2.1 Register variation, World Englishes and CMC

Register is an important driver of variation. However, register analysis is not equally incorporated in every study of World Englishes. Bohmann (Reference Bohmann2019: 32–5) – building on Neumann's (Reference Neumann2014) categorisation – distinguishes three kinds of studies according to the level of incorporation of register: register monolithic; register controlled; and the ones focusing on register as the main object of inquiry. The first two types consider only single or a few features and apply detailed and contextualised attention to them, while the third type takes a feature-aggregate approach (Bohmann Reference Bohmann2019: 34). MD analysis falls under the third category, where the normalised frequencies of a large set of functionally relevant lexico-grammatical features are obtained. Then, co-occurring patterns of these features – called ‘dimensions’ – are discovered using statistical techniques and functionally interpreted in the light of the situational context and linguistic usage of the registers under study (cf. Biber Reference Biber1988; Biber & Conrad Reference Biber and Conrad2019). There have been a few comprehensive MD studies on traditional registers in the context of World Englishes (see section 2.3), but research on online registers has been less common.

In the last few decades, the internet has created new avenues of communication for users. At the same time, this has presented opportunities as well as challenges for researchers in the field of sociolinguistics and World Englishes. At the simplest level, online registers provide a treasure trove of ready-to-use data for small- and large-scale studies (Bohmann Reference Bohmann2019: 35). In particular, interactive and informal online registers, which are mediated but increasingly intertwined with the real-life of netizens, supply vernacular data for the study of sociolinguistic phenomena (Heyd & Mair Reference Heyd, Mair, Lacoste, Leimgruber and Breyer2014: 249). This is especially important for smaller varieties, e.g. Bangladeshi English, where it is hard to collect data in traditional domains. Since online communication has a global as well as a local aspect (Bohmann Reference Bohmann2019: 37), internet-based registers can be used as an alternative to traditional registers. A study of the same set of registers from multiple varieties of English, e.g. those from South Asia, using MD analysis can uncover (communicative) functions associated with English in conjunction with associated features in a holistic and comparative manner across registers and varieties. This can help us to understand the development of these varieties – individually and in the South Asian region as a whole – on the internet. Further, online data can serve as a proxy for offline situations, but with certain caveats. In fact, the use of a multi-register CMC corpus – as opposed to a dataset like GloWbE (Davies & Fuchs Reference Davies and Fuchs2015) where CMC registers are not as clearly defined (cf. Loureiro-Porto Reference Loureiro-Porto2017) – can also improve the World Englishes community's understanding of variation in internet registers, which, in turn, can help clarify whether and to what extent different online registers are valid proxies for offline registers.

2.2 Previous relevant MD studies on online registers

In the following, the findings of relevant MD studies on online registers – i.e. comments, tweets, chats and web forums – are discussed.

Ehret & Taboada (Reference Ehret and Taboada2020, Reference Ehret and Taboada2021) examine Canadian newspaper comments by conducting new MD analyses. They find that newspaper comments are nearer to written registers such as exams, opinions and social letters than spoken registers. In comparison to the online registers of Biber et al. (Reference Biber, Egbert and Davies2015), online news comments are informational, argumentative and also include evaluative discourse. They can range from very short to quite long argumentative and encyclopaedic types of texts (Ehret &Taboada, Reference Ehret and Taboada2021: 8). In contrast, YouTube, ESPN and Yahoo! comments have been found to be involved, personal and opinionated (Titak & Roberson Reference Titak and Roberson2013: 255).

Twitter discourse has been included in a number of MD studies. Titak & Roberson (Reference Titak and Roberson2013) use blogs, emails, newspaper articles, opinion columns, comments (see above) and Twitter/Facebook posts merged in one category. As per the new MD analysis, Twitter and Facebook posts are nominal and informational with very limited narrativity (Titak & Roberson Reference Titak and Roberson2013: 254). Friginal et al. (Reference Friginal, Waugh, Titak and Friginal2018: 358) use the dimensions identified in Titak & Roberson (Reference Titak and Roberson2013) to more closely examine Twitter and Facebook posts related to weather, politics, sports and personal topics. Their analysis shows that generally tweets are less personal, less narrative, less involved and less interactive than Facebook posts (Friginal et al. Reference Friginal, Waugh, Titak and Friginal2018: 358). The authors cite length restrictions in Twitter as one of the possible reasons for this type of discourse. Berber Sardinha (Reference Berber Sardinha, Sardinha and Pinto2014) in his study comparing online registers to non-internet registers using additive MD analysis – where ‘old’ dimensions are applied to a new corpus (Berber Sardinha Reference Berber Sardinha, Sardinha and Pinto2014: 81) – based on Biber's (Reference Biber1988) dimensions demonstrates that online registers related to personal communication like Twitter have similarities to spontaneous speeches, interviews and personal letters. Finally, Berber Sardinha's (Reference Berber Sardinha, O'Keeffe and McCarthy2022: 665–9) analysis of Twitter along with Facebook and Instagram posts reveals that tweets have a comparatively higher score on dimension 2 ‘Informal, interactive and speaker-oriented discourse’.

One of the registers most closely related to text messages, i.e. ICQ chats, has been compared with traditional spoken and written registers using Biber's (Reference Biber1988) dimensions of variation (Jonsson Reference Jonsson2015). The study's findings indicate that ICQ chats are just like face-to-face and telephonic spoken conversations on dimension 1 – i.e. involved, interactive and affective – but they are comparatively less narrative on dimension 2 ‘Narrative versus non-narrative production’. Otherwise, ICQ chats are similar to the above-mentioned registers, in that they have situation-dependent reference, no overt expression of persuasion and non-abstract information on dimensions 3, 4 and 5 (Jonsson Reference Jonsson2015: 292).

Collot & Belmore (Reference Collot, Belmore and Herring1996) conduct an additive MD analysis using Biber's (Reference Biber1988) dimensions on message boards, an earlier form of web forums. The authors note that message boards are highly involved and relatively situation-dependent due to the shared interest of the participants, despite non-shared time and space. In their opinion, the communicative purposes of giving information and discussing specific issues mean that message boards have a non-narrative and highly persuasive discourse. Lastly, they reason that the similarity between message boards and public interviews could be due to the social roles of addressor, addressee and audience (Collot & Belmore Reference Collot, Belmore and Herring1996: 26).

2.3 Previous relevant MD analyses on register and regional variation

MD studies focusing on geographical variation use different combinations of traditional registers or traditional and online registers. The two most comprehensive studies have been conducted by Xiao (Reference Xiao2009), using registers from the International Corpus of English (ICE) – private and public dialogues, scripted and unscripted monologues, unpublished writing such as student essays and personal letters, published writing such as academic writing, popular writing such as magazines, creative writing, press reportage etc. – and Bohmann (Reference Bohmann2019), who uses ICE registers along with Twitter discourse. Xiao's (Reference Xiao2009) MD analysis shows that Indian English is the most elaborate on dimension 1 ‘Interactive casual discourse versus informative elaborate discourse’. On dimension 5 ‘Future projection’, Indian English scores the lowest on all registers (Xiao Reference Xiao2009: 447). Bohmann (Reference Bohmann2019) confirms Xiao's findings, i.e. Indian English is among more informational varieties on dimension 1 ‘Involved versus informational production’, which is true for all mediums, i.e. spoken, written and Twitter (Bohmann Reference Bohmann2019: 109). Similar observations have been made by Hussain's (Reference Hussain2016) study on an incomplete ICE corpus of Pakistani English. Lastly, Kruger & van Rooy's (Reference Kruger and van Rooy2018: 237) analysis of written ICE registers finds that less advanced OC varieties like Indian English generally avoid features related to informality and involvement.

Online registers of OC varieties show similar differences when compared with a native English register. For example, Hardy & Friginal (Reference Hardy and Friginal2012: 159) observe that Philippine English blogs typically deviate ‘from personal and involved stylistic features of informal writing such as blogs written by American authors’. Similarly, Shakir & Deuber (Reference Shakir and Deuber2018, Reference Shakir and Deuber2019) note that Pakistani online registers are informational and restrictive in communicative purposes compared to their US English counterparts.

Overall, whilst there have been comprehensive MD studies on traditional registers across several IC and OC varieties, there has not yet been an examination of multiple CMC registers with a focus on OC varieties from the South Asian region. The present analysis, thus, fills this research gap by conducting an exploratory study by applying MD analysis on four interactive online registers, i.e. newspaper comments, tweets, web forums and text messages, from the six varieties of English (Bangladesh, India, Pakistan, Sri Lanka, UK, USA). Based on previous literature, it is expected that the four South Asian varieties will tend to have formal, literate and informational discourse in these registers.

3.1 Data collection and preprocessing

The data for the present study have been sampled from the SAOnE corpus (Shakir & Deuber Reference Shakir and Deuber2023). The down-sampling was necessary for better preprocessing, especially checking non-standard spellings (see below). Table 1 presents the corpus structure, brief descriptions of the sources, and the sample included in this analysis. In general, I have sampled 50 files per category, which is 10 per cent of the total of 500 texts in most categories. Text messages are an exception as they have only half the word count (500,000 words) with conversations sometimes larger than 50,000 words. These files were sliced into 2,000-word texts and then 50 files were sampled for each country. The total files included in the study are 50 texts * 6 countries * 4 categories = 1,100. There are no data for text messages from the UK and USA.

Table 1. Description of corpus categories in the main corpus with texts included in this study (adapted from Shakir & Deuber Reference Shakir and Deuber2023: 123)

The preprocessing consisted of three steps. The first step was to check for non-standard spellings to improve the grammatical tagging process in the later stages. For this purpose, I have used the spellchecking module syplls in Python to help the grammatical tagger (Shepelev Reference Shepelev2022; Python Core Team 2022). The corrected spellings were checked by the project assistants to reverse any mistakes of the automatic spelling correction. For example, Ambar is a proper noun, which was falsely replaced with Mbar by the spellchecker. The original word was restored in the review process. The data included in SAOnE were also annotated for indigenous content present in the form of single words, multiword expressions and sometimes full posts (i.e. tweets, comments, web forums, text messages). In the second step multiword phrases and clauses of indigenous content present in the South Asian data were removed. These expressions would be tagged either as foreign words (FW) or mistagged by the grammatical tagger. Both of these outcomes were not helpful for the present analysis. The third step was hashtag segmentation. Hashtags in tweets can consist of multiple words which can include nouns and adjectives – e.g. #Trainingprovider – or even clauses – e.g. ‘#ILOVENYC’ (Berber Sardinha Reference Berber Sardinha, O'Keeffe and McCarthy2022: 659). Following Berber Sardinha, I have used the Python module wordsegment (Jenks Reference Jenks2018) to split hashtags – e.g. Training provider, I LOVE NYC – to help the grammatical tagger in identifying word classes.Footnote ³

3.2 Grammatical tagging and MD analysis

Before describing the actual MD analysis and results, I address three issues related to MD analysis. The first issue is related to the selection of features. Generally, the most popular software used for tagging MD related features is BiberTagger (also used in Biber Reference Biber1988). The features tagged by BiberTagger are a combination of lexico-grammatical features introduced in the original 1988 study and more detailed semantic subclasses based on Biber et al. (Reference Biber, Johansson, Leech, Conrad and Finegan1999) and later works such as Biber (Reference Biber2006). Biber's (Reference Biber1988) 67 linguistic features can also be tagged by Multidimensional Analysis Tagger (MAT), which was developed by Nini (Reference Nini2014, Reference Nini, Berber-Sardinha and Pinto2019) and is now available as an opensource tool. This tool has been modified by Le Foll (Reference Le Foll2021a) and released as a separate tagger. The main changes introduced by Le Foll are to make the tagging process more reliable and reduce the need for manual correction as much as possible. For this reason, she excludes features that have very low precision and recall. Some examples include present and past participial clauses, present and past participial WHIZ deletion relatives, that relative clauses on subject and object positions, WH relative clauses on subject and object positions, pied-piping relative clauses, sentence relatives, phrasal and clausal coordination etc. The features listed above have been replaced with a smaller set of features, namely that complement clauses, that relative clauses, WH subordinate clauses, nonfinite -ing forms of verb, nonfinite -ed forms of verb, coordinating conjunctions etc. Additionally, she includes some semantic verb classes from Biber (Reference Biber2006), e.g. activity verbs, communication verbs, mental verbs etc. (see link for the complete list of features included). I have selected Le Foll's feature set because it has over 90 per cent accuracy (Le Foll Reference Le Foll2021b: 34) including on internet-related data, and because manual correction of each feature is a very time consuming task for large number of texts, as for example included here.

I have further expanded Le Foll's (Reference Le Foll2021a) feature set by including Biber's (Reference Biber2006) and Biber et al.'s (Reference Biber, Johansson, Leech, Conrad and Finegan1999) features related to semantic classes of nouns, verbs, adjectives and adverbs. Apart from vocabulary items, there are that, to and WH clauses that can be preceded by different types of nouns, adjectives and verbs, for example that complement clauses preceded by communication verbs. I have modified Le Foll's script ‘Multi-Feature Tagger of English’ or MFTE to include these semantic subtypes in the general categories of nouns, verbs, adverbs, adjectives and that/WH/to clauses. The combination of Le Foll's feature set and my modifications enables me to tag for more than one hundred lexico-grammatical features that are very similar to the output of BiberTagger as they are mostly defined and operationalised following Biber's work. Due to space restrictions I cannot list definitions and examples of all these features here. However, I have put online a supplementary document defining each feature (link). The individual features will be mentioned and contextualised as and when they appear in the analysis below.

The second issue is related to the normalisation unit for the absolute frequencies that will later be used as the input table for MD analysis. Traditionally MD studies use a per-thousand-words normalisation approach for all lexico-grammatical features. However, some authors argue that word-based normalisation may not be suitable for every feature. For example, Evert (Reference Evert2022: 27) points out that passive voice is essentially dependent on total number of verbs instead of total number of words in a text. Moreover, word-based normalisation can create confusion between ‘covariation of features due to situational variation’, which is desired in MD analysis, as opposed to ‘covariation due to grammatical structure’, which is undesired (Le Foll Reference Le Foll2022: 285) – hence prepositions will highly correlate with nouns due to the grammatical constraints of prepositional phrases.

One possible solution for the above is to normalise features, where possible, based on the grammatical feature they depend on mostly – hence prepositions normalised on total number of nouns. Le Foll's (Reference Le Foll2021b) ‘complex normalisation approach’ is one such attempt (see footnote).Footnote ⁴ Though it has its own shortcomings, e.g. it is not always clear which feature is the appropriate normalisation baseline for a feature like adjectives. My experiments with both normalisation methods on the present data show that the former (word-based normalisation) results in slightly larger dimensions with some additional features – especially on the first two dimensions –compared to the latter (‘complex-normalisation’). Otherwise, the interpretation of the dimensions remains the same due to mostly identical features. I adopt the latter for the present study, as both methods result in conceptually similar dimensions for the present data and the latter, to a certain extent, suppresses grammatical covariation.

The third issue is related to the method applied to extract the groups of co-occurring lexico-grammatical features or dimensions. Traditionally, exploratory factor analysis (EFA) has been used in MD studies, including in the pioneering study Biber (Reference Biber1988). The researcher needs to decide the number of factors that should be extracted. Additionally, the factors are rotated to create ‘simple structure’ to help interpret the latent constructs in the data (Bohmann Reference Bohmann2019: 89; Le Foll Reference Le Foll2022: 289). An oblique rotation like ‘promax’ is recommended in MD studies as it allows the factors to be correlated just like everything in language is also correlated with each other (Biber Reference Biber1988: 85). However, orthogonal rotations like ‘Varimax’, which do not allow correlation among factors, have also been used in MD studies such as Grieve et al. (Reference Grieve, Biber, Friginal, Nekrasova, Mehler, Sharoff and Santini2010: 309).

PCA is a similar dimensionality reduction technique that has also been used in MD studies, e.g. Biber & Egbert (Reference Biber and Egbert2016). It has been suggested as an alternative to EFA as it can be more objective by eliminating the need to determine the number of dimensions beforehand (Evert Reference Evert2022: 31). Just like EFA, PCA can also be rotated, as for example Biber & Egbert (Reference Biber and Egbert2016) do. This, however, reintroduces the issue of determining the number of dimensions to be extracted and the composition of resulting components – the first and second components in a two-component rotated solution will be different from the first two components in a three-component rotated solution (Le Foll Reference Le Foll2022: 290). Moreover, the rotated principal components ‘no longer successively account for maximum variance’ (Rencher Reference Rencher1992: 222). Considering the above and my experiments with various combinations of rotated and unrotated PCA solutions, I have used unrotated PCA for the present analysis as the resulting dimensions are easier to interpret.Footnote ⁵

The data have been tagged using Stanford Tagger (Toutanova et al. Reference Toutanova, Klein, Manning and Singer2003) and as the second step by the modified MFTE script as described above.Footnote ⁶ I have utilized psych::principal() in R (Revelle Reference Revelle2022; R Core Team 2022) for PCA. To handle outliers in a better way, I also used z score transformation and then signed-log transformation as recommended by Neumann & Evert (Reference Neumann, Evert, Seoane and Biber2021: 155). Before the application of PCA, I have removed features having 0 occurrences in 66 per cent or more files as well as features with less than |0.1| correlation with all other features, because such features are most probably less relevant to the data under study. Very high correlations can skew the results (Le Foll Reference Le Foll2022: 295) so I have excluded features with very high correlation (>|0.9|) with any other feature. As per Egbert & Staples’ (Reference Egbert, Staples, Berber-Sardinha and Pinto2019: 129) recommendation, features having communalities lower than 0.2 have been removed in this analysis. Kaiser–Meyer–Olkin (KMO) is a measure of sampling adequacy and its value lower than 0.5 means that the data are not appropriate for MD analysis (Egbert & Staples Reference Egbert, Staples, Berber-Sardinha and Pinto2019: 129). It can be calculated for individual features as well for the whole data. I checked for individual features having a KMO less than 0.5 and removed them.Footnote ⁷ The final data table has a KMO of 0.85 (=Meritorious) with 80 linguistic features. Five dimensions of variation were retained in the final PCA as the scree plot flattens after five dimensions in both PCA and EFA (see figure 1). The total variance explained by the model is 40 per cent, which is normal for MD studies: e.g. Biber & Egbert (Reference Biber and Egbert2016: 12) report 42.7 per cent variance accounted for. Dimension scores have been obtained using the regression method, which is the default in the principal() function.

Figure 1. Scree plot of the data (FA and PCA)

Before presenting the results in the next section, a brief explanation of the statistical methods is given here. Dimension scores do not fulfil the assumptions of classic ANOVA, i.e. the within-group variance is not equal and the distributions are not normal. Therefore, a robust regression model was fitted using MASS::rlm (Venables & Ripley Reference Venables and Ripley2002). Then, type III ANOVA tables were generated using car::Anova (Fox & Weisberg Reference Fox and Weisberg2019). Group comparisons were run on the ANOVA object using the emmeans package at 95 per cent confidence level (Lenth Reference Lenth2023). The value of R ² gives an indication of the importance of a dimension as it shows how much variance in the dimension scores of the given dimension can be explained by knowing the text categories (Biber & Egbert Reference Biber and Egbert2016: 12). The text groups in this case consist of register as well as country-based groupings. The R² was generated using the interaction term of country and register in lm() in base R as rlm() estimated higher values.

The figures contain information boxes. Apart from the significance of predictors and R ² values, the information box shows significant emmeans compared to ‘Both’ of the IC varieties or ‘UK’ or ‘US’. ‘Both:rel. group’ indicates that the relevant register in both IC varieties is significantly different from the registers in South Asian countries shown in brackets. Hence line 3, entry 1 in figure 2 shows that Sri Lankan comments are significantly different from both US and UK comments. Indian comments are only significantly different from US comments (line 4 first entry) and so on. Red dashes show overall register means.

Figure 2. Distribution of registers on Dim1 ‘Oral personal (top) versus literate impersonal discourse (bottom)’

Note: Country names are two-letter ISO codes: IN is India; BD is Bangladesh; PK is Pakistan; LK is Sri Lanka; UK is United Kingdom; and US is United States of America. Register codes are as follows: CMT for comments; TWT for tweets; WBF for web forums; and TXM for text messages.

3.3 Dim1 ‘Oral personal versus literate impersonal discourse’

Dim1 consists of 40 variables, of which 10 occur on the positive side and 30 on the negative side, as table 2 reveals. The positive features are related to involved and informal discourse, e.g. pauses and interjections, contractions, pragmatic markers and politeness markers, as in (1). Non-standard spellings and internet slang are used in informal CMC. There are features related to interactive and personal communication like first- and second-person pronouns and imperatives. The features on the negative side are related to impersonal, nominal and literate discourse, which include prepositions, passives, nominalisations, nouns like abstract process, cognitive and place, non-finite verb -ed and -ing forms – which may or may not introduce non-finite clauses as in (2), high ratio of content to function words, high type token ratio and longer words. Features like attributive adjectives and determiners modify nouns and make the noun phrase longer and complex. These nouns are most probably not proper nouns, which do not take determiners in English. Moreover, clauses like that relatives and WH clauses can be used to add further information. Lastly, the presence of third-person plural pronouns indicates that the discourse focuses on third parties. Based on these traits, the first dimension, as in many previous MD studies, can be labelled ‘Oral personal versus literate impersonal discourse’.

1. (1) So yeah [DMA].

   <#>

   . Ah [FPUH].

   <#>

   Ye ye.

   <#>

   . That blue colour uniform.

   <#>

   . When did you [SPP2] change.

   <#>

   Yea [FPUH] I [FPPAll] was in Vidura before!! After o/ l..

   <#>

   Okay [DMA] so that makes you [SPP2] bright.

   <#>

   . Hyatt.

   <#>

   . No [DMA] I [FPPAll] ‘m [CONT] not.

   <#>

   Yeah [DMA] you [SPP2] are.

   <#>

   . How many As did you [SPP2] get for o/ l?

   <#>

   O/ l results?

   <#>

   3as.

   <#>

   Wow [FPUH] nice. (Dim1 Oral personal; LK-TXM162)

2. (2) Start your own [JJATother] domain name [NNABSPROC] and [CC] hosting [VBG] company [NNGRP]! Free [JJATother] billing and [CC] tech support [NNABSPROC] 24/7, actually, it ‘s the [DT] same company [NNGRP] that [THRC] sells godaddy their [TPP3P] reseller program! You literally have the [DT] same products and [CC] support [NNABSPROC], they [TPP3P] just answer the [DT] phone as [INother] “customer support [NNABSPROC]”. But [CC] feel safe your reseller program is managed [PASSAll] by [INother] the [DT] same great support [NNABSPROC], security [NNABSPROC] and [CC] reliability [NOMZ] as [INother] the [DT] big dogs! […] (Dim1 Literate impersonal; PK-WBF015.txt)

Table 2. Dimensions and features of MD analysis^a

^a Features different from Biber (Reference Biber1988) have examples in front of them (see also the supplementary feature description file). Features marked with * are my additions based on Biber (Reference Biber2006) or BiberTagger (using Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999). Semantic verb classes from Biber (Reference Biber2006) are adopted by Le Foll. Cut-off point is |0.295| (rounded to |0.3|). Dimension 1 was inverted for better interpretation.

^b Non-standard spellings were calculated separately before replacing them with standard spellings.

Scores on Dim1 are the most sensitive to the groups of registers and countries included in these data, as shown by the R ² value of 0.59 or 59 per cent – hence a good discriminator of these text categories. As figure 2 shows, text messages have the highest score on the positive side and comments are the most literate register. Tweets are second on the oral side, while web forums are slightly literate. The Bangladeshi tweets and web forums have comparatively higher scores in South Asia due to oral features like filled pauses, non-standard spellings (not shown in the example), first- and second-person pronouns, politeness markers etc. See (3) for forum posts where someone is being wished a happy birthday and is answering someone else's post, and (4) for a couple of tweets (fan addressing two famous twitter accounts):

1. (3) ahhh [FPUH] I [FPPAll] see now [TIME] … Happy Birthday.

   <#>

   Happy Birthday ……….

   <#>

   Thank [POLITE] you [SPP2] very much to all of my [FPPAll] buddies. Your [SPP2] wishes mean a lot to me [FPPAll]. (3 forum posts from BD-WBF336)

2. (4) @X Please [FPUH, POLITE] take care.

   <#>

   @XxxxxxXxxxxxxxx I [FPPAll] love this short film !! That’s [CONT] so inspirational (2 tweets from BD-TWT184)

A comparatively higher presence of the same type of features in the Sri Lankan comments makes them more oral personal compared to the other South Asian comments and the UK and USA. Apart from that, the Indian comments and tweets, and Pakistani and Sri Lankan tweets are significantly more literate in comparison to the relevant categories in UK and USA. Lastly, more South Asian registers have significant differences from their US counterparts (as shown by line 4 in the information box of figure 2) compared to the corresponding UK registers.

3.4 Dim2 ‘Oral elaboration versus informational concerns’

Dim2 has a total of 33 features with 25 on the positive and eight on the negative side. Some features on the positive side point to an oral and informal discourse, e.g. contractions, auxiliary do, that omission, demonstratives, discourse markers, internet slang etc. Features like adverbs (general, stance-related, hedges, emphatics), mental verbs like wonder and appreciate, to infinitives preceded by stance words, adjectives (evaluative, size, relational), modal would etc. can be used to talk about or express opinions/stance about people and things using human nouns and pronouns (quantifying, it). The presence of determiners on the positive side – like the negative side of Dim1 – shows that most of the nouns used here are not proper nouns, hence they require articles and other determiners. For example, a user is expressing their personal opinion about Covid problems in a comment in (6). Thus, the positive side can be labelled ‘Oral elaboration’ like Biber & Egbert's (Reference Biber and Egbert2016) dimension 2. The most important feature on the negative side is general nouns that do not need a determiner, i.e. proper nouns. These long account handles (hence higher average word length) are notable in (7), which is from the account of a Pakistani political worker who tags many party accounts in their tweet. Occasionally such tweets have other features like non-finite -ed verbs, non-finite -ing verbs, imperatives, or time references (e.g. today). The use of time references and imperatives can be seen in (5), which shows an advertisement-related Indian tweet. Hence, the appropriate title for the negative side can be ‘Informational concerns’.

1. (5) Now [TIME] Book [VIMP] reliable and [CC] Safe Cab [NNother] Service across Jharsuguda [NNother] at an affordable price [NNother]. (IN-TWT217)

2. (6) Since the [DT] pandemic started, I do [DOAUX] n’t [XX0, CONT] feel [MENTAL] that much [QUAN] of a [DT] need [NNCOG] to exercise (even [RBother] though I still [RBother] do [DOAUX]), especially [RBother] since school is online, and nobody [QUPR] really [EMPH] sees [MENTAL] more [QUAN] than your face on a [DT] screen. This [DEMO] is a [DT] good [JJEVAL] thing for lots [QUAN] of people [NNHUMAN] who sometimes feel [MENTAL] pressured to have a [DT] “perfect” body, including myself. (Dim2 Oral elaboration; US-CMT333)

3. (7) Many many returns of the day Happy birthday [NNother] chairman sb [NNother] Allah [NNother] bless you @XXxxxxxXxxxxxx [NNother] @XXXXXXXXXXXXXXX [NNother] @XxxXxxxxx [NNother].

The second dimension is rather less sensitive in discriminating the text groupings in these data as R² value based on the interaction of country and register is only 0.26 (26 per cent). The red dash in figure 3 reveals that web forums have the highest overall dimension score compared to other registers. Comments and text messages are also slightly inclined towards the ‘oral-elaboration’ pole of Dim2. Tweets are mostly informational, which is partly due to the presence of proper nouns in the form of @ mentions. Comments, tweets and web forums also show clear regional trends, with the South Asian countries being significantly less oral-elaborative – i.e. having fewer oral and stance-related features – in comments and web forums, and more informational in tweets compared to both native varieties. Text messages are almost in the middle, with Pakistani text messages inclining towards the informational side probably due to communicative purposes like study/work-related information sharing (see section 3.8). The comparatively lower number of stance-related features in the South Asian comments shows that these comments employ different features for this purpose. The communicative purpose of information sharing appears to be more relevant for South Asian tweets as well as web forums.

Figure 3. Distribution of registers on Dim2 ‘Oral elaboration (top) versus informational concerns (bottom)’

3.5 Dim3 ‘Persuasion- and help-oriented versus informational and past-oriented discussions’

In total 21 linguistic features have loaded on Dim3, divided into 13 and 8 on the positive and negative side respectively. The features on the positive side include vocabulary items like abstract and process nouns, technical nouns and nominalisations, which show that technical and abstract things are under discussion. Some of these features are shared with Biber's (Reference Biber1988) dimension 4 ‘Overt expression of persuasion’, including possibility and necessity modals which refer to actions, i.e. what could/might or can/may or will be done and what should or must be done. Conditionals and facilitation verbs like help, require, force in the present data indicate that audience is persuaded, e.g. in comments, or explanation/ guidance is provided or demanded – as in the two forum posts in (8). Communication verbs like threaten, encourage, ask etc. facilitate with persuasion or obtaining help, whereas yes/no questions and human nouns can indicate an interactive discourse. The negative side has nominal features like general nouns, including proper nouns and type token ratio. Past tense verbs indicate the presence of narrative elements. There are also some informal features like contractions, verb particles potentially due to phrasal verbs, along with emphatics, frequency-related adverbs, and clause connectors like still, although, which may suggest some form of discussion, as shown in (9) from Indian comments. Looking at both types of features, this dimension can be labelled ‘Persuasion- and help-oriented versus informational and past-oriented discussions’.

1. (8) I have completed the full course. But, i could [MDPOSSCAll] not answer [COMM] the question [NNABSPROC] no 10. i tried to find the answer [NNABSPROC] by using logic.

   but, i could [MDPOSSCAll] not find the answer [NNABSPROC]. Can [MDPOSSCAll, YNQU] you give me the solution [NNTECH] so that i can [MDPOSSCAll] learn it?

   <#>

   You should [MDNE] be able to get the correct answer [NNABSPROC], if [COND] you use countif. (Dim3 Persuasion- and help-oriented discussions; BD-WBF069)

2. (9) The biggest tell [NNother]- tale [NNother] sign of India [NNother] ‘s western neighbour [NNother] ‘s support for terrorism [NNother] took [VBD] place in 2011. In 2011 mastermind [NNother] of 9/11 terror [NNother] attack was [VBD] eliminated in its territory. Immediately thereafter mastermind [NNother] of 26/11 terror [NNother] attack along with a separatist [NNother] from Kashmir [NNother] valley held [VBD] a prayer [NNother] meeting for the eliminated terror [NNother] mastermind. (Dim3 Informational and past-oriented discussions; IN-CMT380)

As shown in figure 4, the scores on Dim3 are quite sensitive (overall the third most) in discriminating the text categories used in these data, as the variance explained (R²) based on the interaction of register and country is 0.45 or 45 per cent. Text messages have the highest average scores compared to other registers. The top-scoring texts in this register deal with study-related topics and contain features like abstract and process nouns (quiz, attempt, link etc.), modals (will, need, should etc.) and yes/no questions etc. Comments have the highest and tweets the lowest scores, whereas web forums are in-between on the border. In the first two registers, the South Asian countries are significantly different (i.e. more features closer to the ‘persuasion- and help-oriented’ pole) from the UK and USA (i.e. more features closer to the ‘informational and past-oriented’ pole). In terms of forums, Bangladesh, Sri Lanka and Pakistan have significant differences. Also, the spread of the whiskers, especially in Bangladesh and Sri Lanka, shows that there are certain texts where people ask for help on technical topics. Forums related to coding (BD), telecommunication company services (LK) or broadband-related help (PK) appear to be the source of such texts, which have no equivalent in the Indian data included here. Such posts are also less frequent in the UK and USA technology-related forums in these data. The Indian and Pakistani comments also have more past-oriented (i.e. past tense) and general informational features (e.g. general nouns), as shown in (9).

Figure 4. Distribution of registers on Dim3 ‘Persuasion- and help-oriented (top) versus informational and past-oriented discussions (bottom)’

3.6 Dim4 ‘Focus on humans and mental processes versus non-abstract things and activities’

The fourth dimension consists of 14 features equally divided on both sides. The use of present tense, be as the main verb, and human nouns in (10) shows that the focus is on humans. Additionally, mental verbs and WH questions indicate that the participants are engaging with each other or the writer by asking questions, e.g. by mentioning other Twitter users as in (11).

1. (10) Bus drivers [NNHUMAN] are [VPRT, BEMA] a menace on the road. I request [VPRT] the owners [NNHUMAN] to take measures to ensure safety of public on the roads. (Dim4 Focus on humans and mental processes; BD-CMT026)

2. (11) @xxxxx @XxxxxxxxxXXxxxx where [WHQU] are [VPRT, BEMA] you from? (LK-TWT276)

Topical adjectives like economical, political, human etc. are used for rather abstract issues, as seen in (12) taken from Bangladeshi comments:

1. (12) […] This is [VPRT] n't just about being [BEMA] a feminist … this is [VPRT, BEMA] about fighting for the basic human [JJTOPIC] right to education and liberty. […] (BD-CMT021)

The positive side, thus, can be labelled ‘Focus on humans and mental processes’. On the negative side, concrete, quantity and compound nouns refer to non-abstract and technology-related items as observable in (13). Activity verbs and modal verb would are associable to users’ personal experiences or actions, also exemplified in (13). Hedges and verb particles can be attributed to interactive discourse like web forums. In contrast to the positive pole, ‘Focus on non-abstract things and activities’ appears to be an appropriate label for the negative side.

1. (13) […] For water [NNCONC] blocks [NCOMP, NNCONC] I ’d [MDWO] like to use [ACT] Heat [NNQUANT] Killer but they do n't currently have any plans for RTX3090 blocks [NCOMP, NNCONC] so I suspect that I might use [ACT] EK blocks [NCOMP, NNCONC].

   For hardline tubes [NNCONC] and fittings I ’ve previously only used [ACT] […] (Dim4 Focus on non-abstract things and activities; UK-WBF411)

The scores on Dim4 are again reasonably sensitive (overall the second most sensitive) to the text groupings in these data with an R² value of 0.54 or 54 per cent, i.e. variance explained if the text categories are known. Comments and tweets comparatively incline to the positive pole ‘Focus on humans and mental processes’, whereas web forums have more features related to ‘Focus on non-abstract things and activities’. Just like the previous dimensions, the South Asian data are significantly more attracted to one pole of the dimension (‘Focus on humans and mental processes’) in comparison to both IC varieties (or at least one of them) with a reverse trend. This trend is especially notable for the Bangladeshi comments, tweets and web forums. The Indian comments have comparatively lower scores than the other South Asian countries due to the use of concrete and compound nouns; the same is true for the Sri Lankan tweets. The UK web forums have the highest attraction to the negative pole ‘Focus on non-abstract things and activities’, which also makes them significantly different from most South Asian forums.

The general prominence of South Asian tweets on the positive pole (‘Focus on humans and mental processes’) may indicate different communicative preferences of these South Asian users. The use of features like human nouns, topical adjectives, WH questions and mental verbs in, for example, Pakistan is partly due to the use of Twitter by a large number of political activists (self-observation). The same is somewhat true for the Indian and to a lesser extent the Sri Lankan data. In the Bangladeshi data, many Twitter users are fans of foreign celebrities including those from India and South Korea (Shakir & Deuber Reference Shakir and Deuber2023: 134). These users tend to interact with other Twitter users and may employ the features listed above. Hence, the difference in communicative purposes of the South Asian netizens versus those from UK and USA appears to be a trend in the South Asian region but with slight differences in each South Asian variety.

3.7 Dim5 ‘Addressee involved opinion and description’

As shown in table 2, the last dimension has 8 linguistic features with 6 on the positive and only 2 features on the negative side. The features loading on the positive side point to descriptions (predicative adjectives, present tense, be as a main verb) along with expression of personal opinion (mental verbs), as shown in (14). The addressee is also involved either by referring to them as in (14), or in many cases through formulaic expressions, e.g. Thank you, especially in the public Bangladeshi registers. The negative features (past tense verbs and third-person pronouns) indicate a slight tendency of narration, which is visible in (15) from Sri Lankan comments. As figure 6 exhibits, 0.12 or 12 per cent of the variance in the scores on Dim5 can be explained if the text categories are known, which makes it the least sensitive to the text groupings in the data. The most prominent registers are all from Bangladesh on the positive side and Sri Lankan comments on the negative side. Some of them are significantly different from the native varieties, e.g. Sri Lankan comments and Bangladeshi tweets and forums.

Figure 5. Distribution of registers on Dim4 ‘Focus on humans and mental processes (top) versus non-abstract things and activities (bottom)’

Figure 6. Distribution of registers on Dim5 ‘Addressee involved opinion and description (top)’

1. (14) Dear [JJPRother] Ms.Aziz, I also feel [VPRT, MENTAL] like you [SPP2] whenever I watch [VPRT] a movie. I think [MENTAL] it is [VPRT, BEMA] not impossible [JJPRother] for anyone to have a sophisticated command on English if he/ she is [VPRT, BEMA] sincere [JJPRother] about that. The way you [SPP2] have [VPRT] pointed out the obstacles on the way of speaking English fluently, manifests [VPRT] your [SPP2] honest endeavour. You [SPP2] are [VPRT, BEMA] sure [JJPRother] to succeed.

   I wish [VPRT, MENTAL] you [SPP2] all the best. (Dim5 Addressee involved opinion and description; BD-WBF291)

2. (15) Anyway, he [TPP3S] was [VBD] of no use. He [TPP3S] was [VBD] always lying to the brim when the nation knew [VBD] the truth. He [TPP3S] was [VBD] a disgrace to the police department. (Dim5 Narrative tendencies; LK-CMT176)

3.8 Regional variation in South Asian text messages

Text messages are only available for South Asia so they have been described separately. As table 3 exhibits, the Pakistani text messages are significantly different from the other three South Asian countries on Dim2, as are the Bangladeshi text messages on Dim5. Pakistani text messages are also more literate impersonal on Dim1. More literate-impersonal and informational discourse entails that sometimes these chats are used for conveying information; e.g. (16) is from a work-related WhatsApp group where pharmaceutical sales representatives exchange information about orders. The use of general nouns including proper nouns, time references, imperatives and activity verbs is visible in the example.

1. (16) It's already [TIME] delivered [ACT] at FSD [NNother] on 27.02.21 received [VBN, ACT] by Imran [NNother]. Check [VIMP, ACT] with them.

   <#>

   Respected [VBN] sir [NNother]

   Required urgently stock [NNother] for Al [NNother] Aziz [NNother] distribution sargodha [NNother].

   Levopraid [NNother] 50 mg [NNother] 1000 (Dim2 Informational concerns; PK-TXM010)

Table 3. Regional variation in South Asian text messages

The Bangladeshi text messages are significantly more descriptive and addressee inclusive on dimension 5. For example, in (17) participants encourage each other to speak (or write) English by saying that mistakes are a learning opportunity. These texts are from a few group chats in the Bangladeshi data where participants want to practise English because they are preparing for the IELTS exam.

1. (17) […] if you [SPP2] make [VPRT] mistake no problem other member replace your [SPP2] mistake. it will be [BEMA] very [AMP] helpful [JJPRother] for everybody.

   <#>

   However, I wanted [MENTAL] to say, if I/ we Mistake any sentences or words grammatically, who knows [VPRT, MENTAL], will refine my or one's Mistake .

   <#>

   In bengali [JJPRother]

   <#>

   when you [SPP2] make [VPRT] mistake then you [SPP2] can learn [MENTAL] on this mistake. (Dim5 Addressee inclusive opinion and description; BD-TXM096)

3.9 Summary of findings

The resulting dimensions of the MD analysis appear to confirm previous studies (e.g. Biber Reference Biber1988; Biber & Egbert Reference Biber and Egbert2016) in terms of universal dimensions of register variation: oral versus literate discourse (Dim1) and stance/opinion- and persuasion-related features (Dim2 ‘Oral elaboration’ and Dim3 ‘Persuasion- and help-oriented discussions’). There are narrative elements on the negative sides of Dims 3 and 5 but a full-fledged narrative dimension has not emerged. The South Asian data generally have nominal, literate and formal discourse as shown on the negative ends of Dims 1 and 2. The comments, tweets and web forums show regional variation on Dim2 (less oral-elaborative more informational), Dim3 (‘Persuasion- and help-oriented discussions’), and Dim4 (‘Focus on humans and abstract processes’). Dim2 and Dim3 show that the devices used to express opinion and persuade others are present in comments as noted previously (e.g. Ehret & Taboada Reference Ehret and Taboada2021), but the South Asian English comments employ different features (e.g. modal verbs of necessity on Dim3) as opposed to evaluative adjectives and stance to clauses (Dim2) in the IC data. The South Asian web forums on Dim3 have different topics than the UK and USA forums, e.g. technical help, which, as discussed above, may be due to the types of forums included in these data. The comparatively high scores of South Asian tweets on Dim4 (positive pole ‘Focus on humans and mental processes’) are probably due to an abundance of political activist and fan accounts; in the British and American data, in contrast, features like human nouns and topical adjectives are less common.

There are also some individual trends in these South Asian data. The Bangladeshi netizens use features like WH questions, topical adjectives, human nouns etc. in comments and web forums more prominently. The communicative purpose of sharing information is noteworthy in the Pakistani text messages. The Indian web forums are informational and past oriented on Dim3 due to the use of past tense and general nouns. Similarly, the Indian comments are literate and narrative as they use more nouns (general on Dim3; concrete and compound on Dim4), past tense (Dim3 and Dim5) and third-person singular pronouns (Dim5) in comparison to the Bangladeshi and Sri Lankan English comments. The same is partly applicable to the Pakistani comments on Dim3 and Dim4. Finally, the Sri Lankan comments are different from the other South Asian English comments in these data due to oral features like first- and second- person pronouns, discourse markers, etc. (Dim1) as well as narrative features like past tense and third-person pronouns (Dim5).