reasoning is widely used in experimental research. But it is surely the kind of logic that is applied in qualitative research. To support his thesis, Tacq responds to the main arguments that aim at distinguishing qualitative and quantitative approaches. The first argument pertains to the measurement scales, nominal, ordinal, interval, and metric. The first two levels, nominal and ordinal, would characterize the qualitative approach, allowing limited mathematical operations, thus excluding the causal logic implied by quantitative models. While mathematical operations vary depending on the nature of the variables, it does not follow that the causal logic is de facto excluded. The second argument is based on the difference in sample size between the qualitative and quantitative approaches. In extreme cases, qualitative studies will apply to a single case, making causal analysis improbable. Tacq notes that there are few objective criteria for determining the minimum sample size and even the analysis of a single case can make sense, provided it is placed in relation with other single-case studies. The analysis of complex necessary and sufficient conditions is still possible by the counterfactual examination of these conditions. The third argument regards the possibility of statistical tests. Obviously, the power of statistical tests varies greatly depending on the sample size. However, there are a variety of tests that have been produced to validate the results of small samples, and comparison of data with data obtained in other studies is, in itself, a kind of test, even if not statistical. The last argument pertains to the difference between thin and thick analyses. Again, there is no doubt that

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

41

in-depth analysis multiplies the dimensions of the object that can be observed, while the analysis in width multiplies the number of ­individuals observed for a limited number of dimensions. This should not, however, change the argument, especially as there is no reason not to combine qualitative and quantitative procedures at various stages of the research process. The author comes to the conclusion that if we use the counterfactual and conditional approach of INUS’s model and the method of difference at the base of the experimental approach as formulated by John Stuart Mill, there is no principled difference between quantitative and qualitative methods in terms of causal reasoning. We will conclude by showing that the use of measurement is not inconsistent with a qualitative approach. If one refers to the qualitative paradigm, measurement is conceived as a distortion of the research object and would constitute a misleading and unnecessary analysis, precisely because it reduces complexity. However, measurement is one of the research operations that allows at the same time a reduction of the dimensions under study and possibly the production of another order of complexity. We retain the definition proposed by Kaplan (1964, 177): “Measurement, in the most general terms, can be regarded as the assignment of numbers to objects (or events or situations) in accord with some rule.” The properties of the object and their measurability do not exist independently of a theory. The qualitative or quantitative representation of an object depends on the choice of a system of symbolic representation. In the words of Kaplan, “quantities are of qualities and a measured quality has just the magnitude expressed in its measure” (1964, 207). In sum, measure can be applied at various levels of construction of the object. First, it can be applied to any object with an independent material existence, regardless of its nature, size, and complexity, such as individuals, world objects, texts, statements, events. Second, it can be applied to segments or properties of these objects not directly accessible to observation, but arising from research work. Third, the measure may even extend to intangible objects that exist through the work of the mind. This last kind of objects might be a social production (success, wealth, popularity, etc.) or the product of disciplinary knowledge (anomie, social relativism, creativity, etc.).

42 

J. Duchastel and D. Laberge

To resume our earlier discussion, the measuring may indeed appear to be a reduction of information. In the different phases leading to measurement, only certain attributes are deemed relevant to the process. It implies that we waiver the diversity of concrete manifestations, physical or imagined, of one’s research object. This work of abstraction is present in both qualitative and quantitative approaches. It is reflected in the operations of description and categorization of the chosen units. Categorization consists in a double reduction of the object by identifying a particular aspect of the object and allocating an abstract value that can represent it. Giving values to units and their properties follows previous work of reduction and abstraction of the object’s dimensions. In return, measurement may also help restore complexity. It can indeed be a powerful heuristic strategy to rebuild complex representations of aspects or attributes postulated in theory. For example, the construction of indices to represent a concept by adding and weighting indicators leads to the emergence of a form of  complexity non-apparent at the starting point. In the same fashion, multidimensional statistical analysis produces information that was not there from the start (see also Duchastel and Laberge 2011). Discourse analysis is a good example for the use of measurement as part of a mixed methods approach. The different operations of description and analysis of discourse data show that measurement can contribute both to the abstraction of specific dimensions of the object and to the restoration of complexity. Analysis relies on the capacity to identify series of discrete speech units (words, semantically meaningful phrases, broader textual segments, etc.) and to determine a system of categorization (semantic, sociological, argumentative, pragmatic, enunciative, etc.). The researcher remains free to determine whether he will take into account only the units, whatever the type, or if he is interested in their properties. Counting these objects will only give a partial view of the whole. For example, we could learn about the proportion of nouns belonging to a semantic class, the dominant premises of an argument, the relative importance of certain enunciative markers in a political speech, the frequency of speech turns in a conversation, etc. Thus, one can speak of a reductive reading manifested both by a certain selection of aspects of the text and by its representation in a measurement system. But it is also possible to

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

43

speak of a more complex representation of the text by the multiplication of observations and accumulated elements measured. The accumulation of observations and measurements can lead to the construction of indices or increase the size of the analysis. Measurement is then one of the operations available in discourse analysis. It is not inherently incompatible with the qualitative approach.

7

Conclusion

We have shown that discourse analysis is not a discipline but a research practice that is at the confluence of a set of disciplinary and national traditions. The rich heritage of disciplinary, theoretical, and methodological knowledge explains the privileged position of discourse analysis. The very purpose of discourse analysis predisposes it to stay at the frontier of different methodological approaches, which might be called mixed methods. We have shown that the paradigmatic oppositions between qualitative and quantitative approaches, although strongly advocated in the body of scientific literature, have become obsolete in the pragmatic use of mixed methods. We went beyond this pragmatic attitude to defend the thesis that there is indeed a common background in all methodologies, whatever their paradigmatic affiliation. We have shown that we cannot explain without interpreting at the same time, and that the very identification of research units and operations of description and analysis combines, at all times, explanation and interpretation. We further stated that scientific knowledge cannot proceed without applying some reduction procedures, but that the combination of these procedures can lead to a restoration of the complexity of the object. We ended by showing that the logic of causality and measurement, seemingly opposed to the qualitative paradigm, applies to both qualitative and quantitative approaches. Acknowledgements  We are thankful to Beltz Juventa Publishing House  for allowing to reprint this text, which originally was published in the Zeitschrift für Diskursforschung (2014/2).

44 

J. Duchastel and D. Laberge

References Adam, Jean-Michel. 1999. Linguistique textuelle. Des genres de discours aux textes. Paris: Nathan. Althusser, Louis. 1970. Idéologie et appareils idéologiques d’État. La Pensée 151 (juin). Barthes, Roland. 1957. Mythologies. Paris: Seuil. Conein, Bernard, Jean Jacques Courtine, Françoise Gadet, Edward W. Marandin, and Michel Pêcheux, eds. 1981. Matérialités discursives. Actes du colloque de Nanterre (24–26 avril 1980). Lille: Presses universitaires de Lille. Denzin, Norman K., and Yvonna S.  Lincoln. 1994. Handbook of qualitative research. London: Sage. Derrida, Jacques. 1967. L’écriture et la différence. Paris: Seuil. Duchastel, Jules, and Danielle Laberge. 1999a. Des interprétations locales aux interprétations globales: Combler le hiatus. In Sociologie et normativité scientifique, ed. Nicole Ramognino and Gilles Houle, 51–72. Toulouse: Presses Universitaires Du Mirail. ———. 1999b. La recherche comme espace de médiation interdisciplinaire. Sociologie et Sociétés XXXI (1): 63–76. ———. 2011. La mesure comme représentation de l’objet. Analyse et interprétation. Sociologies (Avril). Accessed June 27, 2018. https://journals.openedition.org/sociologies/3435. Fairclough, Norman. 2007. Discourse and social change. Cambridge: Polity. Foucault, Michel. 1969. L’archéologie du savoir. Paris: Gallimard. Gee, James Paul. 2011. An introduction to discourse analysis. Theory and method. 3rd ed. New York: Routledge. Guba, Egon G., and Yvonna S. Lincoln. 1994. Competing paradigms in qualitative research. In Handbook of qualitative research, ed. Norman K. Denzin and Yvonna S. Lincoln, 105–117. London: Sage. Hall, Stuart. 2009. Representation, cultural representations and signifying practices. London: Sage. Haroche, Claudine, Paul Henry, and Michel Pêcheux. 1971. La Sémantique et la coupure saussurienne: Langue, langage, discours. Langages 24: 93–106. Harris, Zellig. 1952. Discourse analysis. Language 28 (1): 1–30. Jakobson, Roman. 1963. Essais de linguistique générale. Paris: Minuit. Jefferson, Gail. 2004. Glossary of transcript symbols. In Conversation analysis: Studies from the first generation, ed. Gene H. Lerner, 13–31. Amsterdam: John Benjamins Publications.

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

45

Kaplan, Abraham. 1964. The conduct of inquiry. Methodology for behavioral science. New York: Chandler Publishing. Kelle, Udo. 2001. Sociological explanations between micro and macro and the integration of qualitative and quantitative methods. Forum Qualitative Social Research 2(1). https://doi.org/10.17169/fqs-2.1.966. Accessed June 27, 2018. Mackie, John L. 1974. The cement of the universe. A study of causation. Oxford: Oxford University Press. Mayaffre, Damon. 2007. Analyses logométriques et rhétoriques des discours. In Introduction à la recherche en sic, ed. Stéphane Olivési, 153–180. Grenoble: Presses Universitaires De Grenoble. Meunier, Jean-Guy  1993. Le traitement et l‘analyse informatique des textes. Revue de Liaison de la recherche en informatique cognitive des organisations (ICO Québec) 6 (1–2): 19–41. Molino, Jean. 1989. Interpréter. In L‘interprétation des textes, ed. Claude Reichler, 9–52. Paris: Editions De Minuit. Osgood, Charles E. 1959. The representational model and relevant research methods. In Trends in content analysis, ed. Ithiel de Sola Pool, 33–88. Urbana: University of Illinois Press. Paillé, Pierre, and Alex Mucchielli. 2008. L’analyse qualitative en sciences humaines et sociales. Paris: Armand Colin. Paveau, Marie-Anne. 2012. L’alternative quantitatif/qualitatif à l’épreuve des univers discursifs numériques. In Colloque international et interdisciplinaire Complémentarité des approches qualitatives et quantitatives dans l’analyse des discours?, Amiens, France. Pêcheux, Michel. 1975. Les vérités de la Palice, linguistique, sémantique, philosophie. Paris: Maspero. Pires, Alvaro P. 1982. La méthode qualitative en Amérique du nord: un débat manqué (1918–1960). Sociologie et société 14 (1): 16–29. Rastier, François. 2001. Arts et sciences du Texte. Paris: PUF. Ricœur, Paul. 1981. Hemeneutics and the human sciences. Essays on language, action and interpretation. Cambridge: Cambridge University Press. ———. 1986. Du texte à l’action. Paris: Seuil. Tacq, Jacques. 2011. Causality in qualitative and quantitative research. Quality and Quantity 45 (2): 263–291. Zienkowski, Jan. 2012. Overcoming the post-structuralist methodological deficit. Metapragmatic markers and interpretative logic in a critique of the bologna process. Pragmatics 22 (3): 501–534.

46 

J. Duchastel and D. Laberge

References of Figure 2.1 Althusser, Louis. 1970. Idéologie et appareils idéologiques d‘État. La Pensée 151 (juin). Austin, John L. 1962. How to do things with words. New York: Oxford University Press. Barthes, Roland. 1957. Mythologies. Paris: Seuil. Benveniste, Emile. 1966. Problèmes de linguistique générale. 1. Paris: Gallimard. Benzecri, Jean-Paul. 1973. L’analyse des données: l’analyse des correspondances. Paris: Dunod. Berelson, Bernard. 1952. Content analysis in communication research. New York: Hafner Publications. Berger, Peter, and Thomas Luckman. 1966. The social construction of reality. A treatise in the sociology of knowledge. New York: Anchor Books. Dubois, Jean. 1969. Énoncé et énonciation. Languages 4 (13): 100–110. Foucault, Michel. 1969. L’archéologie du savoir. Paris: Gallimard. Garfinkel, Harold. 1967. Studies in ethnomethodology. Englewood Cliffs, NJ: Prentice Hall. Guiraud, Pierre. 1960. Problèmes et méthodes de la statistique linguistique. Paris: PUF. Harris, Zellig. 1952. Discourse analysis. Language 28 (1): 1–30. Hjelmslev, Louis. 1947. Structural analysis of language. Studia Linguistica 1 (1–3): 69–78. Holsti, Ole R. 1969. Content analysis for the social sciences and humanities. Reading, MA: Addison Wesley. Lacan, Jacques. 1966. Écrits. 2 vols. Paris: Seuil. Lasswell, Harold D., Daniel Lerner, and Ithiel de Sola Pool. 1952. The comparative study of symbols. Stanford, CA: Stanford University Press. Lévi-Strauss, Claude. 1949. Les structures élémentaires de la parenté. Paris: PUF. Mead, George H. 1934. Mind, self, and society. Chicago: University of Chicago Press. Muller, Charles. 1968. Initiation à la statistique linguistique. Paris: Larousse. Pêcheux, Michel. 1969. Analyse automatique du discours. Paris: Dunod. Sacks, Harvey. 1972. An initial investigation of the usability of conversational data for doing sociology. In Studies in social interaction, ed. David Sudnow, 31–74. New York: Free Press. Schütz, Alfred. 1967. The phenomenology of the social world. Evanston, IL: Northwestern University Press.

  Beyond the Quantitative and Qualitative Cleavage: Confluence… 

47

Searle, John. 1970. Speech acts. An essay in the philosophy of language. Cambridge: Cambridge University Press. Stone, Philip J., Dexter C. Dunphy, Marshall S. Smith, and Daniel M. Ogilvie. 1966. The general inquirer. A computer approach to content analysis. Cambridge, MA: MIT Press.

Part II Analysing Institutional Contexts of Discourses

3 The Academic Dispositif: Towards a Context-Centred Discourse Analysis Julian Hamann, Jens Maesse, Ronny Scholz, and Johannes Angermuller

1

Introduction

In discourse, meanings are realised and established among members of a social community. Discourse is a meaning-making practice, which operates with gestures, images, and, most importantly, with language. From a discourse analytical point of view, texts and contexts, utterances and their The authors thank Johannes Beetz, Sixian Hah, and one anonymous reviewer for their comments on previous versions of this contribution. They are also very grateful to Marie Peres-Leblanc for improving the design of the visualisations.

J. Hamann (*) Leibniz Center for Science and Society, Leibniz University Hannover, Hannover, Germany e-mail: [email protected] J. Maesse Department of Sociology, Justus-Liebig University Giessen, Gießen, Germany e-mail: [email protected] R. Scholz • J. Angermuller Centre for Applied Linguistics, University of Warwick, Coventry, UK e-mail: [email protected]; [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_3

51

52 

J. Hamann et al.

uses in certain social structures and institutional settings always need to be studied in conjunction. Language does not make sense outside the contexts in which it is used. And it is difficult to account for contexts without language referring to them. Discourse analysts, therefore, deal with the problem of how texts are articulated with contexts, i.e. the social configurations, which are not simply represented but also constituted in discursive practices. Discourse analysts deal with the relationships between texts and social contexts in many ways. While for some approaches social contexts are closely tied to discursive practices, others conceive of social context as existing independently of discursive practices. Pragmatic approaches to discourse, which look at the contextualisation of utterances, can be cited as an example of the former. To analyse discourse pragmatically means to ask how utterances, through the cues, markers, and instructions they provide, evoke certain (aspects of ) contexts. The latter, by contrast, is testified by structural or power theoretical approaches to discourse. In the latter view, for the social meaning of texts to be understood, one needs to relate to the wider social and historical horizon in which they are used. Discourse analysis, in that view, would produce a partial picture of social meaning-making practices if it ignored aspects of social reality that are not shown in texts. In this contribution, we will make the case for a context-centred discourse analysis by citing findings from our research on academic discourses in the ERC DISCONEX project.1 Dispositif is a Foucauldian concept that tries to capture and link the heterogeneous textual and non-­ textual context conditions in which discourses emerge. Operationalisations of the concept pay attention to hierarchies and power structures in institutions, architectural order, regulations and laws, and scientific statements, and so on that together form a network  The concept for the information system from which we draw our examples was developed within the research project ‘Discursive Construction of Academic Excellence’, funded by the European Research Council and led by Johannes Angermuller. We are grateful to the whole ERC DISCONEX team for allowing us to present a part of their research ideas to which all four authors have contributed in various stages. For more information see: http://www.disconex.discourseanalysis.net. 1

  The Academic Dispositif: Towards a Context-Centred Discourse… 

53

of dispositional elements that the construction of meaning in discourses depends on (for an introduction: Raffnsøe et al. 2016). A dispositif analysis integrates an analysis of language use with an analysis of the aforementioned context elements. In this contribution, we will give an example of how to account for a dispositif with the help of quantitative methods. We apply these methods to short texts from institutional webpages. This first analysis is part of a larger study in which we combine an analysis of academic texts with a dispositif analysis of academic careers and institutions. This larger study aims to show that academic publications need to be seen against the background of the researchers’ social positions in the institutional dispositif, as well as the symbolic dynamics within their scientific communities. We will apply these theoretical considerations with two specific empirical questions in mind: How can we account for the discursive construction of social order in the academic discipline of sociology? How do social relationships of difference and inequality that constitute the sociological field go together with the linguistic organisation of the field, that is, a specific distribution of words and expressions across the social space? The first section will discuss why we should integrate context data into our discourse analysis, and it will elaborate on the different types of social contexts that sociologists relate to in their discursive practices. In the second section, we will ask how sociological insights into the social context can inform a discourse analysis of the discipline of sociology. Specifically, we will discuss possible (statistical) categories that help to operationalise contexts of academic discourse. Our approach is illustrated in the third section, which outlines an empirical analysis of the research interests that UK full professors in sociology mention on their homepages. We demonstrate how these data can serve as an entry point for a context-centred analysis of academic discourse. The fourth section ties the argument back to the theoretical background of the dispositif approach and reveals the methodological framework that guides the analysis and offers pathways of interpretation. The contribution closes with a discussion of our proposed approach in the fifth section.

54 

2

J. Hamann et al.

Integrating Sociological Data into Discourse Analysis

In this section we will outline a methodological argument that explains why analyses of discourse need an explicit and systematic analysis of social and historical contexts. The combination of linguistically informed discourse analysis and sociologically informed context study seems to be useful because texts, language, gestures, and pictures, on the one hand, as well as social and institutional rules and structures, on the other, refer to and constitute each other.

2.1

The Problem of Text and Context

The question of text and context has been the subject of a great deal of controversy. In line with socially minded linguists, who have long insisted on the systematic empirical observation of real linguistic and discursive practices (Bhatia et al. 2008; Blommaert 2005; Sarangi and Coulthard 2000), we will make the case for a sociological take on social and historical contexts. We will cite and elaborate Foucault’s concept of dispositif in order to seize the social context as an institutional arrangement of linguistic practices and non-linguistic practices, rules, and structures in a larger social community. While text and talk can be analysed with the classical instruments of discourse analysis (from pragmatics to corpus analysis), the dispositif is analysed with the help of sociological methods (such as interviews, questionnaires, statistical analysis, ethnography). With the concept of the dispositif, we make the case for sociological perspectives on discursive practices as embedded in institutional power arrangements (the notion of dispositif has been the object of debate in France, where the term originated, and Germany: Angermüller 2010; Angermuller and Philippe 2015; Bührmann and Schneider 2007, 2008; Maesse and Hamann 2016; Maingueneau 1991; Spieß et al. 2012). The dispositif approach encompasses power and social structures (Bourdieu 2010), the nexus of power and knowledge (Foucault 1972), as well as institutionally organised processes of interpretation (Angermüller 2010). It takes CDA perspectives further, in that it pleads for studying social

  The Academic Dispositif: Towards a Context-Centred Discourse… 

55

context empirically and systematically. Rather than leaving claims about social inequality to intuition, it asks for methodological symmetry: contexts should be studied just as systematically as texts. Moreover, for the dispositif analyst, context is not a social reality waiting as it were behind texts. It is an arrangement of social practices that not only represent the social reality, but by representing it constitute it. Our plea for a combination of linguistic discourse analysis and sociological dispositif analysis responds to a demand for addressing questions of power and inequality in discourse research. While we agree with those interactional and praxeological discourse analysts who criticise structural approaches to discourses for subsuming the material under preconceived theories, we do think that power and inequality are questions that discourse analysts often need to address. Social structures can have an important effect on meaning-making practices in society without being referred to explicitly, or even implicitly, in a given text or discursive practice. Indeed, how could one understand the implicit political messages of an academic publication, an OECD report, a New York Times article, or the performance of a stand-up comedian if one filtered out the institutional setting and the broader power field in which it circulates? Yet, claims about broader institutional, social, and historical contexts cannot be produced in an ad hoc way; they necessitate systematic theoretical and empirical scrutiny. We enter the debate on discourse and context from both a sociological as well as a linguistic point of view. Discourse analysis does not only include the study of linguistic characteristics of texts, symbols, icons, gestures, and other forms of expression. It also requires the systematic empirical study of social contexts in which these linguistic forms orchestrate interpretation and create social meaning. However, the differentiation between text and context is one possibility among others that seeks to come to terms with this problem in discourse analysis. We take the opposition of text and context as an analytical starting point in order to argue for a methodological differentiation between the analysis of linguistic qualities (of, for example, texts) and the sociological study of contexts (e.g. social structures and institutions). While ‘discourse’ is understood as the combination of text and context, we keep both dimensions apart for analytical and heuristic reasons that become relevant in

56 

J. Hamann et al.

practical research processes (Angermüller 2004, 2006; Hamann 2014; Maesse 2010, 2015a).

2.2

Levels of Discourse Analysis

The discourse analytical process usually takes place at three different levels of empirical investigation, as outlined in Table 3.1. The first level deals with problems that are first and foremost linguistic in nature and located on the text level of discourse. At this stage, qualitative and quantitative methods are applied to analyse the formal rules that make linguistic forms (spoken words, written texts, gestures, as well as pictures) readable in respective social contexts. Thus, the analysis of argumentation, deixis, categorisations, polyphony, co-occurrences, and so forth requires the study of small utterances as well as large textual corpora. The particular choice of method depends on the research question or on corpus characteristics. Nonetheless, the linguistic level must not be confused with the social contexts in which language appears in order to create meaning(s). After the linguistic level, a sociological level emerges, which cannot be studied with linguistic methods. At this point, the discourse analytical Table 3.1  Three levels of analysis Level

Example

Analytical goal

Language

A book, a media corpus, an utterance, a corpus of utterances

Social context

An academic discipline, situations such as a workshop, the national Higher Education System ‘Class struggle within academia’, ‘identity formations of researchers’, ‘functional differentiation of academic systems’, ‘face in academic interactions’, etc.

Studying linguistic forms’ deixis, co-occurrence, attribution, etc. with quantitative and qualitative methods from discourse analysis Studying the social and institutional rules as well as social conventions

Theoretical interpretation

Give explanations of data and build theoretical models of what is happing in society

  The Academic Dispositif: Towards a Context-Centred Discourse… 

57

process moves from the first, the linguistic, level, to the second level of investigation: social context(s) (Table 3.1). This switch from one level to another is required in qualitative as well as in quantifying approaches to textual materials. Large data sets and corpora as well as small utterances neither speak nor interpret themselves. As is illustrated in Table 3.1, the linguistic and sociological levels of discourse analysis are complemented by a third level: theoretical interpretation. Taken by themselves, neither linguistic nor contextual data are interpretations. Furthermore, the combination of discourse analytical data with data from sociological analysis is not an automatic and natural procedure either. Interpretations do not emerge from the data; they are made by those who interpret them. This is where the significance of the theoretical level of analysis comes into play. Researchers can mobilise theories and paradigms for the interpretation of data and they can build new theories and explanations on the basis of data interpretations led by theories. Whereas positivistic approaches follow a data-theory determinism, we suggest giving theory its place in research processes as a tool for data interpretation and as a result of those processes. Theory is simultaneously the creative starting point and the result of every research process. It helps to make sense of data. While the theoretical level will be addressed in the fourth section of this contribution, let us briefly return to the contextual level. The three levels of analysis can be observed in various types of discourse analysis. For example, if pragmatic discourse analysts ask how utterances are contextualised through markers of polyphony (Angermuller 2015), their analysis does not stop at the textual level. An example from economic expert discourses (Maesse 2015b) can show how an analysis of meaning-making on the micro level (of utterances) can be combined with the study of institutional contexts on the macro level. The following exemplary statement was uttered by the economist Joseph Stiglitz. What is needed is a macroeconomic theory based on theories of imperfect competition and imperfect information. (Stiglitz 1984, 355)

The author provides us with linguistic material that can be studied with respect to the deixis of person, time, and place. The utterance thus

58 

J. Hamann et al.

refers the reader to the context in which it is uttered by the locutor (i.e. Stiglitz). To make sense of this utterance, the reader will need an understanding of the locutor and their context. It will, therefore, be important to know that Stiglitz has been awarded a Nobel Prize, is an established academic at several Ivy League universities, a popular commentator of economic policies and chief economist of the World Bank. Yet, knowledge about the social context is not only important for this type of linguistic micro-analysis of utterances. As will become clear in the following, more structural approaches, too, articulate linguistic and sociological levels of analysis and integrate them into a theoretical explanation.

2.3

Accounting for Social Context

There is a variety of qualitative and quantitative methods that can be used to account for a multitude of social contexts of discourse (Table 3.2). In sociology, quantitative methods are typically used to portray aspects of social reality from a macro-sociological perspective that is interested in large-scale distribution. For instance, there is a tradition of comparing societies by analysing their social structures. Scholars draw, for example, on large data sets comprising birth rate, education duration and attainment, family size, income level, professional occupation, and income differences. The first impression from a discourse analytical perspective is quite similar: data on social or institutional structures are not very helpful when we are actually interested in discursive practices in the production of knowledge, meaning, and subjectivity in different cotexts and contexts. We argue, however, that socio-structural data are indeed important for a better understanding of the social contexts and institutional power relations in which discourses emerge. Some relevant social contexts may betray more stable structures. They spread out over vast areas of the social world and their influence can only be detected on the macro level. Among these ‘hard facts’ of social life one can discover, for example, stable institutional structures in which a discourse on educational reforms is embedded (Maesse 2010) or the socio-structural backgrounds of a university

  The Academic Dispositif: Towards a Context-Centred Discourse… 

59

Table 3.2  Four ideal typical dimensions of social context for the analysis of discourses Type of Social relations Institutions and Epistemic Forms of social context organisations resources practice Rankings, tacit Reading, writing Universities, Example Academic books/papers/ professorships, knowledge community about certain articles, funding networks, presenting institutions, organisations, teaching papers at big scientific publishers, relations, conferences/ theories and organisational editorial informal circles, methods, boards, hierarchies being involved ideological commissions, between in email knowledge, political deans and communication, knowledge parties, professors or and so forth professors and business firms, about administrative political and PhD economic offices, and candidates, organisations bodies relations between politicians, media journalists, and academics

population that have an impact on how certain academic ideas are discursively constructed (Hamann 2014). Furthermore, it can be worth looking at the institutional context and practices of text production that reveal information on the influences on choices of topics, arguments, and positions in, for example, election manifestos (Scholz 2010). Data that can be analysed with quantitative methods are one suitable route to assess the structures in which a discourse is embedded. Such an approach enables us to analyse social phenomena that are spread over a relatively large geographical space, such as national higher education systems, and which concern relatively large groups, like disciplines or a population of professors. Furthermore, we are able to trace developments of social contexts that encompass discourses over relatively long periods of time. We can account for a large quantity of data in order to get an overview of the research object analysed before we actually start interpreting language-related phenomena in the discourse.

60 

J. Hamann et al.

Accounting for social contexts of discourses will take into consideration four ideal typical and usually interrelated dimensions and, depending on the research question, their development over time: social relations  (actors and their relations), institutional frameworks (organisations, bodies, entities and their rules), epistemic resources (ideas, concepts, tacit and biographical knowledge), and forms of social practice (field rules and their interrelations with other field rules) (Table 3.2).

3

Dispositif as a Heuristic Concept for a Context-Centred Discourse Analysis

We suggest the concept of dispositif as a theoretical framework that can guide our interpretation. It is crucial for our argument that academic discourse—that is, not only the research interests stated that we will discuss below but also, for example, theories, schools, subdisciplinary specialisations—and the practices that constitute them—for example, publications, talks, citations, mentoring, societal engagement—are not taking place in a social vacuum. Rather, academic discursive practices are embedded in social and institutional contexts. These contexts range from very specific settings in specific departments, archives, or laboratories (Knorr Cetina 1981; Grafton 2011) to very broad and durable structures on a national or even global scale (Frank and Gabler 2006; Fourcade 2009). In addition, academic discursive practices are enforced and established in power relations. Power relations are expressed, for example, on the micro level, where uncertain funding may influence scholarly practices, and thus encourage, or discourage, specific research interests (Morris and Rip 2006), on the meso-level, where departments that attract the majority of research funding have the resources to shape the image and development of a discipline (Hamann 2016a), or on the macro level, where there is a social expectation that research has an impact on society (Maesse 2015b). After discussing methodological aspects, we now turn to the theoretical question: how can we combine the study of linguistic material and of social and institutional contexts and how can we conceptualise the multifarious relations between discursive practices and their contexts marked by power and inequality?

  The Academic Dispositif: Towards a Context-Centred Discourse… 

3.1

61

Three Aspects of Academia as a Dispositif

Foucault’s notion of dispositif provides a valuable starting point for this endeavour. The term dispositif describes the ensemble that discourses form with elements as heterogeneous as institutions, practices, rules, and even architectural forms. In short, the focus of interest is powerful apparatuses that relate elements of “the said as much as the unsaid” (Foucault 1977, 194). This makes a dispositif a vital heuristic concept for the question of the relation between discourses and contexts. Foucault’s dispositif concept is well suited to account for social dynamics in the academic world. Academic  power relationships are often organised through a formal status hierarchy. But many are indirect and emerge from spontaneous interactions among academics. While language plays an important role in producing and reproducing the academic social order, any linguistic expressions that academics use are characterised by a certain interpretive openness. Their meanings depend on the contexts in which they are used. While meaning production is by nature excessive and ‘recalcitrant’, it is contained by institutional arrangements such as the academic dispositif. Meanings are homogenised, interpretations are smoothed out, ideas are domesticated through the interrelated practices, rules, and resources of the academic dispositif (Angermüller 2010, 90–96). Conceptualising academia as a dispositif thus emphasises three aspects: power effects of closure and sedimentation, heterogeneous and ­overlapping contexts, and the discursive circulation of signs between academia and other fields (Maesse and Hamann 2016).

3.1.1  Power, Closure, Sedimentation The dispositif concept reminds us that power is more than mere interpretative efforts describing power as an “open, more-or-less coordinated […] cluster of relations” (Foucault 1977, 199). It emphasises effects of closure and sedimentation that are also part of the academic world. There are many examples where meaning-making is domesticated and controlled in the academic world, think of the rhetoric of excellence and competition as well as discourses of inclusion and equality, the pressures for exter-

62 

J. Hamann et al.

nal funding and on university admissions (cf. Zippel 2017; Münch 2014; Friedman et al. 2015; Kennelly et al. 1999).

3.1.2  Fields as Heterogeneous Arenas Conceptualising academia as a dispositif emphasises a second aspect: academic contexts are complex and heterogeneous arenas that overlap with other arenas. Foucault (1977, 199) calls for “a grid of analysis which makes possible an analytic of relations of power”. However, an analytical grid for the power relations that encompass discourses and contexts was not systematically developed by Foucault. Thus, we would like to draw attention to a framework that already exists for the analysis of social power in specific arenas: Bourdieu’s field theory of symbolic production (Bourdieu 1988, 2004). Combining Foucauldian and Bourdieusian approaches allows us to account for the various effects of closure, stratification, and sedimentation that we highlighted in the previous paragraph. It also provides a hypothesis for the relations between discourses and contexts by suggesting a structural similarity between the positions within a field and the statements made from these positions (Fig. 3.3). The main reason for combining Foucauldian and Bourdieusian approaches is, however, that it allows for a sophisticated approach to institutional and social contexts of discourses. Rather than remaining diffuse and opaque, contexts are elevated to a main object of investigation. Understanding them as fields enables us not only to grasp the disparity and asynchrony of different contexts, but also to relate various social logics, structures, and practices to one another (cf. Bourdieu 1983, 1996). Our dispositif approach links both perspectives: the field perspective emphasises the dimension of social and institutional power while the discourse perspective incorporates linguistic practices of interpretation and translation.

3.1.3  Discursive Circulation and Interpretation Conceptualising academia in this way further underscores a third aspect. It highlights that discourses play an important role because they give

  The Academic Dispositif: Towards a Context-Centred Discourse… 

63

social actors the opportunity to act in an open field, as well as enabling discursive circulation throughout many fields between academia and society. Academic discourses span academic and other societal contexts because they consist of signs that can be interpreted differently in various contexts. In linguistic terms, signs are material carriers of multiple meanings. They open up a multitude of meanings that are subsequently closed again by specific interpretations. This is why academic discourses can be simultaneously embedded into interrelated academic and other societal contexts. For example, a journal article in a science discipline might represent a discovery that is translated into a patent in the judicial context, and into research cooperation with enterprises in the economic context. In the social sciences and humanities, other contextual trajectories are more likely. A publication could be interpreted as a theoretical position that matches a certain climate in the political context (Lamont 1987; Angermuller 2015), or as an insight that is picked up by media outlets that interpret the sign in terms of expertise (Maesse 2015b). Signs that are interpreted as symbolic academic capital in the academic context may traverse into judicial, economic, media, or political fields, where different  actors charge them with different meanings. The research that emerges from these positions would have to traverse from the academic into a number of other societal contexts. In order for discursive utterances in Foucault’s sense and capital in Bourdieu’s sense to make sense and mean something, the material sign as a carrier of meaning has to coincide with the meaning it generates in a specific context. The material form of utterances is a prerequisite to generating meaning, just as the form needs a context in order to refer to something meaningful. Our dispositif theory does not describe the circulation of meaning (or capital) since meaning is the product of using language. What circulates is not meaning, but text and talk that are charged with meaning in various contexts (cf. Beetz 2016, 82–88). The semiotic substantiation of our nexus of discourse—­dispositif—field allows us to treat discourses as distinct social entities that are neither above nor subordinated to their contexts, but embedded in them. Our heuristic toolkit will now allow us to understand, first, how elements of academic discourse—for example, a journal article that presents

64 

J. Hamann et al.

a particular approach—are embedded in academic and other contexts, for example, when they are picked up by researchers from other subfields, or when they influence political debates. This requires, second, that we understand how these discursive elements are produced in the context of an academic field, equipped with a specific academic meaning and re-­ interpreted when they travel into other fields.

4

 ow to Use Sociological Data H in Discourse Analysis?

So far, we have outlined the importance of the linguistic, contextual, and theoretical levels of discourse analyses. In the following two sections, we will explain how socio-structural sociological data about institutional contexts of discursive practices can be articulated with discourse analysis in order to study the discursive construction of social order, here for example in an academic discipline. We will draw on empirical examples from an information system that has been built for this purpose in the ERC DISCONEX project. The information system aims to provide sociological information about institutional and epistemic contexts of academic discourses. With this goal in mind, we have produced quantifiable data that can inform us about larger social structures in which discourses emerge. Such an analysis of the macro-structures of the social context helps to decide and justify which particular cases of discourse production should be chosen for a more thorough analysis with qualitative methods. Therefore, we do not understand quantifying methods in terms of a positivistic approach to discourse, but as a means to explore and discover. For our discourse analytical perspective, quantifying methods are a useful heuristic tool, which is followed by a more fine-grained analysis (Kleining 1994). In the first part of this section, we will give an example of a data set that informs us about social contexts of a discourse. Second, we will explain how we use such data for the purposes of discourse analysis. Thereby, we will further develop the theoretical framework for a contextcentred discourse analysis.

  The Academic Dispositif: Towards a Context-Centred Discourse… 

4.1

65

 apping Social Contexts with Statistics: Actors, M Institutions, Organisations

To map the contexts of discursive practices, our first step is to assess the relevant actors within and outside academia. The social contexts of academic discourses consist of, for example, researchers, institutions like universities, publishers, and funding agencies, as well as disciplines, research groups, and networks. In a broader sense, these are all actors that, in one way or another, participate as social entities in an academic discourse, no matter whether they are individual or collective, human or non-human. Hence, the first step in our analysis is to identify the discourse actors that are relevant for the discourse we want to study. This can be done via a systematic investigation of the institutional structures of our research object. In our case we catalogued all national higher education and research institutions that are relevant to the academic discourse in a particular research field together with the full professorships in each institution. In addition, we also tried to account for higher education policies effecting the classification of universities. In the UK case, classificatory instances that are part of the higher education dispositif include, for example, such groups as the Russell Group and European Research Universities, and also governance instruments like the Research Excellence Framework, with its highly influential focus on research excellence and societal impact (Hamann 2016a). The importance of these classificatory instances notwithstanding, our approach in the project is more focused on the individuals in higher education institutions, the way they position themselves, are positioned by others and the career trajectories they have followed. There are numerous other methods to identify the actors or participants of a particular discourse. For a­ cademic discourse, citation analysis has become the preferred approach in order to map the structures of scientific communities (e.g. Estabrooks et al. 2004; Xu and Boeing 2013), and concept-mapping has been applied to identify how different actors in a cross-disciplinary context conceptualise their research area (Falk-­ Krzesinski et al. 2011). Below, we will illustrate how a mapping of the positions of discourse participants can be produced with correspondence

66 

J. Hamann et al.

analysis by contrasting the distribution of all the words used in their texts. Taking the example of academic discourse, the mapping of contexts should not only enable us to identify relevant actors, it should also help to locate researchers institutionally and epistemically based on their professional activities. We distinguish, as suggested in earlier studies (Angermuller 2013; Whitley 1984), between an institutional world where researchers can occupy posts in the organisational hierarchy of academia and an academic world where researchers can occupy positions in academic communities and become references over knowledge horizons. In Bourdieu’s (1988) terms, this means that, in order to build up a meaningful, intelligible position in academia, a researcher needs to draw on two types of capital. S/he needs institutional academic capital in terms of worldly power that is created, acknowledged, and accumulated, for example, in committees, councils, and on boards, and s/he also needs symbolic academic capital in terms of prestige, recognition, and accolades that are awarded by peers. In this chapter we will focus on a synchronic view of the academic world. In addition, however, we are interested in diachronic developments of researchers’ locations. This view of trajectories is first and foremost concerned with researchers’ biographies—if possible going back to their social origins, tracing their careers by systematically collecting information on researchers’ educational pathways and the institutions they have gone through and the positions they have held in the course of their careers (cf. Hamann 2016b). For the quantitative analysis of such data, Sequence Analysis is one route to follow (Blanchard et al. 2014).

4.2

Challenges and Potentials of Context Data

We propose to analyse data pertaining to institutional backgrounds in which a discourse is produced, and to the biographical data of discourse participants that can actually be held responsible for the texts produced within a discursive formation. Both institutional and biographical context data can be analysed from a diachronic and synchronic perspective, which results in different challenges and potentials.

  The Academic Dispositif: Towards a Context-Centred Discourse… 

67

Biographical information on researchers’ administrative positions, teaching, media activities, and non-­academic as well as research-related activities, for instance, allows us to delineate how these various activities influence particular research and academic discourse in general. Furthermore, it might be interesting to see which activities become important at which points in a career (cf. Ćulum et al. 2015). By collecting data on publications and collaborations, we can find out more about the social networks within the community of a particular research field. Also, such data help to study institutions through the prism of their researchers: how is an institutional framework reflected in the writing of its researchers? Which institution has the most diverse research interests, with the greatest range of transdisciplinary research interests, with the most cited researchers, with the most and largest grants, and so on? One objective of exploring the institutional background can be to identify hubs for particular research interests and theoretical approaches. With methods like correspondence analysis, we are able to find out more about the transdisciplinary contacts between research fields and disciplines. Additionally, it is possible to observe trends in research fields from a diachronic perspective. However, these potentials notwithstanding, institutional structures, research infrastructures, affiliations and career progress differ substantially in different countries (Angermuller 2017; Paradeise et  al. 2009). Hence, developing a grid of categories accounting for the institutional and social reality of the academic world in various countries is anything but straightforward. For our example, it is important to note that France has a more centralised academic system, the system in the UK is less centralised, though many resources are still concentrated in Oxford-, Cambridge-, and London-­based institutions, while German higher education and research institutions have more autonomy and are usually dependent on individual states [Länder]. Moreover, in order to account for the context of the academic discourse in different countries, the institutional working conditions under which research is conducted can add an important analytical dimension. For instance, in contrast to Germany or the UK, there is an institutional distinction between research and teaching activities in France. This

68 

J. Hamann et al.

French peculiarity does not exist in many other countries. Research is organised by discipline in trans-institutional laboratories according to particular fields of research. If they are not working for the National Scientific Research Centre (CNRS), researchers have to teach at a university. There is a national body (National Council of Universities [Conseil National des Universités—CNU]) that classifies these researchers and allocates their work to one of the currently 87 sections corresponding to fields of research. After completing a PhD, a researcher pursuing an academic career would normally try to become qualified by the CNU, and then apply for a permanent civil service position. Their counterpart in Germany would have to be happy with a renewable contract. In case they do not qualify for a professorship after 12 years, they risk their academic career coming to an end (Law on Temporary Employment Contracts in Sciences [Wissenschaftszeitvertragsgesetz]). Through the prism of the dispositif approach, we consider such institutional differences as part of the power structures establishing the social context in which processes of text production in academic discourse play out. An analysis of these institutional social contexts will help to discover and interpret socially induced features on the textual level of academic discourse.

4.3

Examples of Statistical Categories

A context-centred study of academic discourses can be based on publicly available information about the positions, careers, and activities of researchers and their institutions. For smaller research projects, such data can be collected in spreadsheets. For more complex studies, however, it might be worthwhile creating a system into which data can be fed manually (e.g. from online CVs) and automatically (e.g. from literature databases). In the case of the DISCONEX project, we created an information system that comprises socio-institutional and biographical data. All data were taken at face value, as presented in their sources. Hence, categories that were used in the data sources are reproduced in the database. In this way we hope to account for the concepts and labels that the actors believe to be important in the social and discursive arena in which they aim to position themselves. Of course, in research practice, this ideal has some

  The Academic Dispositif: Towards a Context-Centred Discourse… 

69

pitfalls that must be tackled. For instance, universities, degrees, and even countries change names. In order to ensure the continuity of the reference labels in the information system, old and new names have to be linked to each other. This is by no means an innocent step because we are intervening in the game of names and their meaning that is at the heart of questions on discursive construction, which we are actually interested in. The quantitative analysis of these data aims to reveal aspects of the social structure and dynamics of research fields. Why would discourse analysts be interested in such questions? The answers help to understand the social and institutional context to which discourse participants must respond implicitly or explicitly if they want to produce a meaningful statement in a particular discourse. We assume that discourse practices relate—in one way or another—to these context conditions. Thus, a discourse analysis that integrates context data can enrich the interpretation of textual data with information about institutional and societal conditions that are usually not obvious merely by studying text data. What we propose here is a first step towards the systematic acquisition and study of such data whose results would still have to be articulated with the (quantitative and qualitative) text analytical methods that are widely used in discourse analysis. In terms of a context-centred perspective, this will help to better understand why a certain type of research, colluding with particular statements and narratives, occurs in particular institutions and locations. In this sense, we could integrate into the analysis of academic discourses, for example, the impact of research and funding policies on the research landscape, and the topography of a field of research. In order to conduct integrative studies, we propose to collect ‘hard data’ (a) on institutions, (b) on individuals and their career trajectories in those institutions, and (c) on the professional networks of these individuals. A. The following information about departments as well as the research interests of the senior research staff allows locating research hubs: 1. Institutions or groups of institutions (e.g. League of European Research Universities, Russell Group) with a high number of researchers in a field;

70 

J. Hamann et al.

2. Mappings of research interests on the institutional, disciplinary, and national levels; 3. The publication types and journals in which researchers with particular research interests publish predominantly—for example, one might find that German economic sociologists publish more books than journal articles when compared to their British colleagues; or that corpus linguists from the University of Lancaster publish their articles in a particular journal that differs from the publication preferences of their colleagues at the University of Birmingham. Furthermore, the enquiry into biographies can help to understand whether particular institutions and/or disciplines prefer individuals with a particular social or career background. By systematically recording education and career trajectories, we might, for instance, identify those institutions, disciplines, or research fields in which researchers have typically a migrant, working-class, or elite background. B. The following information on education and career steps collected from the CVs of research staff allows identifying typical and atypical career trajectories in different fields, disciplines, and countries: 1. Ages and biographical backgrounds of researchers in a field (in future projects this might be completed with information on the social backgrounds of researchers); 2. The average time spent in each career stage in a respective field and country; 3. The average time needed to become a professor in a certain field, discipline, and country; 4. The number of institutional positions held until professorship is reached in a certain field, discipline, and country; 5. The typical and less typical activities of researchers in a particular research field until achieving professorship; 6. The institutional origins of senior researchers in a field. In terms of a discourse analytical perspective that focuses on institutional contexts, this information helps to illustrate the similarities and differences

  The Academic Dispositif: Towards a Context-Centred Discourse… 

71

in career trajectories between different fields of research as well as countries, relations between career progress (seniority) and particular activities (number and quality of publications, acquisition of research funding, administrative activities/positions in the higher education institution) or the conditions for a change of research interests in a career, among other things. C. Finally, information on collaboration in research projects and publications can be used to study the professional networks of individuals in institutions. With the help of social network analysis, we could investigate such questions as: 1. Co-authorship; 2. Who publishes in the same journal, or with the same publisher? 3. What are typical or less typical text types that are published in a discipline or a particular field of research? 4. Citation analysis: social networks of citing researchers in a particular field. –– Who are the most cited researchers in a field of research? –– Do they come from a certain type of institution? –– Relation between seniority and citation frequency: are researchers in senior positions cited more than junior academics? –– How important are early career contacts in citation patterns? In terms of a context-centred discourse analytical perspective, this information facilitates a social network analysis mapping the social and institutional contexts of academic discourses. Once we have collected a data set according to the research interests outlined between A1 and C4, we can produce data queries that correspond to these interests. A results report is basically a spreadsheet containing variables and data, which will then have to be analysed and visualised with statistical software. In the following section, we will give an example that focuses on the research interests of full professors in UK sociology departments.

72 

5

J. Hamann et al.

 xample: Research Interests as an Entry E Point for a Context-Centred Analysis of Academic Discourses

Based on the analytical framework developed in the previous sections, the current section will outline what a quantifying exploration of social contexts could look like. In doing so, we will address the empirical questions posed in the introduction: How can we account for the discursive construction of social order in the academic discipline of sociology? How do social relationships of difference and inequality that constitute the sociological field go together with the linguistic organisation of the field, that is, a specific distribution of words and expressions across the social space? The aim of this section is to demonstrate a method that can help to get an idea about the macrostructure of academic discourse in UK sociology based on research interests. Ideally, this analysis of the institutional context of academic discourse and its structure should help to explain hierarchies and power relations. The division between central and periphery institutions that this section reveals on the level of research interests could help to explain the manifestation and sedimentation of symbolic and/or economic capital that these institutions accumulate in evaluation procedures, such as the Research Excellence Framework or various other rankings—an analytical step that we allude to in the last section of this text. As an example of how to use quantitative methods as a heuristic tool for a qualitative analysis of discourse data, we explore one dimension of academic discourse, the research interests that individual full professors display on their homepage profiles in sociology departments. In a more thorough analysis, other data sets could be integrated depending on our research question. For the sake of illustration, we only represent one of the numerous methods that could be applied. In a second step, which we won’t be able to present here, we could further investigate the similarities in academic discourse of those institutions grouped by our method in one cluster. With this exemplary study we aim to identify the research hubs of various research fields in a given discipline and country—in this case, UK sociology. With the help of correspondence analysis, we are able to create maps that present those institutions on whose webpages we find

  The Academic Dispositif: Towards a Context-Centred Discourse… 

73

similar research interests and keywords close to each other, whereas differences in terms of research interests are represented by greater distance on the map.

5.1

The Corpus and Method

For this study, we compiled a trial corpus with texts of research interests and keywords that full professors at 76 UK sociology departments present on their institutional webpages. There are more than 90 sociology departments in the UK, but not all of them have full professors on their staff. We consider full professors to be preeminent stakeholders in academic discourse and therefore the analysis of their data is the starting point of our study. The corpus was partitioned in such a way that we could compare research interests on the institutional, disciplinary, and national levels. With a size of 11,980 tokens, our corpus of UK sociologists is quite small. However, it is big enough to present the method and its potential for future research. The corpus has not been lemmatised and also includes all grammatical words. Our choice is based on the assumption that different grammatical forms of content words and grammatical words themselves have a particular influence on the meaning construction for what we want to account for in our analysis. We analyse our data set with correspondence analysis. This is a statistical method to simplify complex multivariate data by grouping entities under investigation according to corresponding features. In classical empirical social research, the method has been used to group  actors according to similar occupations and dispositions (Bourdieu 2010), in discourse analysis the approach has been used to group speakers according to corresponding features in their language use. In this sense, it is a powerful method to discover similar language use of different speakers in particular time periods by taking into account the complete vocabulary of a given corpus and comparing it according to the partitions introduced. In this study, we took the partition ‘institution’ and we contrasted the distribution of all word tokens used to express research interests on the website of a given department of sociology in the UK with it. To the extent that the method takes into account the entire vocabulary of a

74 

J. Hamann et al.

corpus, it is a corpus-driven (Tognini-Bonelli 2001) heuristic approach for the analysis of language use. It was first developed in France in the 1970s within the lexicometric approach to discourse (see Chap. 5). Applying correspondence analysis to text corpora is a way of visualising similarities and differences in language use by projecting words into an at least two-dimensional visualisation (Chap. 6 in this volume, and Bécue-Bertaut 2014; Benzécri 1980; Husson and Josse 2014; Lebart and Saporta 2014; Salem 1982). There is now a range of lexicometric tools that offer correspondence analysis, for example, Lexico3, Hyperbase, and Iramuteq. However, for this text we used the software TextObserver, developed at the CEDITEC laboratory, University of Paris-East. This tool provides maximum flexibility in terms of the integration and exclusion of corpus parts into the analysis, which can be helpful to understand the impact that certain corpus parts have on the visualisation. Correspondence Analysis applied to text corpora is based on a table of the distributions of word token frequencies (rows) in different corpus parts (columns). In our case, one can find in the first column of this table all the word tokens of the corpus. All other columns are named after the location of the university from which the texts of the corpus originate. The rows of these columns contain the frequencies of the word tokens they refer to in the texts of a particular university. In simple terms, the method can be described as an algorithm grouping together those words with similar distributions across different universities (columns). Figures 3.1 and 3.2 serve purely didactic purposes—they are not readable as such. However, they demonstrate how the words of a given corpus are taken into account and visualised according to their distribution profiles. By distribution profiles, we mean a similar high or low distribution of a group of the same word tokens across different institutions (columns). The word tokens of such profiles are positioned close to one another. In Fig. 3.1, the words in blue refer to keywords used by professors in UK sociology departments. The blue dots refer to the positions of word tokens calculated with an algorithm comparing the distribution of all word tokens across the columns. In Fig. 3.2, the words in red refer to the locations of the universities from which the texts were taken (column names). Their positions are calculated based on the positions of the blue dots. The names of columns with similar distribution profiles are placed

  The Academic Dispositif: Towards a Context-Centred Discourse… 

75

Fig. 3.1  Correspondence analysis of keywords and research interests of UK sociologists—rows only represented

close to one another, whereas those with very different distribution profiles are more distant from one another. The axes are situated alongside the highest concentrations of similar characteristics (here similar frequencies of the same words in different institutions). Deciphering the meaning of the axes is part of the interpretation process, which often needs further analytical steps using different methods. To demonstrate how the visual can be interpreted, we have chosen to make it more readable by removing certain word tokens from the representation. In Fig. 3.3, we have kept those content words that we understand have particular importance for occupying a position in the research field of sociology. This is simply for the sake of demonstrating the method. A more systematic analysis would have to be more explicit about the

76 

J. Hamann et al.

Fig. 3.2  Correspondence analysis of keywords and research interests of UK sociologists—rows and columns represented

words that have been removed from the visualisation (but not from the analysis). Moreover, the data set should be completed with more and longer texts. However, regardless of these limitations, we think that the potential of the approach will become obvious. When looking at a location in the system of coordinates, we must consider that both axes represent dominant information concerning a certain variable. Finding the answer to the question as to which variable that might be is part of the researcher’s interpretation. In the case of textual data, the concept of ‘variable’ would have to be understood as the semantic realm triggered by the referential meanings of certain words. However, the interpretation of a correspondence analysis based on textual data is never straightforward because it is based on lexical items, not semantic units. The problem is that the same semantic can be expressed

  The Academic Dispositif: Towards a Context-Centred Discourse… 

77

Fig. 3.3  Correspondence analysis of keywords and research interests represented on the websites of professors in sociology departments in the UK

with different words or variations of words. In this sense, the method tends to analyse the style of speech that has a semantic dimension, even though it cannot be reduced to a pure analysis of semantic worlds. Hence, the distances represented on the axes could originate from solely lexical or solely morphological differences, or a mixture of the two. In this sense, we have to be prudent with the claims we make on the semantic level. In order to verify our claims, we can use concordance analysis, or other methods that allow direct access to textual data. Regardless of the outcome of such verification, we can be sure that the method reveals dominant tendencies in the style of speech within texts from the different corpus parts introduced (here, texts from different institutions). Thus, Fig. 3.3 will help to understand which institutions are closely related to each other in terms of the research interests of their professors. At this point we have to acknowledge that the visualisation seems to

78 

J. Hamann et al.

imply that the authors of the texts—UK full professors in sociology— represent the research interests pursued at the particular universities they work for. Moreover, one has to consider that there is no direct correlation between the distance of keywords and their frequencies in the text of an institution. For example, in Fig. 3.3, poverty is relatively close to Bangor (2nd quadrant). However, the word form poverty does not occur in texts taken from the University of Bangor—but it does occur in the text taken from the University of Bath, which is displayed closest to Bangor. These two are located close to one another because of the similarities of their distribution profiles throughout their texts. Yet, this does not mean that the vocabulary in the texts of both institutions is identical, and so particular words can be missing in one or other part. Hence, some words, such as poverty, do not occur in both text collections. Thus, we should concentrate interpretation of the visual on the similarities of institutions and less on particular cases of vocabulary contributing to the visualisation: the texts used at Bangor are most similar to those used at Bath and Belfast. The keywords displayed in Fig. 3.3 can help to give us a general picture. For the interpretation, we should try to identify general common features in the words displayed in a particular realm of the visualisation. In trying to understand Fig.  3.3, we should also direct our attention towards the meaning of the axes. In order to interpret these, we can look at the most extreme cases on both axes. The most extreme cases on the x-axis are the University of Aberdeen and the University of the West of England (UWE Bristol). On the left side of the x-axis we see notions such as societies, European, Europe, digital, death. On the right side of the x-axis there are notions such as governance, power, failure, education, employment, technologies, practices, and others. Ideally, we would now be able to relate the visualisation to the social world—the academic world that they were taken from. Once we integrate the complete list of all word forms contributing to the position of a data point (not represented for reasons of readability), we can interpret the x-axis as follows: on the left side of the x-axis we find institutions at which researchers present their work in terms of an external perspective of their discipline: they refer predominantly to different topics of the discipline, internationality, and interdisciplinarity on their websites (psychology, geography, Europe, international, etc.). In contrast, on the right

  The Academic Dispositif: Towards a Context-Centred Discourse… 

79

side, researchers present their work from more of an internal perspective of their discipline. They emphasise the way they do sociology (comparative, ethnographic, empirical, theoretical), whereas the international aspect remains largely unaddressed by focusing on national British questions (Britain, nationalism). While on the left side we have more problemdriven topics, such as death, crisis, poverty, welfare, on the right side we find, with some exceptions (illness, failure), more applied topics that ask for affirmative solutions (communication, education, governance, employment) and stress technological and/or scientific aspects of society, such as corporate power, regulation and governance of health technologies, regenerative medicine, science and technology studies. Most universities are situated around the origin of the coordinates, which in comparison with the rest of the corpus represents the corpus parts with the least specific distribution of the vocabulary. We can also say that researchers at these institutions use vocabulary that is used most frequently—and thus may represent the mainstream of UK sociology departments. Such terms include culture(s), gender, identity, racism, but also references to time, such as current and historical. The fact that these institutions are situated in the centre could also mean that they use terms that can be found on both sides of the x-axis and also the y-axis to a similar extent. Meanwhile the y-axis seems to represent a continuum between research interests emphasising a rather critical attitude at the bottom and research interests that are presented in a rather concise and neutral style at the top.

5.2

Limits of and Obstacles to Interpretation

With regard to the limitations of our interpretation, we have to admit that texts from the genre that we are looking at are by their nature very short. Thus, the outliers in Fig.  3.3 may also result from the fact that there are only very few researchers, and hence particular research interests that differ greatly from the ‘mainstream’ situated around the origin. A bigger corpus might help to sustain the hypothetical results that we generated at this stage. These limitations notwithstanding, the visual illustrates a method that could be used to integrate the data on institutions into a discourse analysis.

80 

J. Hamann et al.

The advantage of using this somewhat imperfect corpus is that the corpus parts are of a size that we can manage to read. We gain a better understanding of the method by simply reading the closest and the most distant texts (Bangor, Bath, Belfast, and Aberdeen versus Glasgow and Bristol [UWE]). The disadvantage of the relatively small corpus size is that changes in the visual might be quite substantial if we add or remove texts of researchers from these institutions. At any rate, we do not claim that these visuals represent a positivistic depiction of ‘the reality’. Rather, through the prism of correspondence analysis, we get a vague idea about hidden relations that are not visible in academic texts, the aim being to find out about relations that could be relevant either on other levels from subsequent investigations with other variables, or on the discursive level itself. Suppose that we somehow had ‘complete’ data, we could relate these results to studies that map, for instance, the level of access of UK institutions to research funding in general, or to funding for research on particular topics. This would allow us to cluster institutions with a similar level of access to research funding, and subsequently analyse to what extent these clusters match the maps that we produce based on research interests. We could also include other data, for example, on the permanence and duration of positions across institutions, disciplines, and countries, in order to investigate the impact of such variables on academic discourse in the short and long terms. Given adequate data, the analysis of social contexts becomes a powerful supplement to discourse analytical approaches. This section has demonstrated an exemplary starting point for such an undertaking. The remaining question is connected to the third level of analysis (Table 3.1), the theoretical interpretation of linguistic and social context data. In the following section, we will suggest a theoretical framework that integrates the linguistic and sociological dimensions of discourse analysis.

6

 he Heuristic Potential of the Dispositif T Approach

The linguistic and sociological analyses of data that we have sketched out in the previous sections can help us find out about the institutional and discursive structures of research fields. But neither the linguistic material

  The Academic Dispositif: Towards a Context-Centred Discourse… 

81

nor the collection of sociological data can account for the social organisation of academia. In this section, we propose a dispositif theoretical approach in order to go beyond the opposition of micro and macro social structure and discursive practice. The dispositif analysis we propose would read statistical data pragmatically and reflexively. We take them as a starting point for further investigations into an empirical object that resists simplification. We point to three aspects of academia as a dispositif: (a) we emphasise a rather structuralist notion of power that yields effects of closure and sedimentation in academia, (b) we emphasise that academic contexts are complex and heterogeneous arenas that overlap with other arenas, and (c) we emphasise that discourses play an important role because they give social actors the opportunity to act in an open field, as well as to enable discursive circulation through many fields between academia and society. As highlighted by the following sections, all three aspects are addressed by the dispositif concept. Let us illustrate the heuristic potential of a dispositif theory that guides the analysis of academic texts and contexts. Coming back to our empirical example of full professors in sociology in the UK (cf. Sect. 5), the three aspects of our dispositif approach generate the following analytical perspectives: First, we have argued for a rather structuralist notion of power that emphasises the effects of closure and sedimentation (cf. Sect. 3.1.1). What (more) can we take from Fig. 3.3 if we follow this argument? The specific distribution of sociology departments in terms of the research interests of their professors might tentatively be interpreted in terms of a centre and a periphery. Departments at the centre of the field, including many London-, Oxford-, and Cambridge-based institutions, and Warwick, could represent a thematically coherent core of ‘top’ departments. Anchoring this assumption with additional data would enable us to test whether these departments are also ‘competitive’ in terms of funding. Professors at departments on the periphery appear to be pursuing alternative research strategies that do not represent the ‘core interests’ of the field. Second, the dispositif theoretical framework introduces fields as a main object of investigation, thus allowing for a systematic account of different contexts that overlap with each other (cf. Sect. 3.1.2). Following

82 

J. Hamann et al.

this argument, the distribution in Fig. 3.3 can be understood as a field, as a power-soaked arena in which professors and departments struggle for scarce resources. A sophisticated notion of contexts as fields allows us to investigate phenomena such as the specific resources that the actors in Fig.  3.3 are actually competing for. It also allows us to consider—and distinguish—other contexts that overlap with the field of UK sociology and influence it with their specific logics. Interdisciplinary research interests could then be interpreted as influences from other academic fields, while applied research topics that demonstrate an interest in, say, health technologies, could be interpreted as intrusions from respective industry sectors that overlap with the academic sphere. The third aspect of our dispositif framework (cf. Sect. 3.1.3) highlights discursive circulation through academic and social fields. We argue that this circulation is possible because academic discourses consist of signs that can be interpreted differently in various contexts. Coming back to our empirical example of full professors in sociology in the UK, this insight allows us to follow the products of specific research interests throughout different contexts, and to study how they are interpreted differently in their respective contexts. For example, a research interest in ‘poverty’, as displayed in Fig. 3.3, might not only result in a publication in an academic journal, but also in policy advice. While the journal article would be interpreted in an academic context, and thus in terms of its contribution to the state of research, the policy advice that results from the research would be interpreted in a political context, and thus in terms of its practical feasibility and party political implications. Eventually, one would be able to track the discursive circulation even further and study not only how poverty research turns into policy advice, but how policy advice in turn circulates in the judicial field where it is interpreted in terms of new welfare policy legislation.

7

Conclusion

We have highlighted some shortcomings of text-centred conceptualisations of context and pointed out the necessity for a more systematic integration of social contexts and a theory-based interpretation of discourses.

  The Academic Dispositif: Towards a Context-Centred Discourse… 

83

Taking the example of academic discourse, we have identified sociological aspects that, from our point of view, are of importance for a more comprehensive understanding of how discourses shape the social and are shaped by them. Correspondence Analysis was presented as a possible starting point that can help to articulate sociological data with discourse analysis. In the example presented, we produced a map of the field of academic discourse in terms of institutions and research interests. In a follow-up study, we could translate keywords and research interests into one language in order to create a cross-country mapping of research interests related to institutions. This would help to better understand the interconnectedness of research fields across national borders. We hope to be able to interpret such visuals more thoroughly once we have a deeper knowledge of the material and the institutional context we are studying. In order to better understand how dispositifs influence academic discourse, we would also have to take into account other variables producing other maps of the academic world. We would have to look at funding infrastructures, social networks, the social backgrounds of researchers and their career trajectories. Such explorations would help, for instance, to identify centre and peripheral institutions in terms of research interests. They would help us gain better insights into the social dynamics of academic discourse. In the long term, we could also study how these mappings change over time: Are there certain institutions that display consistently a stable conglomeration of research interests? How and depending on which conditions do research interests change over time? Embedded in a social theory framework, the operationalisation of the dispositif concept will help to capture non-linguistic aspects of discourse that can complete the linguistic data and enrich the analysis of discourse.

References Angermüller, Johannes. 2004. Institutionelle Kontexte geisteswissenschaftlicher Theorieproduktion: Frankreich und USA im Vergleich. In Wissenschaftskulturen, Experimentalkulturen, Gelehrtenkulturen, ed. Markus Arnold and Gert Dressel, 69–85. Wien: Turia & Kant.

84 

J. Hamann et al.

———. 2006. L’analyse des textes à la croisée de diverses traditions méthodologiques: les approches qualitatives et quasi qualitatives. In Méthodologie qualitative. Postures de recherche et variables de terrain, ed. Pierre Paillé, 225–236. Paris: Armand Colin. ———. 2010. Widerspenstiger Sinn. Skizze eines diskursanalytischen Forschungsprogramms nach dem Strukturalismus. In Diskursanalyse meets Gouvernementalitätsforschung. Perspektiven auf das Verhältnis von Subjekt, Sprache, Macht und Wissen, ed. Johannes Angermüller and Silke van Dyk, 71–100. Frankfurt/Main: Campus. Angermuller, Johannes. 2013. How to become an academic philosopher. Academic discourse as a multileveled positioning practice. Sociología Histórica 2: 263–289. ———. 2015. The moment of theory. The rise and decline of structuralism in france and beyond. London: Continuum. ———. 2017. Academic careers and the valuation of academics. A discursive perspective on status categories and academic salaries in France as compared to the U.S., Germany and Great Britain. Higher Education 73: 963–980. Angermuller, Johannes, and Gilles Philippe, eds. 2015. Analyse du discours et dispositifs d’énonciation. Autour des travaux de Dominique Maingueneau. Limoges: Lambert Lucas. Bécue-Bertaut, Mónica. 2014. Distributional equivalence and linguistics. In Visualization and verbalisation of data, ed. Jörg Blasius and Michael Greenacre, 149–163. London and New York: CRC. Beetz, Johannes. 2016. Materiality and subject in Marxism, (post-)structuralism, and material semiotics. London: Palgrave Macmillan. Benzécri, Jean-Paul. 1980. Pratique de l’analyse des données. Paris: Dunod. Bhatia, Vijay, John Flowerdew, and Rodney H.  Jones. 2008. Advances in discourse studies. New York: Routledge. Blanchard, Philippe, Felix Bühlmann, and Jacques-Antoine Gauthier. 2014. Advances in sequence analysis. Theory, method, applications (Life Course Research and Social Policies, 2). Heidelberg: Springer. Blommaert, Jan. 2005. Discourse: A critical introduction. Cambridge: Cambridge University Press. Bourdieu, Pierre. 1983. The field of cultural production, or: The economic world reversed. Poetics 1983 (12): 311–356. ———. 1988. Homo Academicus. Cambridge: Polity Press. ———. 1996. The state nobility. Elite schools in the field of power. Cambridge: Polity Press.

  The Academic Dispositif: Towards a Context-Centred Discourse… 

85

———. 2004. Science of science and reflexivity. Cambridge: Polity Press. ———. 2010. Distinction: A social critique of the judgement of taste. London: Routledge & Kegan Paul. Original edition, 1984. Bührmann, Andrea D., and Werner Schneider. 2007. Mehr als nur diskursive Praxis?  – Konzeptionelle Grundlagen und methodische Aspekte der Dispositivanalyse. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 8 (2): Art. 28. ———. 2008. Vom Diskurs zum Dispositiv. Eine Einführung in die Dispositivanalyse, Sozialtheorie: intro. Bielefeld: Transcript. Ćulum, Bojana, Marko Turk, and Jasmina Ledić. 2015. Academics and community engagement: Comparative perspective from three European countries. In Academic work and careers in Europe: Trends, challenges, perspectives, ed. Tatiana Fumasoli, Gaële Goastellec, and Barbara M.  Kehm, 133–150. Dordrecht: Springer. Estabrooks, C.A., C. Winther, and L. Derksen. 2004. Mapping the field: A bibliometric analysis of the research utilization literature in nursing. Nursing Research 53 (5): 293–303. Falk-Krzesinski, Holly J., Noshir Contractor, Stephen M. Fiore, Kara L. Hall, Cathleen Kane, Joann Keyton, Julie Thompson Klein, Bonnie Spring, Daniel Stokols, and William Trochim. 2011. Mapping a research agenda for the science of team science. Research Evaluation 20 (2): 145–158. Foucault, Michel. 1972. The archeology of knowledge and the discourse of language. London: Tavistock. Original edition, 1969. ———. 1977. The confession of the flesh. In Power/knowledge: Selected interviews & other writings, 1972–1977, by Michael Foucault, ed. Colin Gordon, 194–228. New York: Pantheon. Fourcade, Marion. 2009. Economists and societies: Discipline and profession in the United States, Britain and France, 1890s–1990s. Princeton, NJ and Oxford: Princeton University Press. Frank, David J., and Jay Gabler. 2006. Reconstructing the university. Worldwide shifts in academia in the 20th century. Stanford, CA: Stanford University Press. Friedman, Sam, Daniel Laurison, and Andrew Miles. 2015. Breaking the ‘class’ ceiling? Social mobility into Britain’s elite occupations. The Sociological Review 63 (2): 259–289. Grafton, Anthony T. 2011. In Clio’s American atelier. In Social knowledge in the making, ed. Charles Camic, Neil Gross, and Michèle Lamont, 89–117. Chicago and London: University of Chicago Press.

86 

J. Hamann et al.

Hamann, Julian. 2014. Die Bildung der Geisteswissenschaften. Zur Genese einer sozialen Konstruktion zwischen Diskurs und Feld. Konstanz: UVK. ———. 2016a. The visible hand of research performance assessment. Higher Education 72 (6): 761–779. ———. 2016b. ‘Let us salute one of our kind’. How academic obituaries consecrate research biographies. Poetics 56: 1–14. Husson, François, and Julie Josse. 2014. Multiple correspondence analysis. In Visualization and verbalisation of data, ed. Jörg Blasius and Michael Greenacre, 165–183. London and New York: CRC. Kennelly, Ivy, Joya Misra, and Marina Karides. 1999. The historical context of gender, race, & class in the academic labor market. Race, Gender & Class 6 (3): 125–155. Kleining, Gerhard. 1994. Qualitativ-heuristische Sozialforschung. Schriften zur Theorie und Praxis. Hamburg-Harvestehude: Fechner. Knorr Cetina, Karin. 1981. The manufacture of knowledge. An essay on the constructivist and contextual nature of science. Oxford: Pergamon. Lamont, Michèle. 1987. How to become a dominant French philosopher: The case of Jacques Derrida. The American Journal of Sociology 93 (3): 584–622. Lebart, Ludovic, and Gilbert Saporta. 2014. Historical elements of correspondence analysis and multiple correspondence analysis. In Visualization and verbalisation of data, ed. Jörg Blasius and Michael Greenacre, 31–44. London and New York: CRC. Maesse, Jens. 2010. Die vielen Stimmen des Bologna-Prozesses. Bielefeld: Transcript. ———. 2015a. Eliteökonomen. Wissenschaft im Wandel der Gesellschaft. Wiesbaden: VS. ———. 2015b. Economic experts. A discursive political economy of economics. Journal of Multicultural Discourses 10 (3): 279–305. Maesse, Jens, and Julian Hamann. 2016. Die Universität als Dispositiv. Die gesellschaftstheoretische Einbettung von Bildung und Wissenschaft aus diskurstheoretischer Perspektive. Zeitschrift für Diskursforschung 4 (1): 29–50. Maingueneau, Dominique. 1991. L’Analyse du discours. Introduction aux lectures de l’archive. Paris: Hachette. Morris, Norma, and Arie Rip. 2006. Scientists’ coping strategies in an evolving research system: The case of life scientists in the UK. Science and Public Policy 33 (4): 253–263. Münch, Richard. 2014. Academic capitalism. Universities in the global struggle for excellence. New York: Routledge.

  The Academic Dispositif: Towards a Context-Centred Discourse… 

87

Paradeise, Catherine, Emanuela Reale, Ivar Bleiklie, and Ewan Ferlie, eds. 2009. University governance. Western European Perspectives. Dordrecht: Springer. Raffnsøe, Sverre, Marius Gudmand-Høyer, and Morten S.  Thaning. 2016. Foucault’s dispositive: The perspicacity of dispositive analytics in organizational research. Organization 23 (2): 272–298. Salem, André. 1982. Analyse factorielle et lexicométrie. Mots – Les langages du politiques 4 (1): 147–168. Sarangi, Srikant, and Malcolm Coulthard, eds. 2000. Discourse and social life. Harlow: Longman. Scholz, Ronny. 2010. Die diskursive Legitimation der Europäischen Union. Eine lexikometrische Analyse zur Verwendung des sprachlichen Zeichens Europa/Europe in deutschen, französischen und britischen Wahlprogrammen zu den Europawahlen zwischen 1979 und 2004. Dr. phil., Magdeburg und Paris-Est, Institut für Soziologie; École Doctorale “Cultures et Sociétés”. Spieß, Constanze, Łukasz Kumięga, and Philipp Dreesen, eds. 2012. Mediendiskursanalyse. Diskurse – Dispositive – Medien – Macht. Wiesbaden: VS. Stiglitz, Joseph E. 1984. Price rigidities and market structure. The American Economic Review 74 (2): 350–355. Tognini-Bonelli, Elena. 2001. Corpus linguistics at work. Amsterdam: Benjamins. Whitley, Richard. 1984. The intellectual and social organization of the sciences. Oxford: Clarendon Press. Xu, Yaoyang, and Wiebke J. Boeing. 2013. Mapping biofuel field: A bibliometric evaluation of research output (review). Renewable and Sustainable Energy Reviews 28 (Dec.): 82–91. Zippel, Kathrin. 2017. Women in Global Science: Advancing Careers Through International Collaboration. Stanford: Stanford University Press.

4 On the Social Uses of Scientometrics: The Quantification of Academic Evaluation and the Rise of Numerocracy in Higher Education Johannes Angermuller and Thed van Leeuwen

1

Introduction

Corpus approaches have a long tradition. They have recourse to computer-­ aided tools which reveal patterns, structures, and changes of language use that would go unnoticed if one had to go through large text collections ‘manually’. If such research is known for rigorous, replicable, and ‘rational’ ways of producing scientific claims, one cannot understand its success without accounting for the role of non-academic actors. Scientometrics, also known as ‘bibliometrics’, is a type of corpus research which measures the scientific output of academic researchers and represents citation patterns in scientific communities. Scientometrics is a J. Angermuller (*) Centre for Applied Linguistics, University of Warwick, Coventry, UK e-mail: [email protected] T. van Leeuwen Centre for Science and Technology Studies (CWTS), Leiden University, Leiden, The Netherlands e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_4

89

90 

J. Angermuller and T. van Leeuwen

social sciences field studying human communication, in particular processes through which knowledge is codified through texts and published in various forms (books, chapters in books, journal articles, conference papers, etc.). We choose to use the terms ‘scientometrics’ and ‘bibliometrics’ in a more or less synonymous way although we know in the field some scholars wish to distinguish between the terms (Hood and Wilson 2001). Whereas sciento or biblio relates to the elements that are studied in the field, metrics relates to the fact that quantification plays an important role in the field. Scientometrics typically aims to represent developments in science and technology through indicators. In this contribution, we will reflect on scientometrics as an example of the social effects that the quantifying instruments of corpus research have today. Scientometrics responds to specific social demands and occupies a niche which ‘mainstream’ corpus analysis is not always aware of. While corpus analysis is known for its rigorous methods, scientometrics reminds us that the tools of corpus research have become important in some applied fields with a specific empirical expertise. We consider it as part and parcel of ‘numerocratic’ practices that constitute the world of academic research. A field of research since the 1970s, scientometrics has been propelled by non-scientific demands, such as evaluation and policy-making in higher education. Governments and other institutional bodies often rely on scientometric knowledge in institutional decision-making. Therefore, scientometrics is a textbook example of a field of research which responds to power relationships within the academic field as well as outside. An example of such power relationships at work could be found in the university rankings that have gained currency over the last decade. Such rankings often mobilize scientometric knowledge in order to compare individuals and groups, organizations, and systems within and across countries. Within university systems, rankings have an impact on higher education policies which try to ‘learn the lesson’ in order to achieve better ratings. As rankings are presented as objective facts, especially in media discourse, they increasingly organize perceptions of universities in the global academic market place. While academic governance and institutional decision-making increasingly relies on scientometrics, we do not support the idea that

  On the Social Uses of Scientometrics: The Quantification… 

91

s­ cientometrics promotes the agenda of neoliberalism or managerialism as such. Rather, our question is how to account for scientometrics as a social practice and to reflect on its non-scientific conditions, purposes, and effects. Scientometrics impacts on society and society impacts on scientometrics. If scientometrics wants to strengthen its scientific ambitions, it cannot but gain from reflecting on the ways its claims are made and established socially. The objective of this contribution is both theoretical and historical. Inspired by Michel Foucault’s governmentality theory, we will discuss scientometrics against the background of a ‘numerocratic’ dispositif emerging since the eighteenth century. And by unpacking the social practices constituting the scientometric field, we will come up with a critical account of its history. This paper consists of four parts. First, we will start with a theoretical discussion of the ‘numerocratic’ dispositif of power-knowledge in which scientometrics needs to be placed. Second, we will give a historical account of its emergence and ask how it relates to the rise of numerocratic practices of governing academics. In the third part, we will have a closer look into the indicators and rankings scientometrics produces and discuss their social effects. In the conclusion, we will develop critical perspectives on the uses and abuses of scientometrics in the numerocratic dispositif.

2

 cience, the Social Order, and the Rise S of the Numerocratic Governmentality

Social scientists study social life. Their objective is to understand what drives and constrains people in doing, thinking, and saying certain things. And some social scientists study the social life of social scientists themselves. Thus, in the area of Science and Technology Studies, a number of research strands have emerged which deal with the practices and structures, the ideas and institutions that characterize scientific communities and higher education institutions. Post-war sociologists of science looked into the historical circumstances and social conditions in which scientific knowledge emerged (Merton 1962). While they understood science as an

92 

J. Angermuller and T. van Leeuwen

institution of liberal and democratic societies (even though the Soviet Union, too, produced scientific achievements), a certain emphasis was put on the cultural norms and shared values that the science system needs to produce knowledge that is considered as legitimate knowledge (such as a professional academic ethos, the social recognition of scientists and their work, a certain autonomy of higher education institutions, etc.). If their question was to account for the rise of science as a social system, the 1970s marked a turning point when more critical and constructivist epistemologies entered the scene and challenged the nature of scientific knowledge more radically. Against this background, members of the Edinburgh School made the case for ‘symmetric’ approaches (Bloor 1976). Critically interrogating the implied hierarchy (‘asymmetry’) of older, institutionalist approaches, they called for explaining ‘true’ knowledge socially just as one would explain ‘false’ knowledge (rather than asking how ‘true’ scientific knowledge was discovered). The plea for symmetry has inspired a great deal of qualitative researchers and ethnographers (e.g., in Laboratory Studies). These studies focus on the situated practices and contingent processes in which knowledge claims are produced and turn into ‘objective facts’ (Knorr Cetina 1981; Latour and Woolgar 1979). It is against this background that we want to bring in Michel Foucault’s perspective on science as power-knowledge (Foucault 1980). Defending a consistently ‘symmetric’ understanding of scientific knowledge, Foucault puts special emphasis on the intrinsic intertwinements of knowledge and power. Power is needed to make knowledge ‘true’. And knowledge is needed to exercise power in society. The question then is how people are placed in a hierarchical social order through scientific knowledge and how large populations are put under control. Foucault deals with science from several angles, which can approximately be related to the problem of archaeology, genealogy, and governmentality. First, some of his works deal with historically changing configurations of proto-scientific and scientific ideas, most systematically in The Order of Things (2002), where he deals with the emergence of economic, biological, and linguistic ideas in the early modern and modern period (sixteenth to nineteenth centuries). If the label of ‘archaeology’ has sometimes been attached to this early work, a focus on ideas can also be

  On the Social Uses of Scientometrics: The Quantification… 

93

observed in later publications such as the Birth of Biopolitics (2008), which traces the development of ordo- and neo-liberalism in economic theory. Second, in some of his other books, the most well-known being Discipline and Punish (1995), Foucault looks into practices of exercising power and disciplining individuals in various historical moments. By investigating institutional practices of the early modern nation-state, he asks what is the knowledge that legitimates them. These works have sometimes been situated in his ‘genealogical phase’ even though one clearly sees traces of these interests in earlier works such as Birth of the Clinic (1973). And third, in his posthumously published lectures on governmentality, published under the title Security, Territory, Population (2007), Foucault’s interests shift towards the terrain of political sociology which traces the changes in governmental practices and political ideas during the eighteenth century. Crucially inspiring Governmentality Studies, this work investigates how direct modes of exercising power (i.e., order and command) were replaced by more abstract power technologies operating from a distance through a system of rules, regulations, and incentives. From a governmentality point of view, large populations are placed into hierarchical social orders through governmental and non-­ governmental practices. Rather than expressing a sovereign will, these practices are dependent on dominant ideas of ‘good governance’ which are circulated through discourse. Informed by the legal, administrative, and social scientific knowledge circulating at the time and embedded in institutional power structures, governmental and non-governmental practices constitute the dispositif of governmentality which aims at coordinating the members of large populations and giving them their place in the social structure. For Foucault, liberalism and neoliberalism are types of ‘governmentality’ operating by means of ‘freedom’ where subjects are governed from a distance rather than by order and command. Against the background of Foucault’s governmentality theory, one can point out the social forces and constraints working on a societal level: knowledge is not only constructed under conditions of social inequality but also used to legitimate and sometimes even to construct and reinforce inequalities. The social and the relationships that make up the social, therefore, are not only represented but are constituted through discursive practices.

94 

J. Angermuller and T. van Leeuwen

Foucault mainly deals with the seventeenth and eighteenth centuries, when sovereignism in Western Europe showed signs of crisis. This was a time when the seeds for a ‘liberal’ governmentality were sown, which conceived power as a technology of coordinating populations from a distance. In the transition from sovereignism to liberalism, governmental practices as the legitimate exercise of power changed their form. They were no longer directed from one subject to another. They discovered the ‘population’ as a new object of governmental practice. As a consequence, there was an increasing demand for new governmental knowledge about how to run the state and the economy. Specialists, experts, bureaucrats, administrators, and policy-makers, who were no social scientists, yet were needed to devise the rules, standards, and procedures that would allocate goods through ‘free’ markets and coordinate decision-making processes through the nation-state. They were the officials, regulators, and technocrats who made sure that economic and political activities could develop within certain limits. For Foucault, therefore, both the modern nation-­ state and the free market economy testify to a new regime of governing the social. Informed by a certain type of social knowledge, these practices aim to represent the social space and, through representation, also contribute to constituting it. Embedded in a dispositif comprising both governmental and non-governmental practices, they are not only based in the government or a ministry but they also mobilize a host of private and corporate agents which have the social as their object (e.g., mass media, insurances, demographers). While Foucault stops short of an analysis of the more contemporary moment, some of his observations of proto-liberal governance in the eighteenth century can be extrapolated to the way the liberal governmentality worked in the nineteenth century and to the way neoliberalism works today. It is common to the regimes of governmentality since the eighteenth century that they applied a ‘numerocratic’ logic to the social. These regimes can be qualified as ‘numerocratic’ since they aim at governing the social through numbers and other standardizing codes. The effects of ‘governing by numbers’ are typically seen in social arenas where goods and resources are distributed by means of ‘free’ price-setting mechanisms. Thus, markets have come to be seen as arenas in which values are constructed through signalling and networking practices. Yet numbers are

  On the Social Uses of Scientometrics: The Quantification… 

95

also used in the administrative domain of the nation-state where decision-­ making practices increasingly follow the ‘democratic method’: everybody counts and is counted in votes and elections. Numerocracy has given birth to an arrangement of practices which aim at ‘governing’ large social populations by applying numbers and other standardizing codes to the heterogeneity of the social. Numerocratic governmentalities have subjected various social arenas to practices of calculation, quantification, and standardization. In these practices, the social is not only screened, measured, and overseen but it also becomes a reality that can no longer be ignored. One can think of the effects that the introduction of unemployment statistics has had on modern nation-­ states: large social groups are now defined as having work or not and the ‘well-being’ of the country is expressed through statistics. Many economic policies are based on such quantifying knowledge and welfare programmes have been created to reallocate resources among groups. Applying numbers, indicators, grids, and scales to the social is not a neutral operation but a practice that constitute order. Through these numeric and standardizing codes, the various elements that constitute the heterogeneous terrain of the social can be made commensurable within orders of value (a bus ticket is the equivalent of three oranges; writing a novel is the equivalent of two Master’s degrees). What is more, hierarchies are constructed between the elements of the social (writing a novel is worth more than a Master’s degree which is worth more than a bus ticket which is worth more than an orange). As a result of numbers representing social relationships, the heterogeneous relationships of the social are turned into the hierarchical structures of society. And this is why the usage of numeric, standardizing and classifying codes to compare things and people does not just render a given reality. They do not represent the social without intervening in it. And when they are applied to the multifarious ties and numerous relationships of the social, they can bring forth the structures that characterize society. Foucault emphasizes the discontinuity of the historical process. History is seen as a flow of many lines, with old ones terminating and new ones beginning. Yet number-based practices of governing the social have characterized all governmentalities since the eighteenth century ­ (Angermuller 2013a; Angermuller and Maeße 2015). The social has since

96 

J. Angermuller and T. van Leeuwen

been subject to numerocratic practices and increasingly so. Numerocracy is a dispositif of power-knowledge which comprises numerocratic practices that are located inside the state apparatus and outside it. Numerocratic practices typically mobilize a background of power relationships, that is, institutions, bureaucracies, rules, and laws (monthly unemployment statistics need a bureau, department, or ministry which provides resources and gives credence to them). Yet, as they represent power, these practices can constitute power relationships (e.g., between those who are categorized as ‘real’ job-seekers and may legitimately claim social benefits and those who do not count as such). But while numerocratic practices produce and reproduce structures of social inequality, they also produce and reproduce the knowledge that is needed to structure the social (cf. Bourdieu 1992). Unemployment statistics build on administrative and proto-scientific expertise (e.g., on who counts as ‘working poor’, what is a regular work contract, what are relevant parameters measured by statisticians…). Yet, they also create, update, or change the existing knowledge about the state of the economy in a country (some people may no longer be counted, new professions may be recognized, etc.). Such knowledge is often both controversial (which is why the numbers from national and international agencies often differ and numbers are often challenged from various stakeholders) and complex (it may combine sophisticated academic theories of socio-economic development with simple administrative reporting and accounting techniques). Numerocratic practices can operate with numbers (e.g., one can measure the level of income of the unemployed), but also with more general codes (e.g., taxonomies, grids, and scales). Different categories of unemployed people may be projected onto social populations, which may organize the perception of the ‘neediness’ of individuals or groups (as a result, a ‘single mom’ may claim more benefits than an ‘asylum-seeker’). Numerocratic power is at work when members of a population are classified according to hierarchical scales, which can be numeric (e.g., salary scales), ordinal (e.g., professional status) or binary (e.g., nationals and immigrants), mixes of these or all at the same time. Numerocratic practices are discursive in that they have recourse to numeric signifying codes but not all discursive practices are numerocratic. While discursive practices mobilize all kind of codes, including natural languages, numerocratic practices are generally not

  On the Social Uses of Scientometrics: The Quantification… 

97

restricted to ‘natural language’ and also includes standardizing codes such as numbers, benchmarks, indicators, lists, grids, scales, which represent the social and give it a structure. Numerocracy thus designates a dispositif of numerocratic, discursive, and non-discursive practices that classify members of social populations and assign a value to them. Proto-liberal, liberal, and neoliberal governmentalities crucially rely on numerocratic arrangements to coordinate large populations. In proto-­ liberal governmentality (eighteenth century), numerocratic mechanisms (such as the ‘market’) are used to coordinate economic activities. In liberal governmentalities (nineteenth century and early twentieth century), the transition to the market in the economic sphere is complete and numerocratic principles are increasingly extended to the realm of political decision-making and administrative procedures (see the rise of the nation-state and the advent of parliamentary democracy). In neoliberal governmentality (since the late 1970s), the economic and political spheres having been subsumed, numerocracy now discovers the sphere of culture, media, and education. Thus, while we follow Michel Foucault’s explorations of post-sovereignist regime of governance which started to take shape in the eighteenth century, we point out the increasingly important role of ‘big numbers’ in many domains which are not economic or political (Desrosières 1998; Ogien 2010; Porter 1994). If both capitalist markets and democratic nation-states crucially rely on ‘big numbers’ for the distribution of economic goods and for political decision-making, such numbers become ever more important in the sphere of education and higher education (Angermuller 2017) by the last third of the twentieth century. The rise of scientometrics goes hand in hand with the expansion of numerocracy as a social technique. From this point of view, therefore, corpus research not only contributes to the production of knowledge but is also deeply involved in practices of exercising power across large populations. It is part and parcel of what can be called a numerocratic regime of power-knowledge which has subjected more and more areas of the social to the logics of governing by numbers (Miller 2001). In the neoliberal governmentality since the late 1970s, the relation between knowledge and power has been redefined. If older ­governmentalities (until nineteenth-century liberalism) had mobilized knowledge as an ‘external’ resource as it were (i.e., produced by experts

98 

J. Angermuller and T. van Leeuwen

who were exempt from the constraints of the system they helped put in place), the production of knowledge is now fully integrated into the governmentality. The specialized knowledge and the administrative expertise of the agents of governmentality now become the object of numerocratic innovation. The large and growing social arena of educationalists and researchers is now subsumed to numerocracy. Neoliberalism, in other words, heralds the governmentalization of education and higher education. It will now be possible to discuss scientometrics as a field which emerges in a situation of changing historical circumstances. Scientometrics testifies to the growing importance of numerocratic practices for the constitution of social order in Western societies since the eighteenth century in general and the advent of these practices in the higher education sector since the post-war era more particularly. Yet while one can observe a growing demand for scientometric knowledge, scientometrics has always had to grapple with a tension between applied and more academic research orientations. Is scientometrics subordinate to normative political goals or does it engage in fundamental social research in order to reveal how research and researchers work? Also, specialized researchers in scientometrics cannot but acknowledge the explosive growth of scientometric data produced by corporate actors such as Thomson Reuters, Google, or Elsevier. Therefore, it remains an open question how the increasing amount of scientometric data which is now circulating in the academic world impacts on knowledge production and decision-making in academia. To what degree is scientometric knowledge linked with practices of governing academics? To obtain responses to these questions, we will have a closer look at the directions that scientometrics has taken as a field.

3

 he Emergence of Scientometrics T as a Field

Scientometrics (or bibliometrics) as a social science is a relatively young field. Its origins reach back the second part of the twentieth century (Hood and Wilson 2001). It comprises quantifying methods from social research and uses numbers to account for structures and changes of many

  On the Social Uses of Scientometrics: The Quantification… 

99

texts and changes over time. While contemporary scientometrics has recourse to computers and often uses digital texts, corpus research goes back to medieval times. It started with exegetical and philological practices of studying sacred and literary texts, for example, through word lists and concordances. The development of scientometrics after the world war had much to do with managing academic libraries on tight budgets and therefore keeping statistics on what articles and books were requested and how long they were kept. Those figures justified paying for other subscriptions or buying new books. At the time, the term ‘bibliometrics’ was used for that practice.1 If scientometrics can be counted as an established methodology in the social sciences today, it has never been a neutral scientific method which generates ‘pure’ knowledge. The practice of scientometrics has been highly responsive to technological achievements (such as data-processing tools) and dependent on the legal framework of knowledge production (notably copyright laws). And by no means is it a purely academic practice. In today’s knowledge-based economy, the text-­ processing practices of scientometrics have become crucial if one thinks of corporations such as Google and Facebook or governments monitoring digital communication. Initially, scientometrics was embedded in a much broader scientific discipline, the field of Science and Technology Studies (STS), where both qualitative and quantitative methods were used against a broad variety of academic backgrounds, ranging from sociology of science, philosophy of science, and history of science. As a relatively new branch of this field, the quantitative study of science started in the 1960s and involved a number of disciplines, such as library and information science, with some of the early pioneers coming from the natural sciences and engineering, such as computing, mathematics, statistics, and even physics. As a relatively new academic discipline, there was little space for housing scientometrics in existing schools and faculties. Scientometrics has found a more hospitable environment in applied research institutes at the intersection of academic and policy-oriented research. Many scientometricians have a first-hand experience of the marketization and commodification of academic research, which they see in more or less critical terms. And to the  We thank an anonymous reviewer for that idea.

1

100 

J. Angermuller and T. van Leeuwen

degree that the field is part and parcel of numerocratic practices, one can understand that there is a tendency in scientometrics that posits the investment into numerocratic governance as a norm to other researchers (Burrows 2012; Radder 2010). In the late 1980s, scientometrics became autonomous from the larger STS field and developed their own professional organizations (such as ISSI, and somewhat later, ENID) with separate scientific conferences (the ISSI cycle, next to the S&T cycle) and dedicated journals (such as Scientometrics, JASIST, Research Policy to name a few). The academic field of scientometrics has broken off from the more qualitative and theoretical strands in STS.  It has always put strong emphasis on empirical, strongly data-driven empirical research and focused on electronic data of various types. Just like statistics in the nineteenth century, scientometrics testifies to the numerocratization of the social. To some degree, statistics is subservient to civil servants, technocrats and administrators who carry out censuses, create standards in certain areas and devise regulative frameworks of action. Scientometrics, too, is an academic field which is tied to the rise of ‘numerocratic’ techniques of exercising power, which aim to govern large populations through numbers, standards, benchmarks, indices, and scales. All of this did not happen in a sociopolitical vacuum. After the economic crises in the 1970s and 1980s (Mandel 1978), ending a long phase of prosperity, the political climate in Europe and the USA changed and neoliberal approaches prevailed. With Reagan and Thatcher in charge in the USA and the UK, economic policies implemented austerity programmes, which meant budget cuts in various sectors of society, including higher education. The political ideology of neoliberalism was practically supported by New Public Management (NPM). NPM proposes to organize the public sector according to management techniques from the corporate sector (Dunleavy and Hood 1994). However, science has been mainly evaluated through peer review and for a long time quantitative information on higher education was limited. In the USA until the 1980s, for example, the reports by the National Science Foundation were the only source of quantitative data on the higher education system as a whole. Yet, in a period of austerity, more justification was needed for spending taxpayer’s money and policy-makers were less prone to distributing

  On the Social Uses of Scientometrics: The Quantification… 

101

resources on the basis of the internal criteria of the sector. Peer review– based evaluation gives little control to policy-makers and governing agencies over how money is spent. Therefore, one can observe a growing demand for simpler mechanisms of evaluation and decision-making. Examples of these type of simplistic indicators used for evaluation and decision-making will be introduced in Sect. 4. Scientometric measures first appeared in science policy documents in the 1970s, when the US National Science Foundation integrated research metrics in its annual national science monitor, which gives an account of the US science system. Scientometric results were then used to describe research activity on a macro level in the USA and other countries. It took another 20 years before large-scale scientometric reports of national science systems were produced in Europe. It was in the Netherlands in the early 1990s that the first national science and technology monitor appeared (NOWT, reports covering the period 1994–2014). The last series of these reports were produced in 2014 (WTI 2012, 2014). In this series of reports from the Netherlands, the national Dutch science system was compared internationally with other EU countries. Indicators have been devised with the aim to represent technological performance, based on the number of patents or revenue streams. These reports also contained information on the sector as a whole (e.g., the relationship between the public and private sector, other public institutions, hospitals, etc.) and on the institutional level (e.g., comparisons between universities). These analyses typically broke down the research landscape into disciplinary fields and domains with various levels (countries, sectors, institutions). In France, the government also initiated a series of national science monitor reports, produced by the Observatoire des Sciences et Technologies (OST, reports covering the period 1992–current). In France, the reports contained an internationally comparative part, as well as a national regional part. These reports are produced by a national institution financed by the government while in the Netherlands the reports were produced by a virtual institution with various participating organizations from inside and outside the academic sector. Various countries organize now such indicator reports, combining metrics on the science system with indicators on the economic system, innovation statistics, and so on.

102 

J. Angermuller and T. van Leeuwen

In the USA, such monitoring work has often been done by the disciplinary associations, whereas in Germany research institutes in higher education studies have played an important role in producing numerocratic knowledge. The first country in Europe to start developing a national system of research evaluation was the UK, which launched its first research assessment exercise in 1986. From this first initiative in the mid-1980s, the UK has seen periodic assessments of its research system (Moed 2008), with varying criteria playing a role in the assessment. This has been accompanied with a continuous discussion on the role of peer review in the assessment. In particular, it was held against approaches based on metrics solely (see, for example, the 2007 CWTS report and The Metric Tide report of 2015). The UK research assessment procedure evaluates the whole national system at one single moment by cutting the scientific landscape into units of assessment. An important element in the UK research assessment procedure is that it links the outcomes of assessment to research funding. Overall, one can conclude that the UK research assessment exercises tend to be a heavy burden for the total national science system. By organizing this as one national exercise, every university is obliged to deliver information on all the research at one single moment, across a large number of research fields. Many senior UK scholars get involved in the assessment of peers. Other countries in Europe also initiated national assessment systems. Finland, for example, was one of the countries initiating such a system shortly after the UK. Outside Europe, Australia has implemented a system which follows the UK model in many respects. The Netherlands initiated their research assessment procedure in the early 1990s, which is still in place, albeit with a changed design. In the Netherlands, the periodical assessment of research was institutionalized from the early 1990s onwards. Until 2003, assessment was organized under the supervision of the VSNU, the association of universities in the Netherlands (Vereniging van Samenwerkende Nederlandse Universiteiten). Their so-called chambers, consisting of representatives from the disciplines, decided how to design the research assessment process, which includes the question whether research metrics were appropriate for the research evaluation in the respective fields. In a wide number of fields, data from advanced scientometric analysis have been applied to complement peer review in research assessment procedures

  On the Social Uses of Scientometrics: The Quantification… 

103

(e.g., biology, chemistry, physics, and psychology). After 2003, the initiative to organize research assessment was put in the hands of the university boards, which meant that it was no longer a national matter. The Royal Academy of Arts and Sciences in the Netherlands has carried out studies that have influenced the recent revision of the standard evaluation protocol. The focus is now no longer only on research output and its impact, but it also considers, like in the UK, the societal relevance of academic work (‘impact’). It remains to be seen to what extent the system can still rely on peer review and what will be the role of scientometric measures in the future. While some other countries are building up national frameworks of assessing research quality, national evaluation schemes are still an exception (Table 4.1). One can see few tendencies in federal states like in the USA, where evaluation tends to be done on an institutional level, or Germany, where there are no plans for another national evaluation scheme after a ‘pilot study’ of the Wissenschaftsrat evaluated the research excellence of two disciplines (chemistry and sociology) in 2007. The French AERES carries out a ‘light-touch’ evaluation compared with the UK.  AERES regularly visits French research groups (laboratoires) and only examines publication records and not the publications themselves. Such national evaluation schemes can be highly consequential. The measurement of research performance can be done in a very explicit manner, as is the case in the UK, where the outcomes of the most recent research assessment exercise has direct consequences for money flows between and within universities, or Italy, where a similar model as applied in the UK, has been adopted over the last years. Here research performance is linked up to a nationwide research assessment practice, with intervals of five to eight years between the exercises. In other countries, research assessment has crept in more silently as some research funding is distributed on an annual basis, based upon last year’s output numbers, in which productivity in internationally renowned sources (journals and book publishing houses) provide higher rewards as compared to publishing in locally oriented journals and/or book publishing houses (Flanders, Norway, and Denmark). In France, negative assessments can lead to research clusters being closed down, whereas the German evaluation of 2007 has shown no immediate effects and little lasting influence. The

Significant impact on academic recruitment in many fields

Production of Performance statistics of statistics institutions and and the whole sector indicators are produced

Effects on academics

Researchers, departments, and institutions Evaluation of Peer review publications Scientometrics is not used officially through Allocation of Yes funding?

Unit of evaluation

UK REF 2013 Subdepartmental Initially departments and units groups, now only institutes/departments Peer review Peer review No peer review Advanced scientometrics Scientometrics is only when it fits the field not used No direct funding No direct effects No direct linking of assessment of research Scientometrics is effects but and research funding not used future of groups officially can be questioned Academics may be No effects known Only indirect and implicit effects on internal counted as university policies ‘non-publishing’ but their job security is not at stake Some statistics On the national level Ad hoc statistics are produced monitoring of the whole are produced on system, on the institutional institutional level periodic level assessment of research within institutions

Research groups (laboratoires)

France AERES

Germany Wissenschaftsrat 2007

Netherlands VSNU Protocol, currently the SEP (Standard Evaluation Protocol) protocol

Table 4.1  National evaluation schemes in some Western higher education systems

 NA

 NA

 NA

 NA

 NA

USA no systematic, national assessment implemented

104  J. Angermuller and T. van Leeuwen

  On the Social Uses of Scientometrics: The Quantification… 

105

more ambitious schemes collect publications of the researchers and have them read by assessment panels (e.g., UK and Germany). In all such evaluations, one can observe heated debates over the use of quantitative and computer-aided tools in the evaluation of the evaluation files and publications. Yet while more and more aspects of research activities are quantified (such as research revenue, publications, theses) and research activities are assessed by means of factors, indices, benchmarks and national evaluation schemes have so far stopped short of using the quantifying tools of scientometrics to rate individual research publications. In fact, while it is difficult to assess the precise impact of scientometric methods on academic decision-making, one can observe persistent conflicts between evaluators and types of evaluation (such as ‘manual’ peer review and ‘automatic’ quality assessment). What are the effects of the growing expertise that scientometrics can offer? While academic governance has become more professional, the gap between managers and decision-makers on the one hand and the academic disciplines on the other may be growing. Such a gap may be conducive to numerocratic practices in academic governance. Indeed, large disciplines organized in large institutional units (e.g., business and medicine) give more credence to quantitative indicators such as journal impact factors. And especially in non-Western countries, where resources are less abundant and managers sometimes have less training in the disciplines, quantitative signals of academic quality (e.g., publications in ‘internationally recognized’ journals) often play a more important role than in institutions where decision-makers are recognized as experts in their ­disciplinary fields and have a more intimate understanding of what is perceived to be good research. Scientometrics centres around the codified scientific knowledge production as can be measured through publications in various types of communication, especially in journal articles (for a number of historical reasons, see Wouters 1999). The Science Citation Index (SCI) forms the main basis for the scientometric method applied in basic research and in applications supporting research assessment. The SCI has been developed by Garfield Associates, which was to become part of a larger company, started by Eugene Garfield in 1958—the Institute for Scientific Information (ISI), located in Philadelphia. Garfield was an entrepreneur

106 

J. Angermuller and T. van Leeuwen

who obtained a PhD in structural linguistic in 1961 from Pennsylvania. The SCI database started to track the communication and referencing in natural science literature. The SCI was launched in 1963, followed by the Social Sciences Citation Index (SSCI) in 1973 and the Arts and Humanities Citation Index (AHCI) in 1978 (Klein 2004; Cawkella and Garfield 2001). Over time, an increasing number of journals have been covered by the ISI indices, which now count references in well over 11,000 journals from dozens of disciplines over the last decade. New developments now implicate Web-based counting practices, such as the number of occurrences of an author’s name and the number of ‘hits’ or ‘reads’ of texts posted on the Web. Both Web of Science and Scopus now make extensive use of this form of alternative research metrics, also called altmetrics. As always, one has to be careful with these numbers, as the meaning of these new research metrics and their statistical robustness are not yet clear. With ISI, Garfield became the “undisputed patriarch of citation indexing” (Cronin and Atkins 2000, 1). Garfield’s work has been crucially influential in the creation of scientometrics as an academic field. Yet he was also successful commercially. In 1992, ISI and its citation indices were sold to the Thomson Reuters Corporation for $210 million. In October 2016, Thomson Reuters completed the sale of its intellectual property and science division and transferred citation indexing to a new company called Clarivate Analytics. Given the role of Garfield’s commercial activities in the scientometric field, one wonders how academic and economic dimensions relate to each other. Is it a field that has emerged because specialized academic expertise in STS has been applied and commercialized outside the academic world? Or is citation indexing an activity of huge economic impact which has given birth, accidentally as it were, to an academic field?

4

 cientometric Indices and Rankings S as Social Practices

In the slipstream of Garfield’s commercial activities, impact indicators have been developed and taken up in the academic sector. An indicator that Garfield developed in the 1950s and 1960s is the Journal Impact

  On the Social Uses of Scientometrics: The Quantification… 

107

Factor (JIF), with first publications on the indicator in 1955 and 1963 (Garfield 1955; Garfield and Sher 1963). Journal citation statistics were included in the Journal Citation Reports (JCR), the annual summarizing volumes to the printed editions of the SCI and the SSCI. In the growing higher education sector, in which more and more journals appeared on the scene, the JIF became a tool used by librarians for managing their collections. When the JCR started to appear on electronic media, first on CD-ROM, and later through the Internet, JIF was more frequently used for other purposes, such as assessments of researchers and units, which was always sharply criticized by Garfield himself (Garfield 1972, 2006). The JIF has been included in the JCR from 1975 onwards, initially only for the SCI, later also for the SSCI. For the AHCI, no JIFs were produced. From the definition of the JIF, it becomes apparent that JIF is a relatively simple measure, is easily available through the JCR, and relates to scientific journals, which are the main channel for scientific communication in the natural sciences, biomedicine and parts of the social sciences (e.g., in psychology, economics, business and management) and humanities (e.g., in linguistics). While the ISI indices cover a number of features in journal articles, they focus mostly on the references cited by the authors. These references are taken to express proximity and influence between citing and cited people. On the receiving end, the question is to whom they relate. Here references are considered as citations. Citation theory argues that value is added when references become socially recognized citations, objectifying as it were the researchers’ social capital (see Wouters 1999; Angermuller 2013b; Bourdieu 1992). The value that members of scientific communities add to works can be objectified through normalized measures. Thus, by processing and counting references the scientometrician does not only reflect a given distribution of academic value but she or he also adds value to references. Other types of information used in the analysis relates to the authors and co-authors, their institutions and cooperations, the journals in which their papers come out, their publishing houses, information on the moment of publishing, information on the language of communication in the publication, meta-information on the contents, such as keywords, but also words in titles and abstracts, as well as information on the classification of the publications in various disciplinary areas.

108 

J. Angermuller and T. van Leeuwen

The fact that these indices mostly focus on journal publications has been widely criticized. It has been argued, for instance, that certain disciplinary areas have been put in a disadvantageous position (e.g., history, where monographs are more important, or new disciplinary and transdisciplinary fields which don’t have established journals yet). Moreover, it needs to be recalled that the three indexes tend to over-represent research in English as the main language of international scientific communication (van Leeuwen 2013; van Leeuwen et  al. 2001). Areas in which English-language journals are not standard outlets for research tend to become peripheral in citation indexes. As a result, Western (i.e., US and Western European) journals in the natural, life, and biomedical sciences had long been given a certain prominence, which has been called into question only after the rise of the BRIC countries (Brazil, Russia, India and in particular of China). Other geographical regions of the world are now better represented in the ISI, although the bias towards English has remained intact as many of these countries follow models of the English-­ speaking world. The rise of these indices took place when some scientometric indicators such as the JIF and the h-index started to be used in evaluation practices throughout the science system. The JIF was originally designed by Garfield for librarians to manage their journals in their overall library collection and for individual researchers in the natural sciences to help them decide on the best publication strategies (Garfield 1955, 1972, 2006; Garfield and Sher 1963). The JIF has been used in a variety of contexts, for example, by managers who evaluate whole universities (often with a more formal registration of research outputs of the scholarly community) and by individual scholars to ‘enrich’ their publication lists while applying for research grants and for individual promotion or job applications (Jiménez-Contreras et al. 2002). Yet indicators such as the JIF are not neutral as they can bring forth the realities they represent through the numbers. There are no indicators that have been universally accepted to represent quality of research or the performance of researchers. As they have been the object of controversial debate, a number of serious flaws with the JIF have been pointed out which disqualify the indicator for any use in science management, let alone for evaluation purposes. Thus, it is cal-

  On the Social Uses of Scientometrics: The Quantification… 

109

culated in ways that in about 40% of all journals JIF values are overrated (Moed and van Leeuwen 1995, 1996). Another issue is that JIF values do not take into consideration the way the journal is set up. Journals, for example, that contain many review articles, tend to get cited more frequently as compared to normal research articles. Therefore, review journals always end up on top of the ranking lists. A third issue relates to the fact that JIF values do not take into consideration the field in which the journal is positioned. Reference cultures differ, as do the number of journals per field. This means that fields with a strong focus on journal publishing, and long reference lists, have much higher JIF values as compared to fields where citations are not given so generously. A fourth reason relates to the fact that citation distributions are, like income distributions, skewed by nature. This means that a JIF value of a journal only reflects the value of few much-cited articles in the journal while most have lower impacts. This creates a huge inflation in science, given the practice mentioned above, in which scholars tend to enrich their publication lists with JIF values, which say nothing about the citation impact of their own articles. Moreover, JIF values tend to stimulate one-indicator thinking and to ignore other scholarly virtues, such as the quality of teaching, the capability to ‘earn’ money for the unit, the overall readiness to share and cooperate in the community. The h-index was introduced in 2005 (Hirsch 2005). It is to assess an individual researcher’s performance by looking at the way citations are distributed across all publications of that person. If one takes the output in a descending order (by number of received citations), the h-index represents the number of citations received equalling the rank order (i.e., if somebody has published five articles, cited 20, 15, 8, 4, and 3 times, the h-index is four). Due to the simplicity of the indicator, it has been widely adopted and is sometimes even mentioned to justify hiring and firing decisions as well as the evaluations of research proposals in research councils. The problems with the h-index are manifold. First of all, a number of issues related to JIF also apply on the h-index, one of them being the issue of the lack of normalization, which makes comparisons across fields impossible (van Leeuwen 2008). A next issue is the conservative nature of the indicator, it can only increase, which makes the h-index unfit for

110 

J. Angermuller and T. van Leeuwen

predictions. The next set of issues relates to the way this indicator is calculated. Depending on the database, calculations of the h-index can differ significantly. In many cases, authors and their oeuvres cannot be determined easily. A final set of issues relates to a variety of more general questions, such as the publication strategies chosen by researchers (putting your name on every single paper from the team or be more selective), the discrimination against younger staff and the invisibilization of scholarly virtues. Indicators such as the JIF and the h-index are nowadays easily available for everybody. They are readily used by people in research management and science policy, government officials, librarians, and so on. Even though precise effects are difficult to prove, these indicators often play a role for decision-making in grant proposal evaluation, hiring of academic personnel, annual reviews as well as promotion and tenure decisions. If they are applied in a mechanistic way without reflecting on their limits, such indicators can go against values which have defined the academic ethos, for example, the innovation imperative, the service to the community, the disinterested pursuit of ‘truth’. While the JIF and the h-index testify to numerocratic practices within academic research, university rankings are an example for the numerocratization of higher education in the broader social space. University rankings were initially intended to help potential students to select a proper university matching their educational background. Yet these rankings have turned into a numerocratic exercise that now assesses many aspects of university performance. Since 2004, with the launch of the so-called Shanghai ranking, universities have been regularly ranked worldwide, which has contributed to creating a global market of higher education. As a result, these league tables can no longer be ignored by managers and administrators, especially in Anglo-American institutions, which highly depend on fees brought by international students (Espeland and Sauder 2007). With the exception of the Leiden ranking, which is based entirely upon research metrics, university rankings usually contain information on educational results, student-staff ratio, reputation, and research performance. All prominent university rankings, including the ARWU Ranking (Academic Ranking of World-Class Universities, aka Shanghai Ranking), the Times Higher Education university ranking and

  On the Social Uses of Scientometrics: The Quantification… 

111

the QS Ranking, include scientometric information in their ranking formula. Many of the problems that one can observe with scientometric analyses, such as coverage and language, find their way into university rankings (see van Leeuwen 2013; van Leeuwen et  al. 2001; van Raan et  al. 2011a, b). In the European context, the performance of French, German, Italian and Spanish universities is seriously underrated because their output contains local language output (see van Raan et al. 2011a, b). Some researchers, particularly from Germany, have therefore started to publish in English. And since university rankings tend to reward sheer size policy-makers have cited the need to do well in international university rankings to push through reforms such as the German Exzellenzinitiative or the recent clustering of French institutions. Generally speaking, we can point out various types of scientometric studies, which testify to the different levels of numerocratic practices objectifying the social through numbers. A macro-level scientometric analysis provides scientometric information on the level of countries or groups of countries (e.g., the EU, the OECD countries, etc.); a meso-­ level analysis concentrates on institutions such as universities, publicly funded research organizations, corporations and larger units within these institutions (faculties, research groups, etc.). And, finally, one can see micro-level analyses that deal with small research groups, programs, and individual researchers. This is a level to which scientometric analysis is extended more often nowadays. The requirements of data collection vary according to the level of analysis. Raw data for macro-level analyses is relatively well organized for some countries if one takes citation indices as a point of departure. One can distinguish between the outputs of different European countries relatively easily and most work here involves questions such as the grouping the four countries England, Scotland, Wales and Northern Ireland within the UK or the creation of labels such as the EU (as it has been progressively enlarged) or the OECD. On the meso-level, data handling requires significant knowledge of the science system on the national level (e.g., how academic hospitals are related to universities in the same cities), what name variations exist of a university and its various locations, how are other publicly funded research organizations grouped on the national level), and in numerous cases, data handling needs to achieve unification and bring the many variants of an

112 

J. Angermuller and T. van Leeuwen

institution under one umbrella name. On the micro level, the challenges of scientometric analysis are the greatest. In the first place, micro-level analysis often, if not always, needs to involve those who are the object of the analysis so as to define research groups, projects, programmes, and publications realized by these units. Some sort of certification or authorization is required without which the outcomes of the study would lack legitimacy. In the second place, on the level of the individual researcher, the problem of homonyms and synonyms plays an important role. As one single person can publish under various names in the international serial literature (e.g., by using various initial combinations, sometimes one of the first names is written in full, etc.), these occurrences have to be brought back to one single variation. Likewise, one name variation can hide various persons, due to the occurrence of very common names (in the English language area names like Brown or Smith), in combination with one single initial but many Chinese scholars mean formidable challenges to citation indices. Scientometric data handling requires information on the full names of individuals, the field in which people have been working and also about their career track. One can try to collect such data. However, ideally, one should consult the authors as they are those who know best. Verifying publications not only increases the validity of the outcomes of the scientometric study but also adds to the transparency of the process. In order to critically reflect on how numerocracy works in and through scientometrics, one needs to understand how the empirical data for indicators and rankings are collected. With respect to this phase of data collection, we can cite for one moment the work of Pierre Duhem, a French natural scientist and philosopher of science. In his so-called law of cognitive complementarity, he holds that the level of accuracy and the level of certainty stand in a complex relationship with each other (Rescher 2006). If we follow Duhem, the analysis of the macro level can teach us about the performance of a country in a particular field but it cannot tell us anything about any particular university active in that field, let alone about any of the research programmes or individual scholars in that field. Vice versa, while analyses on the micro level can instruct us about individual scholars and the research projects they contribute, it cannot inform us about the national research performance in that same field of research.

  On the Social Uses of Scientometrics: The Quantification… 

113

Even at the meso-level, where we would expect the level of certainty and accuracy to be easier in balance, the world remains quite complicated: an overview of the research in, say, a medical centre in the field of immunology does not relate in a one-to-one relationship to the department of immunology in that centre, as researchers from various departments might publish in the field of immunology, such as haematologists, oncologists, and so on. This tension between the level of certainty and accuracy exists at any moment and influences the range and reach of the conclusions that can be drawn from scientometric data.

5

 onclusion: For a Critical Reflection C of Scientometrics as a Numerocratic Practice

Over the last few decades, higher education has gone through profound changes. While the sector has grown hugely and in many areas continues to grow, higher education is under conditions of austerity and both taxpayers and private sponsors expect more and more justification as to how money is spent on academic activities. Academic systems have been under pressure to turn away from ‘personal’ modes of evaluation (such as peer review) towards more managerial and neoliberal approaches. Scientometrics responds to these demands by subjecting research output of individuals, institutions, and academic systems to numeric analysis and representation. It was the objective of this paper to place the rise of scientometrics in the context of ‘numerocracy’, a regime which uses numbers to place people into social networks, groups, or structures. Against this background, scientometrics signals the advent of numerocracy in the higher education sector. In higher education, numerocratic governance aims at monitoring and controlling large academic populations through numbers. Such approaches typically define arenas where academics are allowed to engage in ‘free’ competition over scarce resources and constitute hierarchies of ‘excellence’ among researchers. Scientometrics can provide the tools that objectify social inequalities and justify hierarchical relationships

114 

J. Angermuller and T. van Leeuwen

between researchers. By measuring research impact and citation patterns, scientometricians numerocratic practices not only legitimate but also create unequal distributions of resources and prestige. Thus, scientometrics can to naturalizing and depoliticizing social struggles among the participants of the academic ‘game’. Numerocracy theory provides critical tools to reflect on the social effects of such practices, which the actor does not always control. While it acknowledges the part of the actors in producing and reproducing social inequalities (Espeland and Stevens 2008), it places scientometric research in the larger context of a knowledge-power dispositif (see Chap. 3 in this volume) which has applied numerocratic techniques to a number of arenas and populations since the eighteenth century. Scientometrics has met with widespread suspicion if not outright criticism from various corners of the academic community for participating in power games. If scientometricians are perhaps the first ones to be aware of the political uses and abuses of the expertise it can offer, we want to insist on the dynamics of power in which scientometric research finds itself entangled. As a field, scientometrics is part and parcel of a system which creates, reinforces, and legitimates social inequalities between academics. How should scientometricians act in the light of political effects of their work? How should they deal with the fact that their knowledge is not always used in ways that is in accord with their goals and values? One solution could be to distinguish between descriptive (or basic) and normative (or applied) research in scientometrics. Descriptive scientometrics would put the emphasis on revealing differences in behaviours and practices of academics. It would draw on methods of social research to study social order in the world of science empirically. The goal of normative scientometrics, by contrast, would be to facilitate institutional decision-making. Through indicators such as the JIF or the h-index, it could claim, for instance, to assess the ‘performance’ of researchers and to measure the ‘quality’ of research output. Scientometricians would then have the choice to take on one or the other role. They would do scientometrics either as an empirical social research (and perhaps integrate it back into STS) or they would produce applied knowledge for decision-­ makers (following Garfield, whose work has crucially contributed to the ISI Web of Science indicators).

  On the Social Uses of Scientometrics: The Quantification… 

115

Yet if we take seriously the idea of scientometrics as an aspect of a numerocratic dispositif, the distinction between descriptive and normative research is difficult to maintain. For neither practice can stand by itself. While descriptive research uses numbers to represent social realities, its representations cannot but partake in the constitution of such realities. Just like any social research that aims to make a difference, scientometrics cannot simply render a given state of the social world without intervening in it. And how could there be normative scientometrics which does not at the same time claim to be descriptive (and the other way round)? Even more fundamentally, one can ask to what degree applied research is in control of the political effects that it has in the social world? Is applied research always affirmative of the values that motivate its sponsor to give resources for it? Does not ‘descriptive’ research that wants to make a positive difference in the world precisely need to follow the ‘applied’ route and seek to work with agents and stakeholders outside its scientific community? There is probably no general response to these dilemmas which many scientometricians are confronted with. And there is no way to find responses without accounting for the specific circumstances in which researchers struggle to find solutions to the problems they deal with in their everyday lives. While it is difficult to assess the impact of academic research on the social world, it is even more difficult to determine the precise effects that scientometric knowledge has on actions and decision-­ making. Over the last couple of years, ethnographic studies on the role of research metrics on shop floor level shows the degree of penetration of those metrics in various aspects of academic life, for example, in decision-­making on publication strategies, on promotion procedures, the role of research metrics in research grant procedures, and so on (Rushforth and de Rijcke 2015). It is all too easy to fall prey to the myth of a scientometric conspiracy in higher education. Research has always been and will always remain a creative social practice, which will never be totally dominated by benchmarks, indicators and algorithms. Scientometrics, therefore, should be critical of its role in the numerocratized dispositif of higher education and it cannot but gain from a critical reflection on the fragile and preliminary nature of scientific knowledge more generally.

116 

J. Angermuller and T. van Leeuwen

Today, numerocratic technologies are used in many areas of social life and the rise of scientometrics more generally testify to the increasing demand for scientifically established numbers. Yet with the spread of the digital medium, numbers have become common in various areas of social life. Consumers use numbers when they choose a hotel or a mobile phone provider. Citizens use numbers in the democratic debate to determine what clickworthy news is. And numbers even have become crucial in the match-making and dating business. By taking scientometrics as an example of the emerging numerocratic dispositif, we invite to critically reflect on the social effects that numbers have in many areas of social life today.

References Angermuller, Johannes. 2013a. Discours académique et gouvernementalité entrepreneuriale. Des textes aux chiffres. In Les discours sur l’économie, ed. Malika Temmar, Johannes Angermuller, and Frédéric Lebaron, 71–84. Paris: PUF. ———. 2013b. How to become an academic philosopher. Academic discourse as a multileveled positioning practice. Sociología histórica 3: 263–289. ———. 2017. Academic careers and the valuation of academics. A discursive perspective on status categories and academic salaries in France as compared to the U.S., Germany and Great Britain. Higher Education 73 (6): 963–980. Angermuller, Johannes, and Jens Maeße. 2015. Regieren durch Leistung. Zur Verschulung des Sozialen in der Numerokratie. In Leistung, ed. Alfred Schäfer and Christiane Thompson, 61–108. Paderborn: Schöningh. Bloor, David. 1976. Knowledge and social imagery. London: Routledge & Kegan Paul. Bourdieu, Pierre. 1992. Homo academicus. Frankfurt/Main: Suhrkamp. Burrows, Richard. 2012. Living with the h-index? Metric assemblages in the contemporary academy. Sociological Review 60 (2): 355–372. Cawkella, Tony, and Eugene Garfield. 2001. Institute for scientific information. In A century of science publishing, ed. E.H. Fredriksson, 149–160. Amsterdam: IOS Press. Committee on the Independent Review of the Role of Metrics in Research Assessment and Management. 2015. The metric tide. Report to the HEFCE, July 2015. Accessed June 27, 2018. http://www.hefce.ac.uk/media/

  On the Social Uses of Scientometrics: The Quantification… 

117

HEFCE,2014/Content/Pubs/Independentresearch/2015/The,Metric, Tide/2015_metric_tide.pdf. Cronin, Blaise, and Helen Barsky Atkins. 2000. The scholar’s spoor. In The web of knowledge: A festschrift in honor of Eugene Garfield, ed. Blaise Cronin and Helen B. Atkins, 1–8. Medford, NJ: Information Today. CWTS. 2007. Scoping study on the use of bibliometric analysis to measure the quality of research in UK higher education institutions. Report to HEFCE by the Centre for Science and Technology Studies (CWTS), Leiden University, November 2007. Desrosières, Alain. 1998. The politics of large numbers: A history of statistical reasoning. Cambridge, MA: Harvard University Press. Dunleavy, Patrick, and Christopher Hood. 1994. From old public-­administration to new public management. Public Money & Management 14 (3): 9–16. Espeland, Wendy Nelson, and Michael Sauder. 2007. Rankings and reactivity: How public measures recreate social worlds. American Journal of Sociology 113 (1): 1–40. Espeland, Wendy Nelson, and Michell L. Stevens. 2008. Commensuration as a social process. Annual Review of Sociology 24: 313–343. Foucault, Michel. 1973. The birth of the clinic: An archaeology of medical perception. London: Routledge. Original edition, 1963. ———. 1980. Power/knowledge: Selected interviews and other writings 1972–1977. Edited by Colin Gordon. New York: Pantheon. ———. 1995. Discipline and punish: The birth of the prison. New York: Vintage Books. ———. 2002. The order of things. An archeology of the human sciences. London: Routledge. Original edition, 1966. ———. 2007. Security, territory, population. lectures at the college de France. Basingstoke: Palgrave Macmillan. Original edition, 1977/78. ———. 2008. The birth of biopolitics. Lectures at the Collège de France, 1978–1979. London: Palgrave Macmillan. Garfield, Eugene. 1955. Citation indexes to science: A new dimension in documentation through association of ideas. Science 122 (3159): 108–111. ———. 1972. Citation analysis as a tool in journal evaluation. Science 178 (4060): 471–479. ———. 2006. The history and meaning of the journal impact factor. JAMA 295 (1): 90–93.

118 

J. Angermuller and T. van Leeuwen

Garfield, Eugene, and Irving H. Sher. 1963. New factors in the evaluation of scientific literature through citation indexing. American Documentation 14 (3): 195–201. Hirsch, Jorge Eduardo. 2005. An index to quantify an individual’s scientific research output. Proceedings of the National Academy of Sciences of the USA 102 (46): 16569–16572. Hood, William W., and Concepción S. Wilson. 2001. The literature of bibliometrics, scientometrics and informetrics. Scientometrics 52 (2): 291–314. Jiménez-Contreras, Evaristo, Emilio Delgado López-Cózar, Rafael Ruiz-Pérez, and Victor M.  Fernández. 2002. Impact-factor rewards affect Spanish research. Nature 417: 898. Klein, Daniel B. 2004. The social science citation index. A black box—With an ideological bias? Econ Journal Watch 1 (1): 134–165. Knorr Cetina, Karin. 1981. The manufacture of knowledge. An essay on the constructivist and contextual nature of science. Oxford and New York: Pergamon Press. Latour, Bruno, and Steve Woolgar. 1979. Laboratory life. Princeton, NJ: Princeton University Press. van Leeuwen, Thed N. 2008. Testing the validity of the Hirsch-index for research assessment purposes. Research Evaluation 17 (2): 157–160. ———. 2013. Bibliometric research evaluations, web of science and the social sciences and humanities: A problematic relationship? Bibliometrie  – Praxis und Forschung, 1–18. Accessed June 27, 2018. http://www.bibliometrie-pf. de/article/view/173. van Leeuwen, Thed N., Henk F. Moed, Robert J.W. Tijssen, Martijn S. Visser, and Ton F.J. Van Raan. 2001. Language biases in the coverage of the science Citation Index and its consequences for international comparisons of national research performance. Scientometrics 51 (1): 335–346. Mandel, Ernest. 1978. The second slump. London: Verso. Merton, Robert K. 1962. Science and the social order. In The sociology of science, ed. Bernard Barber and Walter Hirsch, 16–28. Westport, CT: Greenwood. Miller, Peter. 2001. Governing by numbers. Why calculative perspectives matter. Social Research 68 (2): 379–396. Moed, Henk F. 2008. UK research assessment exercises: Informed judgments on research quality or quantity? Scientometrics 74 (1): 153–161. Moed, Henk F., and Thed N. van Leeuwen. 1995. Improving the accuracy of institute for scientific information’s journal impact factors. Journal of the American Society for Information Science 46: 461–467.

  On the Social Uses of Scientometrics: The Quantification… 

119

———. 1996. Impact factors can mislead. Nature 381: 186. NOWT.  Wetenschaps- en Technologie Indicatoren, by the Netherlands Observatory of Science & Technology (NOWT). Report to the Dutch Ministry of Science and Education, 1994, 1996, 1998, 2000, 2003, 2005, 2008, 2010, 2012, 2014. Ogien, Albert. 2010. La valeur sociale du chiffre. La quantification de l’action publique entre performance et démocratie. Revue française de socio-économie 5: 19–40. OST. 1992. Observatoire de Science et Technologie (OST). Edition 1992, 1994, 1996, 1998, 2000, 2002. Science & Technologie Indicateurs. Paris. Porter, Theodore. 1994. Trust in numbers. The pursuit of objectivity in science and public life. Princeton, NJ: Princeton University Press. van Raan, Ton F.J., Thed N. van Leeuwen, and Martijn S. Visser. 2011a. Non-­ English papers decrease rankings. Nature 469 (1): 34. ———. 2011b. Severe language effect in university rankings: Particularly Germany and France are wronged in citation-based rankings. Scientometrics 88: 495–498. Radder, Hans. 2010. The commodification of academic research. Science and the modern university. Pittsburgh, PA: University of Pittsburgh Press. Rescher, Nicolas. 2006. Epistemetrics. Cambridge: Cambridge University Press. Rushforth, Alex, and Sarah de Rijcke. 2015. Accounting for impact? The journal impact factor and the making of biomedical research in the Netherlands. Minerva 53: 117–139. Wouters, Paul. 1999. The citation culture. PhD thesis, University of Amsterdam, the Netherlands. WTI. (2012) 2014. Wetenschaps-, Technologie en Innovatie Indicatoren (WTI). Report to the Dutch Ministry of Science, Culture & Education, 2012, 2014.

Part III Exploring Corpora: Heuristics, Topic Modelling and Text Mining

5 Lexicometry: A Quantifying Heuristic for Social Scientists in Discourse Studies Ronny Scholz

1

Introduction

Most discourse analytical studies have a common interest in patterns of how knowledge is (re-)produced, (re-)distributed, and controlled through social practices of language use. Discourse analysts have developed numerous methods to demonstrate how meaning is constructed. However, the reason why a particular text or textual sequence in a given corpus was chosen to be analysed often remains arbitrary. Distinguishing hermeneutic from heuristic methods, this contribution introduces a systematic quantitative methodology guiding the analyst’s choices of texts and textual sequences. The text emphasises the heuristic strength I am thankful to Malcolm MacDonald for his helpful comments on earlier versions of this text. Additionally, I want to thank André Salem for the numerous personal tutorial sessions and discussions of the software Lexico3 with which most of the analyses in this text have been conducted.

R. Scholz (*) Centre for Applied Linguistics, University of Warwick, Coventry, UK e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_5

123

124 

R. Scholz

of lexicometric methods, providing the researcher with numerous perspectives on the text collections she or he is analysing. On the basis of these perspectives, the analyst can develop hypotheses about discourse phenomena driven by the corpus data, which can be verified and falsified in the hermeneutic research phase. At least since the 1970s, lexicometry, not without polemic, has developed into one of the prominent methodologies in French discourse analysis, mainly in order to analyse the language use in political discourses (Bonnafous and Tournier 1995; Tournier 1975, 1993, for an introduction; Mayaffre 2016; Mayaffre and Poudat 2013; Scholz and Mattissek 2014). Lexicometric methods are especially suited for the study of recurrent language use patterns assuming that a high frequency of a particular  use of language reflects the typical way knowledge is structured in society or a part of it. These patterns are normally studied with a number of quantitative methods in text collections (corpora) of varying size. When analysing society through language use, we have to operationalise the world in a particular way. Discourse analysts gain their data through a chain of procedures through which they reduce the complexity of the world (see also Chap. 2 in this volume). It is only after having transcribed, edited, and enriched actual language use with some metadata that quantitative and qualitative methods can be used to analyse and interpret discourses. Lexicometric methods allow describing discourses in terms of concrete and ‘differentiated’ language use which can be explained with reference to historical and social situations as well as to individuals and social groups that are embedded into power relations. Computational methods are one way to account rigorously for these relations, to the extent that they can be translated into a machine-readable language. The concept of ‘differentiated use’ [usage différentiel] refers to an understanding of the sociolinguistic diversity (historical situation, social groups, communication situation, and genre) in relation to the language system. With contrastive measurements, one can distinguish the predominant topics, text types, genres, and subgenres (Mayaffre 2005). Stylistic characteristics as well as rhetoric and argumentative effects can be described and juxtaposed. Furthermore, the distribution of particular patterns of language use or similarities between texts can be measured (Mayaffre

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

125

2016). Even if there is a substantial overlap with corpus-assisted discourse studies in the Anglo-Saxon world, both traditions have ignored each other for a number of different reasons, which I will not elaborate here. Therefore, one objective of this chapter is to present the French quantitative approach to a novice readership. Notwithstanding this introductory impetus, experts in the field will also find some useful ideas for their research. In the following section, I outline the historical context in which the lexicometric approach developed. Thereafter, in the third section I explain theoretical aspects that are important for the understanding of a quantitative heuristic approach. In the fourth section, I remind the reader of the criteria that should be taken into account when building a corpus to be analysed with lexicometric methods. The fifth section is the main part. Here, I will demonstrate how exhaustive methods that take into account the whole vocabulary of a corpus simultaneously can be used to take a variety of angles onto the data. I will illustrate how quantifying methods can be used to explore the lexicosemantic macro-structures in a discourse. Important discursive aspects can be discovered without a prior interpretation of texts. Furthermore, I will show how lexicometric methods can be used to reduce a large corpus to a selection of prototypical texts, which then can be interpreted with various methods.

2

 ome History of the Lexicometric S Methodology

Influenced by textual philology, stylistics, statistics, and computer sciences, the lexicometric approach has been developed in France since the 1970s as a computer-assisted methodology for the analysis of language use in political discourses. Early precursors of the approach can be found in quantitative linguistics related to names such as George Yule (1944), George Zipf (1929, 1935) Gustav Herdan (1964, 1966), or Pierre Guiraud (1954, 1960). Influential was Charles Muller’s contrastive statistic model to describe lexico-grammatical differences between the comic and the epic parts of the drama L’illusion comique by Corneille (Muller 1967). Ever since, based on Benzécri’s multivariate statistics (Benzécri

126 

R. Scholz

1963, 1969, 1982), methods for exhaustive, contrastive analysis of corpora have been developed (Lafon 1984; Lebart and Salem 1988; Lebart et al. 1998). The term ‘lexicometry’ (lexicométrie) had been used by Maurice Tournier (1975) and his team at the Laboratoire de lexicologie et textes politiques (Ecole Normale Supérieure de Saint-Cloud) since the 1970s in order to mark the methodological independence of the approach focusing on lexical statistics. Meanwhile, other terms such as textométrie and stylométrie have been introduced highlighting different aspects of linguistic investigation. Lately Mayaffre (2007) has advocated the ‘logometric approach’ aiming for an automated analysis of all textual levels comprising all corpus methods and exceeding the simple analysis of word forms by including lemmata, grammatical structures, and a possible recourse on the original text source. Like Pêcheux’s discourse analysis, lexicometry aims for a thorough analysis of political language use. However, even though both approaches developed in a similar intellectual environment with a similar objective, their proponents encountered each other with distance and suspicion in the first instance. Pêcheux and his disciples (Pêcheux 1982; Pêcheux et al. 1979) whose work is based on Harris’s (1952) distributional approach disputed that counting words cannot help to uncover the hidden inherent ideological permeations of texts—being an important aspect of French discourse analysis at the time (Scholz and Fiala 2017). In contrast, lexicometric scholars criticised Pêcheux’s approach for its oversized theoretical superstructure and for an incoherent methodology when trying to link non-linguistic and linguistic aspects of language use (Demonet et al. 1975, 38). They argued that by founding the analysis on statistical results, in the first instance, the researcher can get rid of his or her ideological presumptions that cannot be easily controlled when interpreting texts. However, since then the mutual scepticism has faded over time due to the development of sophisticated methods allowing quantification of the co-­ text (Heiden and Lafon 1998; Martinez 2012; Salem 1987), retracing of syntactic structures (Fiala et  al. 1987) and direct access to the textual material (Fiala 1994).

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

3

127

 pistemological Rupture and Heuristic E Methodology

Lexicometry is a data-driven approach to (mostly political) language use in which the analyst applies a range of statistical algorithms onto textual data in order to unveil its lexicosemantic macro-structures from a synchronic and diachronic perspective. Historical or sociological categories such as time of text production or authorship can be introduced into the text corpus as metadata prior to the analysis. They are used to partition the corpus and allow typical and untypical language use to be highlighted in the resulting parts of the corpus through the exhaustive contrasting of word frequencies. Thus, ‘exhaustive’ means that all word tokens of a corpus are taken into account at once. Applying these methods, the researcher alienates himself or herself temporarily from the textual material in the first instance for the benefit of an explorative phase. The decontextualized representation of text statistical results can trigger new ideas about relations between different corpus parts referring to authors and time periods that might stay hidden if one limits the analysis to methods of text interpretation. The strength of these quantifying methods is their capability to map a corpus and help the researcher discovering the most dominant lexicosemantic features—even before he or she starts reading the texts. Against this backdrop I want to suggest using quantifying methods as a sort of machine-led reading technique that guides the analyst to statistically salient text sequences, which then can be analysed with qualitative interpretative methods. In this sense, interpretations concerning the construction of meaning in context are postponed to a later point in research (Scholz 2016; Scholz and Angermuller 2013; Scholz and Ziem 2015). Moreover, I argue that by postponing the interpretation in this way, we can systematically induce epistemological ruptures. Based on such ruptures we can put into practice Foucault’s “project of pure description of discursive events1” in their original neutrality, which puts aside the inherent continuities of knowledge formations triggered when allocating an utterance to a certain author, period, text type, and so on (Foucault 2003,  Italic in original.

1

128 

R. Scholz

400). Foucault draws from Bachelard’s epistemology, according to which different ‘worlds of reality’ are the result of different scientific descriptions of the world. New descriptions of the world only have become possible because they have ruptured with the everyday-life perception and thought of the world. Against Husserl’s phenomenology Bachelard states: ‘The world in which one thinks is not the world in which one lives’ (Bachelard 1962, 110). The precondition for scientific thought is an epistemological rupture with the habitual way of thinking (Diaz-Bone 2007). If we understand lexicometric methods as a way of putting epistemological ruptures into practice, it is because they allow us to ignore  the interpretation of utterances in prefabricated categories. Based on word frequencies these methods render apparent continuities and discontinuities of discursive practices not based on the assumption that texts of different authors, periods, and text genres must be different—but by verifying, if, on the statistical level, such differences can be confirmed or not. Lexicometry can help to get a grip on a large amount of data representing discourses. They allow a critical stance by providing a quantifying heuristic of these data, which then can be analysed thoroughly with various methods used in lexicometric and other approaches of discourse analysis. Normally a combination of lexicometric methods with qualitative methods of discourse tackling text and particular utterances on the micro level will give most robust and telling results. Heuristic methodologies can be distinguished from hermeneutic methodologies. Whereas heuristic methods trigger research discoveries based on new perspectives onto the research material, hermeneutic methods help interpreting the meaning of this material. The combination of both activities is crucial for analysis. However, in social research the heuristic aspect sometimes is neglected in favour of an abundant interpretation within a particular theoretical framework. Lexicometry meets the criteria that Kleining has developed for heuristic methods in qualitative social research: Openness of the researcher for new unexpected results, a provisory research object, a maximum of structural variation of perspectives, an analysis focusing on common features (Kleining 1994, 22–46 and 178–198). Lexicometry is a methodology that allows in a given corpus exploring textual data through a variety of operations helping to ­produce hypotheses and narrowing down the research material based on

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

129

statistical salience. Moreover, the different methods inform the interpretation of the textual sequences that have become relevant throughout the first phase of investigation (Glady and Leimdorfer 2015).

4

 ompiling a Text Collection C for a Lexicometric Analysis

When working with lexicometric methods, we normally work with textual data of all sorts which have to be compiled according to particular criteria. First of all, we need to have a basic understanding of our research object. There are numerous rather elaborate definitions of what a discourse is. However, for our purposes, we chose a very general definition that has been advocated for by German discourse linguists. Within this approach, discourse was defined from a perspective of research practice as a ‘virtual text corpus’ containing all kinds of texts concerning a certain discourse topic. To analyse a discourse, a researcher must compile a ‘concrete text corpus’ constituting the object of research which contains a representative number of texts taken from the virtual corpus (Busse and Teubert 2014, 344). Representativeness of our data is an important criterion to produce valid hypotheses of general value about the research object, a given discourse. Notwithstanding the importance of aiming for representativeness of the research data I would say that this is an ideal typical requirement. Because it is rather difficult to decide which texts are representative of a discourse before we actually start the analysis,  instead I would like to highlight that it is important to keep the issue of ‘representedness’ in mind throughout the whole research process: What, who, and how are the texts under investigation representing? To what extent do the data structure and the context, in which the data were produced, influence the statistical results? Finally, we have to be aware of the fact that, in terms of statistical algorithms, we are applying a rigid apparatus on a more or less arbitrary collection of data. And in this sense, the choice of texts predetermines the results. If we are not reflecting on what our data are actually representing, there is always a risk of producing artefacts. For instance,

130 

R. Scholz

visuals can perfectly represent the result of a particular method we applied, but at the same time, it is possible that they do not represent a discourse in general and hence the social reality we were actually interested in. This becomes obvious when we add texts from a particular discourse participant and leave out texts from another participant. By omitting or overemphasising particular aspects of a discourse, we are actively influencing the outcome of our analysis. However, experimenting with different datasets and methods in this way is also an important way of validating our research results. Furthermore, in order to avoid the production of artefacts, it is important that our research corpus is homogenous in terms of text register (Lee 2001). Therefore, all parts of a corpus should contain a similar amount of texts produced in the same or similar contexts. This is necessary because we want to make sure that our rigorous methodical apparatus compares features of language use on the same ontological level. For example, if our corpus contains only public speeches for one discourse participant and only interview data for another participant we would not be able to determine if the result is caused by the different text types or by a different language use of a speaker. As in discourse analysis we are predominantly interested in the latter aspect, we should compile our corpus from texts coming from more or less homogeneous sources of text production. Before we can start the examination of the corpus, we have to make sure that the different metadata, such as author, date of text production, and text type, are allocated to the data material in a way that computer software is able to partition the corpus according to these variables. The partitioning of the corpus is a way to account for the influence of the context of text production onto a discourse. For instance, contrasting corpus parts built from different discourse participants presupposes that there might be different ways, in which different participants speak/write concerning a certain topic and subsequently position themselves in the discursive field; corpus parts created from different periods of text production are based on the presumption that the vocabulary and therefore the discourse on a certain topic evolves over a period of time, and so on. Metadata are one way to integrate sociological research categories into our analysis. In this sense, we can account for different actors and ­institutions, their spheres of activity, and social spaces. When we, in a

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

131

second step, analyse the language use within these sociological categories, we then can refer the use of particular words to the construction of meaning within a particular social context (Duchastel and Armony 1995, 201). Such an analysis aims at an investigation of topics, utterances, and positioning practices structuring the organisation of the symbolic field and the exercise of power in contemporary societies (Duchastel and Armony 1993, 159). It is the contrastive analysis of actual language use with reference to sociological metadata with which we aim to investigate the meaning construction entangled in social and societal relations such as work, class, gender, race, and poverty.

5

Examples of the Empirical Analysis

In the following section of this text, I shall be illustrating some lexicometric methods. The corpus used for the following exemplary studies was compiled in the context of a larger project on crisis discourses in the German press since the 1970s, led by Martin Wengeler and Alexander Ziem and funded by the German Research Foundation. The corpus for the financial crisis 2008 was compiled from press articles taken from the five largest national newspapers and magazines with the intent to cover the German political spectre: the Bild (populist conservative daily tabloid), the Frankfurter Allgemeine Zeitung (FAZ—conservative), the Süddeutsche Zeitung (SZ—progressive liberal), Die Zeit (weekly, centre-left liberal), and the weekly magazine Der Spiegel (centre-left). The research period was determined according to the intensity of the media coverage of the crisis starting in September 2008 with the bankruptcy of the Lehman bank and ending in April 2009 with the G20 summit on financial market issues. The search term used to identify relevant articles was Finanzkrise (financial crisis). The abundance of texts found with this search term prevented the use of other search terms that would have led to an unmanageable number of texts. All identified texts were read rapidly and double-checked for their relevance to the crisis issue. The total number of press articles included in the corpus was 3,814 with a total of 2,324,914 word tokens.

132 

5.1

R. Scholz

 xploring and Discovering with Exhaustive E Statistical Methods

As outlined in Section 3, exhaustive methods take into account all word tokens of a given corpus at once and contrast them alongside a corpus partition that the researcher introduces to the corpus-based on the origin of these texts. The advantage of exhaustive methods when aiming for a quantifying heuristic is that they allow us to reorganise the textual data based on statistical algorithms and not on our interpretation. Such an inductive, corpus-driven (Tognini-Bonelli 2001) approach to the data develops hypotheses about the semantic structure of a particular discourse on the basis of the data and not based on presumptions about the discourse. On the basis of these methods, I will give some examples, of how to retrieve macro-structures from a corpus representing a particular political discourse.

5.1.1  Correspondence Analysis One powerful method to discover similarities in language use between different speakers and time periods is correspondence analysis. This method takes into account the complete vocabulary of a given corpus and compares it according to the partitions which have been introduced. Applying correspondence analysis to text corpora is a way of visualising differences and similarities in language use by projecting words into at least a two-­ dimensional visualisation (Bécue-Bertaut 2014; Benzécri 1980; Husson and Josse 2014; Lebart and Saporta 2014; Salem 1982). There is a range of lexicometric software solutions that offer correspondence analysis, for example Lexico3, Hyperbase,  TextObserver and Iramuteq.  The method is based on a matrix table containing the frequency distributions of each word (rows) in different corpus parts (names of columns). The first column of this table contains all words (types) of the corpus. All other columns are named after the part of the corpus they refer to. The rows of one  column contain the frequencies  of all  word tokens in a particular part of the corpus (for example all token frequencies of the words used by a particular author across all his texts contained in the corpus). Put simply, the method can be described as an algorithm

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

133

grouping together those words with similar distributions across different columns into a profile. By distribution profiles we mean a similar high or low distribution of a group of same tokens across columns. In the visual representation distribution profiles are represented as clusters. Based on these distribution profiles the algorithm calculates the distance between the lexis of each column. As a result, the name of each column can be located in a two-dimensional diagram. Column names with similar distribution profiles are placed close to one another, whereas column names with very different distribution profiles are placed distant from one another. Correspondence Analysis produces more than two axes. The axes are situated alongside the highest concentration of similar characteristics across distribution profiles. The most dominant features of the distribution profiles are represented on the first two axes—less dominant features on the remaining axes. The significance of the difference between the characteristics of distribution profiles decreases with the growing number of axes. We only concentrate here on the first two axes. Deciphering the meaning of the axes is part of the process of interpretation which often needs further analytical steps with different methods. To interpret a visual, we need to get an idea what extreme positions in this visualisation stand for. The word tokens whose distribution profiles provide the basis for this representation are not represented in Fig. 5.2 as the abundance of overlaps would interfere with the readability of most of the represented words. Therefore, only column names are shown. In the current examples they refer to particular months (Fig.  5.1) or names of interviewees (Fig. 5.2).

Lexical Proximity over Time In Fig. 5.1, we have excluded the texts of September 2008 from the calculation as they were extremely distant, which had accumulating effects on all other data points—when excluding the extreme values we get a more detailed picture of the rest of the data. The extreme position of the September 2008 gives us a hint that the vocabulary to describe the financial crisis during the first month of the research period is

134 

R. Scholz

2009/02 2009/03

2008/10 2008/09

2009/04

2009/01 2008/11 2008/12

Fig. 5.1  Correspondence analysis of the German press corpus on the financial crisis 2008 in the partition ‘month’ (Representation of column names only)

Pofalla/CDU Müntefering/SPD

Glos/CSU

Seehofer/CSU

Kauder/CDU

Merkel/CDU

Steinmeier/SPD Leyen/CDU

Schäuble/CDU

Enzensberger

Köhler/B. -Präsi.

Steinbrück/SPD Lagarde/Frankreich Lundgren/ Schmidt H./SPD Juncker/EU Prof. Schweden Weischenberg Merkel/Sarkozy

Barroso/EU

Fig. 5.2  Correspondence analysis of the sub-corpus German press interviews on the financial crisis 2008

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

135

substantially different from the rest of the research period. To understand this extreme position better, we would need to use other methods to find out which words in particular are overrepresented during this month in comparison to the rest period—an analysis which is described further below in this text. We also see that the last three months of the research period are placed together within the same quadrant but in inverse order—February 2009 is the most distant data point whereas April 2009 is closest to the origin of coordinates. The inverse order is a hint that something unusual must have happened in the development during these three months—a phenomenon which, once again, we would have to analyse in more detail with other methods. The fact that the last three months are situated in the same quadrant shows that the vocabulary did not develop as dynamic as during the first months of the research period.

Lexical Proximity between Discourse Participants Another way of using this method is to contrast the language use of different speakers in order to find out which speakers use the most similar vocabulary. Figure 5.2 represents the similarities and differences in the vocabulary of discourse participants that have been interviewed concerning the financial crisis 2008 in the German press throughout the research period. For this analysis, we have created a sub-corpus of the above press corpus on the financial crisis compiled from 28 interviews of 19 interviewees and a report from the G20 summit in Washington (15–16 November 2008) written by the German Chancellor Merkel and the French President Sarkozy. The sub-corpus was compiled with the intention of getting to know better the positions of individuals that were chosen by the journalists to be experts in this discourse. The visualisation shows that the positions of these ‘experts’ cannot be ordered according to their political party affiliation. With the rather intellectual figures Enzensberger, Köhler, Schmidt, and Weischenberg on the right and numerous politicians on the left, the x-axis seems to represent a continuum between ­societal aspects and party politics. Accordingly, the y-axis seems to repre-

136 

R. Scholz

sent a continuum between international/European (at the bottom) and national politics (at the top). In this sense the correspondence analysis can help to find out which social dimensions dominate the lexis in a given corpus.

5.1.2  Descending Hierarchical Classification Another exhaustive method is the descending hierarchical classification (DHC) (Roux 1985), which was further developed by Reinert (1983) in the ALCESTE2 software (also available in the open source software Iramuteq). The method creates thematic classes from statistically significant co-occurrences of word tokens by the following procedure: (A) Based on punctuation signs and word counts, the corpus is partitioned into contextual units. (B) The algorithm measures which words are more likely to co-occur together in one contextual unit. These words are put into the same class of words. (C) Starting with two classes, the algorithm creates up to 10 subclasses by dividing the word classes of a superior class. Each class contains word forms that have a high probability of cooccurring in the same contextual unit. Therefore, words of the same class are considered as belonging to the same ‘semantic world’. The more classes are created, the more they  become consistent. The algorithm stops once a class is stabilised when no statistic significant evidence can be found to create a new subclass. Assigning a word class to a certain topic is a result of the researcher’s interpretation and therefore sometimes could be challenged—especially in cases when a class is constituted of words that cannot be assigned to only one but multiple topics. Figure 5.3 represents six classes that have been formed with the DHC method. The percentages refer to the total of contextual units being classified in a corpus—the two classes referring to the largest part of the corpus being the classes 1 and 5. The most overrepresented words are represented with an increased font size at the top of the list. As words co-­ occur with a variety of other words, a word can occur in more than one  ALCESTE stands for ‘Analyse des Lexèmes Cooccurrents dans un Ensemble de Segments de Texte’, which means analysis of co-occurring lexemes in a totality of text segments. 2

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

137

Fig. 5.3  DHC in the German press corpus on the financial crisis 2008 (Analysed with Iramuteq)

class (e.g. nicht).3 Whereas class 1 contains words that seem to refer to a rather technocratic discourse describing the macroeconomic context of the financial crisis, class 5 contains words which are above all names of banks that were involved or affected by the financial crisis. Class 3 contains words referring above all to effects in the real economy, and class 2 to the social market economy as part of a discussion on the crisis of the predominant economic system criticised from a  viewpoint of political economy. Similarly, class 4 refers to Marx and his analysis of capitalism which seems to refer in a particular way to various social contexts - Kinder (children), Mann (man), Frau(en) (woman/women). Contrary to this discussion of the political system on a more abstract level, class 6 contains words referring to the politics on the national, supra-, and transnational sphere. Without going into more detail, we can see the potential strength  For French and English corpora the software uses more sophisticated dictionaries which exclude functional words for this type of analysis. 3

138 

R. Scholz

of this method to reveal the different semantic worlds and parts of the narrative elements prevailing in a corpus. In this sense it allows us to make assumptions about how semantic units are put into context and subsequently knowledge is constructed.

5.1.3  Semantic Clusters Based on Reciprocal Co-occurrences Another method to analyse corpus data exhaustively is to cluster all reciprocal co-occurrences (Martinez 2011, 2012). Reciprocal co-occurrence is when two terms within a determined space of the text (e.g. a sentence) show the same level of ‘attraction’ for each other. Therefore, we measure not only if the term A is overrepresented in sentences containing the term B (simple co-occurrence) but also if the term B is overrepresented to the same extent in all sentences containing the term A.4  For example, in a given corpus we might find that all sentences stating the word house contain the article the, and the content word insurance. However, if we look into all sentences stating the, we would find that house is not used to the same extent. This would be a case of a simple co-occurrence of house and the. Contrary to this, if looking into all sentences stating the word insurance we might find that house is used to a similar extent as insurance in sentences containing house. As both words co-occur to the same extent in the same delimited space we consider them as reciprocal co-occurrences.5 If different words co-occur to a similar extent, we infer that they are part of the same semantic field. By applying this method to all words of a given corpus we are able to identify groups of words belonging to the same semantic field which can be represented as semantic networks. By manipulating the parameters, we can increase the size of a network (number of words within one cluster) and decrease the total number of networks or decrease the size of the networks and increase their total number. An advantage of this method is that depending on our research question we can narrow down the textual data to a number of keywords that then can be used to identify topic specific text sequences.  The term ‘co-occurrence’ refers to an instance of an above-chance frequency of occurrence of two terms (probability distribution). Instances of a systematic co-occurrence taking into account word order or syntactic relations would be referred to with the term ‘collocation’ in this terminology. 5  Mutual Information score addresses the same issue but with a different algorithm. 4

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

139

Fig. 5.4  The dominating semantic field in the German press corpus on the financial crisis 2008

Figure 5.4 represents the dominant semantic field in the German press corpus on the financial crisis 2008. Reciprocal co-occurrences can be analysed with the software CooCs. We have chosen parameters that produce a maximum of reciprocal co-occurrences in one network when at the same time aiming for a readable representation. Figure 5.4 represents all tokens for which the probability to co-occur in a paragraph containing the node bank is very high. Forming a symmetric matrix of the lexis of a given corpus, it was measured to what extent each type of the corpus is overrepresented in a paragraph containing another type of this corpus—if for instance the token Lehman was overrepresented in paragraphs containing the token bank. To obtain a reciprocal co-occurrence bank is ought to be overrepresented to a similar extent in paragraphs containing Lehman. In Fig. 5.4 one can distinguish a number of different semantic fields enabling us to map the discourse of the financial crisis in the German press. In the centre of the visualisation, we find the token bank. One can see that, for instance, words referring to German politics are interconnected with the rest of the network via particular lexical elements: Merkel is connected to tokens referring to the inter- and transnational political

140 

R. Scholz

field such as Sarkozy, Barroso, EU, G[7], IWF, Weltbank (world bank) and others. Steinbrück is linked to Hypo Real Estate and Finanzminister to tokens referring to US American political actors such as Paulson, Bush, Obama, Kongress (Congress). Based on this map we could go into more depth by analysing in more detail the textual context in which these dominant lexical elements are used in the corpus.

5.2

Typical Language Use

Once we have explored our data with exhaustive methods we usually start deepening our analysis with additional methods.  Keyword analysis is helpful to find out more about the language use that is typical for a particular speaker or a particular time period. This can be done with an algorithm calculating which word tokens are overrepresented in a particular part of the corpus when compared to all the other parts (see ‘partition’ in Sect. 3. and 4.). Based on the total number of tokens in the corpus, the total number of token in one part, and the frequency of each token in the whole corpus the algorithm calculates the expected frequency of each token in the part investigated. If the observed frequency in this part is higher than the expected frequency then this token is overrepresented and can be considered as belonging to the typical vocabulary of this part of the corpus. In the software Lexico3, which we used to run most of the analyses presented in this text, the algorithm is the same as the one which is used for the calculation of the co-occurrences (see also Sect. 5.1.2).

5.2.1  Uncovering Discourse Dynamics To find out about lexicosemantic dynamics in a discourse, we can calculate the typical language use in subsequent time periods. Figure 5.5 illustrates the dynamics in the above-mentioned corpus on the financial crisis 2008. On the basis of the list of the vocabulary overrepresented in each month, we have created groups of words referring to the same content category. For instance, all words referring to a bank of issue have been added to the group of words named accordingly. This category had been created because such words had occurred recurrently in the list of the overrepresented vocabulary for September 2008. In this way we have

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

141

55 45 35 25 15 5 -5 -15

Sep 08

Oct 08

Nov 08

Dec 08

Jan 09

Feb 09

Mar 09

Apr 09

-25 -35 -45 -55 financial crisis

economic crisis

countries

banks

banks of issue

ministry of finance

German government

global actors

financial products

government measures

Fig. 5.5  Over- and under-represented groups of words referring to discourse participants and discourse objects (partition ‘month’)

created the categories: countries, banks, ministry of finance, German government, global actors, government measures, and financial products. financial crisis and economic crisis refer to the word forms as such and not to the name of a group of words. Figure 5.5 shows that in the first month of the research period words referring to country and bank names as well as to banks of issue and ministries of finance are strongly overrepresented. In October 2008 we see that the word financial crisis is predominantly used in the press. If we interpret September and October together we can hypothesise that before the term financial crisis can be used as a common reference in a discourse, it first needs to be constructed with regard to an institutional (banks), geographical (countries) and political (banks of issue, ministry of finance) location. We also can observe that the polity level increases with the ­duration of the crisis. Whereas in the beginning the ministries of finance are represented in relation to the financial crisis, it is later the German government and the global actors that are represented as being in charge of

142 

R. Scholz

solving the crisis. Furthermore, throughout the duration of the crisis the term financial crisis loses importance in favour of the term economic crisis. Moreover interestingly, the government measures against the crisis— overrepresented in January 2009—seem not to be discussed together with the origin of the crisis, the financial products, which are overrepresented in November, March, and April—always together with the global actors—such as the International Monetary Fund (IMF). Figure 5.6 summarises the interpretation of Fig.  5.5: The discourse representing the financial crisis in the German press refers on the one hand to the national political sphere and on the other hand to the international political sphere. Both spheres are structured differently in terms of discourse participants and their represented actions. Whereas on the international level we can find the international political actors deliberating about the origins and the actors responsible for the crisis, on the national level we can observe political action against the effects but not against the origins of the crisis. In this sense, crisis politics seems to be divided into national political action against crisis effects, which are

Fig. 5.6  Summary of the macrostructure of the financial crisis press corpus

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

143

treated separately from their origins on the international level where political action is missing. Of course the value of such an analysis is first and foremost heuristic and needs to be substantiated with other discourse analytical methods. However, this example shows the potential that these methods have for the exploration of large text corpora and the discovery of discourse phenomena regarding their lexicosemantic macro-structures that could not be uncovered systematically by simply reading and interpreting the texts of a corpus.

5.2.2  Reducing the Amount of Textual Data In the last section of this chapter, we will illustrate briefly how we can use quantitative methods of corpus analysis in order to reduce systematically the amount of textual data that can subsequently be analysed with qualitative methods. Based on the above mentioned sub-corpus of press interviews, we have calculated the specific vocabulary for each interviewee. With the help of the resulting lists of keywords we were able to identify prototypical text sequences for each interviewee.

Fig. 5.7  Map of text sections of interviewees displaying prototypical sentences of Angela Merkel’s interviews

144 

R. Scholz

Figure 5.7 is a screenshot of the corpus map function in Lexico3. On the  left-hand side, we see the list of keywords  from interviews of the German Chancellor Angela Merkel when compared to the vocabulary of all other interviewees. On the right-hand side, we see a map of the corpus. Each box represents one sentence of the corpus. The map is divided according to the partition interviewee (interloc). Most boxes are shaded in blue. These boxes contain words that are overrepresented in Merkel’s interviews (list in the left). The darker the shade, the more of these words are contained in one sentence. We can see that two boxes in the Merkel part are highlighted in dark blue. As these are the two sentences that contain the highest number of keywords, we consider these sentences as prototypical sentences for Merkel’s interviews. In order to enable the reader to make more sense of the identified prototypical sentences, I have added the enclosing sentences into the sequence: Merkel: There have been a number of years in which the fund has barely had its classic role to play—supporting countries that have experienced serious economic and financial difficulties. Therefore the savings program was decided. However, if we now assign new tasks to the IMF to monitor the stability of the financial markets, we must also equip it properly. […] With our stimulus package aiming to stabilize the economy, we immediately provide effective support for investment and consumption. We are building a bridge between businesses and citizens so that in 2009 the consequences of the global crisis will be absorbed and the economy will rise again in 2010.6 (Interview by Süddeutsche Zeitung, 14 November 2008)

The German original is kept in the footnote highlighting those words in bold which are contained in the list of the overrepresented vocabulary.  Merkel: Es gab jetzt eine ganze Reihe von Jahren, in denen der Fonds seine klassische Rolle—die Unterstützung von Ländern, die in ernste wirtschaftliche und finanzielle Schwierigkeiten geraten sind—kaum noch ausüben musste. Deshalb wurde das Sparprogramm beschlossen. Wenn wir dem IWF aber nun neue Aufgaben bei der Überwachung der Finanzmarktstabilität übertragen, müssen wir ihn auch ordentlich ausstatten. Mit unserem Paket zur Stabilisierung der Konjunktur geben wir dagegen sofort wirksame Hilfen für Investitionen und Konsum. […] Wir bauen damit Unternehmen und Bürgern eine Brücke, damit 2009 die Folgen der weltweiten Krise aufgefangen werden und es 2010 wieder aufwärts geht. 6

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

145

We could now start analysing this quotation with qualitative methods of discourse research—just to name a few: the analysis of argumentation topoi (Wengeler 2015; Žagar 2010), of cognitive frames (Scholz and Ziem 2013, 2015; Ziem 2014) and of metaphors (Kuck and Scholz 2013) which come into play in Merkel’s rhetoric. With these rhetorical features Merkel constructs a particular knowledge about the crisis and her political position to it—all aimed at convincing the reader from the rightfulness of the political steps that she and her political party proposes against the crisis. We could also take a political science perspective and pay more attention to the political level to which Merkel allocates the political responsibility for the crisis. Contrary to Merkel whose crisis discourse has a strong international orientation (proposing, for instance, a stronger role for the IMF), this international element seems to be missing from the discourse of her political rival, the then Minister of Finance Peer Steinbrück (SPD). His prototypical sentences refer to national economics: Steinbrück: The depth of the recession will not be known until afterwards. […] Spiegel: If one sees with which efforts the recession is countered abroad, one can get the impression that you are quite passive—or just stubborn. Steinbrück: I am not stubborn, I obey economic reason.7 (Interview by Spiegel, 1 December 2008)

Similar to Steinbrück the prototypical sentences of the conservative politicians Horst Seehofer (CSU) and Ronald Pofalla (CDU) refer rather to national politics. The importance of national politics has to be understood in the context of the national parliament elections (Bundestagswahl) which were held in September 2009. In the vocabulary of both p ­ oliticians the term Bundestagswahl is overrepresented next to tax reduction (Steuersenkungen) in Seehofer’s interviews and strong references to other parties of the Bundestag in Pofalla’s interviews.  Steinbrück: Wie tief die Rezession ausfällt, wird man erst hinterher genau wissen. Spiegel: Wenn man sieht, wie man sich im Ausland gegen diese Rezession stemmt, dann muss man den Eindruck bekommen, dass sie ziemlich passiv sind. Oder einfach nur stur. Steinbrück: Ich bin nicht stur, ich gehorche der ökonomischen Vernunft. 7

146 

R. Scholz

Contrary to the national framing of the crisis we find strong references to the supranational level in the interviews with the then President of the European Commission, José Manuel Barroso. However, he reemphasises the political role of the nation-state and does not take the opportunity to argue for the increased political influence of the European institutions: SZ: Do we need a European economic government, as France’s President Sarkozy calls for? Barroso: After the meeting of the Heads of State and Government on 7th November, we in Europe agree that we should better coordinate national activities but not harmonize everything. If, for example, Poland decides an economic program, this affects Germany and certainly also vice versa.8 (Interview by Süddeutsche Zeitung, 14 November 2008)

Even though these text sequences could be analysed in more detail, we want to end this section by having a look again at Fig. 5.2, in which we tried to find out about the dimensions dominating the discourse on the financial crisis in the German press interviews. With the help of the cited text sequences, our hypothesis about the meaning of the axes represented in Fig.  5.2 can be confirmed. Where the y-axis represents a continuum between the international and the national political sphere (Barroso versus Pofalla), the x-axis represents a continuum between political and sociocultural aspects of the crisis—displaying the writer and former Marxist Hans Magnus Enzensberger at the extreme right of the x-axis in whose interview ethical questions concerning morality are addressed: Spiegel: Have the bankers failed morally? Enzensberger: It goes a bit too far to hold especially bankers accountable for morality. […]

 SZ: Brauchen wir eine europäische Wirtschaftsregierung, wie sie Frankreichs Präsident Sarkozy fordert? Barroso: Nach dem Treffen der Staats—und Regierungschefs am 7. November sind wir uns in Europa einig, dass wir nationale Aktivitäten besser koordinieren, aber nicht alles vereinheitlichen müssen. Wenn etwa Polen ein Wirtschaftsprogramm beschließt, wirkt sich das auf Deutschland aus und sicher auch umgekehrt. 8

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

147

Spiegel: The financial debacle caused a profound crisis of the so-called real economy. Enzensberger: It is incomprehensible to me why the whole world is so surprised. This is a bit like in England. If it is snowing in winter, the English are quite taken aback, because entire regions sink into the snow, as if winter were not a periodically recurrent fact. Likewise every boom is followed by a crash. This is of course very uncomfortable.9 (Interview by Spiegel, 3 November 2008)

In this section, we have given an example of how to reduce the amount of textual data systematically according to the distribution of the overrepresented vocabulary of interviewees which we have used to identify prototypical text sequences in their interviews. Other criteria might be used to identify text sequences that are relevant for our research question. For instance, elsewhere  we have analysed textual sequences containing the terms growth and prosperity in order to compare metanarratives used in the petrol crisis 1973 and the financial crisis 2008 (Scholz 2016). In another study we have used topic specific words identified with the semantic cluster method (see Sect. 5.1.2) in order to locate topic specific press articles within the financial crisis corpus dealing with questions of social and regulatory policies (Kuck and Scholz 2013). The idea of this section was to illustrate that a mixed-method approach to language use can be a powerful tool, on the one hand, to guide the qualitative analysis of discourses and, on the other hand, to explain quantitative research results. In this sense a combined use of quantitative and qualitative methods can lead to a mutual fertilisation of research results.

 Spiegel: Haben die Banker moralisch versagt? Enzensberger: Es ist ein bisschen viel verlangt, dass ausgerechnet die Banker für die Moral zuständig sein sollen. […] Spiegel: Aus dem Finanzdebakel erwächst eine tiefgreifende Krise der sogenannten Realwirtschaft. Enzensberger: Es ist mir unbegreiflich, weshalb die ganze Welt davon so überrascht ist. Das ist ein bisschen wie in England. Wenn es dort im Winter schneit, dann sind die Engländer ganz verblüfft, weil ganze Regionen im Schnee versinken, so, als wäre der Winter nicht ein periodisch wiederkehrendes Faktum. Genauso folgt jedem Aufschwung ein Absturz. Das ist natürlich sehr ungemütlich. 9

148 

6

R. Scholz

Conclusion: Strengths and Limitations

In this chapter I have presented the lexicometric approach to discourse. I have emphasised the heuristic strength of this approach in helping to find out about the underlying lexicosemantic macro-structures in text corpora which are not identifiable by simply reading texts. The various quantitative methods allow the researcher to take a number of different perspectives onto the research material guided by statistical algorithms. With the help of these algorithms, the researcher is able to map a discourse according to its lexicosemantic features, to develop hypotheses about the dominant language use by different discourse participants and in different time periods (discourse dynamics). Moreover, the researcher can use these methods to discover new discourse elements that he or she might have missed by limiting the research on qualitative methods. The particular strength of the approach is that it allows a continuous movement back and forward between quantitative and qualitative methods guiding the interpretation process and enriching mutually research results gained with different methods. The multitude of perspectives on the corpus material arising from these different perspectives renders the lexicometric approach a truly heuristic quantitative apparatus. Of course an apparatus comes with all the shortcomings that rigid methods can have in research. First, the results depend on the kind of data that they are based on—changing the composition of a corpus, especially in small corpora—can change the results substantially. Therefore, prudent questioning of what claims concerning a discourse can be made based on the text corpus used is as important as experimenting with different corpus compositions in order to ensure the validity of the results. The examples used in this text might only partly illustrate the potential use  of quantitative text processing methods for a classical sociological investigation. However, these methods have been used previously to analyse questionnaires or interviews (Leimdorfer and Salem 1995). For the analysis of such data, especially the exhaustive methods (Sect. 5.1), can help get an overview of its dominant content and structure. With regard to a social constructionist approach in sociology, it should have become clear how much the methodology presented here can help us understand better the relationship between the social and textual context in language

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

149

use. This then can be used to investigate the construction of social categories such as race, gender, or class (Leimdorfer 2010). Even though the lexicometric approach has not yet been used extensively in sociological research, this chapter should help to integrate more quantitative research on language use into the social sciences.

References Bachelard, Gaston. 1962. La philosophie du non. Essai d’une philosophie du nouvel esprit scientifique. Paris: PUF. Original edition, 1940. Bécue-Bertaut, Mónica. 2014. Distributional equivalence and linguistics. In Visualization and verbalisation of data, ed. Jörg Blasius and Michael Greenacre, 149–163. London and New York: CRC. Benzécri, Jean-Paul. 1963. Course de Linguistique Mathématique. Rennes: Universitée de Rennes. ———. 1969. Statistical analysis as a tool to make patterns emerge from data. In Methodologies of pattern recognition, ed. Satosi Watanabe, 35–74. New York: Academic Press. ———. 1980. Pratique de l’analyse des données. Paris: Dunod. ———. 1982. Histoire et préhistoire de l’analyse des données. Paris: Dunod. Bonnafous, Simone, and Maurice Tournier. 1995. Analyse du discours, lexicométrie, communication et politique. Mots  – Les langages du politique 29 (117): 67–81. Busse, Dietrich, and Wolfgang Teubert. 2014. Using corpora for historical semantics. In The discourse studies reader. Main currents in theory and analysis, ed. Johannes Angermuller, Dominique Mainguenau and Ruth Wodak, 340–349. Amsterdam: John Benjamins. Original edition, 1994. Demonet, Michel, Annie Geffroy, Jean Gouazé, Pierre Lafon, Maurice Mouillaud, and Maurice Tournier. 1975. Des tracts en mai 1968. Paris: Colin. Diaz-Bone, Rainer. 2007. Die französische Epistemologie und ihre Revisionen. Zur Rekonstruktion des methodologischen Standortes der Foucaultschen Diskursanalyse. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 8 (2): Art. 24. Duchastel, Jules, and Victor Armony. 1993. Un protocole de description de discours politiques. Actes des Secondes journées internationales d’analyse statistique de données textuelles, Paris.

150 

R. Scholz

———. 1995. La catégorisation socio-sémantique. Actes des Secondes journées internationales d’analyse statistique de données textuelles, Rome. Fiala, Pierre. 1994. L’interprétation en lexicométrie. Une approche quantitative des données lexicales. Langue Française 103 (Sep.): 113–122. Fiala, Pierre, Benoît Habert, Pierre Lafon, and Carmen Pineira. 1987. Des mots aux syntagmes. Figement et variations dans la Résolution générale du congrès de la CGT de 1978. Mots – Les Langages du Politiques 14 (1): 47–87. Foucault, Michel. 2003. On the archaeology of the sciences: Response to the epistemology circle. In The essential Foucault. Selections from the essential works of Foucault 1954–1984, ed. Paul Rabinow and Nikolas Rose, 392–422. New York: New Press. Original edition, 1968. Glady, Marc, and François Leimdorfer. 2015. Usages de la lexicométrie et interprétation sociologique. Bulletin de Me´thodologie Sociologique 127: 5–27. Guiraud, Pierre. 1954. Les caractères statistiques du vocabulaire. Paris: PUF. ———. 1960. Problèmes et méthodes de la statistique linguistique. Paris: PUF. Harris, Zellig S. 1952. Discourse analysis: A sample text. Language 28 (4): 474–494. Heiden, Serge, and Pierre Lafon. 1998. Cooccurrences. La CFDT de 1973 à 1992. In Des mots en liberté, Mélanges Maurice Tournier, ed. Pierre Fiala and Pierre Lafon, 65–83. Lyon: ÉNS Éditions. Herdan, Gustav. 1964. Quantitative linguistics. London: Butterworths. ———. 1966. The advanced theory of language as choice and chance. Berlin: Springer. Husson, François, and Julie Josse. 2014. Multiple correspondence analysis. In Visualization and verbalisation of data, ed. Jörg Blasius and Michael Greenacre, 165–183. London and New York: CRC. Kleining, Gerhard. 1994. Qualitativ-heuristische Sozialforschung. Schriften zur Theorie und Praxis. Hamburg-Harvestehude: Fechner. Kuck, Kristin, and Ronny Scholz. 2013. Quantitative und qualitative Methoden der Diskursanalyse als Ansatz einer rekonstruktiven Weltpolitikforschung. Zur Analyse eines internationalen Krisendiskurses in der deutschen Presse. In Rekonstruktive Methoden der Weltpolitikforschung. Anwendungsbeispiele und Entwicklungstendenzen, ed. Ulrich Franke and Ulrich Roos, 219–270. Baden-­ Baden: Nomos. Lafon, Pierre. 1984. Dépouillements et statistiques en lexicométrie. Paris: Champion. Lebart, Ludovic, and André Salem. 1988. Analyse statistique des données textuelles. Questions ouvertes et lexicométrie. Paris: Dunod.

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

151

Lebart, Ludovic, André Salem, and Lisette Berry. 1998. Exploring textual data. Dordrecht: Kluwer. Lebart, Ludovic, and Gilbert Saporta. 2014. Historical elements of correspondence analysis and multiple correspondence analysis. In Visualization and verbalisation of data, ed. Jörg Blasius and Michael Greenacre, 31–44. London and New York: CRC. Lee, David. 2001. Genres, registers, text types, domains, and styles: Clarifying the concepts and navigating a path through the BNC jungle. Language Learning & Technology 5 (3): 37–72. Leimdorfer, François. 2010. Les sociologues et le langage. Paris: Maison des sciences de l’homme. Leimdorfer, François, and André Salem. 1995. Usages de la lexicométrie en analyse de discours. Cahiers des Sciences Humaines 31 (1): 131–143. Martinez, William. 2011. Vers une cartographie géo-lexicale. In Situ, 15. Accessed July 1, 2018. http://journals.openedition.org/insitu/590. ———. 2012. Au-delà de la cooccurrence binaire… Poly-cooccurrences et trames de cooccurrence. Corpus 11: 191–216. Mayaffre, Damon. 2005. De la lexicométrie à la logométrie. L’Astrolabe. Accessed July 1, 2018. https://hal.archives-ouvertes.fr/hal-00551921/document. ———. 2007. Analyses logométriques et rhétorique du discours. In Introduction à la recherche en SIC, ed. Stéphane Olivesi, 153–180. Grenoble: Presses Universitaires de Grenoble. ———. 2016. Quantitative linguistics and political history. In Quantitative linguistics in France, ed. Jacqueline Léon and Sylvain Loiseau, 94–119. Lüdenscheid: Ram Verlag. Mayaffre, Damon, and Céline Poudat. 2013. Quantitative approaches to political discourse. Corpus linguistics and text statistics. In Speaking of Europe. Approaches to complexity in European political discourse, ed. Kjersti Fløttum, 65–83. Amsterdam: Benjamins. Muller, Charles. 1967. Étude de statistique lexicale. Le vocabulaire du théâtre de Pierre Corneille. Translated by Pierre Corneille. Paris: Larousse. Pêcheux, Michel. 1982. Language, semantics and ideology (Language, Discourse, Society Series). London: Macmillan. Pêcheux, Michel, Claudine Haroche, Paul Henry, and Jean-Pierre Poitou. 1979. Le rapport Mansholt: un cas d’ambiguïté idéologique. Technologies, Idéologies, Pratiques 2: 1–83. Reinert, Max. 1983. Une méthode de classification descendante hiérarchique. Cahiers analyse des données VIII (2): 187–198.

152 

R. Scholz

Roux, Maurice. 1985. Algorithmes de classification. Paris: Masson. Salem, André. 1982. Analyse factorielle et lexicométrie. Mots – Les Langages du Politiques 4 (1): 147–168. ———. 1987. Pratique des segments répétés. Essai de statistique textuelle. Paris: Klincksieck. Scholz, Ronny. 2016. Towards a post-material prosperity? An analysis of legitimising narratives in German crisis discourses from 1973 and 2008. French Journal for Media Research [online] 5 (Narratives of the Crisis/Récits de crise). Accessed July 1, 2018. http://frenchjournalformediaresearch.com/index. php?id=614. Scholz, Ronny, and Johannes Angermuller. 2013. Au nom de Bologne ? Une analyse comparative des discours politiques sur les réformes universitaires en Allemagne et en France. Mots – Les Langages du Politiques 102: 22–36. Scholz, Ronny, and Pierre Fiala. 2017. Politolinguistik in Frankreich. In Handbuch Sprache und Politik, ed. Jörg Kilian, Thomas Niehr, and Martin Wengeler, 1163–1199. Bremen: Hempen. Scholz, Ronny, and Annika Mattissek. 2014. Zwischen Exzellenz und Bildungsstreik. Lexikometrie als Methodik zur Ermittlung semantischer Makrostrukturen des Hochschulreformdiskurses. In Diskursforschung. Ein interdisziplinäres Handbuch. Band 2: Methoden und Analysepraxis. Perspektiven auf Hochschulreformdiskurse, ed. Martin Nonhoff, Eva Herschinger, Johannes Angermuller, Felicitas Macgilchrist, Martin Reisigl, Juliette Wedl, Daniel Wrana, and Alexander Ziem, 86–112. Bielefeld: Transcript. Scholz, Ronny, and Alexander Ziem. 2013. Lexikometrie meets FrameNet: das Vokabular der ‘Arbeitsmarktkrise’ und der ‘Agenda 2010’ im Wandel. In Sprachliche Konstruktionen von Krisen: Interdisziplinäre Perspektiven auf ein fortwährend aktuelles Phänomen, ed. Martin Wengeler and Alexander Ziem, 155–185. Bremen: Hempen. ———. 2015. Das Vokabular im diskurshistorischen Vergleich: Skizze einer korpuslinguistischen Untersuchungsheuristik. In Diskurs  – interdisziplinär. Zugänge, Gegenstände, Perspektiven, ed. Heidrun Kämper and Ingo Warnke, 281–313. Berlin and New York: De Gruyter. Tognini-Bonelli, Elena. 2001. Corpus linguistics at work. Amsterdam: Benjamins. Tournier, Maurice. 1975. Un vocabulaire ouvrier en 1848. Essai de lexicométrie. Quatre volumes multicopiés. Saint-Cloud: École Normale Supérieure. ———. 1993. Lexicometria – Séminaire de lexicométrie. Lisbonne: Universidade Aberta. Original edition, 1988. Wengeler, Martin. 2015. Patterns of argumentation and the heterogeneity of social knowledge. Journal of Language and Politics 14 (5): 689–711.

  Lexicometry: A Quantifying Heuristic for Social Scientists… 

153

Yule, George Udny. 1944. The statistical study of literary vocabulary. Cambridge: Cambridge University Press. Žagar, Igor Ž. 2010. Topoi in critical discourse analysis. Lodz Papers in Pragmatics 6 (1): 3–27. Ziem, Alexander. 2014. Frames of understanding in text and discourse. Theoretical foundations and descriptive applications. Amsterdam: Benjamins. Zipf, George K. 1929. Relative frequency as a determinant of phonetic change. Harvard Studies in Classical Philology 40: 1–95. ———. 1935. The psycho-biology of language. An introduction to dynamic ­philology. Boston: Mifflin.

6 Words and Facts: Textual Analysis— Topic-Centred Methods for Social Scientists Karl M. van Meter

1

Introduction

In arguing for systematic textual analysis as a part of discourse analysis, Norman Fairclough stated that ‘[t]he nature of texts and textual analysis should surely be one significant cluster of issues of common concern’ within discourse analysis (Fairclough 1992, 196). As a contribution discourse analysis and a reinforcement of the close association between discourse analysis and textual analysis, I will here deal with texts from several different origins and over a time span extending from the 1980s to now. I will try to explain and show how complex statistical methods such as factorial correspondence analysis (see the TriDeux software of Cibois 2016), both descending hierarchical classification analyses (see the Alceste software (Image 2016) and the Topics software (Jenny 1997)) and ascending hierarchical classification analyses (such as Leximappe-Lexinet (Callon et al. 1991) and Calliope (De Saint Léger 1997)), can be used to K. M. van Meter (*) Centre Maurice Halbwachs, École Normale Supérieure—Paris, Paris, France e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_6

155

156 

K. M. van Meter

produce detailed representations of lexico-semantic structures in large text collections or corpora. As you can see from the dates, these methods and their applications to textual analysis came well before the new term ‘big data’ became popular among scientists and the general public. Interestingly enough, the development of these methods and their initial applications in social science research took place primarily in Paris and were centred around such figures as Jean-Paul Benzécri and Pierre Bourdieu, and their work in the 1970s and 1980s. Indeed, most of the authors cited here and I have worked together in textual analysis and have been associated with Benzécri, Bourdieu, and Callon at the Université Pairs VI (now Université Pierre et Marie Curie), the Maison des Sciences de l’Homme, and the École des Mines de Paris. Here we present three studies that we hope will help today’s social scientists in understanding which concepts and topics dominate a particular discourse in society and the relationship between these same diverse topics both diachronically and synchronically. Furthermore, we will show how semantic and thematic shifts in a given society can be traced over time and which future developments might be more or less probable. These methods—used individually or combined in what is now called multi-method analysis (Van Meter 2003)—can produce unintended and sometimes amazing results (Glady and Leimdorfer 2015). I’ll look at three different cases: 1. In the first case, I will look at how statistical analysis of texts from 1989 produces a geographical map of the then Soviet Union and how those texts and this map help to define the political and economic structures of current-day Russia and its ‘near abroad’. 2. In the second case, I’ll take a synchronic perspective on American, French, and German sociologies by looking at abstracts submitted to the annual conferences of each national association of sociology, and also a diachronic perspective in the specific case of French sociology. The first analysis shows how the discursive field of sociology in each country is structured in terms of topics dealt with, of relationships or ties between these topics, and of the history and culture of each country. In a second analysis, I will provide a diachronic perspective by

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

157

looking at the thematic dynamics over time in French academic ­discourse on sociology from 2004 to 2009 (Van Meter 2009). Such an analysis can trace the evolution of a topic from past to present. It even allows possible projections into the future. 3. With the third case, I try to contribute to the discussion of what poet, academic, and diplomat Pete Dale Scott calls ‘Parapolitics’, or emphasizing the real use of power. In his own words: The investigation of parapolitics, which I defined (with the CIA in mind) as a “system or practice of politics in which accountability is consciously diminished.” …I still see value in this definition and mode of analysis. But parapolitics as thus defined is itself too narrowly conscious and intentional … it describes at best only an intervening layer of the irrationality under our political culture’s rational surface. Thus I now refer to parapolitics as only one manifestation of deep politics, all those political practices and arrangements, deliberate or not, which are usually repressed rather than acknowledged. (Scott 1993, 6–7)

‘Parapolitics’ or deep politics are revealed by looking at the largest possible portion of the international press, by analysing how topics emerge in these texts in certain socio-historical and institutional contexts, and by how them are then reproduced in texts over time and in other contexts. In particular, I shall be looking at semantic structures and relationships referring to international conflicts. For instance, in 2009 Obama promised to close Guantanamo, which is still open today, and our ‘parapolitical’ analysis of the international press shows that the closure of Guantanamo was only an Obama Administration priority during 2009 and not afterward. And during the same period of time in the international press, the dominating image of the then new US president was overshadowed by the representation of the previous president, George W. Bush, and his ‘legacy’—even today the term ‘Bush’ remains closely associated with ‘torture’ and ‘Iraq’ in the international media. In a diachronic perspective, such a study can explain how certain forces are put into motion, how they persist over time and how they reach into the future.

158 

2

K. M. van Meter

 extual Geography of the Last Soviet T Central Committee and Russia Today

Following the development of a science, of the adoption of an invention, of the spread of an epidemic, of the policies of a political party, of the evolution of an international conflict, all this can be done—and has been done—through the study of the texts, the words, and their associations that these processes have left behind. And often with such massive data, ‘following’ such developments is so solidly established that such analysis can open a window on the future: in the international media, as mentioned above, the words ‘legacy’ and ‘George Bush’ have continued for several years now to be closely associated with ‘torture’, ‘Guantanamo’ and ‘Iraq’, and will in all probability continue to be in the future. The geographical words in 1989 in the official biographies of the last Soviet Encyclopaedia drew a nice map of what is today Russia and its ‘near abroad” neighbours. Keywords in American sociology of religion— although usually not geographical—remain very different from those of German and French sociologies of religion and that will continue long into the future. The analysis of keywords and geographical names can also show that important scientific research ‘moves around’ geographically as the names of the authors and their institutional affiliations change (Maisonobe 2015). In the case of the Soviet Union/Russia, together with Philippe Cibois, Lise Mounier and Jacques Jenny (Van Meter et al. 1989), we used the online computer data base, SOVT on the server GECAM in Paris, which at that time provided detailed official biographies of leading figures of the Soviet Union. Since 1982 and the death of Secretary-General Leonid Brezhnev, there were three successive secretary-generals, the last being Mikhail Gorbachev. We examined the biographies of all members of the Central Committee of the Communist Party of the Soviet Union during the period 1981–1987. In that data, there were 503 individuals, including 18 women, and a total of 7844 dated sequences for any average of 15.6 dated sequences per individual. The data included names, demographic information, dates, and descriptive texts. We began our study with the analysis of the corpus of all words used to describe the ­individuals of our population, but quickly settled on using only 100 geographical

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

159

terms, and even then, we used only the 20 most frequent names as active variables with the 80 other geographical names being used only as non-­ active or passive elements that did not enter into the calculations. We used the TriDeux factorial correspondence analysis program (Cibois 1983, 1985, 2016) inspired by Benzécri, which is a hierarchically descending method that successively ‘cuts’ the set of data points along the most statistically significant axes or dimensions, thus producing the following twodimensional diagram based on the two most pertinent factors (Fig. 6.1).

Fig. 6.1  Factorial correspondence analysis of geographical terms in the official biographies of members of the last Central Committees of the USSR

160 

K. M. van Meter

Here we see that Stavropol (STAV), Latvia (LETT), and Moscow (MOSC) are very largely responsible for the positive side of horizontal factor or axis 1. Ukraine (UKRA), Dnepropetrovsk (DNEP), Donetsk (DONE), Kiev (KIEV), and Kharkov (KHAR) are largely responsible for the negative side of axis 1. This is of course statistically the most important factor or axis of the entire population. The second most important factor or axis opposes, on the positive side, Tashkent (TACH), Azerbaijan (AZER) and Gorky (GORK, now Nizhny Novgorod) with, on the negative side, Latvia (LETT). In 1989, with this simple textual analysis of geographical terms, we stated: Summarily interpreted, this analysis reveals that there is a very tight and coherent network centred around Dnepropetrovsk and the Ukraine [and other geographical names that figure prominently in Brezhnevian allies’ biographies], which corresponds with the political power base of Leonid Brezhnev. The most significant cleavage in the entire population is between this Brezhnevian group and the rest of the population. This ‘rest of the population’ is in turn multi-centred with Stavropol (Mikhail Gorbachev’s political power base), Latvia (an economically dynamic Baltic republic) and Moscow (the centre of the state apparatus), distributing the remainder of the names in an arc around themselves. The second most important cleavage (the second axis) seems to be related to economic development with Latvia at one extreme and Uzbekistan and Azerbaijan [in Central Asia] at the other. This is, however, a tentative hypothesis that must be examined further.

Indeed, this hypothesis ‘examined further’ is confirmed with Latvia contributing to NATO operations as of 1994 before becoming a member in 29 March 2004, and Ukraine becoming an independent country and rival to Russia, before losing its eastern portion to Russia recently. And soon after the publication of the above cited article, we learned that the last Central Committee of Uzbekistan died before a firing squad, executed for corruption and incompetence for storing outdoors the entire Uzbek cotton crop that had rotted away while being sold and resold on the international market, thus confirming the ‘cleavage’ due to ‘economic development’ mentioned above. Have the situation and the cleavages

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

161

between Russia and Ukraine (the first axis of the diagram) and between the Baltic republic and the Central Asian republics (the second axis) changed that much since our textual analysis produced the above diagram?

3

Ascending and Descending Methodologies

We referred to factorial correspondence analysis as a hierarchical descending method by which we mean that you start with all the data together in the form an n-dimensional cloud of data points, where n corresponds to the number of variables being used in the description of each data point. In the case of the Soviet biographies, it was 20 geographical terms or variables. The cloud is then cut successively—thus in a ‘descending’ manner—according to specific statistical criteria. In the case of the Soviet biographies, the successive ‘cuts’ where on the basis of statistical contributions to the first factor or axis (Ukraine/Rest of the USSR), then to the second axis (Baltic republics/Asian republics or economically developed/ economically underdeveloped), and so on to the third, fourth and other axes with less and less statistical significance. This distinction between ascending and descending methods is even clearer when it concerns classifications (Van Meter 1990, 2003). One either builds classes or ‘clusters’ by putting together data points with similar characteristics or values for variables, or by ‘cutting’ the entire cloud of data points to successively form the most homogeneous classes possible. Alceste mentioned above, and first developed by Max Reinert (1987), is such a descending method specifically intended for the analysis of corpora of texts and has been used extensively in the analysis of very different sorts of textual data and continues to be extensively used today (Reinert 2003; Image 2016). In the other direction, hierarchically ascending classifications try to put in the same class or ‘cluster’ elements that have the most similar characteristics (Van Meter et  al. 1987, 1992). When the data are units of texts, one of the more useful manners of constructing classes is to put together words or terms that appear the most often together in the units

162 

K. M. van Meter

of text in the data or corpus. We usually refer to this method as classification by co-occurrence of keywords. Of course, one can combine these different types of analysis in a multi-method approach. An example of the use of a hierarchically descending factorial method (using the TriDeux software), a hierarchically descending classification method (using the Alceste software), and the hierarchically ascending classification method (using the Calliope software), all applied to the same data, can be found in Van Meter (2008). Classification by co-occurrence of keywords has proven to be a powerful textual analysis tools in several disciplines. One example is the Leximappe method first developed at the Paris Ecole des Mines by Michel Callon et al. (1991) as a DOS program with a graphic extension called Lexinet. Mathilde de Saint Léger (1997) developed a new Windows-­ based version, Calliope, based on the Leximappe-Lexinet DOS package. We first used the latter system in 1992 in an analysis of the initial ten years of scientific articles published on sociological AIDS research (Van Meter and Turner 1992), describing the S-curve typically found in tracking the development of scientific publications in any new specialty. And again in 1995, Leximappe-Lexinet was used to analyse the same data in detail for thematic clusters and to reveal the role of national policies in sociological AIDS research in the US, Great Britain, and France (Van Meter et al. 1995). Then in 1997, we were able to show the role and trajectory of individual authors, and the evolution of publishing policy of individual scientific journals (Van Meter and Turner 1997), thus showing that this textual analysis tool used several time on the same data base could produce new and often unanticipated results each time.

4

F rench, German, and American Sociologies as Seen Through Congress Presentations

In a comparative study on research agendas in French, German, and American sociologies, we analysed corpora comprising all abstracts of the national sociology association congresses in each of the three countries. For France, we analysed abstracts from the congresses of the French

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

163

Sociological Association or AFS (Association française de sociologie) following its created in 2002. The association 2002 replaced the moribund Société française de sociologie, a scholarly society created by Émile Durkheim himself. At the first 2004 AFS congress, along with several other researchers, we decided to analyse all the abstracts using different methods of textual analysis. This produced a thematic issue of the BMS (Bulletin of Methodological Sociology/Bulletin de Méthodologie Sociologique), the number 85 issue in January 2005. In our contribution to that thematic issue (De Saint Léger and van Meter 2005), we used the new Windows version of Leximappe-Lexinet, Calliope, for the first time. That issue of the BMS also became a book in French, Analyses textuelles en sociologie—Logiciels, méthodes, usages (Textual Analyses in Sociology— Computer Programs, Methods, Practices; Demazière et al. 2006). What did we find? Nothing simpler than showing you the Strategic Diagram of AFS 2004. Some explanation would be helpful in understanding the meaning of the strategic diagram that is formed by a first horizontal axis of ‘centrality’ and a second vertical axis of ‘density’. Here, the classes or clusters are constructed using co-occurrence of keywords. Thus, keywords that appear often together in the different units of text—here, in abstracts of scientific presentations made at a conference—will have a high index of similarity and will be placed together in the same class. The classes thus constructed are then used as the basis for the calculation of ‘in-ties’ (ties between keywords within the same class) and ‘out-ties’ (ties between keywords outside of their class and with keywords in other classes). The ‘in-­ ties’ are used to calculate an axis of ‘density’ (the vertical or second axis) and the ‘out-ties’ to calculate an axis of ‘centrality’ (the horizontal or first axis). With density as a vertical axis, and centrality as a horizontal axis, a two-dimensional strategic diagram is constructed in such a manner that it is possible to place the different classes according their mean density and mean centrality. The strategic diagram, by construction, therefore has a dense and central upper right first quadrant, which we call the ‘mainstream’. The dense but not central upper left second quadrant is called ‘ivory tower’. The not dense and not central lower left third quadrant is sometimes called ‘chaos’ or simply ‘unstructured’. The central but not dense lower right fourth quadrant is called ‘bandwagon’ (Callon et al.

164 

K. M. van Meter

1991). This means that ‘mainstream’ classes (in the first quadrant) will have relatively numerous ‘in-ties’ (co-occurrences of the class’ keywords together in the text units) and relatively numerous ‘out-ties’ (co-­ occurrences in text units of the class’ keywords with other keywords from other classes). ‘Bandwagon’ classes have keywords with numerous ‘out-­ ties’ but relatively fewer ‘in-ties’. And, of course, ‘ivory tower’ classes have numerous ‘in-ties’ but relatively few ‘out-ties’. In Fig. 6.2, one can clearly see the dominant role of the term ‘femme’ (and thus sociology of women), and in Fig. 6.3, you can see the internal structure of the class ‘femme’ and its constituent keywords (including ‘travail’) and the cluster’s ‘in-ties’. The statistical weight of the term ‘femme’ was so dominant that we decided to see what the data base would look like without ‘femme’. So we removed that term and redid the analysis, producing a new strategic diagram (Fig. 6.4). The astonishing similarity between Figs.  6.2 and 6.4 leads to the conclusion that the terms

Fig. 6.2  French Sociological Association (AFS) congress 2004 strategic diagram of all abstracts (femme)

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

165

‘femme’ and ‘travail’ are interchangeable in the structure of the 2004 AFS corpus and, by implication, the sociology of women and the sociology of work (a classic and historically dominant theme of French sociology) have become contemporary equivalents (De Saint Léger and van Meter 2005). This result was then confirmed by the analysis of the 2006 AFS congress and the 2009 congress (De Saint Léger and van Meter 2009). In the case of the 2009 congress, there was a declared central theme: ‘violence in society’. Therefore, in the 2009 Strategic Diagram, the dominant position of ‘femme’/‘travail’ had been taken over by the term ‘Violence’. But by employing the same technique of deleting the dominant term and re-analysing the data, we produced a new 2009 Strategic Diagram that was indeed very similar to the 2004 and 2006 results, and

Fig. 6.3  French Sociological Association (AFS) congress 2004 ‘femme’  (woman) cluster with its keywords, including ‘travail’ (work)

Fig. 6.4  French Sociological Association (AFS) congress 2004 strategic diagram of all abstracts (without femme)

166  K. M. van Meter

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

167

thus confirming the most or less established structure of contemporary French sociology. A similar Calliope analysis was done of the 2004 congress of the Deutsche Gesellschaft für Soziologie (DGS), or German Sociological Association, held in Munich. The entire set of 401 conference abstracts was available after the congress free of charge on the DGS Web site and the results permitted a direct comparison of the structure of contemporary French and Germany sociologies and stressed their similarity (Van Meter and de Saint Léger 2009a). This was followed by the Calliope analysis of 2003, 2004, 2005, and 2006 American Sociological Association (ASA) conference abstracts of more than 8000 presentations (Van Meter and de Saint Léger 2014). The comparison of the ASA results with the AFS and DGS results was particularly surprising concerning the sociology of religion and the relationship between family, education, and work, which is quite distinct in each of the three cases and clearly related to the history and culture of these three nations.

5

L ongitudinal Analysis of World Media on Conflicts—Parapolitics 2006–2012

There has been a great deal of exchange over the past few years between discourse analysis, sociology, and political science. This has even resulted in joint sessions in international conferences and joint publications in which the BMS has played a certain role (Marchand 2005, 2007a). Indeed, Marchand’s work in political science resulted in the publication of a book that cites his contribution to discourse analysis, sociological methodology and to the BMS, in particular (Marchand 2007b). But it is Michel Pinault’s work on the politics and history of French scientific research (Pinault 2006) that cited not only sociological methodology developments and the BMS but used some the same methods we have presented here. In this particular area of politics, Pinault utilized Calliope to analyse the texts and discourses of the major social and political actors, and then applied social network analysis to examine the roles and relationships between those actors in the development of a coherent national

168 

K. M. van Meter

policy for French scientific research, another surprising and well-done multi-method analysis of a complex subject. If political science studies politics and the use of political power as it presents itself on the public scene, ‘parapolitics’ is the study of the use of political power that is not presented on the public scene; in short, ‘the real use of political power’, as we have described it above. The term seems to have been coined by Prof. Scott at the University of California, Berkeley, in the late 1970s or early 1980s. We worked with Scott and with one of his students, Jonathan Marshall, who founded the journal Parapolitics USA, in the 1980s. Other colleagues of Scott founded the French ‘Association for the Right to Information’ (Association pour le Droit à l’Information— ADI) and the journal, Parapolitics, during the same period. The ADI continues to publish books and reviews, although the journal Parapolitics has evolved and changed its title to become—since 1995—the current Internet journal, Intelligence, Secrecy & Power (ISP), which is currently published twice a month (https://intelligenceadi.wordpress.com/ and http://groups. yahoo.com/group/intelligence-secrecy-power-adi/). ADI and ISP information has served previously in the social network analysis (or ‘traffic analysis’) of the long-term cat-and-­ mouse/cops-and-robbers dynamic between law enforcement and suspected offenders (Van Meter 2001). Many different centres of research and information on political developments, particularly on the international level, produce annual reports, yearbooks or ‘tops of the year.’ Some of the better-known examples include the ‘United Nations’ Yearbook’, the Stockholm International Peace Research Institute’s (SIPRI) ‘SIPRI Yearbook’, the CIA’s annual ‘World Factbook’, the Council of Europe’s ‘European Yearbook’. Other examples include the International Institute for Strategic Studies’ ‘Strategic Survey’, the Institut des Relations Internationales et Stratégiques annual ‘L’Année stratégique’, the Swedish National Defence College’s ‘Strategic Yearbook’, the ‘Statesman’s Yearbook’, by Barry Turner, and the ‘World Political Almanac’, by Chris Cook and Whitney Walker. In most cases, these works are compilations of a selected number of reports made during a given year, but they are often accompanied by analyses or opinions attributed to experts in the domain concerned, be it international politics, security, economics, armaments, and so on. The events they tend to emphasis are what we often call ‘yearbook’ events.

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

169

Methodologically speaking, these works usually suffer from a double weakness: the limited sampling basis for the information or data used in the analysis and the reliance upon an ‘expert’ or ‘experts’ to analyse and interpret the data in a non-formalized manner, and often without contradictory evaluation or critique. The consequences of this double weakness in international politics can be seen with the Bush White House’s decision to invade Iraq because of the supposed existence of weapons of mass destruction. While it is extremely difficult—if not impossible—to develop either an exhaustive sample or even a statistically representative sample of publically available reports on international political developments for a given year, there is a tremendous amount of information available in this area as the above-mentioned reports show. Indeed, the problem has more to do with handling and managing the vast quantity of information, and selecting the pertinent and representative information for further analysis. Informal experiments have been run in this domain to show how a report by a major wire service—Associated Press, Reuters, Agence France Presse, etc.—is often ‘rewritten’ into multiple, even hundreds, of media reports which furnish little or no further information than that contained in the original wire service report. The difference often resides in the political, cultural, and ideological choices of words to communicate and to interpret the same initial information. This choice of words, this language, does furnish further information, but at another level: that of the analysis of the ‘rewriting’ and those that carry it out. One of the ADI’s objectives is to document international politics and parapolitics through the systematic wide sampling of the media. For ­several years, the ADI has published in each issue daily entries of media report headlines or titles, along with an indication of the source and eventually additional text from the original report. For example, the extract below is the first six entries of the 43 ADI entries for 31 August 2016: • Death of senior leader al-Adnani caps bad month for ISIS/CNN. The death of one of ISIS’ most prominent figures, Abu Mohammad al-­ Adnani, is one more example of the pressure the group is under in both Iraq and Syria.

170 

K. M. van Meter

• Trump to meet in Mexico with the country’s president/WP. Donald Trump will travel to Mexico City on Wednesday for a meeting with Mexican President Enrique Peña Nieto, just hours before he delivers a high-­ stakes speech in Arizona to clarify his views on immigration policy. • Monsoon Rains, Terrorism Ad Lib Snag Kerry in South Asia/NYT. US Secretary of State John Kerry’s visit to South Asia this week has been filled with serious diplomatic meetings and agreements to boost ties with India and Bangladesh but his motorcade’s struggle with monsoon downpours and an … • North Korea Has Executed a Deputy Premier, Seoul Reports/NYT. Kim Jong-un, the North Korean leader, in a picture issued by his government’s Korean Central News Agency. • Vote on Whether to Remove President Nears in Brazil’s Senate/NYT. Senators debated the fate of Brazilian President Dilma Rousseff into the wee hours of Wednesday, then planned a short break before casting votes that will decide whether to remove her permanently as leader of Latin America’s most … • At least nine dead as Typhoon Lionrock pummels Japan/CNN. Nine bodies were found in a home for the elderly in the town of Iwaizumi in Iwate Prefecture, which suffered flooding following Typhoon Lionrock, police tell CNN. These six reports come from a variety of sources: CNN, WP (Washington Post) and NYT (New York Times). In 2008, we used the entire set of 2006 ADI entries as a data set to be analysed with Calliope, but the size of the text corpus encouraged us to divide 2006 into three successive periods of four months and produce a strategic diagram for each period. Figure 6.5 is the strategic diagram for the first four months of 2006. Clearly, ‘Iran’ is the dominant term, but it was replaced by ‘Iraq’ in the two following 2006 Strategic Diagrams, as would be expected with the massive dominance of the Iraq war in 2006 political and parapolitical developments. The presence of ‘Iran’ in the ‘mainstream’ first quadrant in the first (winter to spring 2006) four-month section shows that the Bush White House intention to attack Iran and Iran’s nuclear program played a major political and parapolitical role during the first part of 2006 before declining significantly towards the end of the year, but this topic did come back as a major theme in middle and late 2007.

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

171

Fig. 6.5  Strategic diagram of the first four months of the 2006 Association for the Right to Information corpus

In the first two 2006 Strategic Diagrams, we previously noted the important position of ‘eavesdropping’ (first diagram) and ‘NSA’ (second diagram) in exactly the same position. The first diagram also includes ‘NSA spy’ and ‘surveillance’ clusters near the origin, while the second diagram has a ‘spy’ cluster near the origin. These clusters are related to the Bush White House program of warrantless eavesdropping on American citizens by the NSA. However, the third diagram included only a ‘surveillance’ cluster in the unstructured third quadrant. Nonetheless, the calculated attractive power of the term ‘spy’ (see van Meter and de Saint Léger 2014 for a description of keyword ‘attractive power’) was fairly constant over the three periods of 2006. With the Iraq war as the overall framework, ISP stated, in late February: ‘Although the Bush White House is plagued with numerous scandals and crises, the three current major ones are Iran and its nuclear program, the ‘domestic’ NSA spying scandal, and

172 

K. M. van Meter

the series of reports on US torture/abuse and calls for the closure of Guantanamo.’ This corresponds with the position of ‘Iran’ and ‘eavesdropping’ in the first diagram, and the continued permanence of the latter throughout the year when ‘Iran’ and ‘torture’ declined importance. There were three notable 2006 ‘yearbook’ events: another invasion of Lebanon by Israel (July), a North Korean nuclear weapon test (October), and a US Democrat Congressional election victory (November). These events are duly situated by the analysis as the respective clusters ‘Israel’ and ‘Lebanon’ (second diagram, 2006-2), ‘North Korea’ (second diagram), and ‘Congress’ (third diagram, 2006-3), but only as passing major events. ‘Israel’, ‘Lebanon’, and ‘Gaza’ figure only in the second diagram and not in the ‘mainstream’ first quadrant and nowhere else during the year. North Korea figures in the ‘mainstream’ first quadrant, but only in the second diagram and nowhere else. ‘Congress’ figures only in the third diagram and in the ‘ivory tower’ second quadrant. However, preparations for those important elections were widely commented on during the preceding second period and account in large part for the presence of the clusters ‘GOP’ and ‘Democrat’ in the ‘ivory tower’ second quadrant and ‘election’ in the ‘mainstream’ first quadrant. But now, in 2016, and with this formal textual analysis of the 2006 media, the very major changes these three ‘yearbook’ events supposedly announced at the time—‘the Israeli invasion will change the face of the Middle East’, ‘a North Korea atomic weapons will change world politics,’ ‘a Democrat Congress will change everything in the US’—and the choice of these three events as the three most notable 2006 ‘yearbook’ events seem rather shallow and without fundamental importance. It is only by following keywords or clusters backward and forward in time that coherent major political developments can be identified, unless of course it is a question of a clearly monumental political development. The case of ‘Hamas’ in the 2006 third period is a good example of both how a political development can be identified and followed with this analysis, and how the international media can ‘skip’ or bias the coverage of a major political development. According to our analysis of 2006, there are, however, two major developments that this data rather clearly designate: in early 2006, the

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

173

international press usually talks about ‘Hamas terrorists’ but, by the end of the year, this has become ‘Hamas militants’; China starts the year as a major player on the international scene but finishes 2006 as one of the most insignificant international actors (Van Meter and de Saint Léger 2008, 2009b). The case of ‘Hamas’ is quite interesting in its implications for text analysis of political developments. If a ‘Hamas’ cluster appears in the third diagram (2006-3), it is because the international media concentrated on the killing of an elected Hamas government official by the Israelis in November 2006. But the truly important political development associated with Hamas occurred in January 2006 when Hamas won the internationally recognized democratic elections in Gaza and the Occupied Territories, thus causing a major change in perspective that should have been widely commented in the press, but was not. Again, in July, Israel detained a large number of elected Hamas government officials, against the wishes of the larger international community. But, again, Hamas does not appear in the second diagram (2006-2), and the international media gave only passing attention to this imprisonment of elected Palestinian officials. It was only in November with the Israeli killing of one such official that the international media finally seemed to pay attention to what was happening. Indeed, since the publication of our study, several major news agencies have confirmed that at a very senior level there was an editorial decision to no longer label all Hamas members as ‘terrorists’ and instead use the term ‘militant’ or ‘activist’, or in the case of this killing, ‘Hamas official’. This major change in 2006 is still with us today and could only be identified by moving back and forth in time over the results of these textual analyses. There have been many other similar instances of such ‘uneven’ or clearly biased coverage of particularly sensitive topics or events that can be discerned by following the formation of certain clusters back in time and also forward in time. Inversely, if a topic or cluster cannot be followed over time, it is very likely a transient event or not a coherent topic. The supposedly major ‘yearbook’ event that was the election of a 2006 Democrat majority in Congress hardly resulted in any memorable political developments. Few people other than Middle Eastern specialists remember the 2006 Israeli invasion of Lebanon or how many times Israel has invaded its neighbours. And North Korea has its nuclear weapon, but

174 

K. M. van Meter

has not used it until today and international politics has been thoroughly preoccupied by other issues since then. But let us look at another less than evident 2006 development that is nonetheless fundamental and whose consequences are still with us today. During the first four-month period of 2006, the keyword ‘China’ was the keyword with the highest statistical attractive power in the construction of the clusters and axes (see Van Meter and de Saint Léger 2009b for a description of this index). By looking at the texts of the first period that included ‘China’, one finds that they often involved international negotiations trying to keep the Bush White House for invading ‘Iran’, which was the dominant term of that first period. China, Russia, and Europe were involved in those negotiations to try to keep Bush and Cheney from starting Word War III. Four months later, Bush and Cheney no longer wanted to invade Iran, but were hinting that an invasion of North Korea would stop the development of an atomic arm in that country. Again there were international negotiations, and this time China was playing the leading role, which largely explains how the attractive power of ‘China’ increased to a maximum during the 2006-2 period. But on 9 October 2006, North Korea detonated a nuclear device and that was the end of negotiations and the attractive power of the keyword ‘China’ fell precipitously as China disappeared from the international media as a major player on the world stage (see Fig. 6.6, ‘2006 keywords’ attractive power over the three four-month periods). Little wonder that China soon became far more aggressive on the world scene and is currently browbeating the entire world concerning its offensive in the South China Sea. Since the publication of the 2006 results, we have also looked at the last two years of the Bush White House—2007–2008 (Van Meter and de Saint Léger 2011)—and, as could be expected, we found some rather intriguing developments. That data was divided into four successive six-­ month periods (2007-1, 2007-2, 2008-1, and 2008-2) and strategic diagrams were produced for each period. However, it was in the evolution of keyword attractive power that the most surprising evolution appears (see Fig. 6.7). Most of the terms decline in the last period (2008-2), but two top terms do increase: ‘UN’ (United Nations) and ‘kill’ (which does not designate any particular institution but does tend to characterize an interna-

Fig. 6.6  2006 keywords’ attractive power over the three four-month periods

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

175

176 

K. M. van Meter

tional increase in violence and instability). But among the other top terms, there is one in serious decline: ‘Bush’. This can be interpreted as indicating that during an international situation of increasing violence and instability (‘kill’ going up), the leader of the world’s most powerful nation was in decline or, as certain commentators stated, ‘had abandoned the helm’ or was too discredited to lead the world. That responsibility was being turned over to the institution ‘on the rise’; that is the United Nations. In short, the Republican government of George W. Bush was being replaced on the international scene by the Republicans’ worst nightmare, an ‘international government’ headed by the United Nations. Although Bush was replaced at the White House by a Democrat, Barack Obama, our analysis of 2009–2012, which was recently published under the title 2009–2012—Obama’s First Term, Bush’s ‘Legacy’, Arab Spring & World Jihadism (Van Meter 2016), confirms that the UN and not the US president continues to lead the world on the current violent international scene.

Fig. 6.7  Dominant 2007-1 terms over the four periods of 2007–2008 (Bush vs. UN)

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

6

177

 onclusions—Press Cycles, Fashionable C Thinking, and Cumulating Information

From the above ‘parapolitical’ publications covering successively 2006, 2007–2008, 2009–2012, and soon 2013 and 2015–2016, it becomes clear that ‘yearbook’ events that preoccupy the international press, and of which we are all well aware, are often based on ‘expert advice’ and current events that come in waves—often called the ‘press cycle of attention’— and each successive wave replaces the preceding one on stage ‘front and centre’ while we forget what happened during the previous period. We can escape these successive waves by keeping the texts they generate and performing a systematic and formal analysis of their content and structure. Often when we present these results during conferences, we ask the auditorium, for example, “What do you remember about 2006? What were the most important events of 2006?” The answers are usually a number of ‘yearbook’ events, because that is what we have in our memory. Textual analysis provides tools allowing to look back at that material and to challenge what you have retained from the successive waves of the media flood. In the domain of scientific publishing, where there are fewer waves of fashionable thinking, these same methods can dig deep into the content to reveal structures that do not appear of themselves, even to careful researchers trying to maintain neutrality when studying, for example, the sociology of education and how to bring up your kids and educate them for the future in France, Germany, or the United States. Remember that in the latter country almost half the population will grow up believing that the Bible explains better the existence of man than Darwinian evolution, while less than five percent believe that in the two former countries, and this is reflected in the textual structures of French, German, and American sociologies. As for Russia, its geopolitical structure today still looks very much like the results of our textual analysis of 1989 in the official biographies with Ukraine isolated by itself, the ‘western fringe’ of developed Baltic republics now part of the Western world, and Central Asian republics in a situation of flux and far from the two centres of power which are Moscow and Saint Petersburg.

178 

K. M. van Meter

Formal textual analysis is probably one of the very few methods available to us for the systematic study of scientific and cultural production in all of these countries and throughout the world, permitting anything approaching scientific neutrality, the possibility of comparative study and accumulating further information in these domains.

References Callon, Michel, Jean-Pierre Courtial, and William Turner. 1991. La méthode Leximappe – Un outil pour l’analyse strategique du développement scientifique et technique. In Gestion de la recherché – Nouveaux problèmes, nouveaux outils, ed. Dominique Vinck, 207–277. Brussels: De Boeck. Cibois, Philippe. 1983. Methodes post-factorielles pour le dépouillement d’enquêtes. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 1: 41–78. ———. 1985. L’analyse des donees en sociologic. Paris: Presses Universitaires de France. ———. 2016. Le logiciel Trideux. Accessed June 27, 2018. http://cibois.pagesperso-orange.fr/Trideux.html. De Saint Léger, Mathilde. 1997. Modélisation de la dynamique des flux d’informations – Vers un suivi des connaissances. Thèse de doctorat, CNAM, Paris. De Saint Léger, Mathilde, and Karl M. van Meter. 2005. Cartographie du premier congrès de l’AFS avec la méthode des mots associés. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 85: 44–67. ———. 2009. French sociology as seen through the co-word analysis of AFS congress abstracts: 2004, 2006 & 2009. Bulletin of Sociological Methodology/ Bulletin de Méthodologie Sociologique 102: 39–54. Demazière, Didier, Claire Brossaud, Patrick Trabal, and Karl van Meter. 2006. Analyses textuelles en sociologie  – Logiciels, méthodes, usages. Rennes: Presses Universitaires de Rennes. Fairclough, Norman. 1992. Discourse and text. Linguistic and intertextual analysis within discourse analysis. Discourse & Society 3 (2): 193–217. Glady, Marc, and François Leimdorfer. 2015. Usages de la lexicométrie et interprétation sociologique. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 127: 5–25.

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

179

Image. 2016. Alceste  – Un logiciel d’analyse et d’aide à la décision simple d’utilisation. Accessed June 27, 2018. http://www.image-zafar.com/Logiciel. html. Jenny, Jacques. 1997. Méthodes et pratiques formalisées d’analyse de contenu et de discours dans la recherche sociologique française contemporaine  – État des lieux et essai de classification. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 54: 64–122. Maisonobe, Marion. 2015. Emergence d’une spécialité scientifique dans l’espace  – La Réparation de l’ADN. Bulletin of Sociological Methodology/ Bulletin de Méthodologie Sociologique 125: 46–64. Marchand, Pascal. 2005. Le grand oral de Dominique de Villepin. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 87: 80–85. ———. 2007a. Un vert, ça va  – Dix verts, bonjour les débats! Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 94: 76–83. ———. 2007b. Le grand oral  – Les discours de politique générale de la Ve République. Brussels: De Boeck. Pinault, Michel. 2006. La science au Parlement. Les débuts d’une politique des recherches scientifiques en France. Paris: CNRS Editions. Reinert, Max. 1987. Classification descendante hiérarchique et analyse lexicale par context – Application au corpus des poésies d’A. Rimbaud. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 13: 53–90. ———. 2003. Le rôle de la répétition dans la représentation du sens et son approche statistique par la méthode ‘Alceste’. Semiotica 147 (1/4): 389–420. Scott, Pete Dale. 1993. Deep politics and the death of JFK. Berkeley and Los Angeles: University of California Press. Trade paper edition, 1996. Accessed June 27, 2018. http://www.peterdalescott.net/B-IV.html. Van Meter, Karl M. 1990. Methodological and design issues. Techniques for assessing the representatives of snowball samples. In The collection and interpretation of data from hidden populations, ed. Elizabeth Y. Lambert, 31–43. Washington, DC: National Institute of Drug Abuse, Research Monograph Series 98. ———. 2001. Terrorists/liberators. Researching and dealing with adversary social networks. Connections 24 (3): 66–78. ———. 2003. Multimethod analysis and stability of interpretation. In Interrelation between type of analysis and type of interpretation, ed. Karl M. van Meter, 91–124. Bern: Peter Lang. ———. 2008. Analyses of a quarter of century of publishing at the BMS. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 100: 6–15.

180 

K. M. van Meter

———. 2009. The AFS and the BMS. Analyzing contemporary French sociology. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 102: 5–13. ———. 2016. 2009–2012 – Obama’s first term, Bush’s ‘Legacy’, Arab Spring & world jihadism. Paris: Harmattan. Van Meter, Karl M., Philippe Cibois, Lise Mounier, and Jacques Jenny. 1989. East meets West—Official biographies of members of the central committee of the communist party of the soviet union between 1981 and 1987, analyzed with western social network analysis methods. Connections 12 (3): 32–38. Van Meter, Karl M., Martin W. de Vries, and Charles D. Kaplan. 1987. States, syndromes, and polythetic classes. The operationalization of cross-­ classification analysis in behavioral science research. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 15: 22–38. Van Meter, Karl M., Martin W. de Vries, Charles D.  Kaplan, and C.I.M.  Dijkman-Caes. 1992. States, syndromes, and polythetic classes. Developing a classification system for ESM data using the ascending and cross-classification method. In The experience of psychopathology. Investigating mental disorders in their natural settings, ed. Martin W. de Vries, 79–94. Cambridge: Cambridge University Press. Van Meter, Karl M., and Mathilde de Saint Léger. 2008. Co-word analysis applied to political science. 2006 international political & ‘parapolitical’ headlines. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 97: 18–38. ———. 2009a. German & French contemporary sociology compared: Text analysis of congress. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 104: 5–31. ———. 2009b. World politics and “parapolitics” 2006. Computer analysis of ADI timelines. Paris: Harmattan. ———. 2011. 2007–2008—The end of Bush and the rise of the UN.  Link analysis of world media headlines. USAK Yearbook of International Politics and Law 4: 1–21. ———. 2014. American, French and German sociologies compared through link analysis of conference abstracts. Bulletin of Sociological Methodology/ Bulletin de Méthodologie Sociologique 122: 26–45. Van Meter, Karl M., and Wiliam A. Turner. 1992. A cognitive map of sociological AIDS research. Current Sociology 40 (3): 129–134.

  Words and Facts: Textual Analysis—Topic-Centred Methods… 

181

Van Meter, Karl M., and William A. Turner. 1997. Representation and confrontation of three types of longitudinal network data from the same data base of sociological AIDS research. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 56: 32–49. Van Meter, Karl M., William A.  Turner, and Jean-Bernard Bizard. 1995. Cognitive mapping of AIDS research 1980–1990. Strategic diagrams, evolution of the discipline and data base navigation tools. Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 46: 30–44.

7 Text Mining for Discourse Analysis: An Exemplary Study of the Debate on Minimum Wages in Germany Gregor Wiedemann

1

Introduction

Two developments have widened opportunities for discourse analysts in recent years and paved the way for incorporation of new computational methods in the field. First, amounts of digital textual data worth investigating are growing rapidly. Not only newspapers publish their content online and take efforts retro-digitizing their archives, but also users interactively react to content in comment sections, forums, and social networks. Since the revolution of the Web 2.0 made the Internet a participatory many-to-many medium, vast amounts of natively digital text emerge, shaping the general public discourse arena as much as they form new partial public spheres following distinct discourse agendas. Second, computational text analysis algorithms greatly improved in their ability to capture complex semantic structures. G. Wiedemann (*) Department of Informatics, Language Technology Group, Hamburg University, Hamburg, Germany e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_7

183

184 

G. Wiedemann

Early approaches of computational content analysis (CCA; Stone et al. 1966) have been criticized from a qualitative research perspective as simple ‘word counts’ which treat character strings as representatives of fixed meaning and largely ignore contexts of their discursive formation. As a reaction, in 1969, Michel Pêcheux sketched the program of an ‘Automatic Discourse Analysis’ (Pêcheux et al. 1995) which was based on the idea to examine the formation of the meaning of words themselves through statistical observation of language use in context with other words. In the following decades, many variants of statistical approaches of computer-­ assisted text analysis originated in the Francophone branch of discourse analysis. The methods such as key term extraction, frequency analysis, and co-occurrence analysis can be summarized by the term ‘lexicometry’ (Dzudzek et al. 2009). During the 2000s, lexicometry and corpus linguistics slowly became part of the method toolbox also in Anglo-Saxon and German social sciences and humanities through integration with (critical) discourse analysis (e.g. Helsloot and Hak 2007; Glasze 2007; Baker et al. 2008; Mautner 2009; Scholz and Mattissek 2014). For some years now, there are a number of advanced algorithms for text analysis from the field of natural language processing in computer science. The algorithms, summarized by the term ‘text mining’, further extend computational capabilities to capture semantic structures in large text collections (Heyer et al. 2006). They provide new opportunities for exploratory studies which can be directly applied in discourse analysis contexts. Further, they allow for automation of certain content analysis steps, which can contribute to research designs where coding of texts is utilized as part of operationalization in a discourse study (Wedl et  al. 2014). For discourse analysis, text mining provides promising answers to the challenge of vastly growing amounts of data in our digital era (Wiedemann and Lemke 2016). In this study, I will introduce two classes of ‘machine learning’ algorithms to demonstrate the capabilities of text mining and what it can contribute to a discourse research design. As an exemplary case, I will investigate the public discourse on the introduction of statutory m ­ inimum wages in the Federal Republic of Germany. The discourse is examined by analysing more than 7,600 articles on the topic published in two German newspapers between 1995 and 2015. I will demonstrate how an ‘unsu-

  Text Mining for Discourse Analysis: An Exemplary Study… 

185

pervised’ machine learning (ML) method such as topic modelling contributes to partition the large corpus thematically and temporally for closer exploration of contained discourse structures. Moreover, I show how ‘supervised’ learning, also known as text classification, can contribute to reveal categorical patterns in the text data and, hence, may support more deductive analysis steps. However, before further introduction of the exemplary study and its results, I will explain the ideas behind ML in more detail in the upcoming section. I also reflect methodologically on characteristics of statistical ML models for text in contrast to lexicometric measures and how both may contribute to reveal patterns of discourse in diachronic studies. But it should be mentioned beforehand that for text mining the same applies as for any other computer-assisted text analysis: the mere use of specific software is not a method on its own. Software, first and foremost, supports human analysts to structure and order data within a methodological setting (Kelle 1997). This is also true for the new algorithms, although, as we will see, they are able to capture more complex structures in very large amounts of texts.

2

From Lexicometrics to Machine Learning

Notorious and ongoing debates on the character of discourse research repeatedly state that there is no single formal method to conduct empirical discourse analysis. Consequently, discourse analysis sometimes is described rather as a specific style of research than a method (Wedl et al. 2014, 540). When it comes to concrete data analysis, discourse studies usually borrow instruments and processing steps from a variety of other methods and methodologies such as grounded theory, lexicometry, or content analysis. For this reason, any discourse research requires its own specific operationalization to produce empirically founded hypothesis about the investigated discourse (cp. Angermuller 2014, 24). Consequently, I do not want to pay much attention to methodological debates about discourse theory or any specific notion of ‘discourse’ in this chapter. I rather concentrate on methodical questions to reveal structural patterns in large textual data sets. Doubtlessly, text mining comprises

186 

G. Wiedemann

useful tools for both lexicometry and content analysis (Wiedemann and Lemke 2016). Therefore, I expect that its capabilities to structure and order data also can make a valuable contribution to discourse studies conducted against a large variety of methodological and theoretical backgrounds. First, let us have a look at the way how advanced text mining algorithms, in particular, ML, proceed to extract knowledge from textual data. In a second step, we compare the characteristics of already established computational approaches such as CCA and lexicometric analyses (Lebart et al. 1998), on the one hand, and ML, on the other hand, to reflect on characteristics of the new approaches and their potential for discourse analysis.

2.1

Supervised and Unsupervised Machine Learning

Generally, two types of ML are distinguished: (a) unsupervised learning, which is purely data-driven to obtain structure originating from the data itself, and (b) supervised learning, which tries to infer relations between given text collections and knowledge represented in text external variables, for example, code categories a text is labelled with. In discourse analysis scenarios, unsupervised learning can be, for instance, automatic clustering of document collections by thematic coherence. For this purpose, topic models provide a set of ‘algorithms for discovering the main themes that pervade a large and otherwise unstructured collection of documents’, which can be employed to ‘organize the collection according to the discovered themes’ (Blei 2012, 77). The basic model Latent Dirichlet Allocation (LDA) takes the single words of a document as features for partitioning the collection into a given number of K topics. In a complex computation, LDA infers two ‘latent’ variables as probability distributions. The first is a topic-term distribution, which encodes information on how probable a term occurs in each of the K topics. The second is a topic-­document distribution, which encodes the share of all K topics in each individual document. Highly probable terms in single topic represent semantic coherences from the collection which can be interpreted as a thematic cluster. Since distribution of these the-

  Text Mining for Discourse Analysis: An Exemplary Study… 

187

matic clusters across all documents is a result of the modelling process, we can assess on their overall share in the collection, as well as in arbitrary slices of the collection such as time periods of single years. Therefore, topic models allow for a broad thematic partitioning to gain insights and as an inductive starting point for further detailed analysis. Moreover, topics can also be evaluated with respect to metadata such as time or authorship to analyse evolvement of themes or distinguish thematic preference of authors. Evans (2014), for instance, analyses around 15,000 newspaper documents to identify topics regarded as ‘unscientific’ in the US-American public discourse. Elgesem et al. (2015) utilize topic modelling to study the discourse on climate change in more than 1.3 million Internet blog articles and compare topic prevalence in different net communities. Supervised learning, in contrast, allows not only to cluster collections for broad thematic coherences, but to create models encoding the association of certain textual features with deductively defined content categories. By this, documents, paragraphs, or sentences can automatically be coded analogue to the coding process in manual content analysis (CA). For this, human coders are trained to identify text parts fitting the definition of a certain category given by a code book. The coder closely reads (samples of ) documents from the target collection and extracts text parts fitting code book definitions. A classification model trained by supervised learning in this regard represents the equivalent of a naive human coder who applies his/her knowledge to code a given text. By incorporating this knowledge in a statistical model instead of employing human coders, supervised ML algorithms can decide whether or not any new given text belongs into the category. This allows for automatic coding of potentially unlimited amounts of texts. Once an entire, big collection is automatically coded, it allows for interesting quantitative evaluations of category distributions and dependency, or filtering into subcollections. Interesting analysis, for example, would be measurement of frequency of categories across time or comparison of certain data subsets such as single publications. Another type of analysis could focus on dependency or co-­ occurrence of categories in single documents. Lemke et  al. (2015), for instance, utilize text classification to code paragraphs in several million newspaper articles to learn about the

188 

G. Wiedemann

evolvement of economized justifications of policy measures in German public discourse. They are able to show that the use of economized justification such as arguments referring to cost and efficiency significantly increased in many policy areas over the last six decades. Wiedemann (2016) employs supervised learning on a corpus of around 600,000 newspaper documents to identify patterns of excluding speech acts targeted against certain actors and ideas in the German political discourse on democracy. He finds that the far-right is more likely to be excluded from the democratic discourse by moral and legal arguments referring to democratic ideals and repressive policy measures of ‘fortified democracy’ such as ban of parties. Leftist politics, in contrast, more often is discredited by equating it to its far-­right counterpart, or devalued by reference to ‘extremism’ as generic concept. These examples prove the ability of ML approaches to capture complex, qualitatively defined categories in large text collections and make them accessible to quantitative analysis steps.

2.2

Quality Assurance and Active Learning

One necessary condition for supervised learning to produce accurate results is that it needs to be provided with sufficient training data, that is, enough manually annotated text examples representative for a certain category. The quality of automatic coding processes can be evaluated analogue to inter-rater reliability known from CA where two or more analysts code the same set of texts and compare their results. Instead of two human coders, reliability in the automatic coding setting can be obtained by comparing human and machine decisions. The share of agreement between human and machine decisions reports the accuracy of the process. Measures such as Krippendorff’s alpha or Cohen’s kappa regularly used in linguistic and social science contexts to compare coding results from two coders apply well. Also the F1-score can be used, a measure of the harmonic mean between precision and recall of the coding, which is more common in machine learning contexts (Dumm and Niekler 2016). Unfortunately, it is not as easy to obtain a sufficient training set which captures all relevant features of a category in a collection and, hence, may lead to poor automatic classification accuracy. A training set of sufficient

  Text Mining for Discourse Analysis: An Exemplary Study… 

189

quality can be guaranteed by a certain selection strategy called active learning. For this, Wiedemann (2016, 125ff) proposes a workflow of selecting documents across the entire time period of investigation and from all prominent topics in a collection, to compile and manually code an initial training set. This initial training set is then extended iteratively in the ‘active learning’ process, where the machine classifier is utilized to choose new training examples. For this, the classifier is trained on the current training set and then applied to the yet unlabelled data. The resulting suggestions of positive category texts are retrieved and revised by the human analyst who accepts or rejects the classifiers decision. The revised and potentially corrected positive and negative examples are subsequently used to extend the training set and train a new model. Every iteration of automatic classification followed by human revision makes the training set more representative for the entire collection, and the classification model becomes more accurate. This process is repeated until classifier accuracy is no longer increasing or has reached a satisfying level.1 A model based on the final training set is then applied to code the entire collection.

2.3

Local and Global Contexts

Since discourse analysts strive for inference on transtextual knowledge produced in countless acts of communication, while at the same time actually close reading single documents, they necessarily oscillate between two levels of analysis. These two levels can be distinguished along with specific notions of context they incorporate for their analysis. On the one hand, they look at single documents to identify supposedly relevant structures for the discourse. Single observations within documents are assessed against the background of their local context, that is, the other knowledge structures referred to in the document. Then, a comprehensive overview of extracted patterns on an intertextual level is approached to identify recurrent patterns and to reconstruct underlying knowledge structures or sort out negligible aspects. By reading through multiple  Accuracy in this scenario can be determined by k-fold cross-validation on the current training set (Dumm and Niekler 2016). 1

190 

G. Wiedemann

documents, the view of the analyst is sharpened for specific textual formations and language regularities contributing to shape a discourse in its very own specific way. Analysis on this level embeds empiric observations from the data in their global context. If one is able to condense these formations and language regularities into some sort of analytic categories, she/he keeps on extracting such patterns from local contexts and relating them to each other on global context level until a saturated description of the discourse can be assumed. This alteration between inductive, datadriven category development and deductive category subsumption is at the core process of knowledge reconstruction in discourse analysis—or, as Wodak and Meyer (2009, 9) phrase it: ‘Of course, all approaches moreover proceed abductively’. This way of proceeding has some analogy to the unsupervised and supervised nature of ML algorithms. They also give researchers the opportunity to combine inductive and deductive steps of analysis into creative workflows. On the one hand, unsupervised ML allows for exploration of patterns buried in large text collections to learn about contents without any prior knowledge. On the other hand, supervised ML provides results for descriptive statistics and hypothesis testing on the basis of deductively defined categories. Also algorithmically, ML has some similarities to abduction to infer optimal knowledge representations for given data sets.2 Unlike humans, algorithms are capable of processing extremely large quantities of textual data sets without getting tired or distracted. At the same time, usually these data sets are the only source they can learn structure from in a statistical way.3 So far, in contrast to humans, they lack common ‘world knowledge’ and experience from outside the investigated text collection to relate observed patterns and draw inference on. In this respect, local  Optimization algorithms in machine learning, such as Expectation Maximization, usually start with random or informed guesses for their initial model parameters. In an iterative process the model parameters are adapted in small steps to better fit the given data. In the end, the model parameters (nearly) optimally describe the data set in some structural way. Not coincidentally, this process resembles the abductive process of knowledge reconstruction from text corpora in qualitative data analysis. 3  Of course, there are already text mining approaches which incorporate text external resources such as comparison corpora or structured knowledge bases. It is the task of further research and development to evaluate on the contribution of such resources for specific research questions in qualitative data analysis. 2

  Text Mining for Discourse Analysis: An Exemplary Study… 

191

and global contexts for computer algorithms need to be defined in a drastically limited manner. Local contexts of observations such as words, statements, or topics are simply defined as all the other words, statements, or topics also occurring in the same context window, for example, the same document. Global contexts of an observation then can be defined as (statistically significant) observations which typically co-occur across documents. Saussurean structuralism allows us to describe specific relations based on these contexts. For instance, the term sun often co-occurs in a local context with shine. When both terms appear together in one sentence, they describe a syntagmatic relation. At the global context level, we can observe that this relation is statistically significant, and that other terms such as light also describe a comparable relation with shine. Because sun and light share a similar global context, they form a paradigmatic relation. The new characteristic which distinguishes ML algorithms from already established methods in lexicometry and corpus linguistics is directly related to their ability to connect the local and global context level. ML approaches are able to link meaning structures from local contexts, for example, words in documents, to meaning on global context levels such as thematic coherence or dependency on external variables. Such links are captured explicitly in statistical models. In contrast, lexicometric measures such as ‘keyness’ of terms or co-occurrence are restricted to descriptions of global contexts. They only provide an aggregated, distant view on a text collection. Formation of (latent) meaning of words in this respect is encoded solely in measures expressing their statistical significance. Typically, such aggregated numbers are decoupled from their constituents, the individual observations in the data. ML, in contrast, relies on probabilistic models of latent meaning retaining the connection between statistically inferred global context knowledge (e.g. thematic clusters) and individual data instances (e.g. documents). Statistical models, once trained on a text collection, can even be applied to new data to observe and measure inductively or deductively derived categories across (sub-)collections. Hence, in ML, context is no longer ignored like in simple word counts, or a mere result of an analysis like in lexicometric analysis. Instead, ML takes an intermediate position between the local and the global view, which allows for switching between both perspectives.

192 

G. Wiedemann

With this, the combination of lexicometry and ML, which I summarize under the term text mining, doubtlessly improves opportunities for combined qualitative and quantitative insights into text collections. Applied in the right manner, it allows for complex, insightful analysis workflows which have an enormous potential to open up discourse research to big data sets, which is of special interest for longitudinal studies covering large time periods or broad, fractionated discourse communities (Abulof 2015).

3

 wenty Years of Dispute: Minimum T Wages in Germany

To further display the potentials of ML for discourse analysis, in the upcoming sections, I sketch an exemplary study on the public debate about minimum wages in Germany using topic models and text classification. Rather than providing a comprehensive policy study, the focus of this chapter is to contribute to the understanding of what these technologies can contribute to discourse analysis.

3.1

Example Study

Jäger (2004) defines discourses as ‘flows of knowledge through time’ and, therewith, highlights the temporal component of the concept. In this respect, longitudinal large text corpora appear as natural source to study discourses. Especially, public dispute on reform in certain policy issues can cover long time periods and evolve rather slowly, shaped by persistent pressure from interest groups against or in favour of a certain measure. Also, there are spontaneous reforms as a reaction to abrupt changes of social, political, or institutional conditions. Studying longitudinal discourse under such conditions can be a challenging endeavour due to the massive amounts of data. The introduction of minimum wages in the Federal Republic of Germany makes an excellent case to study as such a longitudinal discourse. Not only is it interesting to reveal how an exceptionally long-lasting political dispute finally became law. It also may dem-

193

  Text Mining for Discourse Analysis: An Exemplary Study… 

onstrate the striking contribution text mining techniques have to offer for discourse analysis. The introduction of statutory minimum wages was a hot topic for more than two decades in Germany. Since January 2015, wages of at least 8.50 EUR per hour have to be paid throughout the country. The underlying law, the ‘Mindestlohngesetz’, passed the Bundestag in August 2014 after long discussions between political parties, business lobbyists, unions, and the public. This act realized an integral part of the coalition agreement between the two major parties, Christian Democrats (CDU) and Social Democrats (SPD), who formed a grand coalition after the federal election in 2013. Up to this point, the dispute lasted already for several years on the highest political level. First peaks of the debate can be dated back to the mid-nineties. A simple frequency analysis of the term ‘Mindestlohn’ (minimum wage) in German newspapers reveals that the term started to occur in the late 1980s, but the debate did not gain importance before 1995, with peaks in 1996, 2007, and 2013 (Fig. 7.1).

documents

400

200

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

2003

2002

2001

2000

1999

1998

1997

1996

1995

0

year publication

FAZ

FR

Fig. 7.1  Document frequency of articles on minimum wages in two German newspapers

194 

G. Wiedemann

For the exemplary study, I inspect and compare two major German newspapers. Following the ‘distant reading’ paradigm (Moretti 2007), first, I strive for revealing global contexts of the debate with respect to its temporal evolvement. What are major topics and subtopics within the discourse, when did they emerge or disappear, and how are they connected to each other? Can we determine distinct time periods of the debate from the data? The use of topic models in combination with further text mining techniques will enable us to answer these questions. For a substantial analysis, we also want to zoom in from this distant perspective and have a close look on individual units of ­language use shaping the discourse. In a deductive step, I will look for statements and utterances expressing political stance towards the issue. How is approval or rejection of the introduction of minimum wages expressed and justified throughout time? Then, with the help of text classification, we will be able to trace this antagonism between proponents and opponents of statutory minimum wages quantitatively.

3.2

Data Selection

To study the discourse on minimum wages (MW) in Germany, I compiled a corpus of articles from two newspapers, the Frankfurter Rundschau (FR) and the Frankfurter Allgemeine Zeitung (FAZ). Both are widely regarded as ‘quality newspapers’ with nationwide dissemination. Moreover, they represent two sides of the political spectrum. The FAZ, founded in 1949, is generally viewed as a voice for conservative milieus in Germany, while the FR, founded in 1945 covers the liberal, left-wing part of the political spectrum. Therefore, a comparison of both publications may give interesting insights into the political debate, which as a measure for promotion of social justice has been propelled by left-leaning actors mainly. Since both newspapers are published in the Hessian metropolis Frankfurt am Main, possible local aspects of the debate may also be covered in a comparable manner. Articles were selected by their publication date between 1 January 1995 and 31 December 2015, covering the time period from the first larger impacts of the issue in the discourse agenda to the first year after

  Text Mining for Discourse Analysis: An Exemplary Study… 

195

the ‘Mindestlohngesetz’ came into force. This allows tracing the genesis of the policy measure from the first noticeable public demands, over various sidetracks of the discourse and the point when lawmakers in majority supported the measure, up to reflections on actual effects of the enacted law. From the entire newspaper archive, those articles were retrieved which contained the term ‘Mindestlohn*’, where the asterisk symbol indicates a placeholder to include inflected forms, plural forms, and compounds.4 Additionally, articles had to be related to German politics mainly, which could be achieved by restricting the retrieval in the archive databases by provided metadata. This results in a corpus of 7,621 articles (3,762 in the FAZ; 3,859 in the FR) comprising roughly 3.76 million word tokens. Their distribution across time reveals the changing intensity of the public debate. Absolute document frequencies indicate that both publications, although from opposing sites of the political spectrum, cover the topic in a surprisingly similar manner (Fig. 7.1).

4

Inductive Analysis: Corpus Exploration with Topic Models

For the exemplary study, I computed an LDA topic model with K = 15 topics on the entire corpus of newspaper articles. Before the LDA model inference, common stop words were removed, as well as words occurring less than 20 times in the corpus. Remaining terms were reduced to their word stem and transformed to lower case to unify similar word types in the vocabulary of the collection. The analysis is realized in the R statistical programming environment using the topicmodels package by Grün and Hornik (2011), which provides algorithms for LDA and other topic models. Table 7.1 displays the inferred topics by listing their ten most probable terms and their share of the collection in decreasing order. Additionally, a label is attached to each topic, manually determined by interpretation of the semantic coherence of its top terms.

 The search index treats German Umlaute as their ASCII equivalent, such that Mindestlohn* also retrieves articles containing ‘Mindestlöhne’. 4

196 

G. Wiedemann

Table 7.1  Topics terms and shares Share No. (%) 1

10.5

2

9.3

3

8.2

4

8.2

5

7.9

6

7.8

7

7.6

8

6.3

9

5.6

10

5.4

11

5.2

12

5.1

13

4.8

14

4.2

15

3.9

Top 10 terms

Label

polit sozial gesellschaft mensch staat wirtschaft okonom gerecht arbeit marktwirtschaft mindestlohn branch gesetz tarifvertrag euro spd union entsendegesetz muntefering gewerkschaft spd merkel partei koalition muntefering beck sozialdemokrat gabriel steinbruck kanzlerin mann alt leut jahr tag mensch polit gut welt leb spd union cdu merkel koalition csu nahl koalitionsvertrag seehof gabriel euro uhr mitarbeit kund frankfurt stadt arbeit preis stund unternehm prozent arbeitsmarkt arbeitslos institut million euro euro milliard kind arbeitslos arbeitslosengeld rent hartz alt sozialhilf langzeitarbeitslos ausland deutsch unternehm illegal arbeit pol baustell schwarzarbeit gesetz bau gewerkschaft dgb zeitarbeit metall leiharbeit verdi beschaftigt gewerkschaftsbund somm arbeit grun fdp cdu spd partei wahl koch hess hessisch link bau prozent gewerkschaft arbeitgeb ost west branch allgemeinverbind wiesehugel baugewerb europa deutsch deutschland wirtschaft regier frankreich griechenland franzos land milliard link partei linkspartei spd lafontain gysi parteitag berlin pds wahl post zumwinkel pin deutsch wettbewerb brief euro konz tnt verdi

Social market economy

Sector-specific MW

Social democrats

General terms Grand coalition MW implementation Job market Social welfare

Undeclared work Temporary work

MW in Hesse MW in construction sector MW in Europe

Socialist party MW in postal sector

  Text Mining for Discourse Analysis: An Exemplary Study… 

197

In general, evaluation of clustering results and, in particular, evaluation of topic models is not a trivial task. There are two major problems to deal with. First, well-fitted separation of data into clusters can be obtained at different levels of granularity. Second, due to mechanisms of stochastic inference which involve random sampling methods to obtain model parameters from data observation, results are not entirely deterministic (Lancichinetti et al. 2015). To solve the first problem, analysts may desire a data-driven, automatic solution to answer the question: How many topics does my collection contain? Unfortunately, although there are numerous measures to determine cluster quality (Walesiak and Dudek 2015) which may be utilized to suggest a numerically optimal number of clusters, such automatically derived parameters may not fit well to the needs of an analyst. Imagine, for example, the data-driven parameter selection suggests splitting the collection into two topics. This purely would not satisfy the researchers demand for gaining deeper insights into the constituent parts of the discourse, since topics would be very general and hardly interpretable as meaningful facets. In another case, it may suggest a separation into a hundred or more clusters as numeric optimal solution. We certainly would not be able to analyse and interpret such a fine-grained, potentially unstable partitioning. This means that although we can use cluster quality indices as orientation, we need to define reasonable lower and upper bounds to the number of clusters according to our research design. The second problem, variance of modelling results due to the non-­ deterministic nature of the inference mechanism, also needs careful inspection during the analysis process. Choosing a higher number K to separate the collection may lead to an unreliable clustering which cannot be reproduced between repeated runs of the model algorithm. This means, topics which occur in one process of modelling and which may appear as meaningful and interpretable semantic coherence to the analyst cannot necessarily be retrieved again in a second run of the algorithm. To produce valid interpretations from topic modelling, researchers should rely on reliable, hence, reproducible topics only. This certainly needs some experimentation with parameters and familiarity with the overall procedure. Usually, in topic model analysis researchers compute a variety

198 

G. Wiedemann

of models and compare them in a systematic procedure. To select an appropriate model, Evans (2014) suggests three steps: The first criterion for model selection is to utilize numeric measures to determine the model quality. Three measures are customarily used throughout the literature: (1) perplexity, a measure to determine how well the model generalizes to unseen data (Wallach et al. 2009); (2) coherence, measuring how often topic defining terms actually co-occur in documents (Mimno et al. 2011); and (3) reliability, the share of topics we are able to retrieve in a stable manner between repeated runs (Lancichinetti et al. 2015). Since perplexity and coherence highly depend on the alpha-­ parameter of the LDA modelling process, a numeric prior which determines the shape of topic distributions across documents, it makes sense to utilize them as an indicator for selection of an optimal alpha-­parameter. For the newspaper collection, I computed different models, each with K = 15 topics but varying the alpha-parameters, and selected the one with optimal coherence and perplexity. For this, reliability could be obtained, showing that around 81% of the topics can be retrieved stably, which is within acceptable range. In case of low reliability, it would be reasonable to decrease the number of topics K to retrieve more stable topics. As the second step, we investigate the top terms of all topics and evaluate if we are able to assign a descriptive label to them. For the presented model on the MW collection, such a label for each topic is given in Table 7.1. While most of the labels represent very straightforward interpretations, in some cases it is rather hard to come up with a unique label. Topic 4, for instance, mainly consists of general terms not related to any specific theme. To support intuitions from top word lists, we can sample topic representative documents from the collection and read them carefully to obtain more information to base our interpretation on. Overall, most topics can be interpreted and labelled unambiguously. We can identify interesting subtopics in the debate around MW, such as consequences for the job market, relations to undeclared work and social welfare or different forms of implementation such as MW for temporary work or specific sectors. In the third step, we assess whether topic distributions over time fit to our prior expectations about the discourse, for example, if certain events or developments we already know about are represented by the model.

  Text Mining for Discourse Analysis: An Exemplary Study… 

199

Since our model provides topic probabilities for each document, we can aggregate probabilities according the document metadata such as its publication year and normalize them in a range between 0 and 1 to interpret them as proportions. Average topic proportions can be visualized, for instance, as an area plot, which allows a visual evaluation of topic trends over time (Fig. 7.2). To make the plot more readable, topics are sorted in a specific manner. On the single curves of the topic proportions over time, I computed a regression line and sorted the topics according the slope of this line. This results in an ordering where largest increases of topic shares are on the top of the list while topic shares most decreasing over time are located at the bottom. Now we can easily identify specific trends in the data. In the beginning of our investigated time period, the discourse is largely dominated by the issue of MW in the construction sector which were introduced with the ‘Entsendegesetz’ in 1996 to prevent dumping wages, but led to heated discussions on increases of undeTopic

1.00

Grand coalition Social democrats MW in Hesse Job market

Aggregated probability

0.75

Socialist party MW and temporary work Sector specific MW 0.50

MW in postal sector General terms MW implementation Social welfare

0.25

MW in Europe Social market economy MW and undeclared work 0.00 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

MW in construction sector

Date

Fig. 7.2  Area plot of topic distribution over time

200 

G. Wiedemann

clared work. Dispute around the millennial turn was focusing on the question whether statutory wages can even be enforced by executive powers. Then, steadily more industrial and service sectors became subject to the debate. Throughout the 2000s, sector-­specific MWs with an exceptional highlight on the postal sector were preferred above a general MW for all sectors in the entire country. During that time, the topics on social market economy entangled with demands for social justice and concerns for the job market formed a steady and solid background for the debate. In 2013, we could identify a new shift of the debate when a general minimum wage became a central policy objective in the coalition agreement of CDU/CSU and SPD after the federal election. From this year onwards, topics on implementation of MW and possible consequences on the job market increase. From the perspective of quality assurance of the analysis process, these results can be viewed as a successful evaluation, suggesting we were able to obtain a valid model for our research purpose. We selected model parameters according to optimized conventional numeric evaluation measures and were able to label and interpret single topics, as well as their quantitative evolvement over time. But of course, carrying out the entire analysis on distributions and trends of semantic clusters covers only a very distant perspective of the discourse. The method rather provides an overview of the discourse, a suggestion for its separation, and a starting point for further analyses to gain insight through contrasting along data facets such as theme and time. To develop a profound understanding of the data and what is going on in distinct topics at certain points of time, we still need to read single articles. Fortunately, the topic model provides us information on which documents to select. For every topic at any distinct point in time, for example, the striking increase in the topic of sector-specific MW in 2004, we can sample from the most representative documents for that subset of our data to prepare further manual analysis steps. Thus, the close connection between the global and the local context through the model allows for a straightforward realization of demands for ‘blended reading’, the close integration of distant and close reading steps (Lemke and Stulpe 2016).

  Text Mining for Discourse Analysis: An Exemplary Study… 

5

201

 eductive Analysis: Classification D of Propositional Statements

The previous step demonstrated how unsupervised ML contributes to information extraction from large text collections by relying on data-­ internal structures. This allowed us to inductively explore the debate on MW in temporal and thematic manner. It made knowledge emerging from trans-textual coherence of language use explicitly accessible through the latent variables of the topic model. At the same time, this only provided rather broad contextual information. To analyse and understand discourses, we do not strive for the macro view only, but want to reveal how single constituents of utterance contribute to its formation. Depending on the research design, we may derive categories of language use contributing to the production of a discourse directly from the data, or already have concrete categories for the analysis derived from text external sources such as theoretical assumptions or previous studies. In both cases, we can apply a subsuming analysis step where we assign categories to samples of text as representatives of empiric evidence. For closed, well-defined category systems, this step of categorization can be supported by supervised text classification or active learning. For the exemplary study on the minimum wage discourse, I employ a very simple category scheme for demonstration purposes. The goal is to measure approval and opposition of political stance in the newspaper articles to reveal their trends. For this, I hand-coded sentences in articles. I extracted 107 sentences representing approval to the idea of statutory MW, and 102 expressing opposition out of a set of 30 documents. Text classification is realized with a linear support vector machine (SVM) operating on a feature set of unigrams and bigrams of word stems in the sentences. SVM introduced in text classification by Joachims (1998) already almost two decades ago, is still one of the most efficient discriminative classification models for this task. I utilized the LIBLINEAR implementation by Fan et al. (2008). In six iterations of active learning, I extended these sets to 460 approval sentences and 524 sentences expressing opposition. Since both categories are constructed as independent, to each set related sentences, called ‘positive’ set, a contrasting ‘negative’ set

202 

G. Wiedemann

of sentences irrelevant to the category is collected. These ‘negative’ sentences mainly result from corrections during the active learning process. Therefore, they also provide very valuable feature information to the classifier to be able to distinguish category representative units correctly from those not fitting the category definition. A tenfold cross-validation on the final training set shows results around 0.7 for the F1-score (Table 7.2), which can be seen as a satisfying agreement between machine and human coding. With the final TS obtained from active learning, we now are able to classify each of the around 259,000 sentences from the entire collection as approval, opposition, or irrelevant in terms of our category system. Finally, the resulting sets of positive sentences fitting each category can be utilized for quantitative evaluation. Figure  7.3 plots frequencies of documents containing at least one expression of approval or opposition towards the issue of MW in the two publications. Frequencies were ­normalized to proportions with respect to the collection size in the single time slices. Table 7.2  Text classification of stances on minimum wages Approval

Opposition

Example sentence

‘An Mindestlöhnen wird deshalb kein Weg vorbeigehen’. [Hence, there is no alternative to minimum wages.]

Initial training set (TS) Final TS positive Final TS negative Features Cross-­ validation F1 Sentences all Sentences positive

107

‘Wissenschaftler warnen seit langem vor der Einführung von Mindestlöhnen in Deutschland’. [Scientists have long been warning against the introduction of minimum wages.] 102

460

524

1,226

1,621

24,524 0.720

28,600 0.667

259,097

259,097

5,937

2,026

  Text Mining for Discourse Analysis: An Exemplary Study… 

203

Trends, that is, changes of proportions across the entire time frame, appear to be very similar between the FAZ and the FR. We can observe early peaks of support for the idea of MWs in 1996 and around the years 1999/2000. In 1996, an MW was introduced in the construction sector. Around the turn of the millennium, although characterized by a large relative share of approval (Fig. 7.3), the debate remained on rather low level in absolute terms (Fig. 7.1). In contrast, intensity of the debate in absolute counts and the share of approval statements for the policy measure started to increase simultaneously from 2004 onwards. Intensity peaks in 2013 while retaining high approval shares. In this year, MWs became part of the grand coalition agreement as a soon to be enacted law. For expressions of opposition towards MW, we can observe interesting trends as well. Not surprisingly, the overall share of negative sentiments towards the policy measure is higher in the more conservative newspaper FAZ. But, more striking is the major peak in the year 2004, just at the beginning of the heated intensity of the debate. In 2005, there has been an early election of the Bundestag after the government led by chancellor FAZ

FR

proportion

0.6

0.4

0.2

0.0 1995

2000

2005

2010

2015 1995

2000

2005

2010

2015

year category

approval

opposition

Fig. 7.3  Relative frequencies of documents containing stances on minimum wages

204 

G. Wiedemann

Gerhard Schröder (SPD) dissolved the parliament. His plan for backing up the support for aimed social reforms, the so-called Agenda 2010, by renewing his government mandate failed. Schröder lost the election and a new government under Angela Merkel (CDU) was formed together with the Social Democrats as a junior partner. Under the impression of heated dispute about the Agenda 2010, the 2005 election campaign was highly influenced by topics related to social justice. Switching to the mode of campaign rhetoric in policy statements may be an explanation for the sharp drop of oppositional statements to MW. One year earlier oppositional stances peaked in the public discourse presenting MW as a very unfavourable policy measure. Due to their bad reputation throughout conservatives as well as Social Democrats, demands for MWs did not become a major topic in the campaign of 2005. The interesting finding now is that the relative distribution of approval and opposition in the public discourse in that year already was at similar levels compared with that in the years 2012/2013 when the idea finally became a favoured issue of the big German parties. This may be interpreted in a way that statutory MWs actually could have been a successful campaign driver in 2005 already, if Social Democrats would have been willing to adopt them as part of their program. In fact, leftleaning Social Democrats demanded them as a compensatory measure against the social hardship of the planned reforms. Instead, the SPD opted for sticking to the main principle behind the Agenda 2010 to include low-skilled workers into the job market by rather subsidizing low wages than forcing companies to pay a general minimum. It took Social Democrats until elections of 2009 to take stance for the idea, and another four years until government leading Christian Democrats became comfortable enough with it. Over the years 2014/2015, we could observe a drop in both approval and opposition expressions, which may be interpreted as a cool-down of the debate. In addition to trace trends of expressions for approval or opposition quantitatively, we also can evaluate on the used arguments more qualitatively. Since supervised learning provides us lists of positively classified sentences for each category, we can quickly assess their contents and identify major types of arguments governing the discourse. For approval, for instance, statements mainly refer to the need for some kind of social

  Text Mining for Discourse Analysis: An Exemplary Study… 

205

justice. The claim, ‘people need to be able to afford living from their earned income’, can often be found in variants in the data. In fact, low wage policy in Germany led to situations where many employees were dependent on wage subsidies financed by the welfare state. Companies took advantage of it, by creating business models relying on public subsidies of labour to increase competitiveness. In addition to the social justice argument, there are more economic arguments presented, especially in relation to the demand for sector-specific MW. They are welcomed not only by workers, but also by entrepreneurs as a barrier against unfair conditions of competition on opened European markets. Oppositional stances to the introduction of MWs also point to the issue of competitiveness. They are afraid that competitiveness of German industry and services will be diminished and, hence, the economy will slow down. Turning it more to a social justice argument, major layoffs of workforce are predicted. Often the claim can be found that MW are unjust for lowskilled workers because they are preventing them from their entry into the job market. One prominent, supposedly very specific German argument in the debate is the reference to ‘Tarifautonomie’, the right of coalitions of employees and employers to negotiate their work relations without interference by the state. Statutory MW, so opponents claim, are a major threat to this constitutional right. For a long time, German workers unions followed this argument, but steadily realized that power of their organized coalition in times of heated international competitiveness was no longer able to guarantee decent wages for their members. This brief characterization of the main argument patterns identifiable in public discourse could be the basis for a further refinement of the category system used for supervised learning. The active learning workflow applied to this extended system would allow for measuring of specific trends and framings of the debate—for instance, if reference to the argument on ‘Tarifautonomie’ diminishes over time, or if oppositional statements are more referring to threats of increased unemployment in a framing of social justice than in a framing of general damage to the economy. Although we would have been able to identify these argumentative patterns also with purely manual methods, we would not be able to determine easily and comprehensibly on their relevancy for the overall discourse. Certainly, we would not be able to determine on trends of their relevancy over time.

206 

6

G. Wiedemann

 iscussion: From Implicit Knowledge D to Explicit Models

In this exemplary study, I demonstrated how current technologies of text mining and ML can contribute to the analysis of the discourse on statutory MW in Germany over the past twenty years. The investigation of public discourse covered in two major newspapers from opposing sites of the political spectrum revealed interesting coherence between the identified patterns such as frequency of relevant documents as well as trends of stances towards the policy measure. We were able to infer that the overall agenda is determining the entire discourse community in very comparable manner to a large extent, resulting in similar macro-level trends for both publications. Further, we were able to identify changes of thematic focuses of the debate across time in a purely data-driven manner with the help of statistical topic models. With text classification we measured approval and opposition towards MWs from its first peak in the mid-­ nineties up to the enactment as law in 2015. For this, we applied an SVM classifier in an active learning workflow which combines human and machine coding of text in an alternating process to produce TSs efficiently and with sufficient quality. How do these approaches compare to non-computational steps of analysing discourse? Obviously, we were evaluating more surface structures rather than revealing micro-structures of utterance as the base foundation of discursive statements. While hermeneutic-­interpretive investigation of text samples may uncover how speech acts about threats of layoff of employed work or forecasts of a shrinking economy contribute to a negative framing of MW, we simply measured their frequency and embedding in a broader thematic structure formed by regularities in certain vocabulary use. In this respect, computer-mediated discourse analysis rather complements qualitative, more hermeneutic steps of discourse analysis by allowing for assessment of relevancy and identification of trends of structures. The major contribution of text mining technologies including ML is that they are able to make implicit knowledge acquired by the individual discourse analyst during her/his work explicit by encoding it into (statis-

  Text Mining for Discourse Analysis: An Exemplary Study… 

207

tical) models of language use representative for the categories of interest. This process of encoding implicit understanding of a discourse in explicit models has three major advantages for the analysis. First, it allows the assessment of very large data sets and, thus, the quantified evaluation of categories. Second, the explicit connection between local and global context in ML models also allows retrieving examples on the micro level, for example, single documents, which can be regarded as paradigmatic for macro-level structures. Typical d ­ ocuments for a topic cluster or exemplary expressions for classified statements can easily be selected from such analysis for further manual inspection. Third, it allows for a more reproducible analysis which can be subject to scientific argument more easily than purely manual analysis procedures. As well as other approaches of computer-­mediated discourse analysis, results based on ML allow for compliance with demands for inter-subjective comprehensibility and reproducibility of research. At the same time, although explicitly relying on quantification and statistics, text mining approaches remain highly compatible with key assumptions from many discourse theories. Since they do not assume fixed meaning associated with character strings, but strive to ‘learn’ these associations from observations in empirical data, they fit the demand for reconstruction of knowledge structures from combined observations on both the textual and intertextual levels. Despite the algorithmic advancements, the debate on best practices for the integration of the new technologies in qualitative data analysis is still far from being settled (Wiedemann and Lemke 2016). Not much has been said so far, on the integration of various text mining methods into complex analysis workflows. While ML on large data sets clearly shed lights on broader macro-level structures of discourse by revealing topics inductively or category trends deductively, lexicometric methods additionally provide valuable information to better understand production of discourse formation on more fine-grained entities such as key words and collocations. They can be interpreted in qualitative analysis steps as individual events in textual data on the meso and micro levels. However, they only become a pattern on the aggregated macro level of discourse. This interplay between the distinguished levels has been subject to broad theoretic, methodological reflection to situate such approaches beyond the traditional qualitative-quantitative divide (Angermüller 2005).

208 

G. Wiedemann

From the technical perspective too, there is a lot of research necessary to systematically integrate text mining approaches for capturing aspects on the different discourse levels. Beyond broad topics, it would be interesting to get closer to the intermediate semantic structures such as propositional patterns as a basis for a discourse formation. If we acknowledge the advantages of computer-­assisted methods for discourse analysis, we need intensified discussion on best practices and guidance for valid, efficient, and reliable application of state-of-the-art technologies, while at the same time paying respect to matured methodological debates in discourse research which deny the possibility for full standardization of the method (Feustel et al. 2014). In this regard, even when computational methods help us finding patterns and selecting good samples, it largely remains the task of the creative human analyst to link findings in the data on the micro- and macro levels, and to draw the right conclusions in conjunction with her/his knowledge about the sociopolitical contexts. This means that also with nowadays ML methods, we are far away from a purely automatic discourse analysis. But the new technologies offer us the chance not only to generate new findings from very large data sets, but at the same time, to facilitate the access to empiric analysis by pointing into interesting, relevant, and inter-subjectively comprehensible directions.

References Abulof, Uriel. 2015. Normative concepts analysis: Unpacking the language of legitimation. International Journal of Social Research Methodology 18 (1): 73–89. Angermüller, Johannes. 2005. Qualitative methods of social research in France: Reconstructing the actor, deconstructing the subject. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 6 (3). Accessed July 1, 2018. http://nbn-resolving.de/urn:nbn:de:0114-fqs0503194. ———. 2014. Einleitung: Diskursforschung als Theorie und Analyse. Umrisse eines interdisziplinären und internationalen Feldes. In Diskursforschung. Ein interdisziplinäres Handbuch. Band 1: Theorien, Methodologien und Kontroversen, ed. Johannes Angermuller, Martin Nonhoff, Eva Herschinger,

  Text Mining for Discourse Analysis: An Exemplary Study… 

209

Felicitas Macgilchrist, Martin Reisigl, Juliette Wedl, Daniel Wrana, and Alexander Ziem, 16–36. Bielefeld: Transcript. Baker, Paul, Costas Gabrielatos, Majid KhosraviNik, Michael Krzyzanowski, Tony McEnery, and Ruth Wodak. 2008. A useful methodological synergy? Combining critical discourse analysis and corpus linguistics to examine ­discourses of refugees and asylum seekers in the UK press. Discourse & Society 19 (3): 273–306. Blei, David M. 2012. Probabilistic topic models: Surveying a suite of algorithms that offer a solution to managing large document archives. Communications of the ACM 55 (4): 77–84. Dumm, Sebastian, and Andreas Niekler. 2016. Methoden, Qualitätssicherung und Forschungsdesign. Diskurs- und Inhaltsanalyse zwischen Sozialwissenschaften und automatischer Sprachverarbeitung. In Text Mining in den Sozialwissenschaften: Grundlagen und Anwendungen zwischen qualitativer und quantitativer Diskursanalyse, ed. Matthias Lemke and Gregor Wiedemann, 89–116. Wiesbaden: Springer VS. Dzudzek, Iris, Georg Glasze, Annika Mattissek, and Henning Schirmel. 2009. Verfahren der lexikometrischen Analyse von Textkoprora. In Handbuch Diskurs und Raum: Theorien und Methoden für die Humangeographie sowie die sozial- und kulturwissenschaftliche Raumforschung, ed. Georg Glasze and Annika Mattissek, 233–260. Bielefeld: Transcript. Elgesem, Dag, Lubos Steskal, and Nicholas Diakopoulos. 2015. Structure and content of the discourse on climate change in the blogosphere: The big picture. Environmental Communication 9 (2): 169–188. https://doi.org/10.108 0/17524032.2014.983536. Evans, Michael S. 2014. A computational approach to qualitative analysis in large textual datasets. PloS ONE 9 (2). https://doi.org/10.1371/journal. pone.0087908. Accessed July 1, 2018. Fan, Rong-en, Kai-wei Chang, Cho-jui Hsieh, Xiang-rui Wang, and Chih-jen Lin. 2008. LIBLINEAR: A library for large linear classification. Journal of Machine Learning Research 9: 1871–1874. Feustel, Robert, Reiner Keller, Dominik Schrage, Juliette Wedl, Daniel Wrana, and Silke van Dyk. 2014. Zur method(olog)ischen Systematisierung der sozialwissenschaftlichen Diskursforschung. Herausforderung, Gratwanderung, Kontroverse. In Diskursforschung. Ein interdisziplinäres Handbuch. Band 1: Theorien, Methodologien und Kontroversen, ed. Johannes Angermuller, Martin Nonhoff, Eva Herschinger, Felicitas Macgilchrist, Martin Reisigl, Juliette Wedl, Daniel Wrana, and Alexander Ziem, 482–506. Bielefeld: Transcript.

210 

G. Wiedemann

Glasze, Georg. 2007. Vorschläge zur Operationalisierung der Diskurstheorie von Laclau und Mouffe in einer Triangulation von lexikometrischen und interpretativen Methoden. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 8 (2). Accessed July 1, 2018. http://nbn-resolving. de/urn:nbn:de:0114-fqs0702143. Grün, Bettina, and Kurt Hornik. 2011. Topicmodels: An R package for fitting topic models. Journal of Statistical Software 40 (13). Accessed July 1, 2018. http://www.jstatsoft.org/v40/i13/. Helsloot, Niels, and Tony Hak. 2007. Pêcheux’s contribution to discourse analysis. Forum Qualitative Sozialforschung/Forum: Qualitative Social Research 8 (2). Accessed July 1, 2018. http://nbn-resolving.de/urn:nbn:de:0114-fqs070218. Heyer, Gerhard, Uwe Quasthoff, and Thomas Wittig. 2006. Text mining: Wissensrohstoff Text: Konzepte, Algorithmen, Ergebnisse. Bochum: W3L. Jäger, Siegfried. 2004. Kritische Diskursanalyse: Eine Einführung. 4th ed. Münster: Unrast. Joachims, Thorsten. 1998. Text categorization with support vector machines: Learning with many relevant features. In Proceedings: Machine Learning: ECML-98, Heidelberg, Berlin, 137–142. Kelle, Udo. 1997. Theory building in qualitative research and computer programs for the management of textual data. Sociological Research Online 2 (2). Accessed February 28, 2017. http://www.socresonline.org.uk/2/2/1.html. Lancichinetti, Andrea, M. Irmak Sirer, Jane X. Wang, Daniel Acuna, Konrad Körding, and Luís A. Nunes Amaral. 2015. High-reproducibility and high-­ accuracy method for automated topic classification. Physical Review X 5 (1). Accessed July 1, 2018. https://journals.aps.org/prx/pdf/10.1103/ PhysRevX.5.011007. Lebart, Ludovic, André Salem, and Lisette Berry. 1998. Exploring textual data. Dordrecht: Kluwer. Lemke, Matthias, Andreas Niekler, Gary S.  Schaal, and Gregor Wiedemann. 2015. Content analysis between quality and quantity. Datenbank-Spektrum 15 (1): 7–14. Lemke, Matthias, and Alexander Stulpe. 2016. Blended Reading: Theoretische und praktische Dimensionen der Analyse von Text und sozialer Wirklichkeit im Zeitalter der Digitalisierung. In Text Mining in den Sozialwissenschaften: Grundlagen und Anwendungen zwischen qualitativer und quantitativer Diskursanalyse, ed. Matthias Lemke and Gregor Wiedemann, 17–62. Wiesbaden: Springer VS.

  Text Mining for Discourse Analysis: An Exemplary Study… 

211

Mautner, Gerlinde. 2009. Checks and balances: How corpus linguistics can contribute to CDA. In Methods of critical discourse analysis, ed. Ruth Wodak and Michael Meyer, 122–143. London: SAGE. Mimno, David, Hanna M.  Wallach, Edmund Talley, Miriam Leenders, and Andrew McCallum. 2011. Optimizing semantic coherence in topic models. In Proceedings of the conference on Empirical Methods in Natural Language Processing (EMNLP’11), 262–272. Stroudsburg: ACL. Moretti, Franco. 2007. Graphs, maps, trees: Abstract models for literary history. London and New York: Verso. Pêcheux, Michel, Tony Hak, and Niels Helsloot. 1995. Automatic discourse analysis. Amsterdam and Atlanta: Rodopi. Scholz, Ronny, and Annika Mattissek. 2014. Zwischen Exzellenz und Bildungsstreik. Lexikometrie als Methodik zur Ermittlung semantischer Makrostrukturen des Hochschulreformdiskurses. In Diskursforschung. Ein interdisziplinäres Handbuch. Band 2: Methoden und Analysepraxis. Perspektiven auf Hochschulreformdiskurse, ed. Martin Nonhoff, Eva Herschinger, Johannes Angermuller, Felicitas Macgilchrist, Martin Reisigl, Juliette Wedl, Daniel Wrana, and Alexander Ziem, 86–112. Bielefeld: Transcript. Stone, Phillip J., Dexter C. Dunphy, Marshall S. Smith, and Daniel M. Ogilvie. 1966. The general inquirer: A computer approach to content analysis. Cambridge, MA: MIT Press. Walesiak, Marek, and Andrzej Dudek. 2015. clusterSim: Searching for optimal clustering procedure for a data set. http://CRAN.R-project.org/package=clusterSim. Wallach, Hanna M., Iain Murray, Ruslan Salakhutdinov, and David Mimno. 2009. Evaluation methods for topic models. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML’09), 1105–1112. New York: ACM. Wedl, Juliette, Eva Herschinger, and Ludwig Gasteiger. 2014. Diskursforschung oder Inhaltsanalyse? Ähnlichkeiten, Differenzen und In-/Kompatibilitäten. In Diskursforschung. Ein interdisziplinäres Handbuch. Band 1: Theorien, Methodologien und Kontroversen, ed. Johannes Angermuller, Martin Nonhoff, Eva Herschinger, Felicitas Macgilchrist, Martin Reisigl, Juliette Wedl, Daniel Wrana, and Alexander Ziem, 537–563. Bielefeld: Transcript. Wiedemann, Gregor. 2016. Text mining for qualitative data analysis in the social sciences: A study on democratic discourse in Germany. Kritische Studien zur Demokratie. Wiesbaden: Springer VS.

212 

G. Wiedemann

Wiedemann, Gregor, and Matthias Lemke. 2016. Text Mining für die Analyse qualitativer Daten: Auf dem Weg zu einer Best Practice? In Text Mining in den Sozialwissenschaften: Grundlagen und Anwendungen zwischen qualitativer und quantitativer Diskursanalyse, ed. Matthias Lemke and Gregor Wiedemann, 397–420. Wiesbaden: Springer VS. Wodak, Ruth, and Michael Meyer. 2009. Critical discourse analysis: History, agenda, theory and methodology. In Methods of critical discourse analysis, ed. Ruth Wodak and Michael Meyer, 1–33. London: Sage.

Part IV New Developments in CorpusAssisted Discourse Studies

8 The Value of Revisiting and Extending Previous Studies: The Case of Islam in the UK Press Paul Baker and Tony McEnery

1

Introduction1

Discourse analyses often tend to be time-bound. A discourse is observed, its nature characterised and an analysis concludes. This, of itself, is not problematic—analyses have beginnings and ends. Researchers invest the time and effort into a research question as their research programme demands and then move on to their next question. A slightly more problematic situation arises, however, when discourse is described and then assumed to remain static. Such an analysis will background the fact that discourse is dynamic. While we may concede that dynamism in ­discourse may be topic sensitive and that such change may vary in terms of speed and degree, it is nonetheless probably the rule rather than the exception  The work reported on in this chapter was supported by the ESRC Centre for Corpus Approaches to Social Science, grant number ES/K002155/1.

1

P. Baker (*) • T. McEnery Department of Linguistics and English Language, University of Lancaster, Lancaster, UK e-mail: [email protected]; [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_8

215

216 

P. Baker and T. McEnery

that discourse is Protean over time. However, it can be easy to refer to past analyses and assume that they still have contemporary relevance. For example, L’Hôte (2010) studies the language of globalisation in the British Labour party in the period 1994–2005 using a 234,387-­word corpus of documents produced by the party. Yet the study has been cited since as being about ‘the particular meaning of “globalisation” in Labour manifestos’ (Evans and Schuller 2015). Similarly, Kambites (2014, 339) says that L’Hôte’s paper ‘analyses the use of the term “globalization” by successive UK governments and finds that “new Labour discourse is significantly more concerned with the process of globalisation than Conservative discourse”’. Both papers imply that the findings of L’Hôte are not time-bound—but at the very least that possibility should be accepted, otherwise contemporary discourse may be mischaracterised. Nonetheless, the difficulty of extending some studies, especially where they are large in scale, may be so daunting that a full investigation of the ongoing dynamism of a discourse may not be undertaken.2 In this chapter, we will return to one major study, Baker et al. (2013), to explore how, in the period following their study, the discourse that they studied changed, if at all. Baker et al. explored the representation of Muslims and their religion, Islam, in the British national press. They did this by examining 143 million words of British newspaper articles from 1998 to 2009. While their study is an exhaustive account of the period covered, numerous subsequent global events relating to, or involving, Muslims, such as the Arab Spring and the rise of ISIS, mean that the possibility clearly exists that discourse around Islam has changed somewhat since 2009. Additionally, the political context in the UK has changed; there was a change in government in 2010, with a Conservative-­led coalition replacing Gordon Brown’s Labour Party. The popularity of the English Defence League as well as increasing support for UKIP ­suggest that Britain’s social barometer has become more right-leaning after 2009. With both the political context and the world context of reporting about Muslims and  An example of such an extension is the work of Blinder and Allen (2016), who looked at the representation of refugees and asylum seekers in a 43-million-word corpus of UK press material from 2010–2012 in a complementary study to the investigation of the same subject by Baker et al. (2008) using a 140-million-word corpus of newspaper articles covering 1996–2005. 2

  The Value of Revisiting and Extending Previous Studies… 

217

Islam in the UK press having changed, the opportunity presents itself to consider an important methodological issue—how stable might the analysis of a discourse prove to be in the time following the completion of a study?

2

The Approach—Corpus Linguistics

While we situate the analysis undertaken here in the broad field of discourse analysis, the method that we use to approach the discourse is corpus-­assisted discourse studies (CADS; Partington 2003; Partington et al. 2004). CADS, as its name implies, relies on the methods of corpus linguistics. Corpus linguistics (McEnery and Hardie 2012) is an approach to the study of language based on the study of large volumes of attested language use. These collections of language data, called corpora (singular corpus), allow analyses which cycle between large-scale quantitative analyses and more traditional close reading of texts. The data is analysed, manipulated and navigated using programmes called concordance systems. For this study a system called CQPweb (Hardie 2012) was used. The main features of CQPweb used in this study are keywords, collocates and concordancing. Keywords are designed to bring into focus words which are unusually frequent (or infrequent) in one corpus when it is compared to another. Keywords have been shown (e.g. by Baker 2006) to be of particular use in CADS as they often help, for example, to characterise the overall content of the data set, the construction of groups within the data set and rhetorical devices used in the corpus. These are clearly useful things for any discourse analysis to achieve. Using keywords, we can achieve these things on a large scale, relatively objectively and swiftly. Collocation is another helpful procedure—it helps the analyst to look at the explicit and implicit construction of the meaning of a word in context (see, e.g. Taylor 2017). It does this by looking at words which have a strong preference to co-occur with the word in question—so-called collocates. This is determined using a suitable statistic, typically an effect size measure which shows the strength of attraction of the two words in ­question (see Gablasova et al. 2017, for a discussion of various collocation

218 

P. Baker and T. McEnery

measures). Being able to use this tool to look at implicit and explicit meaning as well as using it to contrast any differences between two corpora (e.g. one of broadsheet news stories about a group and another of tabloid news stories about the same group) has obvious applications within discourse analysis, especially in the area of the construction of identities and in and out groups. Given that there is growing evidence that collocates have some root in psychological reality (Durrant and Doherty 2010; Millar 2011), the value to the discourse analyst in using collocation to explore discourse is further strengthened. The final tool is the key tool which mediates between the relatively abstract large-scale analyses provided by keyword and collocation analysis; concordancing allows us to map back from the abstract quantitative analyses to the textual reality on which they are based. Concordancing allows us to navigate back to the examples in context that produce a keyword or a collocate, allowing us to rapidly scan those contexts to understand the finding in a more nuanced way. Alternatively, we may start with concordancing and work up to the more abstract level, exploring whether something we see in one text is unique to it, relatively rare, average or in some ways unusually frequent, for example. CADS uses the tools of corpus linguistics in order to subject corpora, both large and small, to discourse analysis. The subsequent analyses have the benefit of scale, can avoid an inadvertent cherry-picking bias that the exploration of isolated, and potentially atypical, texts may promote and has the advantage that some elements of the analysis are relatively subjective and reproducible. This chapter is an exploration of one discourse in the UK press which will use CADS both in order to illuminate how that discourse has changed, if at all, over time and, at the same time, to demonstrate briefly what the CADS approach can achieve. So in order to explore how stable the discourse around Muslims and Islam was in the UK press, we extended the original study, analysing a corpus of articles about Islam and Muslims from 2010 to 2014, which for convenience we will call Corpus B, making comparisons back to the findings from the original 1998 to 2010 study which was based on a corpus we will call Corpus A.

  The Value of Revisiting and Extending Previous Studies… 

2.1

219

Collecting the Articles

To be able to make a fair comparison, we collected Corpus B using the same method of collecting newspaper articles that we used for Corpus A. This involved gathering articles from the online news database Nexis UK. We collected articles between January 2010 and December 2014, using the same search term used to collect Corpus A. As before we collected national newspapers and their Sunday editions. However, even in the process of data collection we encountered our first example of how the discourse around these topics may have changed—since the collection of Corpus A, there had been changes to the availability of some newspapers. The Business was a weekly newspaper which went out of print in February 2008, so we were not able to collect that newspaper. Additionally, The News of the World stopped publishing in 2011 but was replaced with The Sunday Sun, so we have collected that newspaper as the Sunday equivalent of The Sun. These changes, in themselves, open the possibility of an overall change in discourse. Table 8.1 shows the number of words for both the older and the new corpus, for each newspaper.

Table 8.1  The structure of the two newspaper corpora Corpus A: 1998–2009 Newspaper

Total words

Percentage of the corpus

Business Express Guardian Independent Mail Mirror Observer People Star Sun Telegraph Times Total

577,234 6,419,173 24,344,632 25,591,916 17,216,224 8,067,444 10,264,984 663,192 2,669,466 5,018,404 16,125,825 29,621,874 146,580,368

0.3938 4.379286 16.60839 17.45931 11.74525 5.503768 7.002973 0.452443 1.821162 3.423654 11.00135 20.20862 100

Corpus B: 2010–2014 Total words

Percentage of the corpus

N/A 3,510,462 19,740,632 7,622,731 5,379,219 3,117,405 3,516,404 336,226 1,343,924 4,131,110 12,623,169 18,561,226 79,882,508

N/A 4.394532 24.71208 9.542428 6.733914 3.902488 4.40197 0.420901 1.682376 5.171483 15.80217 23.23566 100

220 

P. Baker and T. McEnery

It should be borne in mind that the broadsheets make up the majority of the data in both corpora because they have longer articles and more articles per issue than the tabloids. For the 2010–2014 corpus, just two broadsheet newspapers, The Guardian and The Times make up almost half of all the data. Adding in the other broadsheets (The Independent, The Observer and The Telegraph), amounts to three-quarters of the data. As a proportion of the corpus, The Guardian (a liberal newspaper) actually contributed more in Corpus B than Corpus A (8% more), so it should be borne in mind that the overall composition of corpus data is different between the two time periods examined, and this may subsequently impact on results found.

3

Number of Articles per Month

Looking at the volume of data produced and the trend over time, two features of the Baker et al. study are shown to be stable—the volume of data increases at points in time where violent events involving Muslims occurs and, over time, the overall trend is for more articles to be written 350 300 250 200 150 100 50

1998-01 1998-08 1999-03 1999-10 2000-05 2000-12 2001-07 2002-02 2002-09 2003-04 2003-11 2004-06 2005-01 2005-08 2006-03 2006-10 2007-05 2007-12 2008-07 2009-02 2009-09 2010-04 2010-11 2011-06 2012-01 2012-08 2013-03 2013-10 2014-05 2014-12

0

Fig. 8.1  Average number of articles about Islam per newspaper per month, 1998–2014

  The Value of Revisiting and Extending Previous Studies… 

221

mentioning Muslims and Islam. Figure 8.1 shows both of these points. The figure combines results from Corpus A and Corpus B to show long-­ term trends in the data. This graph shows a trend line indicating that press interest in Islam and Muslims has continued to increase, particularly since the events of 9/11. The rise of the Islamic State and its invasion of Iraq in 2014 appear to have caused a third large spike in reporting (with 9/11 and 7/7 being two other spikes). But since the end of 2009, there appears to have been a notable upturn in reporting, following a general fall from 2005 to 2009. Yet the overall volumes of data over time, while indicative of a potential change, does not describe or account for it. To look at this issue, we decided to explore the differences between Corpus A and Corpus B using the keywords procedure. We wanted to focus on keywords which indicated a difference between the two periods and then to reflect on how those keywords either meshed with, or indicated a change in, our previous findings. So we contrasted the language of the 1998–2009 articles with those from 2010 to 2014 to identify words which occurred significantly more frequently in one of these periods when compared against the other. These keywords were then used as a way of exploring the continuation or change of findings in our original study of Corpus A.

4

Stability—What Has Remained the Same

Stability in the analysis across the two periods was the exception. Yet there were some broadly stable results. Firstly, the relative frequency of the word Islamic, shown by Baker et al. (2013, 44) to be associated with extremism, retains this association. Secondly, the strong association of Muslims and Islam with conflict in the UK press (Baker et  al. 2013, 62–63) has also been sustained. Thirdly, the ‘horror discourse’ around the veil identified in 1998–2009 is still present in 2010–2014. This was identified in the first study (Baker et al. 2013, 216–217) as a way of framing Muslims in Islam in terms of science fiction or gothic horror—for example by comparing them to movie of literary monsters: and in the original and new study it is easy to find examples such as women wearing the veil as looking like Daleks or Darth Vader. Finally, the original study reported

222 

P. Baker and T. McEnery

that the phrase devout Muslim was negatively loaded; this is still true in the 2010–2015 articles where we found references to devout Muslims described them as cheating (e.g. failing drug tests, having affairs etc.), becoming radicalised, engaging in extremist activity or clashing with ‘Western’ values in some way.3 However, even in these cases, there are some slight changes. For example, reporting of the Islamic State group’s activities has served to intensify the association of the word Islamic with extremism. Similarly, in the 1998–2009 corpus the word forms relating to conflict (see Baker et al. 2013, 59) constituted 2.72% of the corpus. For the 2010–2014 data these words constituted 2.75% of that corpus. So while overall these findings have remained the same, minor details relating to the findings have been subject to flux. However, these changes are minimal by comparison to the major changes that have taken place across the two time periods, hence the bulk of this paper will be devoted to a discussion of these differences.

5

What Has Changed

The dynamic nature of discourse is shown clearly when we consider other findings arising from Baker et al. (2013). In the following subsections, we will consider a range of findings from the first study and show how, in the second study, these features have changed. Throughout, the figures quoted in the tables are normalised frequencies per million words as this allows for an ease of comparison between the two corpora. In the analysis that follows, after an initial focus on which countries are mentioned and the prevalence of conflict lexis in the corpora, we will focus on six key questions, relating to major findings by Baker et al., to illustrate the dynamism of the discourse across the two periods with regard to how, in Corpus B relative to Corpus A: • the words Muslim, Muslims and Islam are used; • Muslim women and men are represented;  For example, ‘A JUDGE yesterday ruled that a devout Muslim woman must remove her full face veil if she gives evidence in court’ (The Sun, January 23, 2014). 3

  The Value of Revisiting and Extending Previous Studies… 

• • • •

223

different branches of Islam are constructed; strength of belief is communicated and used; extremism is constructed; radicalisation is discussed.

These investigations are guided by words which are key when the two corpora are contrasted.

5.1

Country and Conflict Keywords

In terms of location, there has been a shift away from stories about conflicts or attacks in Iraq, Palestine, and America, which are key when Corpus A is compared to Corpus B.  When Corpus B is compared to corpus A, we find instead that in Corpus B Syria, Libya, Iran and Egypt are key. In terms of conflict, as noted, the strong relationship with both Muslims and Islam is relatively stable across the two corpora. However, when the lexis used to realise the presentation of conflict in the two corpora is examined, a clear difference emerges. The top keywords (in descending order) in Corpus A, when compared to Corpus B, are war, terrorist, terrorists, attacks, bomb, bombs, terrorism, suicide, invasion, destruction, raids, and hijackers. Key in Corpus B, when compared to Corpus A are islamist, rebels, crisis, revolution, protesters, protest, sanction, rebel, activists, uprising, islamists, jihadists, jihadist and jihadi. How can we interpret these findings? World events tend to be a major driving force in the contexts that Muslims and Islam are written about—such events align well with news values. We therefore hypothesise that since 2009 references to terrorism have fallen sharply in articles about Islam particularly because large-scale orchestrated attacks like 9/11 and 7/7  in Anglophone countries in particular have been less marked. Many words which directly refer to conflict have also seen sharp falls: war, bomb, raids, destruction, attacks. Yet other words relating to conflict of a principally civil kind have increased, such as crisis, revolution, protests, sanctions and uprising. While stories about armed conflict have not gone away, ­reference to political/civil conflict has risen dramatically. This makes us reflect

224 

P. Baker and T. McEnery

again upon the apparently stable finding linking Muslims and Islam with conflict. While the picture in terms of the frequency of conflict words appears relatively stable, the relative proportions of the different types of conflict words are not stable. Concerns over Iran’s nuclear intentions, and reporting of events around the Arab Spring have replaced the focus on the Iraq war and 9/11. While mentions of al-Qaeda and the Taliban have been reduced, they have been replaced by other groups like Islamic State, Boko Haram and the Muslim Brotherhood. There are also more references in 2010–2014 to rebels, activists, Islamists, protestors and jihadists. So rather than being framed around fear of terrorist attacks, the discourse between 2010 and 2014 is more linked to revolution, political protest and Islam as a political force. The concept of jihad and those engaged in it (while less frequent than some of the other terms) has also risen over time. These changes in turn impact on the frequency of the selection of different items of conflict lexis.

5.2

Muslim, Muslims and Islam

In our original study we were particularly interested in words which appear next to Muslim, Muslims and Islam (as well as Islamic, which has been discussed above) as, if they are frequent enough, they are likely to trigger the concept even before it is mentioned. For example, if a person encounters the phrase ‘Muslim fanatic’ enough times, we hypothesise that they are likely to be primed to think of the word fanatic if they hear the word Muslim by itself. So, as in the initial study, we looked for patterns like Muslim [X], as well as [X] Muslims and [X] Islam.

5.2.1  Muslim The top ten words immediately following Muslim in Corpus A are, in descending order of frequency, community, world, council, women, leaders, countries, cleric, country, men and communities. In Corpus B they are brotherhood, community, women, world, men, woman, communities, convert, countries and council.

  The Value of Revisiting and Extending Previous Studies… 

225

The change, tokened by the contrast between the two lists, is quite marked. We see a strong rise in the phrase Muslim Brotherhood, indicating the salience of stories coming out of the Arab Spring and the uprising in Egypt. In 2010–2014 Brotherhood follows over 1 in every 10 mentions of Muslim. The Muslim council appears to be of less interest to journalists in the later period, as do Muslim leaders and the phrase Muslim cleric. So apart from the Muslim Brotherhood, it appears that there is now less focus on people or groups who are seen as leading various Muslim communities. The term Muslim convert has become more common though, although this term usually refers to stories about Muslim converts who are involved in crime, usually terrorism or militancy, for example: In the feverish atmosphere of Kenya’s war on terror, rumours abound as to the whereabouts of Samantha Lewthwaite, the Muslim convert from Aylesbury who is on the run after a foiled bomb plot. (Telegraph, July 16, 2012)

Yet there is an element of stability in the word—Muslim world, Muslim countries and Muslim community continue to be phrases used frequently to refer to Muslims. By continuing to use terms like Muslim world, Muslim community and Muslim countries, the British press continue to collectivise large numbers of Muslims as living in similar circumstances, implying that they are separate from everyone else. What of the plural form, Muslims? For Corpus A, the top ten right-­ hand modifiers are British, young, Shia, moderate, Sunni, Bosnian, fellow, other, radical and devout. For Corpus B, they are British, young, Sunni, Shia, moderate, Bosnian, fellow, other, radical and devout. This example is of interest as, although the two wordlists appear very similar, rank orderings within the lists have changed, indicating a shift in focus in the discourse. Similarly, the normalised frequencies for some words have changed, even though their ranking has not. For example, the ordering of mention of Sunni and Shia Muslims has been reversed. Also, although references to British Muslims have decreased over time (from Table 8.1 examples per million words in Corpus A to 4.9 in Corpus B), this is still the most common way that Muslims are referred to. Further

226 

P. Baker and T. McEnery

changes in the representation of Muslims emerge when we consider what British Muslims are reported as doing in Corpus B. They are described as alienated, sometimes impressionable and prone to radicalisation: The alienation of many young British Muslims goes deep—some despise society enough to want to harm it. (The Times, February 2011) Maybe after the 7/7 bombings and numerous failed plots to blow up aeroplanes, nightclubs, an airport and a shopping centre it’s not so surprising. Nobody knows the true level of radicalisation among British Muslims. (The Guardian, July 2010)

They are described as having travelled (often to places like Syria to join ISIS): According to our intelligence agencies, some 500 young British Muslims have travelled to Syria to join the Islamic State or other terrorist organisations. (The Mail, August 2014)

They are expected to condemn jihad and terrorism (but a minority are sometimes described as not doing so): While most British Muslims wholeheartedly condemn the killers, we know from bleak experience that a significant minority will tell opinion pollsters they actually endorse what was done (The Mail, May 2013)

While the link through collocation of Muslims to alienation and radicalisation is reported in Corpus A by Baker et al. (2013, 42) travelled is not. This is strongly indicative of a shift in the discourse to focus on the issue of Muslims travelling to join jihadi groups and the wish for British Muslims to discourage such travel by condemning it. In the case of condemn, in particular, the press has shifted its stance from a neutral one in Corpus A, where it discussed the reaction or response of Muslims to terrorist acts, to a situation where it was seeking to shape that response, that is, encouraging them to condemn such acts. A focus on the word young reveals further differences. This collocate is not reported in the analysis of Muslims in Corpus A by Baker et al. (2013,

  The Value of Revisiting and Extending Previous Studies… 

227

42–44) yet in Corpus B it is notable that young and British both often appear together at the same time as modifiers of Muslims. As a result, many of the collocates of young Muslims are the same as those of British Muslims. Those for young Muslims actually show a stronger concern about radicalisation. Young Muslims are described as impressionable, disaffected, rootless, angry and susceptible. They are at risk of being lured, recruited, indoctrinated or brainwashed to commit crimes or jihad. The Prison Service has long been concerned at the spread of radical Islam inside Britain’s jails. Experts say a tiny number of fanatics, most serving long sentences, have huge influence over disaffected young Muslims. (The Sun, May 2013) Cameron said a clear distinction must be made between the religion of Islam and the political ideology of Islamist extremism, but the ‘non-violent extremists’ who disparage democracy, oppose universal human rights and promote separatism were also ‘part of the problem’, because they lure young Muslims onto the path of radicalisation, which can lead them to espouse violence. (The Times, February 2011)

So while Muslims are discussed as a collective group, the most salient pattern is in the context of the radicalisation of young British Muslims. The last word we will consider in this section relates to the belief system that Muslims follow, Islam. In this case looking at the collocates immediately preceding the word is revealing. In Corpus A, the top ten such collocates are radical, militant, political, Shia, insulting, fundamentalist, anti, British, insult and moderate. For Corpus B it is radical, anti, militant, political, Shia, Sunni, moderate, fundamentalist, insulting and British. Note that these lists of collocates are similar, but not the same. One collocate is unique to each list—insult for Corpus A and militant for Corpus B. Also, these lists are ordered from strongest to weakest collocates—hence it is clear to see that the rank ordering of the collocates has changed too. These minor changes, however, mask some real similarities. As with Islamic, the word Islam continues to be associated with extremism, with the words radical, militant and fundamentalist appearing in the top 10 left-hand collocates (although moderate also appears). Other words

228 

P. Baker and T. McEnery

suggest conflict: insulting, insult (which have both decreased over time) and anti (which has increased over time): Dorje Gurung, a Nepalese teacher in Qatar, who was imprisoned for 12 days in May 2013, after a 12-year-old student accused him of insulting Islam, believes the men are likely to be held without any form of information or support. (The Guardian, September 2014)

Note, however, that this extremism has been reinforced in Corpus B by the new top ten collocate militant, which reinforces this discourse of extremism in the corpus focussed upon the word Islam. By the same token, however, the loss of insult from the top ten collocate list for Corpus B perhaps tokens a weakening of the link of the word to insult Islam. Nonetheless, the collocate insulting is still present in the Corpus B top ten collocate list, meaning that the link between Islam and the process of insulting persists, though the reality of insults to Islam sometimes appear questionable according to the narrative voice of some articles. Saudi liberal activist Raif Badawi was sentenced to 1,000 lashes, 10 years in prison and a heavy fine for insulting Islam. In fact, his crime was to establish an online discussion forum where people were free to speak about religion and criticise religious scholars. (The Independent, May 2014)

Overall, Islam is still often constructed in Corpus B as an entity that is prone to feeling insulted, with any insults provoking strong reactions from its adherents.

5.3

Representation of Gender

The terms Muslim women and Muslim men are frequent in the corpus. We found that in the previous study, Muslim women tended to be discussed in terms of the veil, due to a debate which took place in 2006 after comments made about veiling by the then Home Secretary Jack Straw. Muslim men were most often discussed in terms of their potential for radicalisation. How have Muslim men and women been written about since 2010?

  The Value of Revisiting and Extending Previous Studies… 

229

Table 8.2 shows collocates (words which frequently occur near or next to the word or phrase we are interested in) of Muslim women in the two time periods—we only considered content words (nouns, verbs or Table 8.2  Collocates of Muslim women Category

Collocates: 1998–2009

The veil and dress wear, veils, veil, wearing, remove, worn, hijab, full, Straw, cover, Jack, niqab, faces, face, headscarves, veiled, Straw’s, dress, covering, head, headscarf, traditional, burkas, burka, wore, dressed, covered, veiling, burqa, burkha young, men, children, women, Identity words people, girls, old, generation, (age, gender husbands, marry and relationships) Reporting and said, saying, feel, asked, call, feeling told, asking, comments, say, ask, talk, revealed, believe, suggested, called, calling, urged, prefer, hope, know, wants, claimed, speak, question, warned, understand Locations British, Britain, country, world, London, English, western, Bosnian, constituency, society, Blackburn, France, Europe, street, Serb, town, French, Arab, community Freedom rights, allowed, forced, ban, issue, choose, debate, power, support, help, free, allow, row, required, banned, encourage, campaign, choice, freedom Religion Others

Collocates: 2010–2014 wear, wearing, veil, face, veils, burka, full, faces, worn, dress, cover, hijab, veiled, covering, niqab, public, remove, headscarves, burkas, modesty, coverings

young, children, women, men, people, group, marry, girls, woman, said, says, see, feel, talking, think, say

British, Britain, France, French, English, London, world, country, central, western, Arab

ban, rights, right, allowed, law, banned, oppressed, choose, allowing, debate, free

Islamic, faith, religious role, take, way, rape, protect, minority, majority, come, made, designed, swimming, sport, living

230 

P. Baker and T. McEnery

a­ djectives) which occurred 10 times or more. Words that are unique to 2010–2014 are shown in bold. We first examined some of the new collocates of Muslim women. For example, modesty is used to refer to wearing the burka or hijab. Modesty is generally represented as a positive aspect of being a Muslim woman, although some authors cite this point in order to problematise it: For many, the hijab represents modesty and freedom of choice, but we cannot ignore that it is also one of the most contentious and divisive issues of modern times—within the Muslim community as well as outside it. (Guardian, February 16, 2010)

Other authors imply that adherence to modesty does not necessarily mean that a Muslim woman cannot be stylish: She is part of an expanding group of bloggers in London and America, known as hijabistas, who are proving that it is possible to wear a hijab, a symbol of modesty, and be stylish at the same time. (Sunday Times, January 19, 2014)

Swimming referred mostly to stories about swimming pools that ran special sessions for Muslim women only. Such sessions were viewed as problematic and contribute towards a wider discourse that was frequently encountered in the 1998–2010 corpus of Muslims receiving ‘special treatment described as unnecessary’: The Walsall debacle comes six months after The Daily Express revealed how Hull City Council was accused of running Muslim women-only swimming sessions in secret—to the fury of regular baths users. (The Express, July 6, 2010)

The collocate rape most often refers to atrocities that took place in Bosnia in the 1990s: In 1996 a UN tribunal indicted eight Bosnian Serb soldiers and policemen for the rape of Muslim women. (Daily Mirror, June 12, 2014)

  The Value of Revisiting and Extending Previous Studies… 

231

Sport is seen as an area which Muslim women should be encouraged to get involved in more: RIMLA AKHTAR, winner of the Community Award, believes more needs to be done to engage Muslim women in sport at the grassroots level, writes Andrew Longmore. (Sunday Times, December 8, 2013)

Of the 18 times that oppressed collocates with Muslim women, almost all of them cite the idea that Muslim women are oppressed in order to disagree with the notion: ‘The media portray Muslim women as oppressed and subjugated and Islam is often presented as misogynist and patriarchal,’ she said, and her book was intended as an antidote to that. (The Observer, March 16, 2014) People use this idea that Muslim women are oppressed as an excuse for pulling off their head coverings. (The Sunday Times, July 20, 2014)

The collocate designed most often refers to an all-in-one swimsuit for Muslim women, called a burkini, with the following case giving a somewhat unflattering description: TV chef Nigella Lawson has admitted she resembled ‘a hippo’ when she wore a burkini on Bondi Beach. The 51-year-old caused a storm two years ago by donning the all-in-one swimsuit designed for Muslim women during a visit to Australia. (The Sun, February 25, 2013)

Finally, role collocates with Muslim women 12 times of which 8 are used positively in the phrase role model. I want to give something back to the community and be a positive role model for young Muslim women—and for young women in general. (Daily Star, March 25, 2011)

So since 2010 there has been a small but significant increase in positive discourses around Muslim women, particularly in terms of questioning their oppression or discussion of positive female role models. However, the main picture is a continuation of older discourses which focus on

232 

P. Baker and T. McEnery

Muslim women as victims, receiving special treatment, victimisation or problematising their dress. In terms of stability in the discourse, the table above suggests that the main focus around Muslim women—veiling—has not actually changed, with a similar set of words relating to veiling and dress collocating mostly with Muslim women. Other categories in the table tend to be linked to veiling, particularly the one relating to freedom. So, has the debate around veiling changed at all since 2010? In the 1998–2009 study we concluded that the veil was discussed negatively and seen as an ‘issue’, characterised by ambivalence and conflict. Some Muslim women were described as oppressed by the veil, others as demanding to wear it. There were a wide range of arguments given as to why Muslim women should not veil, and they were discussed in metaphors relating to things that glide on water, soft furnishings and (most negatively) frightening supernatural monsters. To examine the 2010–2014 articles, we retrieved cases of the following pattern: insist*/demand*/force*/right/cho* to wear/wearing [the veil or any related clothing item] We only counted cases where the construction uncritically or unquestioningly referred to the veil as a choice, right, demand or imposition. Table 8.3 shows what patterns we found for each newspaper. We can see that the most frequent construction here is of Muslim women being forced to wear the veil. The two negative constructions Table 8.3  Patterns around veiling for Muslim women

Times Telegraph Sun Mail Express Star Mirror Guardian/ Observer Independent Total

Forced to wear it

Right to wear it

Choosing to wear it

Demanding to wear it

10 11 3 4 4 0 5 10

3 1 0 2 0 0 0 2

4 9 2 3 2 0 2 14

5 6 3 4 2 3 1 6

1 48

2 10

2 38

1 31

  The Value of Revisiting and Extending Previous Studies… 

233

(force and demand/insist) are more frequent together than the more positive ones (right or choice). Uncritical descriptions of the veil as a right were relatively infrequent. We note the higher frequency of the veil being described as a choice by The Guardian, although this newspaper also has fairly high representations of it being linked to compulsion as well. Table 8.4 compares proportions of these constructions of wearing the veil to the earlier set of data. Over time, the veil is more likely to be described in negative terms, either as Muslim women being forced into wearing it, or in terms of them demanding or insisting on wearing it. Discussion of the veil as a right appears to have sharply declined, although it is slightly more likely to be described as a choice. We also looked at arguments given for why Muslim women should not veil. This was found by carrying out a search on terms describing the veil, appearing in the same vicinity as the word because. Of the 135 cases of these, 32 gave arguments as to why a Muslim woman should not wear the veil. These are shown in Table 8.5. The argument about the veil (particularly face-covering veils) making communication with the veil-wearer difficult was the most frequently cited. In particular, a court case where a veiled female juror was asked to step down was mentioned, as well as there being references to schoolteachers who veil their faces. I’m with Ken Clarke when he says that women should not be allowed to wear the full-face veil in court because it is difficult to give evidence from inside a kind of bag. (Daily Mail, November 5, 2013, Richard Littlejohn) People are nervous about speaking to burka wearers. That’s because we want direct communication, not just through eye contact but through Table 8.4  Patterns around veiling—change over time (summary)

1998–2009 2010–2014

Forced to wear it (%)

Right to wear it (%)

Choosing to wear it (%)

Demanding to wear it (%)

28.5 37.7

34.8 7.8

25.5 29.9

10.6 24.4

234 

P. Baker and T. McEnery

Table 8.5  Arguments against veiling Argument against veiling

Frequency

It makes communication with the wearer difficult It’s a symbol of oppression of women It’s alien to the culture it’s been worn in It reduces your field of vision It compromises national security It’s a symbol of extremist Islam which is seen as intolerant It overexcites men It’s false consciousness It’s offensive

13 5 3 3 3 2 2 1 1

interesting and sometimes revealing facial expressions. We want to see the lips move. (The Sun, April 10, 2010)

Again, the picture is of stability at one level, change on another. There is still a focus on the veil in the second corpus. However, the main argument against Muslim women wearing the veil has changed from the oppression of women (1998–2009) to a focus on difficulties surrounding communication with the veil-wearer (2010–2014). The increase in arguments relating to difficulties in communication could perhaps be seen as ‘strategic’ by those who oppose the veil, as it does not require its opponents to make claims about Islam’s ideology or attitude towards women. Instead, the argument focusses on a more ‘practical’ concern, which may be difficult to counter as being Islamophobic. This could suggest that opponents of Islam are developing more careful and subtle arguments to support their views. Having explored the changing representation of Muslim women in the UK press, let us now consider Muslim men. Table 8.6 shows collocates (occurring 10 times or more) of Muslim men for the two time periods. Again, those in bold are unique to the last period. Apart from the reporting words, many of the categories used to discuss Muslim men are the same in both periods, with discussion of them as victims of violence and also as guilty of crimes. There is a new category though: abuse. The words grooming, abuse and gangs collocate with Muslim men in stories about the sexual abuse of (often) white girls.

  The Value of Revisiting and Extending Previous Studies… 

235

Table 8.6  Collocates of Muslim men Category

Collocates: 1998–2009

Identity words (age, gender and relationships) Locations

young, boys, women, children, boys, young, women, wives, whore, dating girls, marry, children

Killing

Reporting and feeling Law and order

Radicalisation/ terrorism

Srebrenica, Bosnian, British, Britain, Serb, Serbs, Iraq, London, Asian, town massacre, killed, massacred, slaughter, murder, murdered, slaughtered, died, war, killing, suicide, bombers said, say, believed, feel, says, told arrested, accused, executed, innocent, alleged, trial, law impressionable, disaffected, radicalised, radicalisation, training, terror, radical, Abu

Abuse Other

white, allowed, disgraceful, angry, born, beards, forces, gathered, way, dignity, see, get

Collocates: 2010–2014

Srebrenica, Bosnian, British, Pakistani, Serb, Britain, Asian massacre, killed, murdered, killing, slaughter, massacred, deaths, unarmed, murder

accused, arrested, charged, guilty, convicted, found radicalised

grooming, gangs, abuse, force white, forces

And when Jack Straw condemned the grooming by British Muslim men of Pakistani origin of vulnerable white girls, he was instantly flamed as a bigot. (The Times, January 22, 2011) This was far from a one-off case. Police operations going back to 1996 have revealed a disturbingly similar pattern of collective abuse involving small groups of Muslim men committing a particular type of sexual crime. (Daily Mail, January 10, 2011) The authorities have been just as reprehensible in their reluctance to tackle the sickening exploitation of white girls by predatory gangs of Muslim men. (The Express, May 17, 2012)

Force does not occur in stories about sexual abuse but relates to cases where Muslim men apparently force women to wear the veil.

236 

P. Baker and T. McEnery

We also note that the category of radicalisation/terrorism has fewer words in it in 2010–2014. However, overall, the concept of radicalisation has grown during this period (as shown earlier). What seems to be the case is that it is not as gendered a concept as it previously was. When the British press speak about radicalisation, they talk about young British Muslims, but do not mention gender as much as they did in the past. So radicalisation has become a more gender-neutral subject than it used to be. In this category only one word: radicalised occurs as a collocate with Muslim men more than 10 times. However, this occurs in similar ways to its use in 1998–2009, with fears about Muslim men being radicalised (particularly young Muslim men). The father of the Muthana brothers, Ahmed Muthana, suggested yesterday that young Muslim men were being radicalised at ‘pop-up’ meetings in Cardiff rather than at any mosque or via internet videos. (The Guardian, June 24, 2014) THOUSANDS of lone wolf extremists could launch similar attacks to the Woolwich bloodbath, a senior police officer warned yesterday. And Assistant Met Commissioner Cressida Dick said the threat cannot be eradicated while young Muslim men are radicalised via the internet. (Daily Mirror, December 20, 2014)

In Corpus B, stories about Muslim men in the British press have focussed around them as either victims or perpetrators of crime, with particular focus on the sexual abuse of white girls or the risk of them being radicalised.

5.4

Branches of Islam

The earlier study found that about half the newspapers refer to Islam generally rather than discussing different branches of Islam like Sunni, Shia and Wahhabi. Is there any evidence that this behaviour has changed? Figure 8.2 shows the proportions of times that a newspaper refers to these Sunni and Shia Islam in both Corpus A and Corpus B. It shows the proportions for each newspaper for the two time periods.

  The Value of Revisiting and Extending Previous Studies… 

237

3 2.5 2 1.5 1 0.5 0

Fig. 8.2  Proportion of mentions of different branches of Islam for each newspaper

The first bar shows 1998–2009, while the second shows 2010–2014. It can be seen that The Independent has greatly increased the proportion of times it refers to branches of Islam as opposed to writing more generally about Islam. Six other newspapers have also gone in this direction (although not hugely). However, The Guardian, Telegraph, Express and Star have gone the other way and refer to the branches less than they used to. Generally, a distinction can be made between the broadsheets and the tabloids here, with all the broadsheets referring more often to branches of Islam rather than Islam in general, while the reverse is true of the tabloids. So again, we have some stability. British tabloids continue to paint a simplistic picture of Islam, not usually referring to or distinguishing between different branches like Sunni and Shia, although The Mirror is the tabloid that makes the most effort to do this. On the other hand, all the broadsheets are more likely to refer to branches of Islam as opposed to Islam itself, with The Independent being most likely to do this. Yet within this overall picture of stability, variation by newspaper can be notable, especially with regard to The Independent and The People. The change underlying this apparent stability becomes all the more obvious if

238 

P. Baker and T. McEnery

100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014

Fig. 8.3  References to Sunni, Shia, Sufi, Salafi and Wahhabi over time. Dark grey denotes the proportion of mentions of references to branches of Islam (e.g. Sunni, Shia, Wahhabi); light grey bars denote references to Islam

we consider change over time not by period covered by the corpus, but by year. Figure 8.3 shows how overall references to different branches of Islam have changed since 1998. Since the start of the new collection of data (2010), newspapers have begun once again to increasingly make distinctions between different branches of Islam, as opposed to simply referring to Islam itself. However, such references often relating to fighting between Sunnis and Shias (often in Iraq) and the Sunni uprising in Syria.

5.5

Strength of Belief

This section examines how Muslims and Islam are associated with different levels of belief. Phrases like Muslim extremist and Muslim fanatic were found to be extremely common in our earlier study, and one way of gauging whether representations of Muslims have changed is to examine whether such terms have increased or decreased. We would argue that the presence of such terms, particularly in large numbers, is of concern as

  The Value of Revisiting and Extending Previous Studies… 

239

readers of newspapers would begin to unconsciously associate Muslims and Islam with extremism. Such terms tend to be more frequent in the broadsheets, like The Guardian and The Times, simply because these newspapers contain more print than the tabloids. So simply counting the number of times a phrase occurs in a newspaper, and then comparing across newspapers, is not the best way of ascertaining the scope of the problem. Instead, we have taken an approach which looks at the proportion or percentage of cases a word like extremist occurs next to Muslim in each newspaper. This takes into account the fact that some newspapers may mention Muslims a lot and some not very much at all. Our initial study looked at labels which could be used to refer to people. So we were interested in terms like Muslim extremist or fanatical Muslim. We classified three different types of belief, as shown in Table 8.7. As the table includes words based on labels for people, it does not consider words for related abstract concepts like extremism, radicalism, militancy and separatism. Nor does it consider words related to processes like radicalisation. In fact, as shown earlier, the words extremism and radicalisation were found to be significantly more common in the 2010–2014 articles, compared against the 1998–2009 articles. The term extremism occurs after Islamic 15% of the time, while extremist(s) occurs after Islamic in a further 31% of cases. So before looking in more detail at other terms, it is worth bearing in mind that in articles about Islam the newspapers are more likely to write about the concept of extremism than they used to. Hence even if they have reduced the number of times they talk about Islam or Muslims as extremist, there is evidence that there has been a shift in language towards Table 8.7  Levels of belief Level of belief Extreme

Strong Moderate

Words considered Extremist, extremists, fanatical, fanatic, fanatics, firebrand, firebrands, fundamentalist, fundamentalists, hardline, hardliner, hardliners, militant, militants, radical, radicals, separatist, separatists Devout, faithful, orthodox, pious Liberal, liberals, moderate, moderates, progressive, progressives, secular

240 

P. Baker and T. McEnery

a greater emphasis on the abstract idea of extremism. This may make the articles superficially less personalised, although it does not remove the general focus on extremism. As found with the 1998–2009 data set, extremism is more likely to be associated with the word Islamic, than Islam or Muslim(s). Proportionally, The Star uses the most extremist words next to Islamic in 22% of cases (almost 1  in 4). Compare this to The Guardian which does this 6% of the time (about 1  in 17 cases). The Express is the newspaper most likely to associate Islam with an extremist word (1 in 10 cases), while The Mirror does this least (1 in 42 times). For Muslim and its plural, it is The Express again which has the highest use of extremist associations (1 in 13 cases), and The Guardian which has the least (1  in 83 cases). However, overall in the British press, Muslim(s) occurs next to an extreme word 1 in 31 times, for Islam this is 1 in 21 and for Islamic the proportion is 1 in 8. The picture for the words Muslim and Muslims combined shows that fewer uses of the word Muslims are linked to extremism overall, with the proportion in 1998–2009 being 1 in 19, while it is 1 in 31 for 2010–2014. The People shows the largest fall in this practice, although we should bear in mind that this is based on a much smaller amount of data than for the other newspapers (e.g. The People mentions Muslims less than 500 times overall in the period 2010–2014, compared to The Guardian which has over 20,000 mentions in the same period). However, all newspapers show falls in this practice overall. For the word Islamic, there are also falls in its association with extremism, with the average number of mentions of an extremist word next to Islamic being 1 in 6 in 1998–2009 and 1 in 8 in 2010–2014. The Star and Sun are most likely to link the two words, while it is the least common in The Guardian and its sister newspaper The Observer. The picture for the word Islam is somewhat different, however. Here the average number of mentions of an extreme word near Islam has actually increased slightly, from 1  in 25 to 1  in 21. The practice has become noticeably more common in The Express, although most newspapers have followed suit. Only The Mirror and The Telegraph show a move away from this practice. What of the moderate words? It is The Express, Mail and People which are more likely to refer to Muslims as being moderate, with this practice

  The Value of Revisiting and Extending Previous Studies… 

241

being least common in The Mirror. On average it is Muslims who are more likely to be called moderate (1 in 161 cases), as opposed to the concept of Islam (1 in 271 cases). However, these figures are much smaller than those for the extremist words. For the 2858 mentions of extreme Muslim(s) in the press, there are only 558 moderate Muslim(s), or rather 5 extremists for every moderate. However, in the 1998–2009 articles, there were 9 mentions of extremist Muslims for every moderate, so we can see evidence that moderate Muslims are starting to get better representation proportionally, although they are still outnumbered. As Fig.  8.4 suggests, this is not because moderate Muslims are being referred to more, it is more due to a dip in mentions of extremist ones. For Muslim and its plural, it is The People, Express and Mail which have shown greater increases in mentions of moderate Muslims. However, on average, the number of mentions of moderate Muslims has gone up but only slightly (now 1 in 161 cases). For cases of Islamic occurring next to a moderate word, this was never common, and has actually fallen slightly. Figures are based on low frequencies, however, and as we have seen earlier, the word Islamic shows a 20 18 16

1998-2009 2010-2014

14 12 10 8 6 4 2 0

Fig. 8.4  Summary of all data, comparing proportions of change over time

242 

P. Baker and T. McEnery

very strong preference for extremist associations. For Islam, there is a slight increase in the proportion of times it occurs next to a moderate word, although again, this is very low overall. Finally, let us consider the words devout, faithful, orthodox and pious, and how often they occur next to Muslim(s), Islam and Islamic. These words are more likely to occur next to Muslim(s) (about 1 in 140 times) with only a handful of cases of them occurring next to Islamic and Islam. These strong belief words were barely present in the 1998–2009 articles next to Islam and Islamic too, so we will only consider change over time for the word Muslim(s). What we see are falls for all newspapers, with an average of 1 in 92 references to devout Muslims and similar terms in 1998–2009, dropping to 1 in 140 such references in 2010–2014. Figure 8.4 gives a better sense of the overall picture. We can see, for example, how extremist words are most likely to occur near Islamic, but also how there has been a drop in this association over time, as well as a similar drop in the way that extremist words occur next to Muslim(s). We can also see (small) increases in associations of extremist words with Islam, and moderate words with Muslim(s). So while references to Muslims as extremists have fallen in the British press since 2010, journalists are writing about the abstract concept of extremism much more frequently than they used to. Also, the concept of Islam is more likely to be referred to as extreme than it used to be. Extremism is still a hot news topic, but in this context, there has been a move towards focussing more on the religion rather than on its adherents. The ratio of mentions of extreme Muslims to moderate Muslims in the British press is becoming slightly more equal (from 9 to 1  in 1998–2009 to 5 to 1  in 2010–2014). References to devout Muslims have fallen since 2009, although these are still highest in The Mirror. Relatively, The Star and Express have the most references to Islamic extremists, while The Guardian has the fewest. However, the Express refers to moderate Muslims the most too. The Mirror and People have dramatically reduced their references to Islamic extremists. Given the salience of extreme words in Table 8.8, let us now explore those words in some more detail.

  The Value of Revisiting and Extending Previous Studies… 

243

Table 8.8  Extremism keywords Fundamentalism Fundamentalist Fundamentalists Extremism Radicalised Radicalisation Grooming

5.6

1998–2009

2010–2014

22.4 38.1 27.1 42.8 8.3 6.1 2.7

10.3 18.8 13.0 69.6 22.3 19.9 11.2

↓ ↓ ↓ ↑ ↑ ↑ ↑

Extremism Keywords

There has been a strong decline in the words fundamentalism, fundamentalist and fundamentalists. Such words were found to have been particularly strong in the years 2001 and 2004, and they have not returned. However, the abstract concept of extremism (as opposed to a people-­ labelling word like extremist) is more frequent in Corpus B than in Corpus A, as well as three related words, radicalised, radicalisation and grooming. This indicates a major shift in language around extremism, which is more concerned with the process of becoming extreme, rather than labelling people as such.

5.7

Reasons for Radicalisation

Radicalisation has been mentioned in several sections of this chapter so far, so a closer look at radicalisation and the claimed causes of radicalisation seems appropriate—especially as it points to a major difference between Corpus A and Corpus B and hence a source of dynamism in the discourse over time. References to radicalisation have increased since 1998–2009, with there being almost double the number of mentions of that word in the shorter 2010–2014 data set. The main pattern we see over time is increasing attribution of blame for radicalisation on extremist Islam—in 1998–2009 this occurred in 1 in 3 cases. By 2014 it is 2 in 3 cases. Yet the attribution of blame to government policy has decreased over time, from 1 in 3 cases in 1998–2009 to 1 in 12 in 2014. The invasions of Iraq

244 

P. Baker and T. McEnery

and Afghanistan, while still mentioned, are now seen as almost historical factors attributable to the ‘Labour years’, rather than as being relevant to the present situation. Two of the less frequent explanations for radicalisation found in the 1998–2009 data, ‘grievance culture’ and ‘multi-culturism’, seem to have largely disappeared from the discourse around radicalisation in 2010–2014 (Figs. 8.5 and 8.6). In the pie charts that follow, the different causes of radicalisation presented by the press are shown. The first pie chart shows the relative frequency of causes in the period 1998–2009, the second covers 2010–2014, while the third shows 2014 on its own. Below is a brief key explaining each cause listed in the tables. • Extremist Islam—cases where extremists are described as targeting non-extremists. This is often described as occurring in prisons, schools or universities. • Government Policy—this usually refers to foreign policy, for example, the invasion of Iraq, but in a rarer set of cases can also refer to a view

1998-2009

Multiculturism 4%

Poverty 4%

Grievance Culture 3%

Others 1%

Wars 4%

Alienation of Muslims 14%

Extremist Islam 34%

Government Policy 36%

Fig. 8.5  Claimed causes of radicalisation in the press in 1998–2009

  The Value of Revisiting and Extending Previous Studies… 

Poverty 6%

Others 4%

245

2010-2014

Wars 7%

Alienation of Muslims 10% Extremist Islam 57%

Government Policy 16% Fig. 8.6  Claimed causes of radicalisation in the press in 2010–2014

• • • • • •

that the government has failed to properly tackle extremism at home, or that policies that remove civil liberties are to blame. Alienation of Muslims—this refers to Islamophobia, the view that Muslims and Muslim communities are cut off from others in the UK and that youth in particular are disaffected. Wars—references to conflict abroad, particularly the Israel-Palestine conflict, causing anger in the UK but also, more recently, the conflict in Syria engaging British Muslims. Multiculturalism—blame on a general culture of accepting, even promoting difference in the UK. Poverty—economic reasons, especially cases of Muslims living in areas of economic deprivation. Grievance Culture—the view that Muslims wrongly feel victimised. This is similar to ‘alienation of Muslims’, but here the sense of victimisation is described as misguided. Others—‘one-off’ reasons such as individual Muslims experiencing torture, or more global trends such as liberalising modernisation or capitalism.

246 

P. Baker and T. McEnery

2014 only Others 9% Wars 8% Alienation of Muslims 9%

Government Policy 8%

Extremist Islam 66%

Fig. 8.7  Claimed causes of radicalisation in the press in 2014

These three pie charts alone are sufficient cause to cast doubt on the use of any time-bound analysis to cast light on what happens either before or after the period studied. The results from Corpus A are very different in terms of the proportions with which causes of radicalisation are mentioned. Figure  8.7 shows that one year in Corpus B is a much closer match to the results of Corpus B than the comparison to Corpus A produces; that is, there is some evidence for internal consistency within Corpus B, yet evidence of real change between Corpus A and Corpus B.

6

Conclusion

There is little doubt that the availability of corpus data which has allowed large-scale investigations of discourses in certain genres, especially the press, has been one of the most notable methodological developments in discourse analysis in the past couple of decades. Such analyses, however, are of necessity time-bound—the analysts collect data between two dates. No matter how exhaustive the collection of that data, the capacity of the data and its associated analysis to cast light on the discourse that preceded

  The Value of Revisiting and Extending Previous Studies… 

247

or followed that data is, of necessity, limited. In this chapter we have argued for a new methodological approach—the repeat of major studies of this sort and a comparison across time of relevant results to begin to approach the question of how discourse may shift through large-scale systematic studies. To demonstrate the need for this, we repeated the study of Baker et al. (2013), collecting data for the six years following their analysis. By doing that we were able to show the need for such a study—discourse may indeed shift over time. Some elements remain the same, some shift by a small degree, some shift quite substantially. Each shift has the capacity to expose a driver in discourse in society—for example, how the causes of radicalisation have been shifted in the UK press also has the effect of backgrounding blame for some and foregrounding blame for others. To simply assume that on the basis of the examination of 13 years of newspaper discourse (Corpus A) that we can generalise beyond those 13 years is naïve—yet we do see statements in the literature that seem, at least implicitly, to suggest that this is possible. Of course it may be the case that if we subdivided the periods studied further we may also find that the apparently monolithic nature of the two periods studied are in themselves subject to diachronic shift—indeed Baker et al. (2013), Gabrielatos et  al. (2012) explore shifts within the period covered by Corpus A in this chapter. However, the key point in this chapter is to focus on the usefulness of the methods of CADS for exploring discourse, but also to raise the general point that most discourses are not timebound and are dynamic through time—both within a period studied and beyond it. The methodological innovation that is needed to counteract this situation is simple—discourse must be monitored. Studies should be restaged, and the findings of time-bound analyses should not be assumed to be generalisable beyond the period studied unless there is clear evidence from a follow-on study that such a generalisation is warranted. However, on the basis of the study undertaken, we would predict that static findings between an original study and a follow-on study, especially in a dynamic medium such as press reportage, are likely to be the exception rather than the norm.

248 

P. Baker and T. McEnery

References Baker, Paul. 2006. Using corpora in discourse analysis. London: Continuum. Baker, Paul, Costas Gabrielatos, Majid KhosravNik, Michal Kryzanowski, Tony McEnery, and Ruth Wodak. 2008. A useful methodological synergy? Combining critical discourse analysis and corpus linguistics to examine discourses of refugees and asylum seekers in the UK press. Discourse and Society 19 (3): 273–306. Baker, Paul, Costas Gabrielatos, and Tony McEnery. 2013. Discourse analysis and media attitudes: The representation of Islam in the British press. Cambridge: Cambridge University Press. Blinder, Scott, and Will Allen. 2016. Constructing immigrants: Portrayals of migrant groups in British national newspapers, 2010–2012. International Migration Review 50 (1): 3–40. Durrant, Philip, and Alice Doherty. 2010. Are high-frequency collocations psychologically real? Investigating the thesis of collocational priming. Corpus Linguistics and Linguistic Theory 6 (2): 125–155. Evans, Matthew, and Simone Schuller. 2015. Representing ‘terrorism’: The radicalization of the May 2013 Woolwich attack in British press reportage. Journal of Language, Aggression and Conflict 3 (1): 128–150. Gablasova, Dana, Vaclav Brezina, and Tony McEnery. 2017. Collocations in corpus-based language learning research: Identifying, comparing and interpreting the evidence. Language Learning 67 (S1): 130–154. Gabrielatos, Costas, Tony McEnery, Peter Diggle, Paul Baker, and ESRC (Funder). 2012. The peaks and troughs of corpus-based contextual analysis. International Journal of Corpus Linguistics 17 (2): 151–175. Hardie, Andrew. 2012. CQPweb—Combining power, flexibility and usability in a corpus analysis tool. International Journal of Corpus Linguistics 17 (3): 380–409. Kambites, Carol J. 2014. ‘Sustainable development’: The ‘unsustainable’ development of a concept in political discourse. Sustainable Development 22: 336–348. L’Hôte, Emilie. 2010. New labour and globalization: Globalist discourse with a twist? Discourse and Society 21 (4): 355–376. McEnery, Tony, and Andrew Hardie. 2012. Corpus linguistics: Method, theory and practice. Cambridge: Cambridge University Press. Millar, Neil. 2011. The processing of malformed formulaic language. Applied Linguistics 32 (2): 129–148.

  The Value of Revisiting and Extending Previous Studies… 

249

Partington, Alan. 2003. The linguistics of political argument. London: Routledge. Partington, Alan, John Morley, and Louann Haarman, eds. 2004. Corpora and discourse. Bern: Peter Lang. Taylor, Charlotte. 2017. Women are bitchy but men are sarcastic? Investigating gender and sarcasm. Gender and Language 11 (3): 415–445.

9 The Linguistic Construction of World: An Example of Visual Analysis and Methodological Challenges Noah Bubenhofer, Klaus Rothenhäusler, Katrin Affolter, and Danica Pajovic

1

Visual Analytics in Discourse Analysis

Using quantifying methods often leads to complex and confusing data to be interpreted. This is especially true for corpus linguistic methods if they are not merely understood as a way to produce some frequency numbers, This research was supported by the Swiss National Science Foundation. We thank the reviewers for their comments on an earlier version of the manuscript.

N. Bubenhofer (*) • K. Rothenhäusler Department of Applied Linguistics, Digital Linguistics Unit ZHAW, Zurich University of Applied Sciences, Winterthur, Switzerland e-mail: [email protected]; [email protected] K. Affolter ZHAW School of Engineering, Winterthur, Switzerland e-mail: [email protected] D. Pajovic Department of Computational Linguistics, University of Zurich, Zurich, Switzerland e-mail: [email protected] © The Author(s) 2019 R. Scholz (ed.), Quantifying Approaches to Discourse for Social Scientists, Postdisciplinary Studies in Discourse, https://doi.org/10.1007/978-3-319-97370-8_9

251

252 

N. Bubenhofer et al.

but a way to reveal patterns of language use in large corpora. Following approaches of linguistic discourse analysis or lexicometry, the output of an analysis of corpus data often comes in the form of a list, sometimes long lists (collocates of a lexeme, significant keywords, n-grams etc.) following a specific structure. Networks are an example of such structured lists. Or, to be precise, some structured lists represent elements and associations between these elements that can be turned into a new visual representation commonly being interpreted as a network graph. Apart from the well-known and excessively used network graphs representing social networks, in corpus linguistics, collocation graphs are an example for such visualizations (Brezina et  al. 2015; Steyer 2013; Bubenhofer 2015). The goal of using visualization techniques for complex data is twofold. On the one hand, visualizations help to make the data interpretable. The visualization is successful if it turns the data in a form which makes it accessible for interpretation, more accessible than before. On the other hand, and closely connected to the first goal, the visualization should be an instrument to work with, an instrument that allows ‘diagrammatic operations’ potentially leading to new insights. This diagrammatic perspective stresses the exploratory nature of visualizations, clearly being more than just a display of something that is already known, but rearranging data and mapping specific aspects of it to visual elements, making it accessible alternatively. Diagrams not only come in the form of elaborated visualizations such as network graphs, maps and the whole range of displays of frequency distributions (box plots, bar, line charts, etc.), but start in very inconspicuous forms such as lists, enumerations, tables and so on. This means, of course, that the aforementioned structured list of associated linguistic elements (e.g. collocations) is already of diagrammatic nature. A specific type of list, the key word in context list used in corpus linguistics, actually is an important diagrammatic form working as an index or register. This very old form of rearranging data is an important tool to allow a new perspective on text. It decontextualizes the original arrangement and breaks up the unity of a text. Nevertheless, in corpus linguistics, it is often the case that lists are not sufficient to work with. But turning list data into a more

  The Linguistic Construction of World: An Example of Visual… 

253

complex visualization such as a network graph is just a transformation from one diagram into another. Using visualizations in the aforementioned sense as exploratory tools is the core idea of ‘visual analytics’ (Thomas and Cook 2005; Keim et al. 2010; Zhang 2007; Chen 2008). These methods are built upon a long tradition of handling large data sets statistically (Friendly 2005; Friendly and Denis 2001) and are widely used for data mining, both in academic and in non-academic fields (business, government). In the field of the digital humanities, which may be considered as the broader academic context of quantitative discourse analysis, visual analytics plays an important role. Many examples of visual analytics approaches have been developed and applied. Nevertheless, an in-depth discussion about the practice of using visual analytics in the humanities in general and for discourse analysis in particular is missing: The role such tools play in the research process is largely undiscussed. Often, papers in visual analytics introducing a novel approach mainly tackle questions about technology and visual principles and aim at providing a tool for the so-called expert, the humanist. On the other side of the gulf, the humanist plays with these tools without understanding them entirely and tries to manipulate their parameters until they spit out plausible results. We exaggerate, but in this chapter, we want to address some aspects of applying visual analytics in discourse analysis worth being discussed. To exemplify this discussion, we present our approach of calculating geocollocations to visualize how discourses construct world views.

2

Hermeneutics and Distant Reading

Scholars in the humanities and in some branches of the social sciences focussing on text—we will use the term ‘humanists’—are good at reading, understanding and interpreting texts, in short: doing hermeneutics, or doing ‘close reading’, as Moretti puts it: ‘in all of its incarnations, from the new criticism to deconstruction’ (Moretti 2000, 57). The hermeneutic approach to texts is still an important method in the humanities, also in the digital age where massive amounts of text are available in electronic form. But approaches such as discourse analysis (Foucault 1966;

254 

N. Bubenhofer et al.

Spitzmüller and Warnke 2011) or Moretti’s distant reading try to find a new way of doing hermeneutics. Sometimes, these approaches are misinterpreted as just a new version of traditional hermeneutics, which, as a plus, allows to ‘read’ larger amounts of text efficiently by using fancy query engines, statistics and visualizations. That is only half the truth—or perhaps even completely misleading. What we should be concerned with when doing some form of distant reading is finding emergent phenomena: Phenomena which cannot be explained by the properties of the underlying entities. This helps to ‘understand the system in its entirety’ and gain ‘theoretical knowledge’—for the price of losing the details (Moretti 2000, 57). A price not too high to pay for the advantage of getting emergent concepts that can be interpreted to get ‘the big picture’: of the development of world literature, changes in cultural behaviour, zeitgeist, discourses, just to name a few. Methods of visual analysis aim at making this ‘big picture’ accessible. A good example in corpus linguistics again is collocation graphs showing patterns of language use as an emergent phenomenon: The basis of the visualization is a statistical computation to find words in text co-­occurring more often than we would expect (Evert 2009). Representing the collocational profile of a word as a table already helps to see the emergent patterns of language use, but drawing a graph may make these patterns even more obvious. However, the graph representation is not self-explanatory: Many parameters must be set (e.g. how to reduce the multidimensionality to two or three dimensions?) and mapping principles must be found (e.g. do the lengths of the edges have a meaning?). Using such visualizations for discourse analysis has important implications that need to be discussed—we will touch some of them in the form of the following hypotheses.

2.1

 ool Development and Using a Tool Are T the Same

Developing visual analytic tools means getting one’s hands dirty: Finding a diagrammatic expression for the data, selecting a programming framework, using or developing algorithms, and programming. Normally, in

  The Linguistic Construction of World: An Example of Visual… 

255

the humanities, the programming skills are limited. But not only in the humanities, also in more information technology-oriented disciplines, the people building and using such tools are not the same. This separation between visualization tool developers and so-called experts using them is at the same time comprehensible and fatal. The disciplines dealing with visual analytics developed theoretical foundations and methodological frameworks to solve visualization problems on an advanced level. In consequence, humanists may use an increasing number of complex tools, but they are using them as if they were black boxes. If visual analytics is not just a tool, but a framework to explore data and to find emergent, meaningful phenomena, then building the visualization framework is as well an integral part of the research process (Kath et al. 2015). From choosing the statistics and data aggregation modes, the mappings of data, and graphical forms, to designing an interface, every single step in building the visualization framework demands full attention and reflection of the humanist. How does that influence the interpretation? And more important, how does that influence the research process itself? Software is a ‘semiotic artefact’, something ‘gets “expressed” in software’ (Goffey 2014, 36), for example, the cultural surrounding it is developed in, and its enunciations influence the process of interpretation.

2.2

No Best Practice

The development of visual analytic methods is driven by disciplines having a background in information science and engineering: informatics, data and text mining, computational linguistics, business intelligence and so on. The thought styles (German ‘Denkstil’, see Fleck 1980) of these disciplines demand specific scientific standards to be met, one of great importance being evaluation. Methods must be proofed to be useful by evaluating them by means of a so-called ground truth or gold standard. On that basis, descriptions of best practices are drafted to solve a specific problem using methods that have been evaluated. For visual analytic tools the need for evaluation has also been regarded as crucial (Keim et al. 2010, 131).

256 

N. Bubenhofer et al.

But if it is true that developing and using visualization frameworks goes hand in hand, then, obviously, defining best practices for doing visual analytics in the humanities is very hard (Keim et al. 2010, 54). As the research questions and processes differ enormously, a best practice must be defined for each use case. Indeed, doing so helps researchers to reflect on all the decisions taken in building the visualization framework. Despite the difficulties in evaluating visual analytic tools, several approaches have been proposed. Most interesting from a humanist perspective are more ethnographic-oriented methodologies such as design studies: Sedlmair et al. (2012) are fully aware of the pitfalls of problem-­ driven visualization research. They propose a nine-stage framework that can be followed not only to design and develop a visualization framework for a given problem, but also to reflect and evaluate the whole process. But still, this methodology separates the visualization researcher from the domain expert. In our eyes, it is crucial at least from a humanist perspective to combine these roles: For the humanist, the visualization framework is not just a tool to be applied, but an important part of her research process. There is no decision on the visualization side of the joint research process which is not of interest for the humanist.

2.3

 ogmas of Visualization Science Must D Be Questioned

From a humanist perspective some dogmas of visualization science are questionable. Taking for example Shneiderman’s ‘Visual Information Seeking Mantra’: ‘Overview first, zoom and filter, then details-on-­ demand’ (Shneiderman 1996, 336) or the extended version of Keim et al. (2006, 6): What are the details when visualizing language use? The obvious answer is: The text itself or a text snippet. In the example of the collocation graph, one would expect to have the text snippets available upon which the calculation of the collocations is based. But if we understand collocations as an emergent phenomenon, the actual text snippets will not provide any important information we do not already have. If the researcher has the impression of not getting the full picture by inspecting

  The Linguistic Construction of World: An Example of Visual… 

257

the collocations, the phenomenon they are interested in has not been modelled sufficiently. If it is mandatory to go through the text snippets, the researcher is interested in the single text snippets, not the big picture. A visualization framework with an easy access to the text details traps the researcher in the cage of single text hermeneutics: ‘what we really need is a little pact with the devil: we know how to read texts, now let’s learn how not to read them’ (Moretti 2000, 57). Another issue is related to a more general topos in information science described by Fuller as the ‘idealist tendencies in computing’ (Fuller 2003, 15) or by Geoffey as ‘an extreme understanding of technology as a utilitarian tool’ (Goffey 2014, 21). These topoi lead principles in computing like efficiency and effectiveness, to dominate algorithmic approaches to text understanding: Visual analytics therefore aims at building ‘effective analysis tools’ (Keim et al. 2010, 2), ‘to turn information overload […] into a useful asset’ (Keim et al. 2006, 1). As these goals can be justified for business applications, they can be misinterpreted in the humanities as a faster way to read texts. But instead, the capability of visual analytics in the humanities lies in getting a completely different view on human interaction—seeing emergent phenomena. A visual analytics framework useful for humanists will not provide a compact overview of the data and not merely a more efficient access, but should make the data nourishing for further analyses that were not possible before.

3

Geocollocations

The visualizing experiments we will present now stand against the background we sketched so far. The research questions lie in the domain of corpus linguistic discourse analysis.

3.1

Research Question

We are interested in the way mass media and other mass communication players shape our perception of the world. News articles often deal with countries, cities, regions, continents and so on and attribute values and

258 

N. Bubenhofer et al.

statements to them. We use an approach of discourse analysis using large corpora and data-driven methods to reveal linguistic patterns that are relevant for discourses (see Bubenhofer 2009; Bubenhofer et  al. 2015; Sinclair 2004; Teubert 2006; Felder et  al. 2011; Scholz and Mattissek 2014, and also the contributions in this book). In a discourse analytical perspective, we are interested in space being constructed socially (Glasze and Mattissek 2014, 209). We therefore search for linguistic patterns (frequently co-occurring words) contributing to the discursive construction of space. Our approach to reveal such linguistic patterns is the calculation of words co-occurring significantly often with toponyms. These so-called ‘collocates’ of toponyms should reveal the attributions attached to these entities. Our term for collocations consisting of at least one toponym is ‘geocollocations’. An example of a geocollocation would be ‘Switzerland— banks’ where ‘banks’ is the collocate of the toponym ‘Switzerland’. The world shaped by discourses normally differs from geographic reality regarding proximity (what areas seem to be near or far away?) and positioning (country X seems nearer than Y, although geographically it is not). Not all places and areas are present equally detailed and the world may be divided into parts like ‘south’ or ‘middle east’. Of course, all these conceptions are relative to the geographic position of the players in the discourses. Whereas it is already interesting for discourse studies to see which topics and attributions are attached to places (and what the differences are between different discourses), the more challenging goal is another one: to obtain abstract categories how geography is shaped through language use. We will later give some pointers to this question. Of course, the worlds constructed by different discourses may differ significantly from each other. Therefore the selection of a specific discourse as an object of the study and the underlying data is of crucial importance. But the following case study does not aim at giving a complete study on a specific discourse. The intention of our paper is to present our analytical framework and to discuss some methodological issues working with such a framework.

  The Linguistic Construction of World: An Example of Visual… 

3.2

259

Data and Calculations

The data used for this case study consists of two data sets: (1) A corpus of the magazine ‘Der Spiegel’ and the weekly journal ‘Die Zeit’ from Germany from 1946 to 2010 (640,000 texts, 551 million tokens, Spiegel/ Zeit corpus) crawled from the complete digital archives available online1 and (2) records of the German parliament Bundestag of the legislative period 2009 to 2013 (363,000 contributions, 22 million tokens) compiled by Blätte (2013). The data has been processed with the part-of-­ speech tagger and lemmatizer ‘TreeTagger’ (Schmid 1994, 1999). In addition, named entity recognition (NER) has been applied to the data using the Stanford Named Entity Recognizer (Finkel et  al. 2005) in a version adapted to German (Faruqui and Padó 2010). The recognizer tags not only toponyms but also names of persons, companies and organizations. In our case, only the toponyms were used. In order to calculate the geocollocations, all toponyms above a minimum frequency limit were selected and words (lexemes) co-occurring significantly often with the toponym in the same sentence were calculated. The selection of an association measure is influenced not only by statistical considerations, but primarily by the theoretical modelling of the collocation concept. We used a log likelihood ratio significance testing which is widely used in discourse linguistics to study language usage patterns (Evert 2009). The data set now contains toponyms and their collocating lexemes with frequency information and the level of significance of the collocation. In order to place the toponyms on a map, they have to be geocoded. Although there are several geocoding services available like Google Maps API or Nominatim (OpenStreetMap), the task is challenging because of reference ambiguities (‘Washington’[DC or the state], ‘Berlin’[capital of Germany or city in New Hampshire?]), historicity of the toponyms (‘Yugoslavia’ or ‘Ex-DDR’ do not exist anymore) or the use of unofficial names (‘the States’ for USA, German ‘Tschechei’ instead of ‘Tschechien’, ‘Western Sahara’, which is not officially recognized as a state). Luckily  See http://www.zeit.de/2017/index and http://www.spiegel.de/spiegel/print/index-2017.html (accessed 6 March 2017). 1

260 

N. Bubenhofer et al.

with geonames (http://www.geonames.org), a gazetteer is available that can deal with the last two problems by providing a broad range of names for places, including historical variants, and as a community-based project it is easily extendible on top. However, it is just a list and hence on its own it is of no help for the task of ambiguity resolution. To tackle the latter, we used CLAVIN (Cartographic Location and Vicinity INdexer2), an open source tool, which performs automatic geocoding based on the document context. It employs different heuristics to disambiguate the toponyms identified by the NER: place candidates are ranked according to population size and in a second step a set of places is chosen that are within the vicinity of one another. The precomputed collocations are stored in an Elasticsearch3 index for convenient access of toponyms along with their collocates and to allow for fast visual display in the browser. Most of the filtering possibilities discussed in the following section can also be computed in the backend to further reduce the load of the code running on the client side and enhance the user experience.

3.3

Visualization

The computed list of geocollocations contains several thousand entries and obviously is not suitable to get an overview over the data. In addition, reading the list lets the analyst constantly draw references to the geographical places which might be easy in the case of toponyms like Paris, London, Berlin, Rome, France, UK, Germany, and Italy, but is more difficult with an unordered list of less known toponyms like Madhya Pradesh, Dschidda, Masar-e Scharif, Sibiu. Contextualizing these toponyms by placing them on a map helps to draw the references to the actual places and to semantically enrich them for further interpretation. Also geographical vicinity gets transparent. Hence the most straightforward way to visualize the data is placing the collocates of the toponyms on a map which is how we built our first pro-

 See https://github.com/Berico-Technologies/CLAVIN/ (accessed 6 March 2017).  See https://www.elastic.co/ (accessed 6 March 2017).

2 3

  The Linguistic Construction of World: An Example of Visual… 

261

totype. This may be seen as the most literate translation of the term geo-­ co-­location because words that occur near each other in text are located together on a map and share the same coordinates. This representation is easily understood as it stays very close to the data. Nevertheless, it enables the user to interactively explore the information presented on the map. To facilitate this kind of explorations, we provide a number of visual hints and controls: • In order to see the distribution of a collocate and identify similarities in the inventory of collocates between different toponyms we highlight the word under the mouse pointer everywhere else on the map. • The collocates are super-tagged for nominal, adjectival and verbal word classes (identified as sets of part-of-speech tags), which are assigned colour codes via display styles. • The number of collocations identified for even decently sized corpora quickly overgrows the limits that can be displayed in an orderly fashion on a map. Hence, we introduce three means to further reduce this amount. The first puts a threshold on the significance level for the computed collocations so that only higher levels of significance above the standard minimum (p