2.3.1 Language Proficiency in English, Urdu, and Regional Languages

Overall language proficiency for regional languages, Urdu, and English is shown in Figure 3 in descending order of proficiency. Unsurprisingly, the participants indicate the highest proficiency for the regional languages that are mother tongues, followed by Urdu and English. Additional regional languages have the lowest proficiency. For English proficiency ‘native’ and ‘fluent’ get a combined probability of 43%, while intermediate proficiency has a 32% probability. The final model included gender, residence, education, occupation, mother tongue, and an interaction of gender, occupation, and mother tongue with the four languages as predictors. Marginal effects based on different demographic variables (not shown in the figure) show that education has a positive impact on language proficiency in English, Urdu, and both types of regional languages. Men report slightly higher language proficiency for English as compared to women, but the comparison is non-significant. The proficiency in an additional regional language is significantly better in men as compared to women irrespective of residence or mother tongue, which can be attributed to culturally more independent roles assigned to males, who are more likely to travel across Pakistan and learn other regional languages. Teachers and other professionals report higher proficiency in English, Urdu, and mother tongues in comparison to students. In terms of mother tongues, the speakers of Pahari/Pothwari, Punjabi, and Urdu indicate a higher proficiency in Urdu. The speakers of Balochi, Pashto, Sindhi, and smaller languages, on the other hand, report lower proficiency in Urdu, showing significant differences from some language-speaking groups (see Appendix B). English proficiency remains unaffected by mother tongue-based groups.

Figure 3

Language proficiency in English, Urdu, and regional languages

2.3.2 Use of Mother Tongue in Different Situations

We use a separate ordinal logistic regression model for the situational uses of the mother tongues because a lower number of situations are available for this question due to a spelling error in the two situations related to social media. The final model for situational use of mother tongues includes gender, residence, education, occupation, mother tongue, and situation as predictors of usage frequency, while gender, residence, and mother tongue are also included as interactions with situation. We do not include interactions of grouping variables like age and gender due to zero occurrences in certain categories, which lead to warnings by the MASS:: polr function. The overall usage frequency in Figure 4 shows that mother tongue is more likely to be used with family with a 79% combined probability of ‘always’ and ‘usually’. The usage frequency decreases as the face-to-face (F2F) situations become more formal, as can be seen with higher probability of ‘never’ (24%) and ‘sometimes’ (33%) in the case of strangers (including with shopkeepers and in bazaar). The only internet situation included here is emails, and the participants are highly likely to select ‘never’ and ‘sometimes’ (a combined probability of about 89%). The use of the mother tongue on social media for responses where the spelling mistake mentioned at the beginning of this paragraph was corrected before the survey distribution also shows that a combined probability of 70% (personal social media) and 78% (professional social media) for the two options ‘never’ and ‘sometimes’. Thus, overall the mother tongues are more likely to be used with family and friends.

Figure 4

Usage frequency of mother tongue in different situations

Different participant groups also use the mother tongues significantly differently from each other. In other words, gender, residence, and mother tongue have significant interactions with the five situations (see Appendix C for the contrasts, but we do not include the relevant marginal effects plots due to space restrictions). Apart from family-related situations, women are significantly less likely to use the mother tongue in all other situations when compared to men. These gender-related differences are generally stable across residence and mother tongue-based groups, with no reverse trends found in any subgroup. Residents of AJK, Balochistan, and Punjab are less likely to use the mother tongue with colleagues and class fellows. KPK respondents stand out as they are more likely to use the mother tongue with friends, colleagues/class fellows, and strangers (including with shopkeepers and in bazaar): they select ‘always’, ‘often’, or ‘usually’ in 50% or more cases. Pashto (the main language of KPK and parts of Balochistan) and Sindhi (especially the main language of rural Sindh) speakers are more likely to use the respective mother tongue with colleagues/class fellows and strangers as compared to the other regional languages, but these contrasts are not significant. Urdu native speakers are obviously more likely to use their mother tongue (i.e. Urdu) with friends, colleagues, and strangers. Punjabi and Saraiki speakers are significantly different from Urdu speakers in most of these situations, that is, instead of their mother tongues they most probably use Urdu with friends, colleagues/class fellows, and strangers.

2.3.3 Use of English and Urdu in Different Situations

The responses for situational uses of Urdu and English were pooled in one model to enable better comparisons, where language and situation were used as interaction effects for all other demographic grouping variables. The Analysis of Variance (ANOVA) table based on the model is available in Appendix D, while the marginal effects plot of overall usage for the two languages is shown in Figure 5. As it is clear from the panel ‘Urdu’, the Urdu language is most likely to be used with colleagues/class fellows, a rather formal situation. The probability assigned to ‘always’, ‘usually’, and ‘often’ in friends and family-related situations is higher than the situations involving strangers (which includes shopkeepers and bazaar as well). The use of mother tongue (especially by Pashto and Sindhi speakers as mentioned in Section 2.3.2) in the situations related to strangers may be one reason of an overall lower use of Urdu in this situation (see Section 2.3.3, paragraph 3). In the three internet-related situations, the use of Urdu decreases as the formality increases. Hence, Urdu is more likely to be used in personal social media posts, followed by professional posts, and least likely in emails. This is opposite and significantly different as compared to the usage of English in these three situations. This usage also includes Roman Urdu – a practice of writing Urdu in Roman letters. This kind of Urdu is more likely to be mixed with English words, as it is easier to spell English words in Roman letters in comparison to the native Perso–Arabic script of Urdu.

Figure 5

Usage frequency of Urdu and English in different situations

The use of Urdu among different subgroups also shows variation. The eighteen to twenty-five year olds are significantly more likely to use Urdu in emails in comparison to all other age groups, with men leading women in this age group as well, as discussed later in this paragraph. This ever so slightly higher likelihood of using Urdu could be due to the use of Roman Urdu among the younger generation. However, nothing conclusive can be said based on the current data. A more meaningful contrast is between males and females, where women are more likely than men to use Urdu in more formal F2F situations (friends, colleagues/class fellows, and strangers including with shopkeepers and in bazaar). The first two contrasts (i.e. friends and colleagues/class fellows) are also significantly different in both groups (see Appendix D). On the contrary, males are significantly more likely to employ Urdu in the internet-related settings of professional social media and emails. The residents (especially women) of AJK, Punjab, and in some cases those of GB are more likely to use Urdu in F2F situations, while those of Balochistan and KPK are among the least likely to employ Urdu especially with colleagues/class fellows and strangers (including in bazaar and with shopkeepers). This is in line with the higher likelihood of using the mother tongue by these groups in similar situations, as noted in Section 2.3.2. Similar usage patterns can be assigned to relevant language-speaking groups: a higher likelihood to use Urdu in different F2F situations (e.g. friends, colleagues/class fellows) can be assigned to Punjabi and Pahari/Pothwari speakers (the residents of Punjab and AJK, respectively), while Balochi (one of the main ethnolinguistic groups from Balochistan) are less likely to use Urdu with friends and strangers. Some of these subgroups are significantly different from other subgroups, as shown in the three-way contrasts in Appendix D.

These trends of regional and mother tongue-based groups generally show the historical patterns of (dis)favouring the use of Urdu. The residents of Punjab (Punjabi, Saraiki, and partially Pahari/Pthowari speakers in these data) have been more open to adopt Urdu, even from the era of the British India (Singh Reference Singh1997). Punjabis giving lesser prestige to Punjabi in comparison to Urdu is a well-documented phenomenon, for example, Abbas & Iqbal (Reference Abbas and Iqbal2018: 210). Urdu has been the official language of AJK since 1968 (Anonymous 2002). This factor and many similarities between Punjab and AJK (e.g. the Pahari/Pthowari language-speaking group in these data, their geographical proximity) may also be the cause of similar results here. The historical relationship of Pashto and Sindhi-speaking communities with Urdu has not been as smooth. Pashto has been associated with Pashtun nationalism and a closer relationship of such movements with Afghanistan (where a sizeable Pashto speaking population resides). Even in the British Raj, the imperial officers preferred Urdu on Pashto in a bid to amalgamate the then North West Frontier Province (NWFP) better with the rest of British India (Rahman Reference Rahman1995: 156). After the creation of Pakistan and the inclusion of NWFP (now KPK) in the country, a strong nationalist movement involving Pashto and Pashtun identity has remained throughout Pakistan’s history. On the other hand, Sindhi has been used in formal situations (e.g. courts) since even before 1947 (Rahman Reference Rahman1999: 25). After the partition, the influx of Urdu-speaking Muslim immigrants (muhajirs) from central India in urban Sindh increased the nationalistic sentiment of local Sindhi-speaking population. Sindhi is taught in primary schools, Sindhi and Urdu are co-official languages of the provincial government, and Sindhi is one of very few regional languages in Pakistan in which a master’s degree can be obtained (Rahman Reference Rahman1999). Due to these historical and political reasons, Pashto and Sindhi speakers (and the corresponding provinces of Pakistan) are more likely to prefer their respective mother tongues instead of Urdu even in formal situations.

Overall, the use of English is more likely with friends and colleagues/class fellows in F2F situations (Figure 5, English). One reason could be that the interlocuters in situations related to family and strangers (including in bazaar and with shopkeepers) are less likely to be proficient in English (e.g. parents, other elders, shopkeepers); hence mother tongue or Urdu and more appropriate choices. English is much more likely to be used on the internet, where a probability of only 25% or lower is associated with ‘sometimes’ and ‘never’.

Just like mother tongues and Urdu, various demographic groupings also show variation in (dis)preference of English in the seven situations. In terms of age, younger participants (eighteen to twenty-five) assign higher probability to ‘always’ and ‘usually’ in relation to the use of English in F2F situations related to family, friends, and colleagues/class fellows. This age group is significantly different from the age groups of twenty-six to thirty-five and thirty-six to forty-five at least in family-related situations. This shows an overall lowering barrier of formal language use: English might be taking the place of Urdu, and Urdu of the regional languages (mother tongues) in at least some parts of the society (also indicated by a comparatively higher use of Urdu in formal F2F situations by women as mentioned in Section 2.3.3, paragraph 2). Apart from that, on the internet the use of English increases with age. This is most probably because of a higher level of education that also increases with age in these data (see below in this paragraph). In terms of gender, women are ever so slightly more likely to use English in F2F situations of friends and colleagues/class fellows. Men indicate a very small preference for English usage with strangers (also significantly different from women; see Appendix D), but this contrast requires more in depth investigation (e.g. interviews with those who indicated positive responses). The residents of Punjab (in all three internet settings) and GB (in personal social media) are less likely to use English on the internet. Punjab is also significantly different from some other provinces and regions (e.g. Sindh) in email and professional social media. One reason may be a higher percentage of younger participants belonging to Punjab. Education does not show significant contrasts, but generally higher education (just like age) is associated with a higher likelihood of English usage, for example, MPhils and PhDs grouped in the education group of eighteen year plus. Lastly, Balochi speakers stand out in comparison to most other language speaking groups by showing a high dispreference for English in most F2F situations.

2.3.4 Attitudes towards Various Statements about English, Urdu, and Regional Languages

The attitudes of the Pakistani respondents about various statements on language use and language identity are shown in the form of (dis)agreement probability in Figure 6. Like the previous models, this one also included all demographic variables and one variable related to the nine questions as an interaction effect with all other predictors. The ANOVA table based on the model is shown in Appendix E. The overall comparison in Figure 6 shows that Pakistanis are highly in favour of advantages associated with English and children learning English at school (86% probability for the first two positive statements). About 25% probability is assigned to ‘disagree’ and ‘strongly disagree’ in response to the third statement, the negative statement about PakE, which shows that most of the participants are not aware of PakE as a variety – a conclusion similar to that of Baumgardner (Reference Baumgardner1995). The more general statement that ‘Pakistanis speak differently’ is more likely to be agreed with by a little less than 50%, which shows some general perception about the distinctiveness of the local speech. Statements 5–7 are related to the use of Urdu and show varying attitudes towards Urdu as the official language, and its use with kids and in different situations. These responses are explored in more detail in the next paragraph. The last two statements about the mother tongues show that the participants generally have highly positive attitudes about their mother tongues, which are expected from most language speaking groups.

Figure 6

Attitudes towards various statements about languages

Apart from the overall attitudes, there are also interesting differences among different demographic groups within the predictor variables. In terms of age, the youngest respondents (eighteen to twenty-five and twenty-six to thirty-five) are significantly different from the respective older groups in recognising the distinctive English speech of Pakistanis. Moreover, the same two groups also have higher probability for ‘(strongly) agree’ in response to ‘I prefer Urdu in most situations’, indicating that younger participants prefer Urdu over the regional mother tongues (eighteen to twenty-five are significantly different from forty-five years and older groups; the differences are slightly less pronounced between twenty-six to thirty-five and thirty-six to forty-five old women). Men and women significantly differ from each other on four statements: Women are more likely to (strongly) agree with the first two statements (advantages and learning English); They also prefer English by disputing ‘Urdu should replace English’ more often than men (male responses are slightly more varied across regions); Women indicate lesser preference for the mother tongue by disputing ‘Prefer mother tongue mostly’ more often than men. Other professionals (including journalists and doctors) are significantly less likely to agree with ‘Pakistanis speak differently’ and with ‘Urdu should replace English’ when compared with teachers. This shows that Pakistanis not involved in teaching are less aware of speech differences Pakistanis have, and that they favour English potentially due to its practical use in their respective fields. In terms of language speaking groups Balochi, Pashto, Sindhi, and small language speakers (e.g. Shina, and the ones categorised under Other) are more likely to dispute the statement 7 (‘prefer to use Urdu over mother tongue with kids’), while Punjabi, Saraiki, and Urdu speakers are more likely to agree with it. Balochi and Pashto speakers are also more likely to agree with the statement 8 related to the use of the mother tongue (see significant contrasts in Appendix E).

2.3.5 Summary of Findings

The main trends related to age group, gender, residence, and mother tongue-based groups, and profession in three types of languages (regional mother tongues, Urdu, and English) can be summarised as follows.

The mother tongues are primarily used at home with family members, and are least likely to be used on the internet. KPK residents employ their mother tongue in a wider variety of situations. Pashto (spoken in KPK and Balochistan) and Sindhi (from Sindh) speakers prefer mother tongue even in formal situations like colleagues/class fellows where Urdu may be employed by other language speakers (e.g. Punjabi and Saraiki speakers from Punjab). Women are less likely to use mother tongue outside of home and family. In other words, men use mother tongue in more diverse situations. Most participants say they prefer mother tongue in most situations and also strongly identify with the relevant ethnolinguistic group. Balochi, Pashto, and smaller language speakers are more likely to do so.

The findings related to Urdu show that it is mainly associated with friends and work-related situations, and less likely to be used in professional social media and email. Respondents from AJK and Punjab and the ethnolinguistic groups in these areas (Pahari/Pothwari, Punjabi, Saraiki) indicate more preference for Urdu; for example, some report Urdu as their mother tongue, and others indicate a higher proficiency in Urdu. In contrast, the residents of KPK, Sindh, and Balochistan (speakers of Balochi, Pashto, Sindhi, etc.) indicate a lower proficiency in Urdu, are more likely to disprefer Urdu in formal F2F situations, and dispute statements related to the use of Urdu (‘prefer to use Urdu over mother tongue with kids’). Women are more likely than men to employ Urdu in formal F2F situations. They also indicate Urdu as their mother tongue in AJK and Punjab more often than men. Younger participants indicate a higher preference of using Urdu over the regional mother tongues, even in family-related situations.

Lastly, the use of English mainly occurs in friends and work-related F2F situations and on the internet. The majority of participants does not agree with the statement that there is a PakE. However, a little less than 50% participants show awareness about the English speech differences of Pakistanis. Proficiency in English increases with age and education. Younger participants also indicate the use of English in some F2F situations. Most participants agree with the advantages associated with English and the need to teach English to kids, with women more likely to do so than men. Women also slightly prefer English in formal F2F situations compared to men, and disagree more often with the statement about replacing English with Urdu in official functions.

3.1.1 Consonantal Features

Th-stopping, retroflexion, de-aspiration, and rhoticity are the main consonantal features that reflect substrate influences and are present in educated speech. Th-stopping represents the most invariant marker of PakE: the interdental fricatives /θ/ and /ð/, absent from Urdu, are overwhelmingly replaced by dental stops [t̪] and [d̪]. Gemini transcribed 71% cases of think (Audio 1_think.mp4. Audio of think. Audio file available at https://doi.org/10.1017/9781009665308) with [t̪] and 90% cases of this (Audio 2_this.mp4.. Audio of this. Audio file available at https://doi.org/10.1017/9781009665308) with [d̪] in the YouTube audio corpus, which is consistent with prior documentation of these features (Mahboob & Ahmar Reference Mahboob, Ahmar, Kortmann and Schneider2004; Rahman Reference Rahman1991; Shahid et al. Reference Shahid, Bilal, Mirza and Raza2024).

Retroflexion of alveolar /t/ and /d/ to [ʈ] (as in /ˈbe.ʈər/ (Audio 3_better.mp4. Audio of better. Audio file available at https://doi.org/10.1017/9781009665308)) and [ɖ] (as in /ɖuː/ Audio 4_do.mp4. Audio of do. Audio file available at https://doi.org/10.1017/9781009665308), respectively, has been confirmed through acoustic studies of PakE (Syed, Bibi, & Shafi Reference Syed, Bibi and Shafi2024: 69). It occurs especially intervocalically and across morpheme boundaries. In our YouTube corpus, Gemini identified 5%–30% instances of /t/ and /d/ as retroflex stops depending on their position (start, middle, or end) in words like do, top, end, fast. A majority was identified as dental stops ([t̪] and [d̪]), which was deemed incorrect by the human reviewer. Since previous acoustic studies have confirmed this feature in PakE, the present results need further investigation.

The neutralisation of /v–w/ merges the English contrast in favour of a labial approximant [w] or [ʋ], reflecting Urdu phonology, which is found in 35% examples of very (Audio 5_very.mp4. Audio of very. Audio file available at https://doi.org/10.1017/9781009665308) in our data transcribed by Gemini.

The voiceless plosives /p/, /t/, and /k/ in positions where AmE (General American or GA) and BrE (Received Pronunciation or RP) demand strong aspiration are regularly de-aspirated, a finding supported by acoustic analyses of voice onset time in PakE (Yasir et al. Reference Yasir, Bashir, Jahan, Zamir and Ahsan2021). The variety is fully rhotic, that is, postvocalic /r/ is always pronounced, while /l/ remains clear [l] in all environments; velarized [ɫ] does not occur often (Jadoon & Ahmad Reference Jadoon and Ahmad2022). Gemini found a majority of the example words pronounced in this way, that is, de-aspirated [p] in points (Audio 6_points.mp4. Audio of points. Audio file available at https://doi.org/10.1017/9781009665308), rhotic [r] or [ɹ] in far (Audio 7_far.mp4. Audio of far. Audio file available at https://doi.org/10.1017/9781009665308), and clear [l] in all (Audio 8_all.mp4. Audio of all. Audio file available at https://doi.org/10.1017/9781009665308) – see Table 1 for a comparison with RP and GA.

3.1.2 Vocalic Features

Monophthongisation of closing diphthongs is among the most systematic vocalic processes: goat (/əʊ/ ~ /oʊ/) is almost invariably [oː] (e.g. Audio 10_social.mp4. Audio of social. Audio file available at https://doi.org/10.1017/9781009665308), while face (/eɪ/) is regularly [e] (e.g. Audio 9_education.mp4. Audio of education. Audio file available at https://doi.org/10.1017/9781009665308), though slightly more variable (see Table 2 for a comparison with RP and GA). Monophthongisation of these vowels has been observed in earlier empirical work on PakE as well (Jadoon & Ahmad Reference Jadoon and Ahmad2022; Mahboob & Ahmar Reference Mahboob, Ahmar, Kortmann and Schneider2004; Shahid et al. Reference Shahid, Bilal, Mirza and Raza2024), and also matches typological descriptions of diphthong simplification in new Englishes, for example, Bobda (Reference Bobda2007). The absence of low front /æ/ results in trap being raised or backed to [ɛ] or [a], which is also documented across studies of PakE vowel systems (e.g. Yasir et al. Reference Yasir, Bashir, Jahan, Zamir and Ahsan2021).

For monophthongisation, Gemini identified most instances of the goat and face vowels as diphthongs, which was disputed in all example cases in the human review. Thus, a short analysis of relevant words (e.g. so, make) on the same data was run by calculating F1 and F2 values at onset 20% and offset 80% using WebMaus and the praat-parselmouth library in Python (Jadoul, Thompson, & De Boer Reference Jadoul, Thompson and De Boer2018; Boersma & Weenink Reference Boersma and Weenink2026). The results show that the Pakistani speakers monophthongise both goat and face vowels in a majority of cases (see Figure 19 in Appendix F).

3.1.3 Prosodic and Suprasegmental Features

Three features are notable at the suprasegmental level. Firstly, syllable-timed rhythm characterises all 100+ sources across both audio corpora, with roughly equal syllable duration and minimal reduction of unstressed vowels. This contrasts sharply with the stress-timed rhythm of RP and GA and represents perhaps the most immediately salient characteristic of PakE, which has also been documented in previous perceptual and acoustic studies (Asgher, Kashif Shahzad, & Hanif Reference Asgher, Shahzad and Hanif2023; Jadoon & Ahmad Reference Jadoon and Ahmad2022; Rahman Reference Rahman1991; Shahid et al. Reference Shahid, Bilal, Mirza and Raza2024). Secondly, lexical stress shows a strong tendency towards penultimate prominence in polysyllabic words, which reflects Urdu patterns and is documented consistently across empirical investigations (e.g. Yasir et al. Reference Yasir, Bashir, Jahan, Zamir and Ahsan2021). Thirdly, intonation is characterised by narrower pitch range and frequent level or rising terminals on declaratives, where RP and GA would employ falling pitch (Mahboob & Ahmar Reference Mahboob, Ahmar, Kortmann and Schneider2004).

3.4.1 Productive Patterns with English Suffixes

The use of English suffixes to form new nouns has been reported by Rashid Mahmood (Reference Mahmood2009: 255–257) based on PWE. One of the unique patterns he observes is the suffix -ism to create neologisms like adhocism, red-tapism, landlordism, militantism, and technocratism that use English stems. There are also examples where the stem comes from Urdu, for example, bradarism (i.e. the practice of, e.g., voting collectively as a community based on caste). Additional examples of -isms from SAOnE-PK include:

1. (1) fawadism: behaving like a Pakistani cricketer, Fawad Alam, who focused on his performance instead of PR to get selected in the national team.

2. (2) bhuttoism: The way of Zulfiqar Ali Bhutto and his daughter Benazir Bhutto, two assassinated members of a prominent Pakistani political family.

3. (3) wadera-ism, jagirdari-ism, chaudhri-ism: feudalism

4. (4) patwarism: Based on patwari, a negative political slang referring to supposedly illiterate followers or voters of Pakistan Muslim League (Nawaz), a local political party.

5. (5) mullahism, mullaism: The dominance of religious leaders

More recently it has been observed that Pakistani internet users employ the English suffix -ian to create adjectives that refer to the membership of an educational institute or a (political) organisation. The examples from SAOnE-PK (Tweets) include:

1. (6) insafian: a member or supporter of Pakistan Tehreek-e-Insaaf

2. (7) youthian: a negative slang to refer to the supporters of the same political party

3. (8) xxxxx_nedian: a student of the Nadirshaw Edulji Dinshaw (NED) university

4. (9) ex-uetian: a former student of University of Engineering and Technology (UET)

5. (10) xxxxxxxxcomsian: a student of COMSATS university

Example (11) shows the use of -ian with the name of Iqbal (the national poet and philosopher of Pakistan) to refer to his philosophy.

1. (11) Thus, the Iqbalian ‘self’ is an entity in which the body and the soul have to work together (PWE, Books).

3.4.2 Reduplication

Reduplication is a very common process in Urdu and the regional languages of Pakistan. Word reduplication is used to create emphasis as well as to convey certain particular meaning. The present analysis has been conducted on PWE and SAOnE-PK to identify cases of reduplication and explore the influence of Urdu and the regional languages on the PakE communication. PWE has a very few cases of reduplication. These include adjectives (e.g. bad bad), adverbs (e.g. far far, chupke chupke ‘secretly secretly’), and interjections (e.g. dammit dammit). The examples show that emphasis is the main reason for reduplicated lexical items, whether English or borrowed from Urdu. These items occur in the magazine articles, press reportage, and English short story sections of the corpus. SAOnE-PK has a larger variety of such items due to the informal CMC data and its larger size. In the remainder of this subsection, we describe various types of reduplication in this dataset.

Table 4 presents an overview of the reduplicated lexical items in SAOnE-PK that occur in otherwise English sentences. We also checked if the reduplicative use was attested in the other regional components of SAOnE. Hence, the column ‘English (shared with UK, US)’ shows the items found in SAOnE-PK that were also attested at least once in the UK and/or US data along with the other three SA varieties. The number of tokens and unique types shown in the table, however, are only based on the SAOnE-PK data. This column has the largest number of reduplicated items in terms of tokens and types, with the main categories being adjectives (e.g. deep deep, long long, naughty naughty), adverbs (e.g. much much, really really, very very), and discourse items (e.g. ok ok, please please, yes yes). This shows that reduplication (especially for emphasis) occurs in most varieties of English present in SAOnE.

Furthermore, the column ‘Non-words and interjections’ shows the second highest frequency items (thirty-five types occurring 218 times), which are generally used in CMC irrespective of language, for example, laughter-related interjections (ha ha, he he, lol lol), and other interjections (blah blah, *cough cough*, *clap clap*, eh eh, tch tch, tsk tsk, uff uff). The examples with asterisks around are the CMC way to show the supposed sound or action the writer makes/does while typing the message. More familiar interjections were tagged separately, for example, wow wow is English specific (listed under the column ‘English (unique to PK, South Asia)’), while its equivalent in meaning wah wah is an Urdu borrowed item that is categorised under interjections in the column ‘Borrowed items (e.g. from Urdu)’. Overall, reduplication in SAOnE-PK occurs most frequently in tweets, followed by web forums, text messages, comments, and websites.

The columns ‘English (unique to PK, South Asia)’ and ‘Borrowed items (e.g. from Urdu)’ contain reduplicated items specific to Pakistan (in addition to the SA region). The former column shows ninety-six types (total frequency 141) that are English reduplicated words, but only found in SAOnE-PK or the PK, as well as SA data only. These include auxiliary verbs such as must in (12) or lexical verbs like see see, send send, and so forth. (13) has an adjective reduplicated, while congrats (a discourse item) is used two times in (14). In all of these cases, the purpose is to emphasise or intensify the meaning. In (15), the noun condition is reduplicated as the object of the verb play. This structure also occurs with other nouns: playing Dr Dr, playing Twitter Twitter, playing Kashmir Kashmir. In each case, the meaning is satirical, and the writer wants to emphasise that someone is making a mockery of a serious issue. This construction is a literal translation from Urdu. Similarly, numbers can be reduplicated in Urdu for emphasis, and the same can be seen in (16) with two.

The column ‘Borrowed items (e.g. from Urdu)’ shows fifty-one Urdu reduplicated lexical items that were found in English sentences. Discourse-related words are the most common category, for example, Mashallah ‘Praise be to God’ in (17) is on the same pattern as Congrats in (14), showing the tendency of reduplicating discourse items. The noun parchi (‘receipt’) is reduplicated in (18) to refer to the irregular selection of a cricketer in the national cricket team based on the reference of their father, who is an ex-cricketer, instead of a merit-based selection. Other examples include interjections like chee chee (‘eww eww’), the discourse marker jee jee (‘yes yes’), and negation like nahe nahe (‘no no’). The examples and discussion show that reduplication is a productive process in the data of PakE analysed here.

1. (12) We all must must work together to bring peace in the whole region …. (SAOnE-PK, Comments)

2. (13) we have whatsapp group nd members full there so our honarable honarable admin make fb group […] (SAOnE-PK, Text messages)

3. (14) Congrats Congrats Well played (SAOnE-PK, Tweets)

4. (15) When our reserves were going down during first six months and Asad Umar was playing condition condition with IMF. (SAOnE-PK, Web forums)

5. (16) […] whether she was elected for one term or two two, does not matter. (SAOnE-PK, Tweets)

6. (17) Mashallah Mashallah blue shirt with black (SAOnE-PK, Tweets)

7. (18) Parchi Parchi (‘receipt receipt’), An Ex-Cricketer Father’s Parchi (‘receipt’). (SAOnE-PK, Tweets)

3.4.3 Productive Combinations of Local Nouns and Adjectives in English CMC

The aim of the present subsection is to explore the use of cultural vocabulary in the form of noun + noun and adjective + noun combinations in PakE CMC. These combinations can be Urdu (i.e. both items are Urdu but used in English sentences) and English–Urdu mixed (i.e. Urdu adjectives and nouns in combination with English nouns). Using a frequency list from SAOnE-PK, Google Gemini was asked to classify the words that may be relevant to Pakistan’s history, geography, politics, culture, religion, food, flora, fauna, clothing, and local languages. Then, we used the Multi-Feature Tagger of English (MFTE) (Le Foll & Shakir Reference Foll, Elen and Shakir2023; Le Foll & Shakir Reference Shakir2024) to extract productive two-word combinations tagged as Noun Compound (NCOMP) (i.e. compound nouns) where one item was from the selected vocabulary list. Finally, we selected those combinations that were used in an English sentence, resulting in 460 such bigrams. After manually checking the bigrams and their KWIC, we selected fifty-eight most interesting items, of which some occur only in one combination (e.g. dheet ‘audacious’ only occurs once in dheet people) and others are highly productive (e.g. Eid occurs in dozens of different combinations like Eid offer, Eid prayer, Eid day, Eid vibes, Eid selfie, etc.). At the same time, we classified each bigram according to different thematic categories and language (English, Urdu, mixed).

Figure 7 shows the thematic categories along with the language (Urdu or mixed) and their proportions in different registers of SAOnE-PK. The only locally relevant English word is an abbreviation Computerised National ID Card (CNIC), which occurs in bigrams like CNIC number, CNIC copy, CNIC card, and CNIC condition as in (19). The bigrams with mixed English and Urdu words are more frequent and diverse in terms of themes. One of the most frequent Urdu (also Hindi) words in the cultural category is desi (‘native’ or ‘indigenous’), which occurs with many English words, for example, desi parents, desi food, desi liberals, desi community, desi shampoo. Ghairat (‘honour’) is comparatively more specific to Pakistani culture, occurring as ghariat brigade and ghairat brigades (‘groups(s) of people who like to morally police others’). Another culturally relevant word is rishta (‘marriage proposal’), which occurs in bigrams like rishta process, rishta lady (‘the professional agent who mediates the process’): (20) shows the use of the latter.

Figure 7

Thematic categories of local noun + noun and adjective + noun combinations in SAOnE-PK

Ethnolinguistic words are adjectives referring to different ethnolinguistic groups in Pakistan, for example, Baloch/Baluch (the main ethnolinguistic group from Balochistan), Mohajir (Urdu-speaking immigrants, see Section 2.2), Pashtoon/Pukhtoon (the Pashto-speaking ethnolinguistic community). These are used in bigrams like Baloch people, Baloch tribes, Mohajir identity, Pashtoon belt, and so forth. In (21), the bigram Pukhtoon revenge refers to the tradition of revenge in Pashto speakers. The negative/slang category and political slang are thirty-six in number and are mainly used in Twitter. These include Urdu adjectives like bakwas (‘crap’) player, ghunda (‘rascal’) police, jaahil (‘illiterate’) leader, youthia mentality (mentality of Pakistan Tehreek-e-Insaaf (PTI) supporters), patwari league (Pakistan Muslim League Nawaz), and so forth. The word lifafa literally means envelope; in (22), it refers to journalists that supposedly accept bribe cash in envelopes. The category of ‘other local’ includes adjectives like Lahori (‘someone from Lahore’) as in Lahori blogger, Lollywood (Hollywood-inspired name of Pakistani film industry), as in Lollywood movies, Lollywood way, Lollywood icon, and gora (‘white’) as in (24). Lastly, religious bigrams include examples like haram (‘forbidden’) money, Jihadi (‘religious militancy’) tendencies as in (23), Iftar (‘breaking fast’) time, Nikah (‘marriage vows’) dress, and so forth.

1. (19) Mr.Zaidi introduced CNIC condition for business at just Rs.50,000/- …(SAOnE-PK, Comments)

2. (20) We got the rishta through a rishta lady and they’ve been talking for awhile. (SAOnE-PK, Web Forums)

3. (21) APS Pukhtoon children massacre was your fatal mistake. The famous Pukhtoon revenge will come. (SAOnE-PK, Comments)

4. (22) They All are lifafa Journalists and Media Groups. (SAOnE-PK, Tweets)

5. (23) Such jingoism led us to Jihadi tendencies which became the tenth fault line. (SAOnE-PK, Web Forums)

6. (24) Grew up avoiding street food for questionable hygiene and a gora vlogger comes and calls that shit better than KFC. . (SAOnE-PK, Tweets)

The bigrams that are completely in Urdu mainly consist of religious and food-related nouns. The most common bigram in the former category is Eid Mubarak (‘happy Eid’), which mainly occurs in tweets and text messages. Food-related bigrams (e.g. garam masala, chaat masala, tikka ‘BBQ’ masala, gosht ‘meat’ masala) are mainly present in web forums where cooking recipes are posted by the forum members. In (25), dabbul roti (‘box bread’) refers to sandwich bread that is usually baked in a metal box. Other local Urdu nouns can be used as is in English sentences, for example, Chai wala (‘tea maker’) in (26). The word jalsa (‘political gathering’) is used a few times in comments, text messages, and tweets. In (27), such a meeting in Lahore is mentioned.

1. (25) It is dabbul roti … (SAOnE-PK, Web Forums)

2. (26) You have a monster Chai wala next door … (SAOnE-PK, Web Forums)

3. (27) The Lahore jalsa was a failure […] (SAOnE-PK, Comments)

This subsection has shown that Urdu nouns and adjectives are frequently employed by Pakistani users in informal English communication, and that some of these lexical items are highly productive.

3.5.1 Relative Clauses with That and Which

Leech et al. (Reference Leech, Hundt, Mair and Smith2009: 227) observe a decline in the frequency of which relatives – as in (28) – and an increase in that relative clauses – for example, (29) – in written AmE and BrE using the Brown family corpora from 1960s to 1990s. They attribute the increase of that relatives to colloquialisation (Leech et al. Reference Leech, Hundt, Mair and Smith2009: 234). In their analysis, the tendency of replacing which with that relativiser is more pronounced in the AmE corpora (Brown & Frown), than the British ones (Lancaster-Oslo-Bergen (LOB), Freiburg-LOB (FLOB)). Deuber and Hänsel (Reference Deuber, Hänsel, Wiegand and Mahlberg2019: 54) note that the relatively higher frequency of which relative clauses in BrE (including in the press language) can be a good indicator of a variety’s inclination towards AmE or BrE. Hence, the aim of the present subsection is to determine the frequency of that versus which relative clauses in PakE as compared to AmE and BrE using the written and CMC corpora as described in Section 3.2.

The data were tagged (POS, Named Entity Recognition or NER) and parsed (constituency, dependency) with the stanza (Qi et al. Reference Peng, Zhang, Zhang, Bolton and Manning2020) library. Then, we followed the criteria adopted by Hinrichs, Szmrecsanyi, and Bohmann (Reference Hinrichs, Szmrecsanyi and Bohmann2015: 814–815) and by Deuber and Hänsel (Reference Deuber, Hänsel, Wiegand and Mahlberg2019: 54–55) to extract only restrictive relative clauses using a rule-based approach and a Python script. We excluded clauses that were marked by a comma before the relativiser which, pied piping relative clauses involving which, clauses where the antecedent was that and the relativizer was which, clauses with a human noun (PERSON tag in the NER tags) as the antecedent and the relativiser was that, and where the clause could function as adjunct (i.e. that was replaceable by when, where, how). For the last case, a list of adjunct nouns was used (day, time, year, place, reason, way, date) and the clause was excluded from final count where such a noun was the head of the antecedent NP and the relativiser was that. We also excluded the clauses marked as ‘obl’ or oblique (i.e. adjunct clauses) by the stanza dependency parser. For examples of these cases, please refer to Deuber and Hänsel (Reference Deuber, Hänsel, Wiegand and Mahlberg2019) and Deuber et al. (Reference Deuber, Hackert and Hänsel2022: 284).

As a further step, we checked for complement clauses incorrectly extracted as relatives using Google Gemini 2.5 Pro. This last step removed 1–10% clauses from the total (the highest number was from PWE): (30) shows such a clause identified by our script as a relative clause but flagged by Gemini. The overall accuracy of relative clause identification (based on a 2.5% sample of the data) remains around 96% in both datasets.

1. (28) He pointed out the United States had committed a blunder which would further deteriorate law and order situation in Iraq. (PWE, News)

2. (29) It was clearly a failure that needs to be looked at and, if possible, addressed. (Crown, Press reportage)

3. (30) One can infer from this brief discussion that Saras ingenious literary expression consists in the blending of some deviant patterns into non-deviant patterns in her language. (PWE, Research papers)

Figure 8 shows the per million word frequencies and proportions of the two types of relative clauses in the three varieties. The normalised frequencies of both types of relative clauses are much lower in both modalities of PakE, though the regional difference is less pronounced in CMC. The overall lower frequency of relative clauses might be attributable to the use of phrasal features (i.e. prepositional modifiers) instead of clausal modifiers in NPs, as also noted in Section 3.5.4 (last paragraph). The written English registers (news, editorials, and research papers) indicate that which relatives are the least prevalent in AmE, confirming previous studies on the Brown family corpora. The proportion of which relatives in PakE is higher than in both of the native varieties, and the differences are statistically significant based on overall chi-square comparisons using raw frequencies in both modalities (Written: χ² (2, N = 4496) = 332.04, p < 0.001; CMC: χ² (2, N = 3910) = 383.69, p < 0.001). Subsequently, PakE was also found significantly different when compared with both native varieties one by one (PK vs UK, PK vs US) using the function pairwiseNominalIndependence in the R package rcompanion (Mangiafico Reference Mangiafico2025; R Core Team 2025). Overall, the results show that these PakE registers (both in CMC and written) are different from their counterparts in AmE and BrE in the use of relative clauses.

Figure 8

Relative clauses in written and CMC corpora of the three varieties of English

3.5.2 Be and Get Passive

In the past two decades, corpus-based research on passive voice in AmE and BrE has shown that the be passive, for example, (34), is declining in frequency, more in AmE and a little less in BrE, for example, Leech et al. (Reference Leech, Hundt, Mair and Smith2009: 149, 156). At the same time, another variant of the passive voice, the get passive, as shown in (33), is very gradually increasing in frequency. Bohmann et al. (Reference Bohmann, Müller, Honkanen and Neuhausen2025) have conducted a comprehensive analysis of be and get passive in varieties of English using very large corpora, for example, the Corpus of Historical American English (COHA) and the Global Web Based Englishes (GloWbE) corpus. Their results confirm previous observations that get passive has become more frequent with time (Bohmann et al. Reference Bohmann, Müller, Honkanen and Neuhausen2025: 129). This is especially true for central get passives, that is, the ones that are not formulaic as in example (31) or adjectival in nature, for example as in (32). The construction – though still mainly used in informal contexts – is also being increasingly used in formal contexts (Bohmann et al. Reference Bohmann, Müller, Honkanen and Neuhausen2025: 131). In terms of varieties, AmE leads with the highest proportion of get versus be (Bohmann et al. Reference Bohmann, Müller, Honkanen and Neuhausen2025: 163–164). PakE is among the varieties with the lowest proportion of get passives in their GloWbE data.

1. (31) […] there’s a chance you might actually enjoy it once you get used to it and gain some competency. (SAOnE, UK Web forums)

2. (32) But don’t get carried away. (CLOB, Press editorials)

3. (33) The film’s release also got pushed back from July 2014 to April 2015. (SAOnE, US Websites)

4. (34) However, he was dispatched without grace to Saudi Arabia immediately after landing. (PWE, Newspaper columns)

The aim of the present subsection is to explore the use of be and get passives in the three varieties using the corpora described in Section 3.2. These datasets include three written (editorials, news, research papers) and four CMC (comments, tweets, web forums, websites) registers. Hence, these are relatively more diverse than the web-based data used by Bohman et al. (Reference Bohmann, Müller, Honkanen and Neuhausen2025). We have used MFTE (Le Foll & Shakir Reference Foll, Elen and Shakir2023; Le Foll & Shakir Reference Shakir2024) to extract the frequencies of the relevant tags PASS (be passive) and PGET (get passive) from the corpora. We then used Google Gemini 2.5 Pro to classify the get passive instances according to the six way classification of Collins (Reference Collins1996: 45–49): central as in (33), psychological (e.g. ‘got frustrated’), reciprocal (‘get married’), adjectival as in (32), formulaic as in (31), and complex reflexive (e.g. ‘Don’t get yourself killed’). A 10% sample of the get passives was used to check the accuracy of MFTE and subsequent classification by Gemini 2.5 Pro. The accuracy of passive recognition by MFTE was 98.85%, while the accuracy of Gemini’s classification of central versus non-central get passive was found to be 90%.

Figure 9 shows the words per million frequencies of be and get passives in the three varieties and two modalities. The get passive is slightly more frequent in CMC as compared to the three written English registers. AmE leads in the use of get passive in both modalities, which confirms previous studies. The PakE data trails the AmE and BrE data in CMC in the use of get passive. However, due to a slightly higher frequency of this variant in Pakistani newspaper columns, the frequency is overall higher than in BrE in the written medium. Otherwise, the main variant of the passive voice in PakE is be passives. It is either higher than in both of the native English varieties (in CMC) or near the BrE frequencies (in the three written registers). Deuber et al. (Reference Deuber, Hackert and Hänsel2022: 282) find a similar trend in Indian English. The differences in the use of be versus get passive in the written modality are not statistically significant among the three varieties (χ² (2, N = 15271) = 5.03, p = 0.081). However, these varieties do show significant differences in CMC (χ² (2, N = 11778) = 27.69, p < 0.001). Subsequent comparisons of PakE with AmE and BrE are also statistically significant in CMC – performed using the function pairwiseNominalIndependence (Mangiafico Reference Mangiafico2025). Overall, passive voice and especially be passive appears to be comparatively more common in PakE.

Figure 9

Be and get passives in written and CMC corpora of the three varieties of English

3.5.3 Modal Verbs of Obligation

Shakir (Reference Shakir2025) has conducted a study on the (semi-)modal verbs of obligation must, have to, (have) got to, and need to using the public registers (i.e. comments, tweets, web forums, and websites) in the SAOnE corpus. Like the other three SA varieties (Bangladeshi, Indian, and Sri Lankan Englishes), PakE has a higher preference for must, particularly in comments and tweets. The main use of must is in the root meaning to express personal opinions, referred to as ‘discourse internal obligation’ by the author. Example (35) shows the use of must from the SAOne corpus used by Shakir (Reference Shakir2025). The two native varieties (i.e. AmE and BrE) mainly employ need to for this purpose. PakE (along with the other SA varieties), however, uses need to as well as must. The author attributes the higher usage of must (especially on Twitter) to the presence of political activists who express their opinion using this verb. The number of unique lexical verbs used with must is also significantly higher in PakE (143) as compared to AmE (91) and BrE (72) (Shakir Reference Shakir2025: 10–11). The semi-modal need to has the highest frequency in the Pakistani data (Shakir Reference Shakir2025: 9). The author also finds epistemic uses of the verbs (have) got to and have to, as in (36) and (37).

1. (35) Pakistan must revive PSM. It is serious and major matter. (SAOnE-PK, Comments)

2. (36) Excellent … This has got to be the first one … (SAOnE-PK, Tweets)

3. (37) And look at their achievements, Bangladesh (i know raw was created in ‘68 but there has to be some sort of agency before that). (SAOnE-PK, Web forums)

3.5.4 Noun Phrase Complexity

The past two decades saw the rise of research on NPs to explore syntactic complexity and its impact across registers. Early linguists like Aarts (Reference Aarts1971) identified that ‘light’ (non-modified) NPs bear a specific relationship with spoken texts and subject positions, while ‘heavy’ (modified) NPs are associated with written texts and non-subject positions. NP complexity arises through the addition of modifiers – either pre-modifying or post-modifying the head noun (Quirk et al. Reference Quirk, Greenbaum, Leech and Svartvik1985). Biber et al. (Reference Biber, Johansson, Leech, Conrad and Finegan1999: 578) found that complex (i.e. pre- and post-modified) NPs occur more frequently in academic writing and news reports as opposed to spoken conversations.

Several studies have also been conducted on NP complexity in varieties of English. For example, Brunner (Reference Brunner2014) comparatively analyses British, Singaporean, and Kenyan NP patterns using the respective International Corpus of English (ICE) components. He attributes the differences across the three varieties to the structural uniqueness of each substrate language. Hence, Singaporean English employs straightforward pre-modification reflecting Sinitic head-final typology, while Kenyan English exhibits considerable Bantu patterns favouring more post-modification (Brunner Reference Brunner2014: 37). Brato’s (Reference Brato2020: 391) diachronic analysis of Ghanian English shows that NP complexity increases over time, which is attributable to increased education and proficiency in English, instead of language contact or substrate influence. Schilk and Schaub (Reference Schilk and Schaub2016: 81) analyse ICE components of Hong Kong, Singapore, Jamaica, India, and Canada. Their findings show that syntactic function of the NP (subject vs other) is the strongest predictor of complexity, while variety and register only become relevant in interaction effects. For example, they show that Indian English academic writing shows a strong tendency towards simple NPs (Schilk & Schaub Reference Schilk and Schaub2016: 78).

No analysis of NP complexity is available for PakE, so our aim in this subsection is to conduct a preliminary study on NP complexity in PakE in comparison to AmE and BrE. Since the NP types are highly dependent on registers, we limit ourselves to the three comparable written registers from PWE (i.e. editorials, news, and research papers) and their counterparts from Crown and CLOB. We use Schilk and Schaub’s (Reference Schilk and Schaub2016: 63) definitions to identify four types of NPs:

• Simple: (Det) + Noun, for example, (The) book, UN.

• Pre-modified: (Det) + Adjective(s) + Noun, as in (The) original book.

• Post-modified: (Det) + Noun + Prepositional phrase/relative clause, as in (The) book that you read/of Mathematics.

• pre- and post-modified: (Det) + Adjective (s) + Noun + Prepositional phrase/relative clause, as in (The) old book that you sent me/of Mathematics.

NPs with complex pre-modification (e.g. good A and B, good A and bad B) were also included in the last type, as proposed by Schilk and Schaub (Reference Schilk and Schaub2016: 63–64). The NPs were extracted from the tagged (POS and NER) and parsed (constituency and dependency) data using stanza (Qi et al. Reference Peng, Zhang, Zhang, Bolton and Manning2020) and the SuPar parser (Zhang, Li, & Min Reference Zhang, Zhenghua and Min2020; Zhang, Zhou, & Li Reference Zhang, Zhou and Li2020). The latter tool was selected because it performed slightly better than the built-in parser of stanza in constituency parsing. We used custom scripts written in Python to extract NPs from subject and other positions (e.g. object). NPs in stand-alone prepositional phrases were ignored. After extracting the NPs, we selected 100 random sentences (and the corresponding NPs) from each register category and each corpus. This generated about 420 NPs from each category and corpus, making a total of about 3700 NPs. These were manually checked for accuracy and false positives were discarded. The final dataset includes 3204 NPs.

The results are plotted with respect to the syntactic function (subject vs other) in the three varieties and three registers in Figure 10. The proportions show that there is hardly any difference in the use of NP types among the varieties under consideration. As shown in previous research, syntactic function shows the most visible differences in the use of NP types, where simple NP types are more common in subject position, while complex pre- and post-modified types are more frequent in all non-subject (i.e. other) positions. Nonetheless, two slight variations in PakE registers can be noted. Firstly, Pakistani research papers employ simple unmodified NPs more often when compared with both AmE and BrE, as depicted in the bottom panel ‘Research Papers’ under the panel titled ‘subject’. An example of such an NP is presented in (38). This finding is in line with Schilk and Schaub’s observation about Indian English, another SA L2 variety of English. Secondly, Pakistani news reports show a higher proportion of post-modified NPs in non-subject (i.e. other) positions, which is visible in the panel ‘News’ under the vertical panel ‘other’. As per our random sample, the post-modifiers in these NPs are mainly prepositional phrases, as shown in (39). In our mixed effects multinomial regression model (np_type ~ variety * register_category * np_position), the variables ‘register category’ and ‘NP position’ (subject vs other) are statistically significant predictors of NP type, but variety is not significant.

1. (38) The teacher acts as a facilitator of their learning. (PWE, Research papers, Syntactic function: subject)

2. (39) Speaking on the occasion, Tahir Iqbal said that […] these are essential for […] disseminating the message of peace and fraternity. (PWE, News, Syntactic function: other)

Figure 10

NP complexity according to syntactic function in the three varieties and selected registers

4.3.1 Data Annotation

The posts have been annotated for theme and language through Google Gemini 2.5 Pro. Gemini was not provided with an explanation of the thematic categories. Instead, a 10% random stratified sample (approximately ten posts from each Instagram account) of the selected data was manually annotated for the nine categories listed as follows. Then Gemini was instructed to tag additional data (including the manually annotated posts) using the given scheme:

• Ads: Promotions of film/drama projects (of the celebrity) or of local/international brands and businesses (clothes, beauty products, housing societies, etc.).

• Follower engagement: Asking a question and/or for feedback about a TV show, or otherwise engaging with the followers.

• Hashtags: Just hashtags with no other theme deducible from the hashtags.

• Personal life: Posting about self, family, and friends, including travels or simply selfies with some text.

• Professional life: Posting about career, job, colleagues, professional relationships, drama/film projects (e.g. back the stage updates).

• Politics/social: Posting on local or international political or social issues, including the ones that may affect the actor community.

• Quotations: Motivational quotations or posts that address the audience in general, song lyrics or poetry verses in Urdu or English, or a funny quotation.

• Unclear: Posts with no text including emojis, or posts that are hard to fit in any of the other categories.

• Wishes: On holidays, to colleagues or family members on birthdays, anniversaries, and simple prayers.

The data were also annotated for language using the same model. Code-switching was defined at phrase or clause level (two or more words) to make separation more realistic and hopefully get more accurate chunks of each language.

4.3.2 Validation of Annotated Data

The thematic annotation was completed in multiple runs to improve accuracy. Initially, the Large Language Model (LLM) was given a list of fourteen categories and brief explanations to annotate the data. A random sample of 100 posts indicated about 90% accuracy in theme and language annotation. Later on, a 10% sample (as described in Section 4.3.1) was manually tagged and compared with the LLM-based annotation. The comparison showed a 0.63 Kappa (unweighted for two raters: human vs machine). The accuracy was 69%. In the second run, the human-annotated data were used as examples for re-annotation of the complete dataset. The Kappa for the 10% sample (human vs machine) was 0.89, and the accuracy was 90%. The language annotation resulted in a Kappa of 0.46 in the first run, and 0.51 on the 10% sample in the second attempt. It was mostly mixed language posts that caused disagreement.

To understand the overall trends in the thematic categories, the 10% human-annotated sample and the overall dataset tagged by Gemini have been compared side by side in Figure 15. The bar charts show that Gemini tends to annotate social/political posts and wishes less frequently as compared to the hand-annotated data. The lack of cultural understanding required beyond the post text appears to be the main cause of this discrepancy. However, the three most frequent thematic classes (i.e. ads, personal life, and professional life) have almost similar proportions in both datasets.

Figure 15

Percentage comparison of thematic annotation (machine vs human)

4.3.3 Data Analysis

The data have been analysed for the effect of age, gender, and language on the post theme. The posts annotated by the LLM for English (74%) and Urdu (8%) were retained for the final analysis. Mixed posts (approx. 15%) and unclear cases were removed to reduce incorrectly annotated cases. Multinomial regression model from nnet::multinom was applied to the data, and the final model was selected using the Mass::stepAIC function (Venables & Ripley Reference Venables and Ripley2002) using R (R Core Team 2025). The model shows that all three explanatory variables interact with each other, with the interaction of language and gender being the weakest. The results have been visualised in the form of bar plots with average predicted probability of the eight thematic categories with respect to the selected explanatory variables. These marginal effects have been calculated using the marginaleffects package (Arel-Bundock, Greifer, & Heiss Reference Arel-Bundock, Greifer and Heiss2024) with the option newdata=‘balanced’ in the function avg_predictions to account for unequal subgroups in age, gender, and language.

Figure 16 shows the average predicted probabilities of post theme as per age groups and gender. The likelihood of professional life-related posts increases with age in both genders, while personal life-related posts remain almost steady in females, with a decrease in older males. Example (40) shows a professional life-related post where a female celebrity is talking about an old drama she performed in. A post captioned with my family has been given the category of personal life, as in (41). Posting ads (own drama/file project or a brand’s advertisement) decreases with age. The highest probability appears to be in younger female celebrities. Example (42) shows the advertisement of a clothing brand without clarifying that it is an ad. Example (43) is a much more clearly phrased advertisement of a brand. The promotion of own drama project in (44) is an ad, but it also shows why it is difficult to annotate for language even for a capable model like Gemini 2.5 Pro. This is especially the case when post text is shorter. Lastly, posts about social/political issues are more likely in older age groups in both genders, as shown in (47) and (48).

1. (40) Blast from the past throwback. AROOSA days with the most amazing Fatima Surraya Bajia (May her soul rest in peace Ameen). What a personality. Am so lucky to have worked with such a stalwart. #xxxxxxxxx #xxxxxxxxxxxxxxxxxx #ptv #ptvhome #pakistanicelebrities (female, 46+, professional life)

2. (41) My Family MashaAllah [Family, Support System, Love, Farooqui Brothers, twins, Xxxxxx Farooqui, Xxxxx Farooqui, Bleesed] (male, 30–45, personal life)

3. (42) Dark feminine Wearing @xxxxxxxxxxxxxx (female, <30, ads)

4. (43) Check out the amazing LUXURY LAWN COLLECTION 2022 by @xxxxxxxxxxx well known for introducing remarkable luxurious and demanding articles with high quality fabric. Shop Now: <URL> (female, <30, ads)

5. (44) SHAMI KI SHADI (‘Shami’s wedding’) A tale of love, laughter, and unexpected chaos! Starting from 29th August Every Thursday at 9:00PM Only on @aurlifeofficial Staring @xxxxxxxxxxxxxxxxxx @xxxxxxxx @xxxxxxxxxxxxxxxxxxxx Writer: @xxxxxxxxxxxxxx & @xxxxxxxxxx Director: @xxxxxxxxxxxxxxxxxxxx Makeup Artist: @xxxxxxxxxxxx DOP: @xxxxxxxxxx #xxxxfilmsproduction (male, 30–45, ads) The interaction of age and language to explain the post theme has been depicted in Figure 17. Urdu posts are only a small fraction of the total dataset. Nonetheless, the data show some interesting patterns. Urdu posts have been mostly tagged as quotations, especially in younger people (less than 30), and as political/social commentary mostly in older people of both genders. Example (45) shows Hindi/Urdu song lyrics posted by a male celebrity. Follower engagement-related posts in under thirty males in Urdu are more likely. An example of such a post with a simple question is shown in (46). Two posts in the category of social/political are shown in (47) and (48) in Urdu and English, respectively. Posts related to personal wishes (e.g. a birthday) or holiday wishes (e.g. Eid) can occur both in English and Urdu, but the latter is slightly more favoured. Two such examples are shown in (49) and (50). Furthermore, it can be seen in the figure that professional life and advertisement-related posts are less likely to occur in Urdu; especially the under thirty and thirty to forty-five age groups are more likely to use English for such posts. This shows that even if the content they create is in Urdu (e.g. dramas), they prefer to talk about their professional life and promote the content in English. Lastly, post captions where personal life and family are discussed, as in (41), are also preferably written in English.

6. (45) Andhere ko baahon main leke (‘Taking darkness in its arms’)

   Ujaale ne ghar basaya hay (‘The light has built a home’)

   Churaya tha jo chukaya hay!!! (‘It has paid back what it had stolen’) (male, <30, quotation, Hindi/Urdu Bollywood song lyrics)

7. (46) Kiya bolti hai public? (‘what does the public say?’) (male, <30, follower engagement, Urdu)

8. (47) Kash social media logoon ko Jordta …. (‘Wish social media would unite people’)

   Kash social media kay chaudry pages chalanay kay liay mohabat pahlatay na kay nafrateen. […] (‘Wish the elites of social media promoted love instead of hate.’) (male, 46+, social/political, Urdu)

9. (48) 2020 just keeps getting worse and worse but I cannot comprehend such a great loss. Not just for me but for all of us. Sister Ruth Lewis, your service to mankind has been and always will be one that no words can do justice to. […] (female, 46+, social/political, English)

10. (49) Happy birthday to my very old friend @xxxxxxxxxxx

    You hold a very special place in my heart. God bless you and always stay happy and healthy. […] (female, 30–45, wishes, English)

11. (50) Chaand raat Mubarak […] (‘Happy moon night’: refers to the night before Eid) (female, <30, wishes, Urdu)

To summarise this subsection, we have presented background and analysis of Instagram posts of Pakistani celebrities (mainly actors) from two genders and different age groups. The post themes and language have been automatically tagged through an LLM, which may have inaccuracies as shown by the comparisons in Section 4.3.2. However, the overall trends for thematic categories are approximately the same in the human-annotated sample and the machine-annotated complete dataset. The results show that Pakistani celebrities use English for posts belonging to such themes like ads, personal life, and professional life. Urdu and mixed language posts also occur, but they are only a small portion of their online content. This shows that Pakistani celebrities embrace English as the dominant language on the internet while communicating with their fans within Pakistan as well as with those living abroad, and possibly also with the international audience of Pakistani dramas that may reside in other SA countries (e.g. Bangladesh) or elsewhere.

Figure 16

Average probability of post theme according to age and gender

Figure 17

Average probability of post theme according to age and language

4.4.1 Data and Methodology

The data for the present analysis are based on the SAOnE corpus (see Section 3.2). The tweets category includes one million words in 500 text files (one Twitter user/file; 100+ tweets per user) from each of the six countries. The users’ profiles were manually checked to confirm their country of origin. The data were collected in 2021, so most of the tweets were posted in this year or a couple of years earlier. During the compilation phase of SAOnE and while conducting studies on parts of these data (Shakir Reference Shakir2024; Shakir Reference Shakir2025), it was observed that Pakistani tweets include political discussions. Hence, this dataset is well-suited to explore main themes and discourses in English tweets of Pakistani users.

Lexical Multidimension Analysis (LMDA) proposed by Sardinha and Fitzsimmons-Doolan (Reference Sardinha and Fitzsimmons-Doolan2025: 26–27) is employed for this study. LMDA is inspired by Multidimensional Analysis (MDA) propagated by Biber (Reference Biber1988) and many other studies who followed him. An MDA starts with the selection of functionally relevant lexico-grammatical features (e.g. adverbs and their kinds, verbs and their semantic classes, types of that complement clauses). Normalised per-1,000-word frequencies of these features are obtained from a representative corpus. Then a dimensionality reduction statistical technique like Exploratory Factor Analysis (EFA) or Principal Component Analysis (PCA) is applied to obtain clusters of co-occurring linguistic features. These groups of features are then functionally interpreted by taking into account the data included and the situational characteristics of the register categories in the data. LMDA uses content words (i.e. nouns, verbs, adjectives, and adverbs) and excludes function words like prepositions as well as structural features like clause types. LMDA builds on the concept of keyword analysis – which produces a list of statistically significant words by comparing one corpus to another (Baker Reference Baker2023: 165) – to find clusters of keywords that co-occur and may uncover underlying discourses and themes in a clearer way (Clarke, McEnery, & Brookes Reference Clarke, McEnery and Brookes2021; Sardinha & Fitzsimmons-Doolan Reference Sardinha and Fitzsimmons-Doolan2025: 13–26).

For the present analysis, we identify keywords by comparing the Pakistan tweets sub-corpus from SAOnE to the tweets of the five other countries included in the corpus. The corpus analysis tool AntConc (Anthony Reference Anthony2024) has been used for this purpose. The data are tagged and lemmatised with the stanza (Qi et al. Reference Peng, Zhang, Zhang, Bolton and Manning2020) package in Python (Python Core Team 2025). As a pre-processing step, we remove # and @ from Twitter handles and hashtags, respectively. This way proper nouns can be easily identified by the tagger and later included in the keyword analysis. We use the dispersion-based keyness computation method proposed by Egbert and Biber (Reference Egbert and Biber2019), which is available as an option in AntConc. This way, we are able to identify a list of 1735 lemmas that are used in many text files, and by extension by the Pakistani Twitter users. The normalised per-1,000-word frequencies are calculated for these features and compiled in a table. As a further step before LMDA, we remove 170+ function words based on the list from Dang and Webb (Reference Dang, Webb and Nation2016: 188–189) to keep only the lemmas of content words for further analysis. Additionally, we remove any lexical items with zero frequency in 80% or more documents. We also remove features which have very low correlation (<|0.1|) to any other feature, communalities of less than 0.2 as recommended by Egbert and Staples (Reference Egbert, Staples, Berber-Sardinha and Pinto2019), or a Kaiser–Meyer–Olkin (KMO) value of less than 0.5. Lastly, we apply signed-log transformation to the frequency table to reduce the impact of skewed frequencies (Neumann & Evert Reference Neumann, Evert, Seoane and Biber2021: 155). For dimensionality reduction, we use PCA with no rotation – see Shakir (Reference Shakir2024: 377–379) for a discussion on this. Figure 18 shows that about six dimensions can be extracted through both EFA and PCA.

Figure 18

Scree plot of data (EFA and PCA)

We include four dimensions in our analysis here, as the fourth and sixth dimensions are not interpretable. The final analysis has seventy-five keywords with an overall KMO value of 0.80 (‘Meritorious’), which shows the suitability of the data for EFA or PCA (Egbert & Staples Reference Egbert, Staples, Berber-Sardinha and Pinto2019). The variance explained by the PCA solution is 35%. The cumulative variance explained by the first dimension is the highest, and it drops with subsequent dimensions. Only those lemma keywords are retained which have a communality of |0.295| on either dimension, as shown in Table 5. The table also shows the interpreted thematic labels for each of the four principal components.

4.4.2 Themes and Discourses in PakE Tweets

The first principal component consists of forty-eight keywords. It is the biggest dimension in terms of lemma keywords loaded and the variance explained in the PCA model (8% of the cumulative variance explained). The keywords can be grouped in various categories. There are proper nouns that are either names of Pakistani politicians (e.g. Nawaz, Umar, Khan) or their Twitter account handles (e.g. ImranKhanPTI, MaryamNSharif, BBhuttoZardari). Other proper nouns are related to party names (e.g. ppp, PTI), famous journalist names or Twitter handles (e.g. HamidMirPAK) or place names (e.g. Kashmir, a region disputed with neighbouring India). Some of these proper nouns may also occur as hashtags. Common nouns such as election, leader, govt., corruption, court, nation, minister, law, force, citizen, shame, action, issue, state show that a common theme of politics with various aspects (e.g. corruption) is present in these tweets. The adjectives political, terrorist, innocent, foreign also strengthen this interpretation as they cover various internal and external political and social challenges that may be under discussion in English tweets of Pakistanis. This dimension appears to represent ‘Issues related to Pakistani politics’ on Pakistani Twitter.

As mentioned in Section 4.4.1, these tweets are from 2021 when Mr Imran Khan was still the prime minister of Pakistan. The Twitter accounts active at that time and included in these data are not only PTI-affiliated activists but also from other parties. In (51) and (52) most probably a sympathiser of or a social media account of another political party, Pakistani Muslim League Nawaz (PMLN), is alleging the same corruption charges that would be levelled against Mr Nawaz Sharif (the PMLN leader) by PTI-affiliated and sympathiser accounts. In (53), the governance abilities of Pakistan People’s Party (PPP), which is in power in the province Sindh, are being called into question. Examples (54) and (55) are from a PTI-affiliated account and the same discourse of corruption is now being wielded against PMLN and PPP leaders (Zardari) in Sindh. Along with politicians the tweets may also mention army – which can refer to Pakistan’s army or another foreign army – and other security challenges like terrorism. On the whole, this dimension confirms previous observations that the most dominant discourse or theme on Pakistani Twitter revolves around various political issues, and a sizeable part of it is communicated in English as well.

1. (51) ImranKhanPTI Imran Khan never take any action Janhangir corruption. (PK-TWT135.txt)

2. (52) I condemn Imran Khan PTI chairman to open first offshore company with out name for the purpose to evade Tax and wealth. (PK-TWT135.txt)

3. (53) BBZ Why Afghan influx in Sindh is not checked by Sindh govt. (PK-TWT300.txt)

4. (54) ArifAlvi ImranKhanPTI ImranIsmailPTI Asad Umar: Zardaris Sindh Police criminally assaulting Karachites & enforcing unannounced Curfew. (PK-TWT-025.txt)

5. (55) You bulldozed the artificial majority of corrupt N- League Mafia- that’s like a real roaring tiger of PTI. (PK-TWT025.txt)

The second dimension has eighteen keywords with seven on the positive side and eleven on the negative side. The keywords on the positive side are again from one political party, that is, PPP, and its affiliated accounts, where they tweet about various activities of the Sindh provincial government, for example, (56). On the negative side, the keywords are related to regional and international security, and political issues. Hence, there are country names (US, Afghanistan), foreign groups and political forces (Afghan, Taliban), and adjectives like terrorist, showing that these accounts are more interested in tweeting about international and regional affairs. This is also relevant because the US was preparing to leave Afghanistan in this time period, so this theme emerges as a concern for Pakistan. Examples (57) and (58) present example sentences from the tweets of two different Pakistani accounts. Both of them are either concerned about a Taliban-led government or the future of foreigners in Afghanistan (here most probably referring to the Americans who would soon leave the country). The themes related to terrorism are in line with other such studies on different datasets originating from Pakistan; for example, Gries (Reference Gries2025: 16) finds terror and relevant keywords based on English newspaper data collected around 2020. Combining both the positive and negative sides, the second dimension can be labelled as ‘PPP Sindh vs regional issues’.

1. (56) Spokesperson Sindh Govt murtazawahab1 and Minister Information SyedNasirHShah condolence with PLF G S Sindh Arshad Naqvi on sad demise of his father in clifton karachi. (PK-TWT392.txt)

2. (57) This is interesting: Though Taliban reportedly cancelled inauguration ceremony but Russia it will not attend any inauguration of Taliban govt. (PK-TWT247.txt)

3. (58) These foreign stooge have no future in Afghanistan. (PK-TWT208.txt)

The third dimension has eleven keywords on the positive side and two on the negative side. The positive keywords appear to be non-political (e.g. bless, prayer, Allah, recovery, soul) and rather religious. However, as (59) and (60) show, many such tweets are from the same politically affiliated accounts that may praise their respective leaders or wish them good health. The two negative keywords (Asad, Umar) make the name of a prominent PTI politician Asad Umar. The negative loadings of these keywords suggests that they are absent from the text files where the positive – seemingly less political – keywords occur. Thus, the majority of dimension 3 is related to themes where political issues are not under discussion but prayers and blessings for others (including politicians). Hence, it can be named as ‘Prayers and blessings’.

1. (59) GetWellSoonMaryamNawaz May Allah bless you with a long and healthy life and […] (PK-TWT265.txt)

2. (60) Our Shah Jee always relish our minds & hearts by posting beautiful roses. (PK-TWT225.txt)

The last dimension consists of seven keywords on the positive side (common nouns action, official, citizen, issue, city, road and the address form Sir) and three on the negative side (politician names). The keywords on the positive side along with examples (61) and (62) show that certain Pakistani Twitter users try to engage with public officials to lodge their complaints or plead for help on issues like road repair. Such accounts probably do not or less often tweet about political leaders, which is shown by the negatively loaded politician names (Khan, Nawaz, Imran) that are the hallmark of politically affiliated accounts. This dimension can be interpreted as ‘Citizen issues and pleadings to govt. officials’. It is noteworthy that this tendency to use social media and viral posts is not limited to Twitter. In recent years, smartphone captured videos of incidents (e.g. of police brutality) have been used to get the attention of relevant authorities to seek justice. The language of such videos is mostly not English, but the examples here show that English in also used in this context along with other local languages in Pakistan.

1. (61) Thanks for sharing the information’ many people will be saved from fraudulent messages’ hope action will be initiated against such fake and fraudulent peoples. (PK-TWT459.txt)

2. (62) HyderabadDeputy Dear sir, great initiative by ur kind, one request to ur gdself, plz repair road or make paver block from citizen colny to bhittai nagar road whch z in bad condtion since three years no repair work done. (PK-TWT384.txt)

To summarise, the LMDA has shown that politics and political issues is the main theme of PakE tweets. Pakistani netizens on Twitter appear to be either active members of a political party, paid by them, or are their sympathisers. It is also noteworthy that already in 2021 other political parties had tried to catch up to PTI, the main party who successfully utilised social media to advance their narrative and gained power in 2018. Thus, we have seen the discourse of corruption being used both for and against PTI. The other themes discussed on Twitter are regional and security issues, and lastly some users try to connect with public sector officials and departments to get their local issues solved. This analysis reveals how, on social media, English – along with other languages like Urdu – is employed in communication about a highly local issue, that is, politics, where the interlocutors are also locals. This is despite the fact that Urdu (the national language and lingua franca) and the regional languages are the main vehicles of communication for politics in most other mediums like face-to-face conversations or electronic media.