📄 A Pilot Study on the Collection and Computational Analysis of Linguistic Differences Amongst Men and Women in a Kuwaiti Arabic WhatsApp Dataset

Summary

Kuwaiti Arabic (KA) is a low-resource Gulf dialect with few conversational texts and almost no sociolinguistic work on how men and women write differently. The authors built what they describe as the first gender-labelled corpus of naturally occurring KA conversation. With ethical approval and informed consent, they collected chats from three long-running Kuwaiti book-club WhatsApp groups, anonymised the members, and studied interactional and lexical cues that could later feed a gender classifier. They compared men and women on number of turns, turn length and emoji counts using statistical tests. They then looked qualitatively at which emojis each group used, at words used by only one gender, and at word lengthening. None of the counted features showed a statistically significant gender difference, even though the raw numbers differed noticeably. Emoji choice and elongated words, however, did separate men from women. The work offers a scarce dialect resource and a list of candidate features for future KA gender profiling. It also shows that small samples can make large-looking gaps statistically inconclusive.

Key points

• The corpus has 4479 turns, 2623 from 28 women and 1856 from 14 men, drawn from three Kuwaiti reading-club WhatsApp groups. The chats were exported after the researcher had been a member for 9 months.
• Usernames were replaced by pseudonyms that encode gender, such as USER1F. Access to the data is by request to the first author.
• Raw totals: women used 2144 emojis and 17,388 words; men used 801 emojis and 14,005 words. Per-person spread was wide (women's mean of 94 turns against a median of 29).
• Women produced 0.057 lengthened words per turn and men 0.037, about 1.5 times as often. Women's elongations had more varied functions (encouraging, agreeing, greeting, missing someone), while men's were mostly laughter and greetings.
• Few gender-exclusive dialect words appeared, mostly in women's messages, which the authors attribute to the formal book-discussion register. The authors also note limits: small size, gender imbalance and an unrepresentative register.

Interesting ideas

• I1. Elongated spellings carry a pragmatic function such as laughter, amazement, agreement or greeting, so they should not be treated as noise.
• I2. Measuring a phenomenon as a per-turn rate (once per 18 vs 28 turns) gives a fairer comparison than raw totals when groups produce very different amounts of text.
• I3. The study pairs significance tests with qualitative inspection and shows that sizeable raw gaps can still be inconclusive on small, skewed samples.
• I4. Grouping elongations by communicative purpose turns a surface trick into a signal about what the writer is doing.
• I5. The authors note that a formal register suppressed the very phenomena they wanted to find. This shows that the domain of a sample shapes which features are even observable.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): Normalise forms like "sooo" or "yeees" to their lemma but keep an emphasis/elongation feature on the token so later stages can still see it.
• I2 (learn): Report error or phenomenon rates normalised per sentence or per text when comparing parsers or genres with unequal sizes.
• I3 (learn): Evaluation reports should include significance checks and median-vs-mean views before claiming one parser or rule set beats another on small test sets.
• I4 (frames): An elongation flag could serve as a cue for speech acts and affect (greeting, agreeing, enthusiasm), feeding speech-act and irony detection in crate prag.
• I5 (answer): When building QA or reader tests, check that the source texts actually contain the constructions being tested, rather than assuming one genre stands for the language.