3. Methods and Research Design
Texte intégral
3.1 Corpus-Assisted Discourse Studies
1Corpus-assisted discourse studies (CADS)1 is a relatively nascent approach to discourse analysis, particularly outside linguistics and adjacent disciplines, and indeed draws heavily on the theory and methodology of corpus linguistics (CL). The corpus-assisted part of it concerns the use of computerised corpora to study patterns in language. A corpus is a collection of texts built to be representative of a particular body of discourse; in my case that of two English-language newspapers. If discourse is to be understood as productive and constraining we need to be able to conceive of and study discourse as more than the sum of its parts, to look beyond individual texts and understand how a ‘body of discourse’ in its totality works to produce the social world in which we live. All approaches to discourse attempt to do this, but corpus approaches are uniquely suited to the study of discourses as systematic and patterned, because of their ability to draw on frequency data and to make statistical claims.
2Traditionally, in corpus linguistics, this would be used to compare language use in particular contexts, for instance to compare Indian and British English or 19th and 21st century fiction. The focus here would be on understanding what makes particular discourses linguistically unique. More recently, the applicability of this to discourse studies has led to an emergence of more politically engaged analyses (Partington et al. [2013] and Baker and McEnery [2015] represent major interventions of CL methods into discourse analysis).
3At the heart of corpus linguistics is the idea of collocation. Collocation (as in co-location) is when words occur together with a frequency greater than chance. Collocation enables us to see how language is structured and patterned throughout a discourse. Collocation patterns can tell us how words are used and the semantic associations they hold. Building on the notion of collocation in CADS is the theory of lexical priming developed by Hoey (2005). Lexical priming suggests that words are psychologically primed through regular use to occur (and avoid occurring) with certain other words, with certain semantic associations, in certain grammatical roles, and in certain positions within a discourse (Hoey, 2005: 13). It is this priming that accounts for the naturalness of language, and consequently the sense that certain language use is unnatural even if grammatically coherent.
4This may seem banal, but it is in fact highly significant because it challenges many of the assumptions underpinning classical linguistics, particularly in its assertion that “grammar and semantics are post hoc effects of the way lexical items have been primed” (Hoey, 2005:391). It is significant for discourse analysts because, by studying the ways in which certain words are primed, we can understand how they are used, and the meanings they carry with them. In this sense, words are not innocent; they are learnt through their encounters with other words and become “loaded with the cumulative effects of those encounters” (Hoey, 2004: 386). Studying collocation and priming therefore enables us to study discourse not as a set of statements or texts that have meaning individually, but rather to understand discourse in a thicker sense and to be able to make claims as to how ingrained and how common certain meanings, representations and attitudes are in a discourse. It is therefore about the study of patterns: a study made possible by the large-n approach of CL.
5If discourse is fundamentally more than the sum of its parts (individual texts and utterances), then the ability to analyse a body of discourse as a whole is important, as is the ability to make claims to external validity. CL can obviate the potential in lower-n studies to ‘cherry-pick’ texts and elements in the discourse that most fit their arguments. Importantly however, it is corpus-assisted to the extent that CL methods are only part of the study (Partington et al., 2013: 10). Close reading and the use of methods and tools found in discourse analysis more generally (including elements commonly used in critical discourse studies [CDS]) are also important. Combining methods is therefore a form of triangulation that can help to show what is common, and to catch elements and interpretations that may escape the notice of another approach (Taylor and Marchi, 2018). It is very much in the interest of triangulation that this study is carried out, in the belief that more than one method adds validity and completeness.
6Given the assisted nature of CADS, it is possible to use most discourse-analytical tools alongside it. There were therefore decisions to be made regarding what tools I would use and what textual features I would focus on. Discourse analysis has been popular in IR for some time, originally with those working within critical and post-structural epistemologies, and later with mainstream constructivists. What counts as discourse analysis has varied widely (see Lundborg and Vaughan-Williams, 2015 for an overview). Perhaps the most common approach to date in IR has been ‘predicate analysis’ as developed by Milliken (2005) among others. Predicate analysis looks at the verbs, adverbs and adjectives used to characterise (predicate) subjects. This can be useful to understand how subjects are constructed in discourse and what agency they are accorded. Although not under the rubric of ‘predicate analysis’ I do study predication when I look at the collocations of subjects like Bangladesh, border and migrant.
7Conventional predicate analysis usually looks at how a subject is predicated in each instance in a number of texts. The resulting analysis is likely to be more fine-grained with the potential to yield more detail and nuance. Nonetheless, there are a number of advantages of CADS in the analysis of predication, notably the ability to work with a far larger and potentially more representative set of texts, the ability to infer patterns of predication that may be largely invisible if studied manually, and the ability to make inferences about lexical priming. Moreover, issues of detail and nuance are not as pronounced as might be assumed, because of the use of concordance analysis and because the assisted nature of CADS means I can employ additional tools that can reinsert granularity. Lastly, it is interesting to note that Baker and Levon (2015) compared the results of purely corpus-based and ‘qualitative’ analyses by different researchers on the same texts (one down-sampled from the corpus) and found the corpus approach to be similarly nuanced. In my study, as with most others, I combine both approaches in the interest of triangulation.
8Perhaps the biggest limitation of CADS is also its key strength; if the phenomenon of interest hinges on discrete words or phrases and their patterned use, CL is likely to add value, but for phenomena that do not crystallise around lexical items it is of little use (Mautner, 2016). This is overcome by concordance analysis and contextualisation. Indeed, one of the potential issues of CL methods is the potential to exclude context from the analysis: both the context in which the words or collocates are used and the context of the corpus in relation to corpus-exterior information. There are simple ways to obviate this potential. The first is the use of concordancing; a concordance line is a word in the corpus in its immediate context. Figure 1 shows concordance lines for Bangladeshi, displaying the search word in its context. This can then be clicked on and the entire article can be read. This adds context and granularity back into the study and indeed it is through the iterative use of collocation and concordance analysis that discourses can be studied (I discuss more about using concordance analysis as a plausibility probe in the following chapter). Moreover, by accessing the collocation in context, and by expanding the context, sometimes to the whole article, I can study dynamics more common to CDS approaches.
Figure 1

9The second way I recontextualise my corpus analysis, is to look outside the corpus. The discourse I study is of course a small part of a larger set of discourses, processes, materialities and power relations, all of which are crucial in understanding the discourse. In looking at some of these processes outside the corpus – for instance the development of citizenship law, the materiality of the border and the testing of nationality – I ensure that my analysis is grounded in a broader politics, and am cognisant of the fact that the discourse is a product of this wider context and in turn helps to produce the context in which these other processes and discourses occur.
3.2 The Corpus
10I used LexisNexis to create the corpus, extracting articles with the terms Bangladesh and Bangladeshi from the Times of India (TOI) and the Hindustan Times (HT) in the years 2011, 2013, 2017 and 2019. I chose to build the corpus in this way for two reasons. Firstly, I am able to disaggregate the corpus into four sub-corpora by year. In this way I can understand what has changed in the discourse over these years, and what has remained constant. Secondly, I have chosen these years specifically because I wanted to sample periods of Congress and BJP government. The first two periods were during the most recent Congress-UPA term, while the latter two were during the current BJP-led government. As I shall show, there has been minimal discursive change between these periods, reinforcing the idea that anxiety over Bangladesh plays a central role in more than one form of Indian nationalism, and suggesting that this anxiety is not merely an instrument of party politics. It would have been interesting to conduct a larger diachronic analysis, going back to the independence of Bangladesh and trace the changes in discourse. The archive I used did not contain TOI articles earlier than mid-2010 (HT went to 2004), and to construct a corpus going further back was therefore not feasible in this study. Nonetheless, this span still allows for a meaningful diachronic comparison.
11I chose the TOI because it is the highest-circulating English daily in India, and the most visited news website in the country (Audit Bureau of Circulations, 2019; Alexa, 2019). While The Hindu has the second highest circulation, I chose the HT (third) because LexisNexis did not have archived material for The Hindu before 2014. As English-language newspapers, these publications necessarily cater to an educated readership, though both are relatively populist and could not be described as ‘high-brow’. Furthermore, while English is only understood by around 30% of Indians, it is the second most spoken language in India (MRUC, 2019). English-language newspapers play a large role in Indian discourse due to their imbrication with the urban middle and upper classes. Consequently, the discourse found in English newspapers has power beyond those who read it, in its capacity to influence the wider framing of issues (Jolly, 2016: 4). Importantly however, I would not expect them to differ greatly in their framing from mainstream non-English papers such as Dainik Jagran or Dainik Bhaskar, the two largest papers in the country, particularly regarding issues of ‘national security’ and immigration, which are also highly populist and right-wing. Like most mainstream media in India, both the TOI and HT are relatively right-wing, with the TOI being further to the right than the HT (Subramaniam, 2012).2
12In seeking to understand how Bangladesh might figure in ontological security, there were numerous sources of evidence to choose from. I chose to focus on discourse because it is at least in part through discourse that I would expect ontological (in)security to be narrated and discussed, and because discourse works to create ontological security by structuring the world and our relationship to it. That is to say that I understand discourse to be productive of ontological security; that it is a securitising practice (among others). Ontological security is a difficult thing to study because the link between indicators and theory is not straight forward and will always need to be argued. This difficulty is why I have sought to triangulate my findings in the discourse with other sources of evidence: namely changes in citizenship law, citizenship testing and the practices of border security.
13I chose newspaper discourse in particular for two reasons. Firstly, there is a wide availability of data; there is a lot of it and it can be gathered in a systematic way. Secondly, I wanted to study everyday discourse that could be said to in some way reflect a broader discourse on Bangladesh in Indian society. Newspaper discourse has a cumulative effect in that it is built upon day after day. As Baker (2005: 61-62) notes, “a negative or ambiguous word, phrase or association may not amount to much on its own, but if similar sentiments appear on a regular basis, then the discourse will become more powerful, penetrating into society’s subconscious as the given way of thinking”. I had considered using parliamentary debates but decided against it in this study, partly because the data was not as easily available and partly because I did not consider it to be as closely imbricated with popular and everyday parlance as newspaper discourse.
14The challenge of any discourse analysis is that discourse ‘out there’ cannot be studied in its totality or its naturalness. The discourse that is produced in everyday interactions at tea shops, on the bus or over dinner, etc. is inaccessible, at least in the form required by the discourse analyst. Similarly, of the spoken or written discourse that could be ‘captured’ and analysed, decisions still have to be made about what to include and what not to include. One cannot study discourse in its entirety. The task therefore, is to choose a section of discourse that can be said to most effectively stand in for that wider discourse that we cannot access, and to acknowledge the limitations that this decision necessarily implies. Newspaper discourse is extremely profuse, is engaged with by a large number of members of a discourse community (relative to most other sources of data) and is used in a casual manner (unlike, for instance political speeches that may be either very diplomatic or the complete opposite). Additionally, it is more productive than many other parts of a discourse precisely because it is widespread, incremental, and often authoritative in the sense that it provides information to others and thus helps to frame certain issues. Importantly, though, these framings usually already exist (and corpus-based analysis is ideal for studying these patterns). Much of my corpus is not built from articles specifically about Bangladesh; instead, most of the instances of the search words Bangladesh and Bangladeshi are used in passing, in the context of another issue (e.g. crime, legal change, urban development). This is an advantage when we want to study everyday discourse because comments made in passing are usually indicative of a particular taken-for-granted understanding. Newspapers are also very nationalistic, particularly mainstream Indian newspapers. They are therefore useful places for studying nationalisms and their ontological insecurities.
3.3 Parameters
15A collocation is a pair of words that co-occur together in a corpus with an unexpected frequency. Collocates are built using all of the words proximate to a node (the word searched for). The collocation ‘window’ is the number of words either side of the node from which the collocate list is built. The common default in CADS is five words to the left of the node, and five to the right. This is expressed in notation as L5R5. Sometimes I restrict the window in order to focus on the words immediately preceding or succeeding the node. This can show how the node is modified. For instance, I search for Bangladeshi with an L1R0 window, to see that the adjectives that occur immediately before Bangladeshi are illegal and suspected. Moreover, searching only for adjacent collocates can help to manually build collocation networks (described below, and in chapter four). It should also be noted that lexical priming is most effective in adjacent collocates; language develops a formulaic quality most powerfully in words that are immediately proximate. Formulaic language does extend beyond adjacent collocates, but not as strongly (Vilkaitė, 2016; Gray and Biber, 2013; Molinaro et al., 2010). Thus, if there is a strong collocation in L1 or R1, this is most likely to be evidence of lexical priming, with collocates built from a wider window telling us a lot more about the semantic context of words and their primings, but not so directly about the primings themselves.
16Building collocations involves using a measure of collocation. In general, I use MI3 because, like all MI (mutual information) measures, it focuses on the strength of association between collocates, but overcomes some of the weaknesses of the standard MI measure. MI biases infrequency and exclusivity. It identifies those collocates that are almost always found with the node (e.g. Herring is usually with red, O’war is always with Man, okey is always with dokey). However, these collocates may be very strongly associated with a node, but they are usually also highly infrequent. In the case of Bangladesh, one of the top collocates using MI is $12, because $12 occurs only once in the corpus and happens to be proximate to Bangladesh. However, because it is so infrequent, it tells us nothing about patterns in discourse. Traditionally a frequency threshold has been applied to filter out infrequent collocates. In MI3 (or MI cubed) this is obviated as more weight is accorded to frequency, producing a list of collocates that are both relatively exclusive (in comparison to t-score and log-likelihood for instance) and relatively frequent (in comparison to MI). MI3 has become a popular metric with CADS researchers for these reasons (see Brezina, 2018 for a detailed discussion on measures in corpus linguistics; also Baker, 2014). Unless otherwise stated, I used MI3, with the default statistical and frequency cut-offs of 9 and 5 respectively, with an L5R5 window.
17In addition to concordancing and collocation lists, Lancsbox also enables the production of collocation networks (see Brezina, McEnery and Wattam, 2015). While most collocation analysis looks solely at first-order collocates (the collocates of a given node), collocation networks allow us to visualise the collocates of collocates (which would be second-order, third-order, etc.). This can be useful, because it enables tracing of more complex semantic contexts, as shall be demonstrated in the analysis.
3.4 Discursive Signs of Ontological Insecurity
18Before continuing to the analysis, it is pertinent to briefly suggest what might be expected to be found, and what might constitute signs of ontological insecurity. As discussed in the previous chapter, I expect to see indeterminacy, and an anxiety over this indeterminacy. Concretely, I expect to identify issues of identification and misidentification, where there seems to be a confusion over the nationality of certain people. Signs of this might include words that concern forms of documentation or a lack thereof, or words such as citizenship, nationality, belonging etc. which might bespeak a concern over how to know if a person is Bangladeshi or Indian. Moreover, I would expect discussion of the National Register of Citizens (NRC) to be prevalent, and for this to be concerned in part over the capacity for mistakes in the identification of nationality. Secondly, and relatedly, I would expect to see the prevalence of modifiers predicating nationality and, in doing so, casting doubt on it: for example, ‘she was perhaps a Bangladeshi’ or ‘four potential Bangladeshis were…’, etc. Other modifiers might include alleged, suspected, considered, thought to be, likely, and so on.
19Having thus identified indeterminacy, I would therefore expect to see signs of anxiety over this indeterminacy. It should be noted that I am likely to only find the residue of anxiety over indeterminacy in discourse. I would not expect this to be discussed or accounted for explicitly, given, as Giddens (1991) reminds us, much of what unnerves us is dealt with at the level of practical as opposed to discursive consciousness. Moreover, newspapers are not in the business of offering soul-searching accounts of identity and selfhood. Consequently, ontological insecurity or the anxiety over indeterminacy is unlikely to be declared in so many words. In this context, those other markers become particularly important, though, of course, they are precisely residues or signs. A key sign of this may be the sheer prevalence of discourse on the challenge of identification. This would imply an obsession and a concern with such challenges. Another marker might be a semantic preference for outrage or shock upon finding that Bangladeshis had been found passing off as Indians. Lexis in this vein might include shocked, alarmed etc. More generally, it might suggest that the indeterminacy is a security concern and lead to drives to identify other Bangladeshis who had hitherto passed off as Indian. The concern then is with how many others might be hiding.
20The other side of this, which I would expect to be perhaps more prevalent, is the overcoding of Bangladeshi and Indian populations. Evidence of overcoding might be seen in the denial of ambiguity. This may be found in the way nationality is asserted as if proven. It might also be seen in the way Bangladeshis and Indians are portrayed consistently, in diametrical ways, thus once again reinforcing difference. Conversely, if Bangladesh and Bangladeshis were represented inconsistently, for instance with a high level of semantic variation and collocate competition, this would suggest that overcoding was less prevalent. In line with my reasoning in chapter two, concerning Bangladesh as dystopic and how this may confer a sense of exceptionalism and provide a contrast space for selfhood, I would expect a largely negative portrayal of Bangladesh and Bangladeshis, perhaps explicitly contrasted to India and Indians.
Notes de bas de page
1 Corpus approaches to linguistics go under various names, some positioning it with CDS, others seeking to be less ‘critical’ (see Baker and McEnery, 2015: 5). I have chosen to use the term CADS because it differentiates it from other approaches within CDA, but also acknowledges that the study I will undertake will use eclectic methods, some of which will be based in corpus research, hence corpus-assisted. Following Partington et al. (2013: 6), I see CADS “not as a discipline or field of study but more a methodology, that is, a set of tools and general practices and ways of using those tools for the purpose of language analysis”, notwithstanding the fact that corpus linguistics has fundamentally changed the way we understand language and discourse.
2 For an overview of the Indian media environment, see e.g. Sulehria (2018).

Le texte seul est utilisable sous licence Creative Commons - Attribution - Pas d'Utilisation Commerciale - Pas de Modification 4.0 International - CC BY-NC-ND 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.
The Development of International Refugee Protection through the Practice of the UN Security Council
Christiane Ahlborn
2010
The SWIFT Affair
Swiss Banking Secrecy and the Fight against Terrorist Financing
Johannes Köppel
2011
The Evolving Patterns of Lebanese Politics in Post-Syria Lebanon
The Perceptions of Hizballah among Members of the Free Patriotic Movement
Fouad Ilias
2010
La justice internationale à l'épreuve du terrorisme
Défis, enjeux et perspectives concernant la Commission d'enquête internationale indépendante (UNIIIC) et le Tribunal spécial pour le Liban
Sébastien Moretti
2009
Aut Dedere, aut Judicare: The Extradite or Prosecute Clause in International Law
Claire Mitchell
2009