Elamur Research⌕
☰
home / papers / 2020 / Predicting gender and age categories in English conversations using lexical, non-lexical, and turn-taking features

Predicting gender and age categories in English conversations using lexical, non-lexical, and turn-taking features

Paper · 2020 · Andreas Liesenfeld, Gábor Parti, Yuyin Hsu et al.

Summary

The paper tests whether the social categories attached to speakers (binary gender, and very young versus very old age) can be guessed from how they talk in everyday conversation. The authors draw two subcorpora from the Spoken BNC2014, a collection of about 11.4 million words of informal British English. Using Scaled F-score and Scattertext plots, they first compare which words and phrases each group favours. They then compare non-lexical material that dialog datasets usually discard: minimal particles sorted into five rough form-function classes, laughter, pauses, overlaps and truncated words. Finally, they train a logistic regression classifier on these features, labelling either a speaker within a conversation or a single turn. When the non-lexical cues are added to lexical ones, speaker-level gender accuracy climbs to about 81% and age accuracy to 92%. Turn-level prediction gains much less. The central lesson is that noisy, disfluent transcripts still hold useful signal, and that tokens such as "mhm" or "erm" should be kept rather than cleaned away. The authors also warn at length about the biases built into binary social labels.

Key points

  • Data: Spoken BNC2014 (recorded 2012–2016, mostly 2–4 speakers). The gender subcorpus covers 365 female and 305 male speakers. The age subcorpus covers 56 speakers over 69 and 49 speakers aged 18 or under; middle-aged speakers were left out on purpose to sharpen the contrast.
  • Corpus statistics: female speakers contribute more turns and slightly longer ones (about 9.4 words per turn versus about 9.0). Male turns show a higher type-token ratio.
  • Group tendencies:
    • Male talk favours "eh", "uh", "em", "innit" and "ain't".
    • Female talk favours "mhm", "hm", "huh" and "my God".
    • Young speakers laugh, overlap and swear more.
    • Old speakers produce more truncated words, more turn-management fillers ("er", "erm", "um") and more "oh".
  • Minimal particles are grouped by form into five functions: positive responses/continuers, turn stalling, turn management, repair initiators ("hm?") and change-of-state tokens ("oh").
  • Logistic regression results against a 50% balanced baseline:
LevelTaskLexical onlyLexical + non-lexical
Per conversationGender70.95%80.71%
Per conversationAge89.50%92.00%
Per turn (90/10 split)Gender62.39%63.57%
Per turn (90/10 split)Age67.33%67.74%

Interesting ideas

  • I1. When a corpus has no functional labels for interjections, group them by surface form (nasal vs. vocalic, rising pitch) into a few broad functions; this is a cheap first approximation that can be refined later.
  • I2. Keep disfluency and transcription markers (truncation, pauses, overlaps, laughter) as first-class tokens instead of stripping them, since the paper shows they carry measurable information.
  • I3. Weak per-item evidence becomes strong once it is pooled over many turns by the same speaker; short turns alone rarely contain anything biased toward a label.
  • I4. Balance class sizes, drop speakers with too little talk, and report against an explicit 50% baseline, so that accuracy gains are not artifacts of corpus imbalance.
  • I5. Use a frequency-aware distinctiveness score (Scaled F-score) to find items characteristic of one subcorpus while keeping rare terms visible.

For Mova

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

  • I1 (frames): prag could map "mhm", "hm?", "oh", "erm" to coarse speech-act tags (backchannel, repair request, realization, floor-holding) so that dialog turns get an explainable act label instead of being skipped.
  • I2 (tokens): the en tokenizer could emit typed tokens for fillers and transcription tags, so that later stages can either use them or explicitly ignore them.
  • I3 (world): attributes of a speaker entity, such as stance or politeness, could be inferred by accumulating cues across all of that entity's utterances, recording the provenance of each cue.
  • I4 (learn): evaluations of discourse-level classifiers should use equal-sized classes and a chance-level control alongside the real score.
  • I5 (learn): the same contrast could surface lemmas or constructions over-represented in sentences where a parser or reader fails, pointing to candidates for repair rules.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Hong Kong Polytechnic UniversityHong Kong · China84 works · 4,956 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Andreas Liesenfeld, Gábor Parti, Yuyin Hsu and Chu-Ren Huang. 2020. Predicting gender and age categories in English conversations using lexical, non-lexical, and turn-taking features. Proceedings of the 34th Pacific Asia Conference on Language, Information and Computation.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{liesenfeld2020predicting,
  title = {{Predicting gender and age categories in English conversations using lexical, non-lexical, and turn-taking features}},
  author = {Andreas Liesenfeld, Gábor Parti, Yuyin Hsu and Chu-Ren Huang},
  booktitle = {Proceedings of the 34th Pacific Asia Conference on Language, Information and Computation},
  year = {2020},
  url = {https://aclanthology.org/2020.paclic-1.19/}
}

References

26 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abouelenien, M., Pérez-Rosas, V., Mihalcea, R., and Burzo, M. (2017). Multimodal gender detection. In Proceedings of the 19th ACM International Conference on Multimodal Interaction, pages 302–311. find
  2. Acton, E. K. (2011). On gender differences in the distribution of um and uh. University of Pennsylvania Working Papers in Linguistics, 17(2):2. find
  3. Argamon, S., Koppel, M., Fine, J., and Shimoni, A. R. (2003). Gender, genre, and writing style in formal written texts. Text & Talk, 23(3):321–346. ↗
  4. Baker, P. (2014). Using corpora to analyze gender. A&C Black. find
  5. Bird, S., Klein, E., and Loper, E. (2009). Natural language processing with Python: analyzing text with the natural language toolkit. O’Reilly Media, Inc. find
  6. Couper-Kuhlen, E. and Selting, M. (2017). Interactional linguistics: Studying language in social interaction. Cambridge University Press. find
  7. Herring, S. C. and Paolillo, J. C. (2006). Gender and genre variation in weblogs. Journal of Sociolinguistics, 10(4):439–459. ↗
  8. Honnibal, M. and Montani, I. (2017). Spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1). find
  9. James, D. and Drakich, J. (1993). Understanding gender differences in amount of talk: A critical review of research. In Tannen, D., editor, Oxford studies in sociolinguistics. Gender and conversational interaction, page 281–312. Oxford University Press. find
  10. Jefferson, G. (1988). On the sequential organization of troubles-talk in ordinary conversation. Social problems, 35(4):418–441. ↗
  11. Joshi, C. K., Mi, F., and Faltings, B. (2017). Personalization in goal-oriented dialog. In NIPS 2017 Conversational AI Workshop, 4-9 Dec 2017. ↗
  12. Kessler, J. S. (2017). Scattertext: a Browser-Based Tool for Visualizing how Corpora Differ. In Proceedings of ACL-2017 System Demonstrations, 30 July - 4 August 2017, Vancouver, Canada. Association for Computational Linguistics. find
  13. Lakoff, R. (1973). Language and woman’s place. Language in society, 2(1):45–79. ↗
  14. Liesenfeld, A. (2019a). Action formation with jan-wai in Cantonese Chinese conversation. PhD thesis, Nanyang Technological University. find
  15. Liesenfeld, A. (2019b). Cantonese turn-initial minimal particles: annotation of discourse-interactional functions in dialog corpora. Proceedings of the 33rd Pacific Asia Conference on Language, Information and Computation (PACLIC 33). find
  16. Liesenfeld, A. and Huang, C. R. (2020). Name-Spec Asks: What’s Your Name in Chinese? A Voice Bot to Specify Chinese Personal Names through Dialog. In Proceedings of the 2nd Conference on Conversational User Interfaces, CUI ’20, New York, NY, USA. Association for Computing Machinery. find
  17. Love, R., Dembry, C., Hardie, A., Brezina, V., and McEnery, T. (2017). The Spoken BNC2014: Designing and building a spoken corpus of everyday conversations. International Journal of Corpus Linguistics, 22(3):319–344. find
  18. McEnery, A. and Xiao, Z. (2004). Swearing in modern British English: the case of fuck in the BNC. Language and Literature, 13(3):235– 268. ↗
  19. Oger, K. (2019). A Study of Non-Finite Forms of Anaphoric do in the Spoken BNC. Anglophonia. French Journal of English Linguistics, (28). ↗
  20. Prabhakaran, V. and Rambow, O. (2017). Dialog structure through the lens of gender, gender environment, and power. Dialogue & Discourse, 8(2):21–55. ↗
  21. Schofield, A. and Mehr, L. (2016). Gender-distinguishing features in film dialogue. In Proceedings of the Fifth Workshop on Computational Linguistics for Literature, 16 June, 2016, pages 32–39. find
  22. Skantze, G. (2017). Predicting and regulating participation equality in human-robot conversations: Effects of age and gender. In 2017 12th ACM/IEEE International Conference on Human-Robot Interaction (HRI), 9-11 March, 2017, pages 196–204. IEEE. find
  23. Tannen, D. (1990). You just don’t understand: Women and men in conversation. Morrow New York. find
  24. Tannen, D. (1993). Gender and conversational interaction. Oxford University Press. find
  25. Wolters, M., Vipperla, R., and Renals, S. (2009). Age recognition for spoken dialogue systems: Do we need it? In Tenth Annual Conference of the International Speech Communication Association, 6-10 September 2009. find
  26. Xiao, R. and Tao, H. (2007). A corpus-based sociolinguistic study of amplifiers in British English. Sociolinguistic studies, 1(2):241–273. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liesenfeld-2020-predicting-gender-age-categories-english/