3.1.1 CoSEM Studies Thus Far

In the previous section, we discussed CoSEM’s design and logistics; in this section, we provide an overview of research conducted with the corpus to date. One of our first publications, Gonzales et al. (Reference Gonzales, Hiramoto, Leimgruber and Lim2023a), introduced CoSEM, situating it alongside existing corpora of Singapore English, detailing its construction and 3.6-million-word composition (v.1), and illustrating its analytical potential through preliminary case studies on two CSE features: the Hokkien-derived (bo)jio ‘neg invite’ and clause-final adverbs (CFAs). In conjunction with the larger and more recent CoSEM (v. 5) described earlier, this work established the corpus as a key resource for tracing change in CSE in computer-mediated communication. Subsequent studies have used CoSEM to investigate SFPs – arguably the most well-known feature of CSE and an enduring area of interest in CSE research. Leimgruber et al. (Reference Leimgruber, Lim, Gonzales and Hiramoto2021) showed that gender and ethnicity shape the use of certain particles in CSE, reinforcing the view that SFPs are a particularly sensitive locus of sociolinguistic differentiation in CSE (see also Leimgruber Reference Leimgruber2009; Smakman & Wagenaar Reference Smakman and Wagenaar2013; Botha Reference Botha2018). Building on Khoo (Reference Khoo2012), Hiramoto et al. (Reference Hiramoto, Gonzales, Leimgruber, Lim, Choo, Ngefac, Wolf and Hoffmann2022) traced the development of the novel SFP sia from its origins as a Malay swearword into a marker of youth identity and a ‘coarse but cool’ style, used across a variety of speakers and interactional contexts. Hafiz et al. (Reference Hafiz, Hiramoto, Leimgruber, Gonzales and Lim2025) extended this work, examining the distribution of sia and its variants (sial, siak, and siol) across demographic groups and showing that the use of sia and its variants has spread beyond young male CSE speakers and is increasingly used by younger CSE speakers.

Building on the pilot study in Gonzales et al. (Reference Gonzales, Hiramoto, Leimgruber and Lim2023a), J. J. Lim, Hiramoto et al. (Reference Lim, Hiramoto, Leimgruber and Gonzales2024) revisited CFAs in CSE, showing that already and only have increased in frequency over time while also has remained relatively stable. These results provide evidence for growing morphosyntactic endonormativity (Schneider Reference Schneider2007) in CSE. The authors also demonstrated that younger and male speakers lead in the use of CFAs, reinforcing earlier findings that age and gender are highly relevant for understanding variation in CSE.

The Corpus of Singapore English Messages has also facilitated our investigation of innovative CSE features, such as adverbial confirm and the clause-initial question tag is it. Drawing on data from CoSEM and ICE-SIN, J. J. Lim, Hafiz et al. (Reference Lim, Hafiz, Gonzales, Leimgruber and Hiramoto2024) documented the emergence and formal properties of a novel adverb in CSE – confirm – used to express strong speaker certainty in an utterance. Strikingly, the authors also showed that adverbial confirm was used more frequently in CoSEM over other surface-semantically equivalent adverbials such as definitely and for sure. As the adverbial use of confirm has not been reported elsewhere in the CSE literature, the authors introduce it as an emerging CSE resource and use CoSEM text message data to describe its formal properties and distribution. Similarly, Gonzales et al. (Reference Gonzales, Hiramoto, Leimgruber and Lim2023b) first drew attention to the innovative use of is it in clause-initial position in CSE, documenting not only its formal properties but also showing that this construction is more likely to be used in rhetorical, playful contexts and by younger Singaporean speakers. Building on this work, Hiramoto et al. (Reference Hiramoto, Gonzales, Leimgruber, Lim, Hafiz, Rüdiger, Leukert and Leimgruber2025) showed that the use of clause-initial is it in CSE has increased over time, and drawing on Siegel’s (Reference Siegel2000, Reference Siegel and Michaelis2008) substrate transfer theory, argued that it entered CSE due to contact with local Sinitic languages in Singapore.

In addition to the abovementioned features, CoSEM contains many other playful and innovative CSE forms that may indicate ongoing changes in the variety. Using data from CoSEM, several undergraduate honours theses have investigated features like already and only (Cheong Reference Cheong2016), toh ‘fall, die’ of Hokkien origin (Chee Reference Chee2019), dei, a vocative directed particularly at Indian men (Kaur Reference Kaur2019), and the lesser-studied SFPs bah and wor (Choo Reference Choo2024). Other promising avenues for further study have surfaced in exploratory student projects, such as the use of double SFPs (see Boo, Lee, & Tan Reference Boo, Lee and Tan2023). Together, these studies underscore the richness of the CoSEM dataset and its potential to capture socially salient innovations that have not yet been systematically described in the literature. At the same time, our reliance on CoSEM necessarily constrains the kinds of claims we can make about CSE more broadly. As a text-message corpus, CoSEM privileges relatively young, digitally active speakers and offers only indirect access to speakers’ metalinguistic awareness, social evaluations, and reported usage outside the messaging context. The digital communication landscape has also shifted considerably since we began building CoSEM, with many Singaporeans now using other messaging platforms beyond WhatsApp and Telegram and making it increasingly difficult to capture CSE usage in Computer-Mediated communication. This may be why some of the innovative items identified in the corpus (e.g., toh, dei, and double SFPs) occur infrequently, making it challenging for fine-grained quantitative modelling by age, ethnicity, or gender. However, these features are recognisable and stylistically salient for many CSE speakers. To address these limitations and test whether these features are diffusing beyond the specific cohorts represented in CoSEM, it became crucial to supplement the corpus with an attitudinal and usage survey.

3.1.2 Motivations for the Survey

A central question in sociolinguistics research concerns the relationship between how speakers perceive their own language use and their actual linguistic behaviour. Classic work by Labov (Reference Labov1966, Reference Labov1972, Reference Labov1975) demonstrated that speakers’ self-reported usage of certain linguistic features often diverges from observed usage, suggesting that metalinguistic intuitions may not always be reliable indicators of linguistic behaviour (see also Trudgill Reference Trudgill1972, Reference Trudgill1974a). Such mismatches may arise because speakers’ perceptions are shaped by their attitudes towards certain linguistic features, as well as broader prescriptive norms about language. At the same time, other studies (e.g., Bailey, Wikle & Tillery Reference Bailey, Wikle and Tillery1997; Chambers Reference Chambers1998; Bailey & Tillery Reference Bailey and Tillery1999; Tillery, Wikle & Bailey Reference Tillery, Wikle and Bailey2000; Benson et al. Reference 75Benson, Chik, Gao, Huang and Wang2009; Dollinger Reference Dollinger2012; Dollinger Reference Dollinger2015; Kostadinova Reference Kostadinova2018) have demonstrated that speakers’ perceptions can, under certain conditions and for certain linguistic features, reliably track actual patterns of language use. The supplementary survey to CoSEM allows us to probe how widely these forms are known, how often and in what contexts they are used (e.g., speech vs. texting), and what social meanings speakers attribute to them, including possible associations with particular ethnic groups, genders, or stance styles. In this way, corpus data and survey evidence can be brought into productive dialogue: the corpus provides ecologically valid records of usage in situ, while the survey offers a broader window onto the penetration, social evaluation, and perceived normativity of emerging CSE features among speakers outside CoSEM.

We administered an online survey containing questions about respondents’ demographic info, CSE use (awareness, acceptability, frequency, etc.) (see Section 3.3 for details), and contact information for possible follow-up. The survey was distributed across our social networks in Singapore between 2024 and 2025 and was targeted at adults over eighteen who reported being native CSE speakers and were living in Singapore. While it was not possible to exclude CoSEM participants in the survey entirely, we tried our best to recruit participants who are outside of the CoSEM cohorts by encouraging the survey participants to distribute the survey to a wider range of people in terms of age, race, occupation, education, and language background. This snowball method yielded 230 participants, of which 214 met our criteria and were retained for analysis.

In this section, we focus on comparing survey data on Singaporeans’ self-reported usage of CSE linguistic features with actual usage patterns attested in CoSEM as our initial step of utilising the combination method. To do this, we highlight three salient CSE linguistic features mentioned earlier as our test samples: clause-initial is it (Gonzales et al. Reference Gonzales, Hiramoto, Leimgruber and Lim2023b; Hiramoto et al. Reference Hiramoto, Gonzales, Leimgruber, Lim, Hafiz, Rüdiger, Leukert and Leimgruber2025), adverbial confirm (Lim, Hafiz, et al. Reference Lim, Hafiz, Gonzales, Leimgruber and Hiramoto2024), and the Malay-derived SFP sia (Khoo Reference Khoo2012; Hiramoto et al. Reference Hiramoto, Gonzales, Leimgruber, Lim, Choo, Ngefac, Wolf and Hoffmann2022; Hafiz et al. Reference Hafiz, Hiramoto, Leimgruber, Gonzales and Lim2025). These linguistic tokens were selected due to the fact that they show frequencies in CoSEM substantial enough to support a meaningful comparison. We anticipate that insights gained from these high-frequency items will help guide future analyses of less frequent CSE features by revealing how survey-based perceptions may align with or diverge from corpus-based observations.

It is important to note that the survey respondents and corpus participants do not constitute the same group of individuals. Because both datasets were collected anonymously, it is not possible to determine any overlap between them, nor to directly compare a respondent’s perception of their linguistic behaviour with their actual observed usage. However, this is not our objective. Rather, by examining whether self-reported usage by survey respondents aligns with or diverges from previous corpus-based findings, we aim to contribute to ongoing discussions about the extent to which self-reports can serve as reliable indicators of language use across social variables.

Section 3.2 provides a brief overview of the three linguistic features under consideration. Section 3.3 outlines the survey design and our hypotheses. Section 3.4 presents the results of our analysis, and Section 3.5 concludes with a discussion of how self-reported usage compares with observed patterns in CoSEM, and what this might reveal about language variation in CSE.

3.2.1 Clause-Initial Is It

In addition to its canonical use in standardised English, where is it functions primarily as a question tag seeking agreement from the listener (1), CSE also contains an innovative clause-initial use of is it, which serves mainly as a rhetorical device conveying a playful or teasing affect (2). It is hypothesised that this clause-initial use arose through substrate influence from the clause-initial A-not-A question construction found in the Sinitic languages spoken in Singapore. See Gonzales et al. (Reference Gonzales, Hiramoto, Leimgruber and Lim2023b) and Hiramoto et al. (Reference Hiramoto, Gonzales, Leimgruber, Lim, Hafiz, Rüdiger, Leukert and Leimgruber2025) for detailed discussions of CSE clause-initial is it, including how it differs from the question-tag is it in standardised English.

1. (1) it’s not their problem is it

   <ICE–SIN:S1A–062#89:1:B>

1. (2) wa is it you think i wuliao

   ‘Wow, are you saying that I am boring?’

   <COSEM:18CF59–2829–20CHF–2017>

Gonzales et al. (Reference Gonzales, Hiramoto, Leimgruber and Lim2023b) find that younger speakers are more likely to use clause-initial is it than older ones. However, inadequate evidence of gender and ethnicity effects is found.

3.2.2 Adverbial Confirm

As in standardised English, the lexical item confirm can function as a verb in CSE (3). However, unlike in standardised English, CSE also features an adverbial use of confirm, which expresses strong speaker certainty, as illustrated in (4). J. J. Lim, Hafiz, et al. (Reference Lim, Hafiz, Gonzales, Leimgruber and Hiramoto2024) argue that CSE adverbial confirm represents a case of grammatical replication (Heine & Kuteva Reference Heine and Kuteva2003; Heine & Kuteva Reference Heine and Kuteva2005; Heine et al. Reference Heine, Kaltenböck, Kuteva and Long2021) from Mandarin kending, which similarly functions both as a verb meaning ‘to confirm’ and as an adverb meaning ‘definitely’. See J. J. Lim, Hafiz, et al. (Reference Lim, Hafiz, Gonzales, Leimgruber and Hiramoto2024) for intricate evidence that supports treating CSE adverbial confirm as a strongly subjective speaker-oriented adverb.

1. (3) But let’s wait for <name> to confirm the timing?

   ‘Let’s wait for <name> to confirm the time, shall we?’

   <COSEM:17CF24-111-23CHM-2016>

1. (4) Aiyah I will confirm be late

   ‘Damn, I will definitely be late.’

   <COSEM:18CF47-1636-20CHF-2017>

While previous work has not reported any effect of age, gender, and ethnicity on the use of adverbial confirm, ongoing research in the team indicates that younger speakers, male speakers, and Chinese speakers are more likely to use adverbial confirm.

3.2.3 Sentence-Final Particle sia

A recent addition to CSE’s well-known inventory of SFPs is sia, which contains four variants: sia, sial, siak, and siol. Previous research has suggested that sia functions as an intensifier signalling coarseness (Khoo Reference Khoo2012) and may also be pragmatically employed to index youthfulness and coolness (Hiramoto et al. Reference Hiramoto, Gonzales, Leimgruber, Lim, Choo, Ngefac, Wolf and Hoffmann2022). See Hafiz et al. (Reference Hafiz, Hiramoto, Leimgruber, Gonzales and Lim2025) for a detailed discussion of the pragmatic meaning differences between the four variants of sia. An example of sia is shown in (5).

1. (5) How much honey he buy sia

   ‘How much honey did he buy?!’

   <COSEM:18CF42-1803-23MAF-2016>

Building on Khoo (Reference Khoo2012) and Hiramoto et al. (Reference Hiramoto, Gonzales, Leimgruber, Lim, Choo, Ngefac, Wolf and Hoffmann2022), Hafiz et al. (Reference Hafiz, Hiramoto, Leimgruber, Gonzales and Lim2025) show that younger speakers, male speakers, and Malay speakers are more likely to use sia.

3.2.4 Intermediate Summary

Table 6 summarises the findings regarding the relationship between the use of clause-initial is it, adverbial confirm, and SFP sia, and the speaker’s age, gender, and ethnicity.

First, younger Singaporeans are found to be more likely to use all three linguistic features, reinforcing the idea that these variables represent recent innovations in CSE. However, limited evidence of the effect of ethnicity has been found on the use of clause-initial is it in CoSEM, while Chinese speakers are more likely to use adverbial confirm, and Malay speakers are more likely to use SFP sia. This contrast suggests that adverbial confirm and SFP sia may represent later developments in CSE, while clause-initial is it may represent an earlier development and is therefore broadly shared across Singapore’s major ethnic groups. The hypothesis that adverbial confirm and SFP sia are newer innovations receives some support from gender patterns: inadequate gender effects are observed for clause-initial is it, while males are more likely to use adverbial confirm and SFP sia. Previous research has similarly identified the male gender as a positive predictor of the use of morphosyntactic features in CSE (e.g., Leimgruber et al. Reference Leimgruber, Lim, Gonzales and Hiramoto2021; Lim, Hiramoto, et al. Reference Lim, Hiramoto, Leimgruber and Gonzales2024; Hafiz et al. Reference Hafiz, Hiramoto, Leimgruber, Gonzales and Lim2025).

3.3.1 Structure of the Survey

We designed the survey using Google Forms, titled Survey on Singlish use. The aim of the survey was to investigate the contexts in which CSE is used among Singaporeans, how Singaporeans perceive the use of CSE across different age groups, and respondents’ perceptions of their own use of specific CSE morphosyntactic features. The survey was divided into eight sections, and respondents could only proceed to the next section after completing all mandatory fields in the current one.

Section 1 introduced the survey and included its title, purpose, and the investigators involved in the research (i.e., the authors of this monograph). It also provided an overview of the survey questions and information about data privacy and contact details. Participation in the survey was entirely voluntary, and respondents did not receive any compensation for completing the survey.

Section 2 collected demographic information from respondents, including their age (numerical value), gender (male, female, or option to self-identify), ethnicity/race (Chinese, Malay, Indian, Eurasian, or option to self-identify), nationality (Singaporean or option to self-identify), years of residency in Singapore (numerical value), and social media platforms used (TikTok, WhatsApp, Twitter, Line, Viber, WeChat, Facebook, Messenger, Instagram, Signal, Telegram, SMS, Reddit, and an option to specify others). Respondents could also optionally self-identify their generational membership (Gen Alpha, Gen Z/Zoomer, Gen Y/Millennial, Gen X, Baby Boomers, Silent Generation, Greatest Generation, or option to select ‘No idea’ or to specify others).

Section 3 asked respondents to indicate which social media platforms they frequently use CSE on, with the same options available as presented in Section 2. Respondents were then asked to describe, based on their experience, the distinctive ways in which children, teenagers, young adults (ages twenty-one to fifty), and older adults (over fifty) use CSE. Respondents were also invited to share examples and reflections on whether they perceive there to be any age-related differences in the use of Singlish on social media.

Sections 4 to 7 focussed on respondents’ self-reported frequency of use of specific morphosyntactic features in CSE. For each feature, respondents indicated whether they used it ‘a lot’, ‘occasionally’, ‘very rarely’, or ‘never’. Because the survey covered a wide range of features, the questions were divided across three sections to reduce participant fatigue. Section 4 included a mix of relatively new and older CSE features, as well as a question asking whether respondents believed jio or bojio entered CSE first, whether both emerged simultaneously, or whether they were unsure. Section 5 concentrated on relatively older CSE features. Sections 6 and 7 focussed primarily on SFP usage, with particular attention to double SFPs. To minimise misinterpretation, each feature was accompanied by an example sentence – some drawn from CoSEM – illustrating how the feature might be used in a CSE utterance. The features investigated in each section are as follows:

Finally, Section 8 asked whether the respondents would be willing to be contacted for a follow-up interview, with the option to provide their contact information.

3.3.2 Survey Distribution

The survey was distributed between 2024 and 2025, beginning with the authors’ personal networks and subsequently extending to the networks of their friends and family members. The first responses were collected on 17 May 2024, and the final response was received on 13 March 2025.

3.3.3 Respondent Pool

We collected a total of 230 survey responses. Sixteen were excluded because the respondents were either non-Singaporean or Singaporean individuals who had spent the majority of their lives outside Singapore. Of the 214 remaining responses, 118 (55.1%) belonged to female respondents and 96 (44.9%) belonged to male respondents, creating a slight imbalance in gender distribution. We also assigned the respondents to generational cohorts based on their age: 56 (26.2%) Zoomers (ages 12 to 27 in 2024; 13 to 28 in 2025), 83 (38.8%) Millennials (ages 28 to 43 in 2024; 29 to 44 in 2025), 59 (27.6%) Gen X respondents (ages 44 to 59 in 2024; 45 to 60 in 2025), and 15 Boomers (ages 60 to 78 in 2024; 61 to 79 in 2025). Only one respondent was 80 years of age – above 78 – in 2024, and we opted to group this individual with the Boomers due to the proximity of their age rather than creating a separate category. This brings the total number of Boomers to 16 (7.5%). In terms of ethnicity, 167 (78.0%) respondents were Chinese, 20 (9.3%) respondents were Indian, 13 (6.1%) respondents were Malay, and 14 (6.5%) respondents identified as Other. A limitation of the study is the uneven distribution across generational cohorts and ethnic groups in the respondent pool.

3.3.4 Predictions

As mentioned earlier, this study investigates whether CSE speakers’ self-reported frequency of use of three features – clause-initial is it, adverbial confirm, and SFP siaFootnote ² – aligns with previously reported and ongoing corpus-based findings. To recap, these findings have shown that younger speakers are more likely to use all three features, male speakers and Chinese speakers are more likely to use adverbial confirm, and male speakers and Malay speakers are more likely to use SFP sia. Based on these findings, we predict that younger CSE speakers – which we define as belonging to the Zoomer and Millennial cohorts – will report higher frequencies of use for all three features. We further expect that male speakers will report higher use for both adverbial confirm and sia, while Chinese speakers will report higher use of adverbial confirm and Malay speakers higher use of sia. By testing these predictions, we aim to assess whether self-reported usage matches corpus-based distributions, thereby contributing to broader discussions on the reliability of self-reports as indicators of language variation across social groups.

3.4.1 Self-Reported Frequency of Clause-Initial Is It Usage

A Spearman correlation test between continuous age and recoded frequency revealed a significant association between the age of the respondent and self-reported frequency of clause-initial is it usage (n = 214, r = −0.276, p < 0.001). In other words, as age increases, the less likely it is for individuals to report using clause-initial is it.

The data also display clear generational patterns. To recap, we grouped participants into generations based on their ages in 2024/2025: Zoomers (ages 12/13–27/28), Millennials (ages 28/29–43/44), Gen X (ages 44/45–59/60, and Boomers (ages 60/61–78/79). As seen in Figure 1, younger cohorts (Zoomers and Millennials) are more likely to report using clause-initial is it ‘A lot’, while older cohorts (Gen X and Boomers) are less likely to do so, creating a descending staircase pattern for ‘A lot’ across the age groups.

Figure 1

Self-reported frequency of clause-initial is it usage across generations

Figure 1 Long description

Percentages are displayed along the y-axis. On the x-axis are the four generational groups (Zoomers, Millennials, Gen X, and Boomers). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Zoomers: never (3), very rarely (4), occasionally (9), a lot (40). Millennials: never (3), very rarely (11), occasionally (23), a lot (46). Gen X: never (9), very rarely (11), occasionally (12), a lot (27). Boomers: never (2), very rarely (4), occasionally (7), a lot (3).

Figure 2 shows the self-reported frequency of clause-initial is it usage across gender. Inferential statistics (i.e., Mann–Whitney U-test) revealed a significant difference between females (n = 118) and males (n = 96), U = 4789, p < 0.05, indicating that females are more likely to report using clause-initial is it.

Figure 2

Self-reported frequency of clause-intial is it usage across genders

Figure 2 Long description

Percentages are displayed along the y-axis. On the x-axis are the two genders (Female and Male). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Female: never (11), very rarely (11), occasionally (23), a lot (73). Male: never (6), very rarely (19), occasionally (28), a lot (43).

Finally, the same test was conducted to assess whether there was an effect of ethnicity on the participants’ self-reported frequency of clause-initial is it use. The test did not find any significant difference between Non-Chinese (n = 47) and Chinese (n = 167) respondents, U = 3,684, p = 0.48, despite the descriptive variation shown in Figure 3.

Figure 3

Self-reported frequency of clause-intial is it usage across ethnic groups

Figure 3 Long description

Percentages are displayed along the y-axis. On the x-axis are the two ethnic groups (Chinese and non-Chinese). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Chinese: never (13), very rarely (22), occasionally (38), a lot (93). Non-Chinese: never (4), very rarely (7), occasionally (13), a lot (23).

3.4.2 Self-Reported Frequency of Adverbial Confirm Usage

Turning next to adverbial confirm, a Spearman correlation test between continuous age and recoded frequency revealed a significant association between the age of the respondent and self-reported frequency of use (n = 214, r = −0.18, p < 0.001). This suggests that as age increases, the likelihood of reporting using adverbial confirm decreases.

The data also shows clear generational patterns, as shown in Figure 4. Younger speakers (Zoomers and Millennials) are more likely to report using adverbial confirm ‘A lot’, while older cohorts (Gen X and Boomers) are less likely to do so, creating a descending staircase pattern for ‘A lot’ across the age groups.

Figure 4

Self-reported frequency of adverbial confirm usage across generations

Figure 4 Long description

Percentages are displayed along the y-axis. On the x-axis are the four generational groups (Zoomers, Millennials, Gen X, and Boomers). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Zoomers: never (6), very rarely (9), occasionally (14), a lot (27). Millennials: never (5), very rarely (14), occasionally (28), a lot (36). Gen X: never (10), very rarely (8), occasionally (23), a lot (18). Boomers: never (2), very rarely (7), occasionally (5), a lot (2).

Figure 5 shows the self-reported frequency of adverbial confirm usage across gender. A Mann–Whitney U-test revealed no significant difference between females (n = 118) and males (n = 96), U = 5,567.5, p = 0.82.

Figure 5

Self-reported frequency of adverbial confirm usage across genders

Figure 5 Long description

Percentages are displayed along the y-axis. On the x-axis are the two genders (Female and Male). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Female: never (13), very rarely (19), occasionally (40), a lot (46). Male: never (10), very rarely (19), occasionally (30), a lot (37).

Finally, the same test was conducted to assess whether there was an effect of ethnicity on the participants’ self-reported frequency of adverbial confirm use. The test did not find any significant difference between Non-Chinese (n = 47) and Chinese (n = 167) respondents, U = 3,429.5, p = 0.16, despite the descriptive variation shown in Figure 6.

Figure 6

Self-reported frequency of adverbial confirm usage across ethnic groups

Figure 6 Long description

Percentages are displayed along the y-axis. On the x-axis are the two ethnic groups (Chinese and non-Chinese). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Chinese: never (19), very rarely (29), occasionally (60), a lot (59). Non-Chinese: never (4), very rarely (9), occasionally (10), a lot (24).

3.4.3 Self-Reported Frequency of Sentence-Final Particle Sia

Turning finally to the SFP sia, a Spearman correlation test between continuous age and recoded frequency revealed a significant association between the age of the respondent and self-reported frequency of use (n = 214, r = −0.45, p < 0.001). In other words, as age increases, the likelihood of reporting using sia decreases.

The data also shows clear generational patterns, as shown in Figure 7. Younger speakers (Zoomers and Millennials) are more likely to report using sia ‘A lot’, while older cohorts (Gen X and Boomers) are less likely to do so, creating a descending staircase pattern for ‘A lot’ across the age groups.

Figure 7

Self-reported frequency of sia usage across generations

Figure 7 Long description

Percentages are displayed along the y-axis. On the x-axis are the four generational groups (Zoomers, Millennials, Gen X, and Boomers). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Zoomers: never (2), very rarely (6), occasionally (13), a lot (35). Millennials: never (9), very rarely (20), occasionally (17), a lot (37). Gen X: never (17), very rarely (16), occasionally (12), a lot (14). Boomers: never (3), very rarely (7), occasionally (5), a lot (1).

Figure 8 shows the frequency of sia usage across genders. A Mann–Whitney U-test revealed no significant difference between females (n = 118) and males (n = 96), U = 5,507, p = 0.71.

Figure 8

Self-reported frequency of sia usage across genders

Figure 8 Long description

Percentages are displayed along the y-axis. On the x-axis are the two genders (Female and Male). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Female: never (20), very rarely (25), occasionally (25), a lot (48). Male: never (11), very rarely (24), occasionally (22), a lot (39).

Finally, the same test was conducted to assess whether there was an effect of ethnicity on the participants’ self-reported frequency of use of sia. The test did not find any significant difference between Non-Chinese (n = 47) and Chinese (n = 167) respondents, U = 3,356, p = 0.3, despite the descriptive variation shown in Figure 9.

Figure 9

Self-reported frequency of sia usage across ethnic groups

Figure 9 Long description

Percentages are displayed along the y-axis. On the x-axis are the two ethnic groups (Chinese and non-Chinese). The stacked bars are colour-coded by level of use and include never (black), very rarely (medium grey), occasionally (light grey), and a lot (dark grey). Under the chart is a data table with the following counts. Chinese: never (24), very rarely (39), occasionally (39), a lot (64). Non-Chinese: never (6), very rarely (10), occasionally (8), a lot (23).

3.4.4 Perceived Age-Related Differences in CSE Use on Social Media

Given that younger speakers self-report greater use of innovative CSE features while older speakers self-report less, we also wanted to find out whether respondents perceive age-related differences in CSE usage. One survey item asked participants whether they perceived any age-related differences in the use of CSE on social media, and to describe the differences, if any. Based on the response, we categorised it as ‘Yes’ (i.e., the participant perceived age-related differences in the use of CSE on social media) or ‘No/Unsure’ (i.e., the participant did not perceive any age-related differences). Four participants did not respond to this question and were excluded (n = 210). A Mann–Whitney U-test revealed no significant difference between continuous age and ‘Yes’ vs. ‘No/Unsure’ responses, U = 4,585, p = 0.05. Nonetheless, Figure 10 indicates a clear generational trend: as their age increases, the participants’ responses are more likely to indicate ‘No/Unsure’.

Figure 10

Perception of age-related differences in CSE usage on social media across generations

Figure 10 Long description

Percentages are displayed along the y-axis. On the x-axis are the four generational groups (Zoomers, Millennials, Gen X, and Boomers). The stacked bars are colour-coded by response: No or Unsure in light grey and Yes in dark grey. Under the chart is a data table with the following counts. Zoomers: No or Unsure (21), Yes (35). Millennials: No or Unsure (29), Yes (51). Gen X: No or Unsure (31), Yes (27). Boomers: No or Unsure (10), Yes (6).

Several respondents also highlighted differences in the digital domains where Singaporeans use CSE. As one noted: ‘Facebook generally has an older population, TikTok probably has the youngest. X/twitter is probably a mix of teens + 21–50 since [it’s] kinda old and new at the same time’. These observations suggest that platform choice may correlate with age-based variation in CSE usage – a promising avenue for further investigation.

3.4.5 Intermediate Summary

Table 7 summarises the survey-based findings in this subsection and compares it to previous and ongoing corpus-based findings.

Table 7Corpus-based vs. survey-based findings on the relationship between CSE features and social variables (bold text = corpus and survey results align)

Table 7 Long description

The y-axis shows the corpus and survey findings for three CSE features (clause-initial is it, adverbial confirm, and sentence-final particle sia). The x-axis shows three social variables (age, gender, and ethnicity). Bold text marks cases where corpus and survey results align. The table shows how corpus and survey findings relate to each social variable:

For is it: corpus and survey findings show higher use among younger speakers and no evidence of an ethnicity effect; corpus findings show no evidence of a gender effect, while survey findings show higher use among females. For confirm: corpus and survey findings show higher use among younger speakers; corpus findings show higher use among males and Chinese, while survey findings show no evidence of a gender or ethnicity effect. For sia: corpus and survey findings show higher use among younger speakers; corpus findings show higher use among males and Malays, while survey findings show no evidence of a gender or ethnicity effect.

Across both the data from CoSEM and the survey, age consistently emerges as the most robust predictor of variation in the three CSE features examined herein. In every case, younger speakers show higher rates of use, regardless of whether the data comes from naturally occurring CMC or self-reported usage. In contrast, gender and ethnicity effects are largely confined to corpus data, as most of the self-reported data – apart from the gender effect on clause-initial is it – do not show significant differences.

4.1.1 Data Sources and Data Pre-Processing

The dataset for this study draws from two primary sources: CoSEM (Gonzales et al. Reference Gonzales, Hiramoto, Leimgruber and Lim2023a) and the social media platform X (formerly Twitter).Footnote ⁵ Data underwent preprocessing using both manual and automated tools to facilitate linguistic and sociolinguistic analysis. To examine the distribution and usage of jio, the corpus was queried using R, with regular expressions extracting all occurrences of jio, including related tokens such as bojio and invite. This process yielded 1,745 tokens, which were subsequently reviewed by two proficient speakers of CSE to ensure accuracy and exclude spurious matches (e.g., cases where jio appeared within unrelated strings like emoji only). While CoSEM provided a structured dataset of text messages, no comparable corpus existed for Twitter data at the time of this study. Moreover, the introduction of restrictive and costly API policies in 2023 precluded the construction of a specialised Twitter corpus akin to CoSEM. To approximate CoSEM’s scope and relevance, we leveraged data scraping techniques. Fortunately, in late 2021, a research team member had already conducted preliminary data extraction using Gonzales’ Twitter Corpus Suite (Gonzales Reference Gonzales2024, Reference Gonzales2025b), a Python-based framework. This yielded a dataset of 5,286 tokens of jio (including bojio and invite), distributed over 2010 to 2021, circumventing API-related constraints. As with CoSEM, two proficient speakers of CSE reviewed the extracted tokens to ensure accuracy. Unlike CoSEM, the Twitter dataset lacked sociolinguistic metadata, necessitating computational methods to infer demographic variables. Specifically, we adopted an approach informed by the social construction of age and gender, following Gonzales’ (Reference Gonzales2024) methodology for demographic inference on Twitter. Using Wang et al.’s (Reference Wang, Che, Guo, Liu and Liu2019) deep learning framework, we extracted age and gender information based on profile pictures, self-descriptions, and other textual markers. However, no existing computational tools achieved sufficient accuracy (>60%) for race and nationality inference, leading to their exclusion from the dataset. While this omission limits cross-corpus generalisability for these variables, it does not significantly affect the study’s primary focus on age as a sociolinguistic variable.

The final dataset, integrating WhatsApp (CoSEM) and Twitter (jio) data, comprised 7,031 tokens. These were coded for key linguistic factors, including presence/absence of morphological affixation (e.g., jio-ing, bo jio-ed) and negation (e.g., jio vs. bojio, invite vs. not invite). To streamline annotation, Natural Language Processing (NLP) tools such as nltk and spaCy (Bird, Klein, & Loper Reference Bird, Klein and Loper2009; Montani et al. Reference Montani, Honnibal, Boyd, Van Landeghem and Peters2023) automated the coding of English tokens, while non-English items underwent manual annotation by two proficient speakers of CSE.

Beyond linguistic features, sociolinguistic variables were computationally derived to assess their influence on jio usage. These included measures of sociolinguistic style, sentiment, intensity, and platform-based variation. Following Kiesling et al.’s (Reference Kiesling2018) discussion of sentiment as one operationalisable component of social and interactional meaning in computational work, sentiment was analysed using the sentimentR package (Rinker Reference Rinker2021), which assigns each text a polarity score representing the direction of affective evaluation: positive values indicate approval, enthusiasm, or endorsement, whereas negative values indicate disapproval, frustration, or complaint. In Kiesling et al.’s terms, this measure captures the affective dimension of stancetaking, that is, (positive, neutral, and negative) evaluation towards discourse entities, while abstracting away from the dialogic processes of alignment and investment that emerge across turns and participants. To operationalise this concept, we thus draw on polarity as a metric, which broadly captures what kind of stance is being taken. In contrast, sentiment intensity captures how strongly that stance is expressed, regardless of whether it is positive or negative. Intensity was operationalised as the absolute value of the polarity score, such that both highly positive and highly negative utterances receive high intensity values, while neutral or weakly evaluative utterances receive low values. For example, a strongly negative warning and a strongly positive invitation differ in polarity but may exhibit comparable intensity, reflecting similar levels of emotional investment. Distinguishing between polarity and intensity allows the analysis to test whether jio usage is more sensitive to evaluative direction (e.g., positive vs. negative stance) or to affective strength and engagement more generally.

Stylistic variation was examined along three dimensions. First, platform-based style was categorised based on dataset origin (text message vs. Twitter). Second, formality was inferred computationally following Grafmiller et al. (Reference Grafmiller, Szmrecsanyi, Röthlisberger and Heller2018), incorporating markers such as mean word length, nouniness, passive constructions, and noun-verb ratio. Third, informational versus interpersonal tendencies were measured through linguistic indicators, including subordinating conjunction frequency, stranded prepositions, and mean sentence length. These stylistic features were distilled into two overarching dimensions: formality and informationality/interpersonality. Additionally, POS tagging was applied, as the syntactic environment of jio – particularly adjacent POS categories – was hypothesised to condition its variability.

Both datasets underwent systematic collection and preprocessing to enable cross-platform analysis of jio usage. Prior research has shown that linguistic variation is shaped by platform-specific constraints, including character limits, multimodal affordances (e.g., images and videos), and interactional norms (Davies & Fuchs Reference Davies and Fuchs2015; Bohmann Reference Bohmann and Squires2016). By incorporating text message and Twitter data, this study offers a comprehensive account of the sociolinguistic factors shaping the distribution and use of jio, bojio, and invite in Singapore English.

4.1.2 Data Analysis

Using a curated dataset, we examined variation in the use of jio through a combination of frequency distribution analysis, where we get the proportional distribution of variants and compare them by factors of interest. To enhance generalisability, Bayesian logistic regression analysis was also used, following contemporary approaches in linguistics (Vasishth et al. Reference Nicenboim, Beckman, Fangfang and Kong2018; MacKenzie Reference MacKenzie2020). To assess the robustness of our model beyond the training data, we implemented an 80/20 training-test data split. The training set was fitted using a mixed-effects binomial model, appropriate for the binary syntactic variable under investigation – jio vs. invite. We estimated model parameters via Markov chain Monte Carlo sampling implemented in the brms package in R.

The model was trained on 5,625 observations, with 30,000 iterations per chain across four Markov chains. To mitigate the influence of early stage sampling bias, we applied a burn-in period of 15,000 iterations per chain and set a thinning parameter of 2. Weakly informative priors (normal distribution [0,5]) were assigned to both intercepts and slopes. Testing alternative priors (e.g., uniform, Cauchy [0,5]) revealed no substantial impact on the resulting posterior distributions. Model convergence was verified using R̂ values and Effective Sample Size (ESS) estimates, in line with Vehtari et al. (Reference Vehtari2021: 683). Specifically, we ensured that R̂ remained below 1.01 and ESS exceeded 400 (see ‘Results’ for specific values). We did not conduct a frequentist regression analysis, as Bayesian methods, despite their computational demands, provide more interpretable probabilistic inferences and allow for explicit statements regarding the absence of effects – something frequentist approaches do not permit (McElreath Reference McElreath2020; Levshina Reference Levshina2022).

The final model incorporated a range of predictors (Table 8), selected based on sociolinguistic research on CSE, previous studies, and ethnographic insights into variables influencing (jio) variation. Multi-level categorical variables (e.g., race) were coded using Weighted Helmert coding (Sonderegger Reference Sonderegger2023). Additionally, we specified random intercepts for the POS of adjacent words and for lexeme, though, following Levshina (Reference Levshina2016: 253), only lexemes and users with more than five tokens were retained as individual factor levels. Those with fewer than five occurrences were collapsed into an ‘other’ category to ensure sufficient data representation.

Table 8Factors tested in the model

Table 8 Long description

The table shows predictors used in the analysis, grouped into linguistic and socio stylistic categories. Each predictor is listed with its levels, variable type, and whether it is modelled as a main or random effect. Linguistic predictors are all categorical and include affix presence (absent or present) and negated context (negated or non-negated), both main effects, plus four contextual variables modelled as random effects: L1 POS and R1 POS (part of speech immediately left or right) and L1 Lexeme and R1 Lexeme (the neighbouring words). Socio-stylistic predictors include the following main effects: age (continuous), race (four levels, categorical), nationality (categorical), year (continuous), sentiment (continuous), intensity (continuous), and style (both interpersonal-informational and informal-formal, continuous) and user I D as a categorical random effect.

Following the Bayesian posterior summary, we assessed the predictors influencing (jio) variation using the probability of direction (pd) measure. A predictor was considered to have a meaningful effect on the dependent variable (e.g., likelihood of using jio) if its median estimate was substantively distant from zero or if its credibility intervals excluded zero. The probability of direction (pd) – defined as the proportion of posterior draws that share the sign of the median – serves in this study as a probabilistic measure of effect existence and certainty. A pd value approaching 1 suggests strong evidence for an effect in the indicated direction (positive or negative), whereas a pd near 0.5 implies that the effect may be negligible or non-existent (Makowski et al. Reference Makowski, Ben-Shachar, Chen and Lüdecke2019). For instance, if a predictor’s median estimate in our model is −1.2 with a pd of 0.95, this indicates that 95 per cent of the posterior distribution falls below zero, while only 5 per cent of values exceed zero (Levshina Reference Levshina2016). Given that jio was coded as the reference category, this would imply a 95 per cent probability that the predictor in question reduces the likelihood of selecting the jio variant.

4.4.1 General Introduction

The Bayesian model helps clarify which factors make speakers more or less likely to use (jio) rather than its alternative. Table 10 reports the model’s posterior estimates, with the median values indicating the direction and typical size of each effect (positive values increase the likelihood of (jio), negative values decrease it). Importantly, all predictors show a high probability of direction (pd > 0.95), meaning that the model is highly confident about the direction of each effect.

Table 10Posterior estimates from the Bayesian model showing how different predictors increase or decrease the likelihood of using jio rather than invite.

Table 10 Long description

Posterior estimates from the Bayesian model, divided into fixed and random effects. For both effect types, the median, standard deviation, 89% confidence interval, probability direction, Bayes factor, R-hat and effective sample size are given. The fixed effects include the intercept, affix presence, negation context, sentiment, intensity, age, gender, race, nationality, style (WhatsApp versus Twitter), year, style (interpersonal versus informational) and style (informal versus formal). The random effects are the user ID, as well as the part of speech and the lexeme in the left and right context of the search term.

Note: Multi-level categorical predictors are coded using Weighted Helmert contrasts, which compare each level to the average of the remaining levels; reference levels are shown in boldface. For each predictor, the table reports the median estimate (the most typical effect size), SD (uncertainty around that estimate), the 89 per cent highest density interval (HDI) indicating the most credible range of values, pd (probability that the effect is positive or negative), BF (Bayes factor, indicating strength of evidence for an effect), R̂ (a convergence diagnostic showing that the model has run properly), and ESS (effective sample size, indicating the reliability of the estimates).

A clear quantitative pattern emerges when comparing groups of predictors. Intralinguistic factors have larger effects on average (mean absolute median n = 1.75, SD = 1.11) than socio-stylistic factors (mean n = 1.15, SD = 1.02). This means that grammatical and structural conditions exert a stronger pull on speakers’ choices, while social and stylistic factors still matter but operate more subtly. In practical terms, structure sets the main limits on where (jio) is likely to appear, and social meaning fine-tunes usage within those limits.

4.4.2 Model Predictions and Accuracy

To check how reliable the model’s results are, we tested how well it predicts new data that were not used to fit the model. Specifically, we compared the model’s predictions to the actual choices speakers made in a held-out test dataset (n = 1,406). This comparison is summarised in a confusion matrix (Table 11), which simply shows how often the model correctly predicts each variant and where it makes mistakes. The model predicts speakers’ choices very well: it correctly identifies the observed variant in about 94 per cent of cases (Table 12). This performance is far better than a simple baseline strategy of always guessing the most frequent variant (invite), which would be correct only 59 per cent of the time. Statistical tests confirm that this improvement is not due to chance (p(Acc > baseline) < 0.001; McNemar’s p < 0.001), and the high Kappa value (0.87) indicates strong agreement between the model’s predictions and the observed data beyond what would be expected by chance. In practical terms, these results show that the model has learned systematic patterns in how speakers choose between jio and invite, instead of merely memorising the training data. Its strong performance on unseen data indicates that the predictors included in the model provide a robust account of variation and generalise well beyond the specific dataset analysed here, supporting the validity of the analysis.

Table 11Jio/invite model accuracy – Confusion matrix (n = 1,406 unobserved tokens)⁶

Table 11 Long description

In this matrix, the actual and predicted values for the variants jio and invite are compared: the actual reference values are the columns, and the predicted values are the rows.

4.4.3 Structural/Linguistic Constraints on (Jio)

The analysis reveals that affix presence exerts minimal influence on the selection between the Hokkien-derived variant jio and the English-derived invite, where the use of Hokkien-derived jio is slightly more preferred when (jio) is not affixed (Figure 11a). The lack of a stronger effect is somewhat unexpected, as morphological factors often play a strong role in constraining variation, particularly in contact settings where phonological or morphosyntactic integration can influence borrowing patterns (Szmrecsanyi Reference Szmrecsanyi2006; Gonzales Reference Gonzales2025a). However, in this case, the relative stability of jio use regardless of affix presence suggests that speakers may not perceive affixation as a meaningful constraint on variant selection.

Figure 11

Marginal effects of structural factors on jio usage. The y-axis shows the model-estimated probability that jio is used rather than invite, with 100 per cent indicating a categorical preference for jio. Panels display posterior predicted probabilities as a function of (a) affix presence, (b) negation, (c) sentiment polarity, and (d) sentiment intensity, holding all other predictors constant. Points and lines represent predicted values, and error bars and shaded bands indicate 95 per cent credible intervals.

a

Figure 11a Long description

Figure 11(a) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded areas indicating 95 percent credible intervals. The graph compares affix absence and presence; both show high predicted probabilities with overlapping uncertainty.

b

Figure 11b Long description

Figure 11(b) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded areas indicating 95 percent credible intervals. The graph compares negated and non-negated contexts; both have high predicted probabilities, with greater uncertainty for negated contexts.

c

Figure 11c Long description

Figure 11(c) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded areas indicating 95 percent credible intervals. The graph plots sentiment polarity; predicted probability decreases as sentiment becomes more positive, with wider uncertainty at higher values.

d

Figure 11d Long description

Figure 11(d) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded areas indicating 95 percent credible intervals. The graph plots sentiment intensity; predicted probability increases with intensity and approaches near-maximal levels, with wider uncertainty at low intensity.

Negation shows a nuanced effect on jio usage once model-based predictions are considered. While raw frequencies indicate that jio is more common than invite in negated contexts overall (Table 9), the marginal effects in Figure 11b reveal that non-negated contexts exhibit a higher and more stable predicted likelihood of jio usage, as indicated by the tighter uncertainty interval. In contrast, negated contexts display greater variability in predicted probabilities, suggesting that negation does not uniformly favour jio across interactional settings. This divergence reflects the difference between aggregate distributional tendencies and conditional effects estimated by the model. The wider uncertainty associated with negated contexts likely arises from heterogeneity within the negation category, which subsumes both relatively fixed expressions such as bojio and more compositional constructions (e.g., not jio and never jio). These subtypes differ in their degree of routinisation and pragmatic function, resulting in less consistent model predictions. By contrast, non-negated jio usage appears more regular and contextually stable, yielding higher and more consistent predicted probabilities. These results collectively suggest that negation does not categorically license jio, but rather introduces context-dependent variability, whereas non-negated environments support a more uniform preference for jio in the lexical alternation with invite.

The results for sentiment introduce a gradient social constraint that is implemented as a continuous structural predictor in the model, with jio being strongly associated with negative sentiment and invite emerging as the preferred variant in more positive contexts (Figure 11c). In this sense, sentiment occupies an intermediate position: it is social in its interpretation (indexing affective stance), but structural in its formal treatment as a scalar constraint on variant selection. This pattern is consistent with sociolinguistic work demonstrating that linguistic variants often carry affective or evaluative meanings, linking language choice to stance-taking and social positioning (Kiesling Reference Kiesling2018; Pratt Reference Pratt2023). The association between jio and negative sentiment aligns with its frequent use in complaint-driven discourse, further reinforcing its pragmatic function in expressing exclusion or missed opportunities. However, the increasing variability in jio use at the more positive end of the sentiment scale suggests that speakers are not entirely constrained by sentiment alone, indicating that affective stance sets probabilistic tendencies rather than categorical limits, and leaving room for context-dependent variation.

Finally, the effect of intensity, while statistically significant, appears more subtle after an examination of median values. The trend suggests that as emotional intensity increases, speakers are more likely to use jio (Figure 11d), though the absolute-value effect (media n = 0.25) is weaker compared to negation (media n = −1.31) and sentiment (media n = −3.36). This suggests that speakers may prioritise conveying their attitudes or evaluations over merely reflecting the intensity of their emotions in their linguistic choices. Therefore, the current finding that affective stance exerts a stronger influence than intensity is consistent with previous research in variationist sociolinguistics. Another possible explanation is that intensity interacts with other factors, such as discourse framing or interlocutor expectations, which were not directly tested in the current model. This suggests avenues for future research that could explore whether jio is particularly sensitive to pragmatic or interactional constraints beyond those captured here.

Collectively, these findings reinforce the idea that linguistic variation is shaped by structural constraints, with jio serving not merely as an interchangeable alternative to invite but as a variant whose use is systematically conditioned by negation, sentiment/affect, and to a lesser extent, intensity. While some results, such as the role of negation, were expected based on previous research, the minimal influence of affix presence and the relatively weak effect of intensity suggest that additional factors, such as discourse context and pragmatic function or even socio-stylistic factors, may be at play in shaping jio variation.

4.4.4 Socio-Stylistic Constraints on (Jio)

Figure 12 presents the effects of key socio-demographic factors – age, gender, race, and year– on the variation between jio and invite. The results for age (Figure 12a) reveal a gradual decline in jio use as speaker age increases. Younger speakers exhibit the highest rates of jio usage, while older speakers show a clear shift towards invite. This pattern suggests that jio is more robustly maintained among younger individuals but becomes less frequent with age, aligning with research on age-based variation, where younger speakers tend to favour innovative or localised forms while older speakers conform more closely to dominant linguistic norms (Meyerhoff Reference Meyerhoff2018).

Figure 12

Marginal effects of socio-stylistic factors on jio usage. The y-axis shows the model-estimated probability that jio is used rather than invite, with 100 per cent indicating a categorical preference for jio. Panels display posterior predicted probabilities as a function of (a) age, (b) gender, (c) race (simplified categories), and (d) year of production, holding all other predictors constant. Lines and points represent predicted values, while error bars and shaded bands indicate 95 per cent credible intervals.

a

Figure 12a Long description

Figure 12(a) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded regions indicating 95 percent credible intervals. The graph plots age, showing a gradual decline in predicted probability of jio usage with increasing age.

b

Figure 12b Long description

Figure 12(b) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded regions indicating 95 percent credible intervals. The graph compares gender categories; both female and male groups show high predicted probabilities with overlapping uncertainty.

c

Figure 12c Long description

Figure 12(c) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded regions indicating 95 percent credible intervals. The graph compares simplified race categories; predicted probabilities vary across groups, with wider uncertainty in non-Chinese groups.

d

Figure 12d Long description

Figure 12(d) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability, with error bars or shaded regions indicating 95 percent credible intervals. The graph plots year, showing a gradual decline in predicted probability of jio usage over time, with increasing uncertainty in later years.

From a sociolinguistic perspective, this age pattern can be interpreted in two theoretically distinct ways, depending on how community norms are assumed to behave over time. Under an age-grading account, individuals change their linguistic behaviour across the life course, but community-level norms remain stable: each new generation would be expected to use jio at similarly high rates when young and to reduce usage as they age. Under a lifespan effect account, by contrast, individuals shift their usage in response to changing community norms, such that both individuals and the community distribution change over time. In this scenario, younger cohorts could use jio less than earlier cohorts did at the same age, signalling community-level change rather than stable age stratification. The present data cannot adjudicate conclusively between these interpretations, as doing so would require panel or real-time evidence, but the results nonetheless show that age is a strong and systematic predictor of jio variation, with important implications for its long-term vitality in speakers’ repertoires.

If we examine the change in real time instead of apparent time (Figure 12d), we also find that jio usage shows a steady decline over time. While jio was more evenly distributed with invite in the early 2010s, its use has steadily decreased, particularly after 2017. This downwards trajectory suggests an overall shift towards invite, raising the possibility that jio is undergoing recession in favour of its English-derived alternative. This trend is further corroborated in Figure 13, which provides a year-by-year breakdown of jio and invite use. In earlier years, jio and invite were used at comparable rates, but by 2021, jio had nearly disappeared from the dataset. The accelerated decline of jio in recent years suggests that while it may have once been a stable variant, it is now becoming increasingly restricted in use. This pattern is in line with broader studies of lexical shift in contact varieties, where localised forms may recede due to ongoing globalisation and the increasing dominance of English.

Figure 13

Change in (jio) over time

Figure 13 Long description

Percentages are displayed along the y-axis. On the x-axis are the twelve years, from 2010 to 2021. The stacked bars are colour-coded by value with jio (light grey) and invite (dark grey). Under the chart is a data table with the following counts. 2010: jio (20), invite (24). 2011: jio (251), invite (252). 2012: jio (378), invite (299). 2013: jio (246), invite (268). 2014: jio (357), invite (389). 2015: jio (404), invite (512). 2016: jio (398), invite (517). 2017: jio (386), invite (522). 2018: jio (171), invite (425). 2019: jio (105), invite (423). 2020: jio (113), invite (278). 2021: jio (12), invite (281).

Turning to gender (Figure 12b), the results show no meaningful difference in jio use between male and female speakers, as indicated by overlapping confidence intervals. This finding contrasts with much of the variationist literature, which frequently identifies gender-based differentiation, particularly when a variant carries strong social or indexical associations (Eckert Reference Eckert1989; Gonzales et al. Reference Gonzales, Chan, Zhang and Ng2026). The absence of a significant gender effect suggests that jio is not strongly linked to masculinity or femininity and may function as a broadly available linguistic resource for expressing other identities, which is consistent with research on gender and language contact (Meyerhoff & Birchfield Reference Meyerhoff, Birchfield, Darquennes, Salmons and Vandenbussche2019).

Ethnolinguistic background (Figure 12c) emerges as a more salient predictor, with jio use being highest among Chinese speakers and significantly lower among other ethnic groups. While Eurasian, Indian, and Malay speakers display a more balanced distribution between jio and invite, non-Singaporeans overwhelmingly favour invite. Although the presence of jio among non-Chinese speakers suggests some degree of diffusion, its lower frequency among these groups reinforces its continued association with Chinese linguistic heritage.

In addition to socio-demographic constraints, stylistic factors also play a significant role in shaping jio variation. Figure 14 examines the influence of style, revealing distinct patterns across the informational–interpersonal dimension, which is analytically independent from the formal–informal dimension. To clarify this distinction, two stylistic dimensions, formal–informal and informational–interpersonal, were derived separately based on the primary linguistic parameters identified by Grafmiller et al. (Reference Grafmiller, Szmrecsanyi, Röthlisberger and Heller2018) for English, allowing these dimensions to vary independently rather than being treated as opposites.

Figure 14

Marginal effects of sociolinguistic style on jio usage. The y-axis shows the model-estimated probability that jio is used rather than invite, with 100 per cent indicating a categorical preference for jio. Panels display posterior predicted probabilities across levels of (a) informational style and (b) formal style, holding all other predictors constant. Solid lines represent predicted values, and shaded bands indicate 95 per cent credible intervals.

a

Figure 14a Long description

Figure 14(a) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability; shaded regions indicate 95 percent credible intervals. The graph plots informational style, showing an increase in predicted jio usage as informational style values rise, approaching near-maximal probabilities at higher values.

b

Figure 14b Long description

Figure 14(b) shows model-predicted probabilities of using jio rather than invite. The y-axis ranges from 0 percent to 100 percent and represents predicted probability; shaded regions indicate 95 percent credible intervals. The graph plots formal style, showing a decrease in predicted jio usage as formality increases, with wider uncertainty at higher values.

In highly informational styles (Figure 14a), jio is used at higher rates, with usage increasing as the level of informational content rises. Importantly, this is not a contrast between formal and informal styles: texts can be highly informational while also being formal, and vice versa, as established in multidimensional models of register variation. This suggests that jio is preferred in more communicative settings, where speakers engage in high-information exchanges that prioritise efficient transmission of content rather than interpersonal or presentational concerns. This pattern aligns with Bell’s (Reference Bell1984) audience design framework, which posits that speakers modulate linguistic choice in response to communicative context.

By contrast, formal discourse (Figure 14b) exhibits the opposite trend, with jio usage declining as formality increases. At the highest levels of formality, invite is overwhelmingly preferred, suggesting that it is perceived as the more standard or socially neutral variant. This result is expected given long-standing findings that formal speech styles tend to favour standard or English-derived forms over localised variants (Labov Reference Labov1972). The association between jio and informality suggests that it may be socially marked in a way that invite is not, further reinforcing the idea that jio is undergoing stylistic restriction.

5.2.1 Kena Passives in CSE

The use of the Malay-derived kena passive in CSE was first reported by Bao and Wee (Reference Bao and Wee1999), who also asserted that it was more frequently used than the give passive. Similar to the standard be and get passives, in the CSE kena passive construction, the theme or patient Noun Phrase (NP) is promoted to subject, while the agent NP is optionally expressed as an oblique argument using a by-phrase. The kena passive in both Malay and CSE also usually involves an adversative reading. Examples (16) and (17) are taken from Bao and Wee (Reference Bao and Wee1999), demonstrating this documented use of the kena passive in CSE.

1. (16) The thief kena caught by the police

   ‘The thief was caught by the police.’

   (Bao & Wee Reference Bao and Wee1999: 2)

1. (17) The durian kena eaten already

   ‘The durian got eaten already.’

   (Bao & Wee Reference Bao and Wee1999: 2)

Examples (18), (19), (20), and (21) are taken from CoSEM, and they show how the kena passive is still attested in present-day CSE.

1. (18) Cried bc kena scolded by ms <name>

   ‘He cried because he got scolded by Ms <name>.’

   <COSEM:21DC45-1514-19SGCHFST-2018>

1. (19) More and more ppl kena quarantined

   ‘More and more people are being quarantined.’

   <COSEM:21DC09-9133-42SGCHFGH-2020>

1. (20) I kena cut from the team

   ‘I was cut from the team.’

   <COSEM:19CF10-5176-23SGCHM0A-2018>

1. (21) Tmr my dinner kena cancelled so I can MAHJONG

   ‘My dinner tomorrow got cancelled, so I can play mahjong.’

   <COSEM:18CF42-3081-23SGMAF00-2017>

5.2.2 Kena in Malay

Kena in Malay may function as a passive marker,Footnote ⁸ as seen in (22)–(24). In the Malay kena passive construction, the agent can be omitted as shown in (22) and (23), and may also be expressed as an oblique argument using oleh as shown in (24).

1. (22)

   Aminah	kena	tampar
   Aminah	pass	slap
   ‘Aminah got slapped.’
   (Nomoto & Kartini Reference Nomoto and Kartini2012: 364)

1. (23)

   Dompet	aku	kena	curi	semalam
   purse	1sg.poss	pass	steal	yesterday
   ‘My purse got stolen yesterday.’
   (Nomoto & Kartini Reference Nomoto and Kartini2012: 364)

1. (24)

   Pencuri	itu	kena	tangkap	oleh	polis	semalam
   thief	that	pass	catch	by	police	yesterday
   ‘That thief was caught by the police yesterday.’
   (Nomoto & Kartini Reference Nomoto and Kartini2012: 367)

Note that kena possesses other functions in Malay. First, as shown in (25) and (26), kena can be used to express a deontic reading similar to ‘have to’ in English (see Nomoto and Kartini Reference Nomoto and Kartini2012 for more discussion on potential polysemy between these functions of kena in Malay).

1. (25)

   Nak	kena	bongkar	kat	rumah
   will	deo	ransack	at	house
   ‘(I) will have to ransack the house.’
   <COSEM:18MF03−19411−23SGMAF00−2017>

1. (26)

   saya	kena	bikin	kuih
   1sg	deo	make	snack
   ‘I have to make snacks.’
   <COSEM:18MF03−14392−23SGMAF00−2017>

Kena can also be used as a verb to mean ‘to strike or be struck by’ or ‘to receive or be subjected to’. As a verb, kena is followed by an NP object, as seen in (27) and (28).

1. (27)

   kereta-nya	kena	tiang	eletrik
   car-poss	hit	pole	electric
   ‘His car hit the electric pole.’
   (Dewan Bahasa dan Pustaka Reference dan Pustaka2010: 735)

1. (28)

   kena	penyakit	demam	kepialu
   receive	illness	fever	typhoid
   ‘(He) contracted typhoid fever.’
   (Dewan Bahasa dan Pustaka Reference dan Pustaka2010: 735)

Lastly, kena also serves as an adjective that roughly means ‘suitable, right’, as seen in (29) and (30).

1. (29)

   Warna	baju	Aminah	kena	dengan	warna	selendang-nya
   colour	clothes	Aminah	suitable	with	colour	shawl-poss
   ‘The colour of Aminah’s clothes matches that of her shawl.’
   (Nomoto & Kartini Reference Nomoto and Kartini2012: 366)

1. (30)

   Dia	menjadi	berang	kalau	ada	sesuatu	yang	tidak	kena
   3sg	become	furious	if	exist	something	that	not	right
   di	hati-nya
   at	heart-poss
   ‘She becomes furious if there is something that is not right in her heart.’
   (Nomoto & Kartini Reference Nomoto and Kartini2012: 366)

Of these functions, this section will primarily focus on the use of kena as an adversative passive marker (Bao & Wee Reference Bao and Wee1999), which is well-attested in the CoSEM data set. However, it is pertinent to note that kena as a verb – meaning ‘to be subjected to’ – can also be found in CSE, as seen in the CoSEM examples (31) and (32). In these examples, CSE kena co-occurs with an NP object. This is in contrast with the passive marker kena, in which kena must precede the main verb.

1. (31) wah I kena sore throat

   ‘Damn, I got a sore throat.’

   <COSEM:19SF01-47194-20SGSIFNA-2018>

1. (32) I hope u kana car accident

   ‘I hope you get caught in a car accident.’

   <COSEM:18CF52-15075-23SGCHM00-2017>

It is also worth noting that the ‘have to’ or deontic function of kena can be found, albeit to a smaller extent, in some CSE utterances in CoSEM, as shown in (33) and (34).

1. (33) Yeah kena wait for 2nd round

   ‘Yeah (you) have to wait for the second round.’

   <COSEM:18MF03-11905-24SGMAF00-2017>

1. (34) all kena scan IC

   ‘(They) all had to scan their identity cards.’

   <COSEM:19MM02-206-42SGMAM00-2019>

5.2.3 The Tio Passive in CSE

The tio passive construction was not mentioned in Bao and Wee’s (Reference Bao and Wee1999) study, and appears to be a relatively new feature of CSE. As such, it has not, to our knowledge, been studied in contemporary CSE research. Interestingly, the online dictionary Wiktionary does list tio as a lexical item of CSE, describing one of its functions as that of a passive marker (Wiktionary 2025). In this section, we present a more thorough description of the tio passive in CSE based on our findings from CoSEM.

In the corpus, tio is used similarly to be and get passives, as shown in examples (35) and (36). On the surface, tio is used in place of the standardised English passive markers, be and get.

1. (35) Can not say later i tio scolded again hahaha

   ‘I can’t say, otherwise I’ll get scolded again.’

   <COSEM:17CF24-5711-26SGCHM00-2015>

1. (36) Wah today station games tio squeezed by guys

   ‘Today at the station games, I got squeezed by guys.’

   <COSEM:17CF34-61588-21SGCHF00-2013>

Notably, example (36) shows the tio passive co-occurring with the optional by-phrase, by guys, which contains the agent argument. This is interesting since tioh passive constructions in Singapore Hokkien and Teochew, which we argue in what follows is the source of the tio passive construction in CSE, may not contain an oblique argument identifying the agent.

5.2.4 Tioh in Singapore Southern Min

There are three major varieties of Southern Min spoken in Singapore: Hokkien, Teochew, and Hainanese. Hokkien and Teochew can be reasonably identified as the source of CSE tio, as both varieties contain the phonologically similar passive marker /tioʔ/ (or /tieʔ/ in certain Teochew varieties). In contrast, Hainanese does not contain a phonologically similar passive marker,Footnote ⁹ which suggests that Hainanese is unlikely to be the source of CSE tio. Demographically, Hokkien and Teochew speakers also greatly outnumber Hainanese speakers in Singapore (Department of Statistics Singapore 2020), which further supports the idea that Hainanese has had less of an influence on CSE than Hokkien and Teochew. For these reasons, we will base the following discussion on examples drawn from Singapore Hokkien and Singapore Teochew. As Hokkien and Teochew are closely related varieties, we have tried to pick examples that are intelligible in both varieties. We will henceforth refer to these two varieties as ‘Southern Min’ as a shorthand. The examples discussed here were obtained via consultation with two native speakers of Singapore Hokkien and two native speakers of Singapore Teochew.

In a tioh passive construction such as (37), the passive marker tioh precedes a transitive verb like ‘scold’, and the theme argument appears in the subject position, which precedes tioh. However, unlike the kena passive construction in Malay, the tio passive construction in CSE tio, or the bei passive construction in Mandarin, tioh passives do not allow the expression of an oblique agent. To express the agent in Singapore Southern Min, another passive construction using a marker derived from the verb ‘give’ must be used instead (cf. Bao and Wee’s (Reference Bao and Wee1999) discussion of Cantonese and Hokkien ‘give’ passive constructions).

1. (37)

   ua	tioʔ	me /	phaʔ
   1sg	tioh	scold	beat
   ‘I got scolded/beaten.’

1. (38)

   *ua	tioʔ	i	me /	phaʔ
   1sg	tioh	3sg	scold	beat
   Intended: ‘I got scolded/beaten by him/her.’

1. (39)

   ua	puŋ / khoʔ	i	me /	phaʔ
   1sg	give	3sg	scold	beat
   ‘I was scolded/beaten by him/her.’ (Teochew)

1. (40)

   ua	ho	i	me /	phaʔ
   1sg	give	3sg	scold	beat
   ‘I was scolded/beaten by him/her.’ (Hokkien)

For completeness, it is worth pointing out that there are dialectal differences in whether and how tioh is used as a passive marker in Southern Min more generally. Hing (Reference Hing2019), citing Zhou (Reference Zhou2006: 683), reports that tioh passives are absent in mainland China varieties. However, tioh passives appear to exist in Taiwanese Southern Min (e.g., Lien Reference Lien2001: 191–192; also cited in Ansaldo Reference Ansaldo and Lim2004); Hing (Reference Hing2019, Reference Hing2023) also reports a tioh passive in Penang Hokkien. In both Taiwanese Min and Penang Hokkien, the agent may appear between tioh and the verb, unlike the case in Singapore Southern Min. In other words, the equivalent of example (38) is acceptable in these varieties.

In addition, like Malay kena, Southern Min tioh has other uses: as a deontic modal meaning ‘must’, a lexical verb that might be translated as ‘obtain, get, win’, an adjective meaning ‘correct’, and as an aspectual marker suffixing to verbs, among others (see Lien Reference Lien2001; Hing Reference Hing2019 and references therein). For scope reasons, we will review in detail only the deontic and lexical verb uses, which, as we will see, might explain the emergence of the tio passive in CSE. In the deontic modal use, Singapore Southern Min tioh is often paired with another modal marker ai (in Hokkien, as shown in (41)) (Lien Reference Lien2001; Chappell Reference Chappell, Vittrant and Watkins2019; Hing Reference Hing2019: 152) or ãĩ (Teochew), which variously has desire (‘want, like’), deontic (‘must’), and/or future uses (we tentatively gloss this morpheme as deontic here). However, as (42) shows, ai / ãĩ is not necessary for the modal interpretation.

1. (41)

   ua	tioʔ	ai	khun.
   1sg	deo	deo	sleep
   ‘I must sleep.’ (Hokkien)

1. (42)

   tioʔ!

   deo

   ‘[S/he] must!’ (in response to a question like i tioʔ ai khun? ‘Must s/he sleep?’)

Second, tioh can be used as a lexical verb, with a meaning along the lines of ‘obtain’, ‘get’, and ‘win’. This is a generally well-documented usage of tioh in Southern Min, even outside of Singapore (e.g., Lien Reference Lien2001; Ansaldo Reference Ansaldo and Lim2004; Hing Reference Hing2019). In Singapore Southern Min, the verbal use of tioh is prominently associated with winning games of chance (43) and contracting diseases (44). For thoroughness, it should be pointed out that the same verbal use can also be found in CoSEM.

1. (43)

   tioʔ	bepio
   win	lottery
   ‘to win [a prize] in the lottery’

1. (44)

   i	bo	tioʔ	COVID.
   3sg	neg.perf	get	COVID
   ‘S/he did not get (catch) COVID.’

5.2.5 Summary

This section has compared the use of kena in CSE to the use of kena in Malay, and the use of tio in CSE to the use of tioh in Singapore Southern Min (Hokkien and Teochew). While Malay kena can be used as a passive marker, a marker of deontic modality, and a lexical item to mean ‘to match’ or ‘suitable’, CSE kena may only be used as a passive marker. While tioh in Singapore Southern Min can be used as a passive marker, a marker of deontic modality, and a lexical verb meaning ‘to obtain, get, win’, only the passive marker and lexical verb uses are found with CSE tio.

5.3.1 Identifying Passive Constructions in the Corpus

We first wrote a Python script to extract all lines containing kena, tio, give, be, and get, and their various inflected forms (e.g., was and were) and alternative orthographic forms (e.g., kana for kena). Each message was then slightly preprocessed to remove emojis and standardise characters (e.g., quotation marks), which are known to affect the accuracy of statistical parsing. The preprocessed messages were then parsed with a version of the Stanford Parser (Klein & Manning Reference Klein and Manning2003 and subsequent work) customised for CSE – this was done by training the Stanford Parser on CSE parses manually validated by Bao, Huang, and colleagues in another CSE corpus project (see Lin et al. Reference Lin, Han and Hing2023; Huang et al. Reference Huang, Lin and Han2025). The parser parsed and generated phrase structure trees for all preprocessed lines except those that contain forty words or more, as these tend to create memory issues and cause the parser to crash. We note that lines of this length or greater are very rare, making up 0.3 per cent of all lines.

We then wrote another Python script to check each generated syntax tree. This was particularly important for the be and get tokens, since not all instances of be and get appeared with passives; be, of course, can appear with the present participle (a hypothetical standard English example: They will be leaving). The Python script extracted only tokens of be and get that co-occur with a VP headed by what the parser tagged as a past participle.

Within this initial sample, be and get were by far the most frequent markers. Their extremely high frequencies made a full annotation of them infeasible. As such, we extracted a 10 per cent random sample of the be and get parses. A linguistics-trained research assistant checked these samples and identified tokens that were genuine verbal passives. This was necessary because the parser is not perfectly accurate and might have made mistakes with tagging or parsing. A second reason is that the parser is also not always able to differentiate verbal passives from adjectival passives, as in tired in They are/got tired, which we consider to be adjectives rather than verbs. We excluded tokens that were clearly adjectival, where we can felicitously insert a degree adverb like ‘very’ immediately before the word (e.g., They are/got very tired).

The Python script also extracted all tokens of give co-occurring with a VP or clause as potential give passive tokens. Manual inspection, however, showed these to be parser errors; we did not find any tokens of genuine give passives.

Because the set of kena and tio tokens extracted was relatively small (kena: 673 tokens; tio: 94 tokens), we did not do any random sampling. However, we note that kena and tio, not unlike be, get, and give, also have non-passive uses, in which case they do not occur with verbs. As mentioned earlier, CSE kena can be a verb, as in I kena COVID ‘I got COVID’. To exclude these non-passive uses, the research assistant reviewed the full set of tokens, sorting them based on whether they appeared with verbs, NPs, intransitive, or other frames.

Table 13 lists the count of tokens for be, get, kena, and tio passives with unambiguous verb frames in our final sample. With this, we ran a multinomial regression with age, gender, and race as predictors, and choice of passive marker as the dependent variable, setting be-passive as the reference ‘standard’ category. For thoroughness, we created a reweighted data set that synthetically increased the be and get tokens ten times, so that the number of be and get tokens in this reweighted sample more accurately represents the actual distribution of passive markers. For visualisation purposes, we used the reweighted sample to generate all figures. We also re-ran the regression model on this reweighted sample and obtained qualitatively similar results.

5.3.2 Coding for Sentiment

We coded this dataset for sentiment by referring to only the action or state described by the verb and ignoring all other contextual information. For example, we would consider all tokens of paid as positive, even in contexts like getting paid peanuts or didn’t get paid, which most might consider as describing negative outcomes (note that these are hypothetical examples to illustrate the point). We do so because we judge that kena can be used for a verb with negative sentiment, even when the broader context has a positive sentiment, as illustrated by disqualified in the hypothetical sentence My favourite team didn’t kena disqualified ‘My favourite team didn’t get disqualified’.

All tokens identified were coded as ‘negative’, ‘positive’, or ‘neutral’ by the same linguistics-trained student research assistant and one of the authors (either Mohamed Hafiz or Nick Huang); some examples for the be-passive are shown in (45). We chose to manually annotate for sentiment here, instead of using software (cf. Section 3 on jio). This is because we were interested in the sentiment expressed through the verb itself, and a number of verbs in our sample are of Malay or Chinese origin (e.g., tekan ‘to treat poorly, to bully’, from Malay; buah ‘to scrape’, from Southern Min) or are English words that have acquired novel uses (e.g., arrow ‘to nominate [someone] for an undesirable task’). To ensure accurate annotation, we decided against using software, which is usually targeted at standard English varieties. In order to keep the process of manual annotation straightforward, we decided on a three-way categorical scheme, instead of a finer-grained one. Results reported in Section 5.4 are based on the research assistant’s annotations, but we note that there was a relatively high degree of inter-annotator agreement (Cohen’s kappa of .77).

1. 1. (a)

      Pls slowly do carefully and check if we miss out anything or if it can be improved. (Positive because of improve)

      ‘Please do [the survey] slowly and carefully and check if we missed out anything or if it can be improved.’

      <COSEM:17CM05−8172−23SGCHM00−2016>

   2. (b)

      i am being abused now (Negative because of abuse)

      <COSEM:17CF12−65931−21SGCHM00−2016>

   3. (c)

      I hope the stuff im writing can be absorbed in (Neutral because of absorb)

      <COSEM:17CF34−54907−21SGCHF00−2015>

5.4.1 Relative Frequencies of Passive Constructions

As shown in Table 13, no give passives were found in CoSEM, suggesting that the construction might have fallen out of use for younger CSE speakers. This supports Bao and Wee’s (Reference Bao and Wee1999) prediction. It is also not unexpected, given Bao and Wee’s original observation that the give passive was not as robust compared to the kena passive at the time of their study. Tio passives are the least frequent in CoSEM, which supports our intuition that the tio passive has only recently emerged in CSE.

Table 14 presents the estimates for predictors influencing the likelihood of using a particular passive construction relative to others, with the be-passive serving as the reference ‘standard’ category. We discuss each predictor in turn.

Table 14Estimates of selected predictors from Bayesian models

Table 14 Long description

The y-axis shows the intercept and six predictors with different reference levels for race indicated in bold: age, gender (reference: male), race (reference: Other/Eurasian), race (reference: Indian/Eurasian/Other), race (reference: non-Chinese), and year. The x-axis shows the median estimates and probabilities of direction (p d) of three passive C S E constructions (get, kena, and tio passives). The table shows: 1. negative coefficients with high p d values for age for all three passives, indicating that the use of all passives decreases with age. 2. low p d values (less than 90 percent) for gender for all three passives, indicating limited effects of gender. 3. low p d values (less than 90 percent) for race for all three passives, indicating limited effects of race. 4. negative coefficients with high pd values for year for all three passives, indicating that the use of all passives decreases over time.

Age. Table 14 indicates that there is strong evidence that as age increases, the likelihood of using get, kena, or tio decreases. First, this pattern aligns with trends reported for American English, where the get passive has been increasing at the expense of the be passive (Sneller & Fisher Reference Sneller and Fisher2015). This parallel may provide support for the influence of American English on the English of younger Singaporeans, although the data cannot establish a causal link. Second, the results indicate that the substrate-influenced passives are relatively more popular among younger speakers than older ones. For the kena passive, the finding that younger speakers continue to use it aligns with Bao and Wee’s (Reference Bao and Wee1999) prediction that the kena passive will remain robust in CSE. The tio passive shows a similar trend, which is expected given its relatively recent emergence, likely introduced by younger speakers.

The distribution of passive constructions across age groups, as shown in Figure 15, also reveals some generational patterns. The be passive is the most frequently used across all age groups, with older speakers notably displaying the highest usage, followed by younger speakers, and next, middle-aged speakers. The get passive shows a gradual decline in usage with age. The kena passive is most frequent among middle-aged speakers but is considerably lower among younger and older speakers. Finally, the tio passive appears only marginally in the data, with extremely low frequencies found among younger and middle-aged speakers. The patterns suggest that the use of standardised English be and get passives dominate the use of the substrate-derived kena and tio passives across different generations of CSE speakers.

Figure 15

Percentage of a given passive marker relative to all passives used by an age group

Figure 15 Long description

In all four graphs, the y-axis represents percentages, and the x-axis represents three age groups: younger (18 to 30 years of age), middle-aged (30 to 50 years of age), and older (over 50). The bars in the first graph are for the be-passive and stand at 68.8 for the younger group, 52.6 for the middle-aged group, and 96.2 for the older group. The bars in the second graph are for the get-passive and stand at 26.0 for the younger group, 22.6 for the middle-aged group, and 2.1 for the older group. The bars in the third graph are for the kena-passive and stand at 4.6 for the younger group, 23.3 for the middle-aged group, and 1.7 for the older group. The bars in the fourth graph are for the tio-passive and stand at 0.6 for the younger group, 1.5 for the middle-aged group, and 0 for the older group.

Although our Bayesian results indicate that kena and tio have declined over time relative to be, it is not the case that be has been growing steadily at the expense of kena or tio. Rather, as Figure 16 shows, there is much fluctuation in the use of be. In contrast, kena is used at a low, but comparatively stable, rate. We take this relative stability to support Bao and Wee’s (Reference Bao and Wee1999) prediction that the kena passive will continue to be used in CSE, which is especially notable given that their hypothesis was proposed over two decades ago. As for tio, Figure 16 shows that it is attested as early as 2013, but occurs even less frequently. While tio remains far less established than kena, our current data do not yet allow us to determine whether its usage will increase or decline.

Figure 16

Percentage of a given passive marker relative to all passives in messages from a particular year

Figure 16 Long description

In all four graphs, the y-axis represents percentages and the x-axis represents the ten years from 2012 to 2021. The bars in the first graph are for the be-passive. Overall, be-passives are the most frequent passive construction.

The bars in the second graph are for the get-passive. Get-passives are generally the second-most frequent construction. The bars in the third graph are for the kena-passive. Kena-passives are the third-most frequent, around 10 percent or less, but are attested in all years. The bars in the fourth graph are for the tio-passive. Tio-passives are the least frequent and not attested in certain years.

Gender. Table 14 provides limited evidence for an effect of gender on the likelihood to use kena or tio (pd < 0.9). This is unsurprising, as these passives are relatively ‘neutral’ linguistic features compared to other features in CSE. For example, the SFP sia in CSE has a vulgar origin and may thus show gender-based distributional differences (Hiramoto et al. Reference Hiramoto, Gonzales, Leimgruber, Lim, Choo, Ngefac, Wolf and Hoffmann2022; Hafiz et al. Reference Hafiz, Hiramoto, Leimgruber, Gonzales and Lim2025: 202). The substrate-influenced passives kena and tio lack such connotations, which may explain their relatively balanced usage across genders, as shown in Figure 17.

Figure 17

Percentage of a given passive marker relative to all passives used by a particular gender

Figure 17 Long description

In all four graphs, the y-axis represents percentages and the x-axis represents two genders: Female and Male. The bars in the first graph are for the be-passive and stand at 72.0 for Female and 69.4 for Male. The bars in the second graph are for the get-passive and stand at 24.0 for Female and 24.0 for Male. The bars in the third graph are for the kena-passive and stand at 3.8 for Female and 5.7 for Male. The bars in the fourth graph are for the tio-passive and stand at 0.3 for Female and 0.9 for Male.

Race. Similarly, Table 14 shows that there is limited evidence of a racial effect on the use of kena (pd < 0.9). Figure 18 shows that Malay speakers use kena more frequently than the other ethnic groups, which is unsurprising given kena’s Malay origin. Notably, Chinese and Indian speakers also use kena, indicating that it has become a shared feature across ethnic groups. For tio, the data is too sparse to draw firm conclusions, though it is unsurprising that Chinese participants use it most frequently. Interestingly, Indian participants use tio about as often as Chinese participants, while Malay participants use it least – perhaps because they are more familiar with kena, the Malay-derived passive in CSE.

Figure 18

Percentage of a given passive marker relative to all passives used by members of an ethnic group

Figure 18 Long description

In all four graphs, the y-axis represents percentages and the x-axis represents five ethnic groups (Chinese, Eurasian, Indian, Malay, and Other). The bars in the first graph are for the be-passive and have the following values: Chinese 72.3, Eurasian 42.9, Indian 64.8, Malay 67.0, and Other 66.0. The bars in the second graph are for the get-passive and have the following values: Chinese 22.0, Eurasian 57.1, Indian 32.4, Malay 23.3, and Other 33.0. The bars in the third graph are for the kena-passive and have the following values: Chinese 5.0, Eurasian 0, Indian 2.1, Malay 9.5, and Other 0.9. The bars in the fourth graph are for the tio-passive and have the following values: Chinese 0.6, Eurasian 0, Indian 0.6, Malay 0.3, and Other 0.

5.4.2 Distribution of Sentiment

Table 15 shows the distribution of sentiment for each passive marker in our data set; this is represented visually in Figure 19. These patterns let us answer, given a passive marker, what kind of verb (bearing positive, neutral, or negative sentiment) is most strongly associated with it. Within kena passives, the most common verbs have a negative sentiment, compared with be and get passives (Fisher’s exact test for be: p < 0.01, for get: p < 0.01). Tio also has a stronger negative bias compared with be (Fisher’s exact test p < 0.01), but no more so than with get (p = 0.22). We found no evidence of a clear difference between kena and tio (p = 0.35).

Figure 19

Distribution of sentiment, by passive marker.

Figure 19 Long description

In all four graphs, the y-axis represents percentages and the x-axis represents three sentiment levels (Negative, Neutral, and Positive). The bars in the first graph are for the be-passive and have the following values: Negative 11.1, Neutral 85.3, and Positive 3.6. The bars in the second graph are for the get-passive and have the following values: Negative 35.5, Neutral 60.3, and Positive 4.3. The bars in the third graph are for the kena-passive and have the following values: Negative 60.3, Neutral 39.4, and Positive 0.4. The bars in the fourth graph are for the tio-passive and have the following values: Negative 50.0, Neutral 50.0, and Positive 0.

We may also flip the question around: given a verb with a positive (or neutral, or negative) sentiment, which passive marker is likely to be used? To answer this question, we analysed the reweighted data set, which better represents the high frequency of be and get passives (Table 16 and Figure 20). Given the general preference for be and get passives (Table 13), it is unsurprising that at the sentiment level, be and get are also the most common options, with kena and tio trailing in distant third and fourth positions respectively (for each sentiment, the proportions of be and get both exceed that of kena, all ps < 0.001; the same for tio which is even less frequent than kena). In other words, although both kena and tio have a distinctive adversative flavour, they do not have a monopoly over negative or even neutral sentiment verbs. One plausible reason is that in these situations, speakers might still avoid kena and tio for register and/or sociolinguistic reasons; for instance, speakers might avoid tio in conversations with other speakers who (they suspect) do not use it (see Section 5.4.1).

Figure 20

Distribution of passive markers used to express a particular sentiment

Figure 20 Long description

In all three graphs, the y-axis represents percentages and the x-axis four passive markers (be, get, kena, and tio). The bars in the first graph are for Negative sentiment and have the following values: be-passive 40.1, get-passive 43.6, kena-passive 14.8, and tio-passive 1.5. The bars in the second graph are for Neutral sentiment and have the following values: be-passive 78.4, get-passive 18.8, kena-passive 2.5, and tio-passive 0.4. The bars in the third graph are for Positive sentiment and have the following values: be-passive 71.1, get-passive 28.4, kena-passive 0.5, and tio-passive 0.