10.2.1 Social Actors in Obesity Coverage

In their analysis, Brookes and Baker (Reference Brookes and Baker2021) incorporated Reisgl and Wodak’s discourse historical approach (Reference Reisigl and Wodak2001), which like social actor analysis considers referential/nomination and predication strategies. The former pertains to how individuals are named and referred to, while the latter deals with how they are described and what qualities or characteristics are attributed to them. A nomination strategy might involve using a noun like fatso, while a predication strategy could include adjectives like pathetic or verbs describing certain actions (e.g., wolfing down food) or being the recipient of actions (e.g., labelled as obese).

Additionally, Brookes and Baker (Reference Brookes and Baker2021) utilised Van Leeuwen’s (Reference Van Leeuwen, Caldas-Coulthard and Coulthard1996) social actor representation framework, which offers a system for categorizing the ways that social actors are portrayed in English discourse. For instance, terms like the obese involve physical identification, uniquely defining individuals based on their physical characteristics within a given context. This framework helped researchers identify how references to people with obesity can either include or exclude, personalise or impersonalise, assimilate or differentiate them in the discourse. The researchers uploaded the corpus into the online analysis tool CQPweb (Hardie, Reference Hardie2012), so the analysis described in the following section uses the search facilities associated with that tool, although other tools should allow for similar kinds of searches to be carried out.

10.2.2 Nomination Strategies

The analysis began with the identification of nomination strategies. Once a set of those were found, Brookes and Baker (Reference Brookes and Baker2021) could then more easily find the predication strategies through the consideration of collocates and concordance lines. Identifying nomination strategies was not a simple task, however, and the researchers employed several non-corpus-assisted and corpus-assisted tactics in order to produce a list.

The two main non-corpus-assisted methods involved using introspection and reading samples of articles from the corpus. The researchers asked friends, family members, and colleagues to think of possible terms that might appear in the corpus, in order to triangulate from different perspectives. Subsequent to the completion of the study, the researchers decided to see whether a tool like ChatGPT would have been useful for identifying further terms. ChatGPT initially offered a list of terms that it described as ‘neutral and descriptive’, like person with a higher body mass. Its initial response was lightly chastising, stating: ‘Remember to use language that is respectful and non-stigmatizing when discussing individuals’ weight or body size.’ However, when the purpose of the study was explained in more detail, ChatGPT produced a wider list of words which included stigmatising ones like whale and blob.

The corpus-assisted methods of eliciting nominations involved the examination of frequency and keyword lists. The researchers identified keywords for each newspaper by using ‘the remainder method’ (i.e., comparing a frequency list of articles from a single newspaper against a reference corpus consisting of the remaining newspaper articles they had collected; see Chapter 7). This helped identify terms that tended to be popular in a single paper. For instance, nouns like fatties and flab were prominent in The Star compared to the other newspapers in the corpus. The researchers also tried to identify relatively infrequent nouns in each newspaper by looking at wordlists. For example, when they obtained a frequency list of The Times, they looked at plural nouns that occurred between 4 and 10 times in the corpus. This produced words like bloaters, heavyweights, and overeaters, which had not been identified by any other means.

10.2.3 Collocation Analysis

Once they had an initial set of words, Brookes and Baker (Reference Brookes and Baker2021) identified collocates, focussing particularly on adjectival and verb collocates, as these were most likely to reveal predication strategies. These collocates also provided an additional route for finding new nominations. For example, when examining collocates of slob, the researchers found adjectives like fat, boozy, lazy, and out-of-shape. Conducting searches of these adjectives and examining words which collocated with them produced additional nomination labels like lard-arse, hog, porker, lump, and fatty. As they conducted their analyses, the researchers kept an open eye when reading concordance lines, in order to identify further terms. This meant that they had to regularly update frequency tables as new terms were elicited during the analysis.

The researchers were working with a grammatically tagged version of the corpus, which allowed them to easily expand the analysis to account for both singular and plural forms of nouns (e.g., slob, slobs), as well as comparative and superlative forms of adjectives (e.g., chunky, chunkier, chunkiest) and related verb tenses (e.g., guzzle, guzzles, guzzled, guzzling). Within CQPweb, putting a search word within curly brackets retrieves all its related grammatical forms. For instance, searching for {waddle/V} would elicit cases of waddle, waddles, waddling, and waddled as verbs.

For some terms, searches elicited a large number of false positives (i.e., unwanted cases). For example, the word fat often referred to the amount of fat in food. The researchers could search just for adjectival cases (tagged JJ), using the search query fat_JJ, although this still produced a high number of noun cases that were erroneously tagged as adjectives. Other cases used fat to modify nouns that did not refer to human beings (e.g., fat crisis, fat gene, fat camp). There were 22,578 cases of fat tagged as an adjective, so in order to obtain a reasonable estimation of the number of times the word fat appeared as an adjective which referred to a human being, the researchers carried out a qualitative examination of 10 per cent of concordance lines of fat_JJ, presented in random order (2,258 cases). They then multiplied the number of cases that referred to people by 10. Another set of unwanted cases involved the phrase the obese, which sometimes occurred as a collective noun but also often occurred as part of a longer noun phrase, such as the obese singer. A similar strategy to fat was used to isolate only the cases where the obese was used as a collective. This term only occurred 1,703 times in the corpus, so it was possible to carry out a concordance scan of all cases and remove cases like the obese singer, which didn’t apply.

The researchers also obtained the collocates of a set of nouns that more generally referred to human beings (e.g., people, men, women, children, kids, boys, and girls), as these words sometimes occurred as parts of noun phrases relating to obesity (e.g., obese people, supersize kids). Different collocational statistics are likely to produce different kinds of findings. For example, the log ratio and mutual information statistics tended to produce low-frequency collocates which would point to rare nominations (e.g., super-obese is a top log-ratio collocate of people, occurring 11 times in the corpus). On the other hand, log likelihood foregrounded more frequent collocates like obese, which co-occurred with people 3,714 times.

In order to consider representation beyond simple noun labels, the researchers aimed to find a list of adjectives and verbs that tended to be used to refer to people with obesity in the corpus of news articles. Once they had identified a list of nominations, it was possible to obtain collocates for each one. As there were a large number of terms, comprising both singular and plural forms, a short cut to obtaining lists of adjective and verb collocates was to combine searches on multiple words in CQPweb. For example, researchers can put a number of similar terms within brackets, separating them with the | symbol, to consider them together:

Researchers can then obtain the collocates of these collated terms, which would save time compared with having to conduct separate searches on each word. The researchers checked the resulting concordance lines of collocates to remove false positives. For example, based on the search query provided, the top-10 collocates (using log ratio) are hydroxyl, Peppa, ‘Oi, suckling, guinea, Percy, Daddy, salty, and metabolise. Most of these collocates are modifying non-human cases such as hydroxyl fatty acids and Peppa Pig, and thus would need to be removed completely or at least have their frequencies adjusted.

10.2.4 Predication Strategies

Once Brookes and Baker (Reference Brookes and Baker2021) had removed the false positives, they were able to obtain a list of adjectives used to refer to people with obesity, which was illuminating in revealing typical qualities that are associated with them. The researchers grouped the adjectives into categories which related to attractiveness (e.g., ugly, beautiful), health (ill, unfit), emotional state (depressed, happy), level of activity (lazy), intelligence (stupid), or other qualities (funny, proud). However, the researchers also wanted to identify adjectives that refer more directly to obesity as opposed to other qualities, such as tubby, solid, and expansive. Adjectives such as these can appear as predicative (e.g., he is tubby), and when attributive, they might not necessarily modify nouns relating to obesity but rather occur with general nouns like person or man. Again, a number of retrieval strategies were employed in order to obtain a list of these adjectives: introspection, reading samples, concordance lines, and through prospective collocation searches on relevant nouns or pronouns.

Once they had obtained a list of these adjectives, the researchers categorised them into those which were positive, euphemistic, or directly insulting towards people with obesity. One additional aspect of these adjectives that emerged as the researchers collected them was that some of them seemed to be gender-specific (e.g., voluptuous and curvy were used to refer to women with obesity, while beefy and portly referred to men with obesity), so the researchers noted such cases accordingly. Further, they noted that many of these words were often used alliteratively. For example, of the 125 references to porky, 63 (50.4 per cent) were followed by another word beginning with p (e.g., pooch, pets, PCs, pupils, etc.). Similar alliterative patterns are found with the other euphemistic adjectives (e.g., chubby children, tubby toddlers, lardy lags, hefty hounds, flabby felines, voluptuous vixens). Such cases are more typical of tabloid newspapers, which often use literary devices to make headlines more eye-catching. However, in terms of representing people, it could be argued that this alliterative language can result in a fictionalising effect, encouraging readers to think about the subjects as caricatures and distancing them from thinking about them as actual human beings.

Using collocates, the researchers also obtained lists of verbs that were used to modify references to people with obesity. Verbs were classified into groups which referred to

• food being consumed quickly and in large quantities (e.g., gorge, scoff, cram, shovel);

• ungainly movement (lumber, waddle, jiggle);

• discomfort (sweat, wheeze);

• weight gain (balloon, pile on); and

• difficulty fitting into space (cram, wedge, clog, squeeze).

These verbs were considered further in terms of the metaphors that they implied – for example, the verbs related to eating sometimes represented people with obesity as animal-like (e.g., wolf, swill, pig out). The researchers also considered verbs that positioned people with obesity as patients of actions, as opposed to agents. (The Sketch Engine analysis tool has a Word Sketch function that can make this distinction automatically; see Chapter 9.) This led the researchers to a small set of verb collocates like brand, dub, and shame, which were often used in sympathetic articles about people with obesity, where they were sometimes presented as objects of pity or were the subject of redemptive weight-loss narratives. For example,

This mention of narratives points to a further, more qualitative stage of the analysis, which extends beyond identifying collocates to consider the wider contexts in which people with obesity are involved. This can be a less systematic form of analysis to engage in, and even a concordance line analysis may not be especially productive, requiring analysts to expand concordance lines to read entire texts to obtain a sense of the narrative structure of a particular article. Continuing with the redemptive weight-loss narratives, for example, the researchers found a pattern whereby numerous news articles (particularly in tabloids) told a real-life story about an individual (often an ordinary member of the public, although sometimes a celebrity) who had been bullied in childhood for being overweight and had gone on to lose a lot of weight as a result of a particular diet or exercise regime. The individual was described as having undergone an amazing and positive transformation, both physically and in terms of achieving success in other areas of their life. Such stories were somewhat formulaic, although the researchers also noted what they did not include, such as criticism of the bullies or acknowledgement that many people who lose a lot of weight quickly are likely to regain it after a few years. These narratives are therefore perhaps both inspiring and misleading to readers.

Alongside these narratives, the researchers’ more detailed readings of concordance lines helped them identify two other ways that people with obesity were represented – as subject to ridicule and criminals. For example, they found a number of cases involving stories of convicted criminals who were described as having obesity:

This article observes Huntley’s weight gain, portraying it as deserving public attention. However, reporting on his weight gain is tinged with a sense of delight, which, considering the overall media stance on weight gain, may be interpreted as a form of fitting retribution (with the article labelling Huntley as evil and a monster). The Sun article also implicitly links weight gain and crime, and the researchers also found articles that referred to people whose criminal behaviour was described as being directly related to their obesity:

Such stories collectively contribute to the portrayal of individuals with obesity as clumsy, petty wrongdoers, associating negative traits like greed and laziness with their weight. Consequently, these narratives contribute to depicting people with obesity in a comedic role. While the crimes of individuals like Huntley are not presented as humorous, stories about their weight gain are portrayed more frivolously, demonstrating the scope within which obesity is presented as subject for ridicule. Although such stories are not as prevalent as the weight-loss narratives mentioned earlier in this section, their frequency is sufficient to identify a pattern. The casual association of crime with obesity arguably represents one of the most unfavourable forms of representation in the corpus.

In summary, the case study described in this section indicates how representations can be found in corpora by beginning with the idea of nomination and nouns, and gradually expanding the remit to include predicative and attributive adjectives, as well as verbs which position social actors as agents or patients. We also saw how the analysis was expanded further to consider metaphors, gendered language, and stylistic features like alliteration, and then finally went beyond the lexis to explore narratives. In doing so, the analysis moved from the quantitative to the qualitative. While it cannot be said that every form of representation of people with obesity in the corpus has been identified, what the approach taken here does offer is a representative picture in order to draw meaningful and reasonable conclusions, while also allowing the researchers to provide information about the frequencies of different kinds of representations in comparison to one another.

10.3.1 Hearing the Voice: Interviews with People Experiencing Psychosis

The data for this study were collected by the Hearing the Voice (HtV) team, operating in northeast England and carrying out work to better understand how voice-hearing experiences change over time (Woods et al., Reference Woods, Alderson-Day, Fernyhough, Woods, Alderson-Day and Fernyhough2022). The HtV team interviewed volunteers who had sought support from local early intervention in psychosis services for their voices, and those interviews were transcribed and made available to CASS researchers for (corpus) linguistic analysis. Participants were given pseudonyms, and other personal information was anonymised. Forty individuals took part in semi-structured interviews with the HtV team, describing what their voices are like, when they started, what they do and say, and how they have changed over time. This amounted to 205,941 words of participant data that the CASS team compiled and analysed as a corpus.

One of the objectives of the HtV team was to investigate the personification of the reported voices and how this varies according to degrees of complexity. This aim was predicated on the understanding that voices can variously be attributed attitudes, intentions, and different kinds of identities (even names). They can also manifest as other non-human (demon, birds, bomb) or abstract entities (thoughts, scenario, sensation). The HtV team established a binary classification for manually coding the reported voices in terms of the complexity of personification, according to the following definitions:

Based on these definitions, the CASS research team drew on concepts from literary linguistic theory to develop a characterisation model through which analysts could formalise the documentation of personness in the descriptions of voices by voice-hearers (see Semino et al., Reference Semino, Demjén and Collins2021).

In this chapter, we focus specifically on agency as a component of the HtV team’s definitions for personification and our own view of recording personness (Semino et al., Reference Semino, Demjén and Collins2021). Agency has also been shown to be a key part of other taxonomies developed among clinical psychologists for describing personification (see Wilkinson and Bell, Reference Wilkinson and Bell2016). From a linguistic perspective, the CASS team was interested in a semantic view that allowed them to discuss degrees of agency, as opposed to a strictly binary grammatical view (i.e., who is active and who is passive in a clausal structure; Darics and Koller, Reference Darics and Koller2019). As Darics and Koller (Reference Darics and Koller2019: 219) argue, ‘Clearly, it is more agentive to effect a material change in the world … than to merely become or be something’. To this end, we can refer to the transitivity model and the classification of process types to help us discuss the different ways in which social actors are reported to have agency.

10.3.2 Transitivity

Transitivity is one element of the complex system of linguistic analysis known as systemic functional linguistics (SFL), most commonly associated with Halliday (e.g., Halliday, Reference Halliday1994; Halliday and Matthiessen, Reference Halliday and Matthiessen2014). Transitivity is oriented around meaning insofar as it represents the happenings around us, the states of affairs of the world, and our responses to them. The transitivity framework gives us the resources to describe and evaluate this flow of events ‘as a configuration of elements centred on a process’ (Halliday and Matthiessen, Reference Halliday and Matthiessen2014: 213). These processes involve participants (who enact and are affected by these processes) and circumstances (how, when, and where these processes occur). Each component can broadly be mapped onto clausal elements – that is, the participants are typically realised in the nominal group (it, the voices), circumstances by the adverbial group (today) or prepositional phrase (from next door), and processes in the verbal group (started, will be going).

The various processes through which we can refer to the many different aspects of our experience have been classified into the following process types (Halliday and Matthiessen, Reference Halliday and Matthiessen2014: 214–15):

• Material: construing the outer experience of physical actions

• Mental: representing the inner experience of thoughts and emotions

• Relational: referring to processes of identifying and classifying

• Behavioural: representing the outer manifestations of inner workings such as laughing or sleeping

• Verbal: concerned with saying and expressing meaning

• Existential: recognising the existence or happening of various phenomena

However, Halliday and Matthiessen (Reference Halliday and Matthiessen2014: 218) acknowledge that these are ‘fuzzy categories’; in other words, the boundaries between them are not clear-cut. In response to ambiguities in the classification, alternative taxonomies have been developed and the prevailing schools in this area are those derived from Halliday (Reference Halliday1994), known as the Sydney model (SM) and the Cardiff grammar (CG) model, best described in Fawcett (Reference Fawcett2000) and Neale (Reference Neale2002). The CG model proposes an alternative taxonomy for process types, namely action, mental, relational, influential, environmental, and event-relating. One of the key differences between the models is indicated in the influential category, which captures verbs such start, try, continue, or stop, as well as verbs that denote success and failure (Bartley, Reference Bartley2018). In our discussion, we will take the SM taxonomy as our starting point and refer to pertinent aspects of the CG model as they correspond with the challenges identified through the CASS research team’s investigation of processes described in their interviews with voice-hearers.

10.3.3 Querying References to Voices

Since the CASS team were interested in the variety of labelling terms used to refer to voices, a preliminary task in their analysis was to locate references to voices in the participant responses. A member of the research team read through the interview transcripts and manually tagged references to voices in the corresponding corpus files. This process demonstrated how identifying a referent is based on contextual information, particularly in the case of pronouns (i.e., differentiating when ‘it’ referred to a voice and when it did not) and when the referent is introduced in a preceding turn. For example,

Manual identification of references to voices confirmed that the range of terms used was beyond the team’s introspection and showed that voices could be invoked according to different parts of speech:

• Pronouns: it, she, they

• Nouns: voices, shadow, Roxy

• Determiners: this, some, which

• Verbs as gerunds: commenting, whispering

• Adjectives: for example, ‘there’s one good and one bad’

As such, the procedure of identifying references to voices in the first instance was not amenable to (a simple) corpus query. However, once the researchers had manually tagged references to voices in the data, they were able to run corpus queries for frequency and distribution of voice label types (see Collins et al., Reference Collins, Brezina, Demjén, Semino and Woods2023). This allowed them to determine that there were 9,030 instances of references to voices in the data, expressed as 392 different types. The most common types were it (1952), they (1518), he (732), she (673), them (481), and voices (383).

Furthermore, the tagging of voice referents allowed the CASS team to perform additional queries that highlighted associated qualities and actions through adjective and verb collocates. In particular, the voice labels (i.e., the referents that we tagged) and their collocates directed the researchers to identification strategies that correspond with Van Leeuwen’s (Reference Van Leeuwen2008) social actor network. For example,

• Classification: girl, man, feminine, childlike, old

• Relational: dad, ex-girlfriend

• Physical: big, angular

• Appraisement: evil, clever, useless

These are discussed in some detail by Collins and colleagues (Reference Collins, Brezina, Demjén, Semino and Woods2023), including which kinds of labelling strategies are favoured in cases of minimal and/or complex personification. Here, we consider the reporting of actions, positing the voices as social actors with various degrees of agency.

10.3.4 Classifying Process Types

The investigation of processes associated with voices was guided by the identification of verb collocates of the voice labels that the researchers tagged. The proximity of a referent and a verb does not entail a definite clausal relationship (i.e., we cannot assume that the verb is carried out by or acts upon our node as a participant in the clause). Nevertheless, we can set our collocational window to maximally target subject-verb combinations. Manual checking of the outputs from alternative settings (i.e., three/four/five tokens to the left/right of the node) showed that three tokens to the right of the node were optimal for the precision and recall of processes directly attributable to a voice. The research team set a minimum frequency of one as an association measure in order to capture the full list of verb collocates. They identified 462 different lemmatised verb types, the most common of which were be (1569), say (433), come (280), go (270), do (265), get (252), and tell (230).

The researchers then considered automatically categorising the verb collocate types according to Halliday and Matthiessen’s (Reference Halliday and Matthiessen2014) taxonomy; in other words, collating the verb types that correspond with each process type category (talk as verbal; think as mental, etc.). However, it was problematic to reduce each verb type to a singular meaning that corresponded with just one process type category. Other researchers have found that because it is a semantically oriented task, there is a range of contextual factors, along with different usage patterns, that make it difficult to develop algorithms for automatic process-type classification (Yan, Reference Yan2014). As such, while the CASS team’s observations began with a table of the verb collocates and their frequency, the researchers’ interpretations were based on instances as they occurred in the original context of the interview.

One of the first steps in classifying process types is to identify the main process in a clause. Bartley (Reference Bartley2018: 12) explains that there are differences in how the SM and CG model instruct analysts to approach phrasal elements with a catenative verb. These differences apply to verb phrases such as

• She started talking about them.

• He wants to hurt me.

• They’re just trying to distract me.

In such cases, the SM would tend to focus on talking, hurt and distract as processes, whereas the CG model considers started, wants and trying to be the main processes. Furthermore, verbs such as start and try typify the influential process category in the CG model. Researchers must establish whether they will record one or both these elements of the process and decide which model they want to refer to, if they are looking to (quantitatively) describe patterns according to these top-level process types. Choosing either the SM or the CG model will generate very different views of the (same) data.

In the analysis of the voice-hearer interviews, while the researchers might collectively address the category ‘material processes’ based on the prevailing meaning of terms such as control, fight, and hurt, it is important to acknowledge the significance of the extended verbal group (trying to hurt), given its clear relevance to the discussion of capabilities and intentions – in other words, agency (see the definitions provided in Section 10.3.1). There are other instances in which the researchers focussed on the particular elements of a verbal phrase, based on its relevance to the inquiry into agency. For example, Collins and co-authors (2019: 48) discuss the qualitative differences in how verbs such as stop are used: the intransitive form (e.g., the voices stopped) demonstrates very limited agency; however, in the non-finite complementation clause ‘they’ve stopped me from doing so much’, the voice-hearer is explicitly attributing a negative outcome to the voice, which is shown to have the capacity to ‘make something happen’.

The CASS research team found that the capacity of the voice(s) to communicate was of particular significance to their investigation, with communication verbs such as terms say, tell, and talk appearing as high-frequency collocates of references to voices. Bartley (Reference Bartley2018: 13) is critical of the CG model for its classification of communicative processes (under the mental cognition category), as this seems to understate the significance of the action of deliberately transferring information to other sources – which has direct relevance to our discussion of the agency of voices. Verbal processes are treated as a distinct category in the SM and defined as covering ‘any kind of symbolic exchange of meaning’ (Halliday, Reference Halliday1994: 140). Among the terms that are included in this category, it is possible to further differentiate levels of agency and impact. For example, Collins and colleagues (2019: 48) discuss how respond and answer indicate the capacity for voice-hearers to participate in dialogue with their voices. This is particularly significant when the unidirectional nature of threats and shouting is often reported as a source of distress for voice-hearers, when they have limited options for affecting the interpersonal and communicative dynamic.

Similarly, there are certain mental processes that describe different levels of intent and subsequently – different degrees of surveillance or antagonism that can be a source of distress for voice-hearers. For instance, some voices are reported to judge, embarrass, hate, or reassure the voice-hearer, demonstrating their capacity to ‘want/think/feel’ (Alderson-Day et al., Reference Alderson-Day, Woods, Moseley, Common, Deamer, Dodgson and Fernyhough2021). Bartley (Reference Bartley2018: 5) explains that the subcategorisation of the CG model offers more detail – compared with the SM – for distinguishing ‘the act of consciously perceiving something and doing so intuitively’, which we can see in the comparison between saw versus looked and hear versus listen. This distinction can be particularly useful in capturing the perceived intent behind the voice ignoring the voice-hearer, for example.

10.3.5 Considerations for Investigating Transitivity

In this exploration of voice-hearer accounts, we have established that the concepts of agency and positing participants as social actors are of interest to researchers working in disciplines beyond linguistics, but that language-based theories offer frameworks for documenting dimensions of identification and agency in the labelling strategies and description of processes. While there are established frameworks that are predicated on shared principles for how agency is encoded in language, we have seen that there are different approaches to categorising processes. This means that researchers will need to critically reflect on which aspects they are particularly interested in capturing, particularly if they are looking to provide quantification, since fundamental steps such as identifying the main process in a clause will have implications for what is categorised and how. The different practices for documenting processes and related participants show how this is an analytical procedure that is difficult to automate; nevertheless, we have shown how wordlists and collocation analysis (following annotation) can help in identifying the most common terms used to denote participants and processes. Subsequently, we have considered some of the complexities in how processes are described. Given these complexities, processes that appear in the data warrant close, contextualised examination – particularly for the purposes of discerning degrees of agency attributed to different social actors in the text.