Elamur Research⌕
☰
home / papers / 2025 / What did you say? Generating Child-Directed Speech Questions to Train LLMs

What did you say? Generating Child-Directed Speech Questions to Train LLMs

Summary

Language models trained on adult speech addressed to children (Child-Directed Speech, CDS) often perform better than models trained on the same amount of other types of text. The authors test whether one of the typical features of CDS, a high proportion of questions, can be transferred to other data. To do this, they took the non-CDS portion of the Strict-Small corpus from the BabyLM 2025 challenge: BNC, Gutenberg, OpenSubtitles, Simple Wiki, and Switchboard. CHILDES was left unchanged as a reference. The model gpt-5-mini wrote short child-style questions, together with answers, for fragments of these texts. Each dataset was then randomly reduced so that the total volume stayed within 10M words. Several GPT-Wee configurations were trained on these data and compared. It turned out that the quality and form of the generated questions depend strongly on the properties of the source text. The added questions improve results on tasks involving interrogative constructions. At the same time, the overall BLiMP score declines, probably due to reduced syntactic diversity. The work shows that artificially adapting data toward CDS has both benefits and risks.

Key points

  • Proportion of questions in the source data: in CHILDES, 20.54% of sentences are questions, and their mean length (MLU) is 4.92 words. CHILDES is followed by OpenSubtitles (17.74%), BNC (15.15%), Gutenberg (7.91%), Switchboard (4.05%), and Simple Wiki (0.08%). That is, the non-CDS data contain more questions than the authors expected, but these questions are longer and structured differently.
  • How the questions were generated: gpt-5-mini was asked to act as a reading assistant for a 5–6-year-old child. For each fragment, it wrote five simple questions shorter than 10 words and provided answers to them. After reduction, each dataset differed from the original by no more than 10 words.
  • Which models were compared: a baseline model trained on the original data, a model with all sources modified, and five models, each with only one source modified. Ten runs were conducted for each configuration.
  • BLiMP results: the baseline model achieved the best overall score, and the model with all sources modified achieved the worst. Modifying a small file (Switchboard) was less harmful than modifying a large one (Gutenberg). No single source yielded a consistent improvement across all tasks. On the subtask wh_vs_that_with_gap_long_distance, the baseline model scored about 7.62%, and on wh_vs_that_no_gap_long_distance it scored 98.68%.
  • Resource and limitations: the data and models are published in the GitHub repository NLPlabMSU/BabyLM_Questions. Because of its built-in safety restrictions, gpt-5-mini initially refused to generate questions for OpenSubtitles. The authors worked around this by stating in the prompt that the request was made for research purposes.

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

Fields: Linguistics & languages

Topics Language models & pretraining Psycholinguistics & language acquisition

Cite

Cite as Whitney Poh, Michael Tombolini and Libby Barak. 2025. What did you say? Generating Child-Directed Speech Questions to Train LLMs. Proceedings of the First BabyLM Workshop.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.babylm-main.18

@inproceedings{poh2025what,
  title = {{What did you say? Generating Child-Directed Speech Questions to Train LLMs}},
  author = {Whitney Poh, Michael Tombolini and Libby Barak},
  booktitle = {Proceedings of the First BabyLM Workshop},
  year = {2025},
  doi = {10.18653/v1/2025.babylm-main.18},
  url = {https://doi.org/10.18653/v1/2025.babylm-main.18}
}

References

35 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Igor Bascandziev, Patrick Shafto, and Elizabeth Bonawitz. 2021. The sound of pedagogical questions. In Proceedings of the annual meeting of the cognitive science society, volume 43. find
  2. BNC Consortium. 2007. The british national corpus, xml edition. find
  3. Peter Brodsky and Heidi Waterfall. 2007. Characterizing motherese: On the computational structure of child-directed language. In Proceedings of the annual meeting of the cognitive science society, volume 29. find
  4. Bastian Bunzeck and Sina Zarrieß. 2023. GPT-wee: How small can a small language model really get? In Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning, pages 35–46, Singapore. Association for Computational Linguistics. child directed speech. Cognit… find
  5. Lucas Charpentier, Leshem Choshen, Ryan Cotterell, Mustafa Omer Gul, Michael Hu, Jaap Jumelet, Tal Linzen, Jing Liu, Aaron Mueller, Candace Ross, Raj Sanjay Shah, Alex Warstadt, Ethan Wilcox, and Adina Williams. 2025. Babylm turns 3: Call for papers for the 2025 babylm workshop. Preprint, arXiv:2502… find
  6. Emily N Daubert, Yue Yu, Milagros Grados, Patrick Shafto, and Elizabeth Bonawitz. 2020. Pedagogical questions promote causal learning in preschoolers. Scientific reports, 10(1):20700. find
  7. Baxter S Eaves Jr, Naomi H Feldman, Thomas L Griffiths, and Patrick Shafto. 2016. Infant-directed speech is consistent with teaching. Psychological review, 123(6):758. find
  8. Lukas Edman and Lisa Bylinina. 2023. Too much information: Keeping training simple for BabyLMs. In Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning, pages 89–97, Singapore. Association for Computational Linguistics. find
  9. Martin Gerlach and Francesc Font-Clos. 2020. A standardized project gutenberg corpus for statistical analysis of natural language and quantitative linguistics. Entropy, 22(1):126. find
  10. Akari Haga, Akiyo Fukatsu, Miyu Oba, Arianna Bisazza, and Yohei Oseki. 2024. BabyLM challenge: Exploring the effect of variation sets on language model training efficiency. In The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning, pages 252–261, Miami, FL, USA. A… find
  11. Michael Y. Hu, Aaron Mueller, Candace Ross, Adina Williams, Tal Linzen, Chengxu Zhuang, Leshem Choshen, Ryan Cotterell, Alex Warstadt, and Ethan Gotlieb Wilcox, editors. 2024. The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning. Association for Computational Li… find
  12. Philip A. Huebner, Elior Sulem, Fisher Cynthia, and Dan Roth. 2021. BabyBERTa: Learning more grammar with small-scale child-directed language. In Proceedings of the 25th Conference on Computational Natural Language Learning, pages 624–646, Online. Association for Computational Linguistics. our card
  13. Maureen A Callanan and Lisa M Oakes. 1992. Preschoolers’ questions and parents’ explanations: Causal thinking in everyday activity. Cognitive development, 7(2):213–233. find
  14. Anishka Jean, Emily Daubert, Yue Yu, Patrick Shafto, and Elizabeth Bonawitz. 2019. Pedagogical questions empower exploration. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 41. find
  15. Thea Cameron-Faulkner, Elena Lieven, and Michael Tomasello. 2003. A construction based analysis of Jaap Jumelet, Lucas Charpentier, Michael Hu, and Jing Liu. 2025. Babylm_2025. OSF. find
  16. Pierre Lison and Jörg Tiedemann. 2016. OpenSubtitles2016: Extracting large parallel corpora from movie and TV subtitles. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 923–929, Portorož, Slovenia. European Language Resources Association (EL… our card
  17. Brian MacWhinney. 2000. The CHILDES project: The database, volume 2. Psychology Press. find
  18. Jessica L Montag, Michael N Jones, and Linda B Smith. 2015. The words children hear: Picture books and the statistics for language learning. Psychological science, 26(9):1489–1496. find
  19. Jessica L Montag, Michael N Jones, and Linda B Smith. 2018. Quantity and diversity: Simulating early word learning environments. Cognitive science, 42:375– 412. find
  20. Aaron Mueller and Tal Linzen. 2023a. How to plant trees in language models: Data and architectural effects on the emergence of syntactic inductive biases. arXiv preprint arXiv:2305.19905. find
  21. Aaron Mueller and Tal Linzen. 2023b. How to plant trees in language models: Data and architectural effects on the emergence of syntactic inductive biases. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11237–11252, Toronto, C… find
  22. Nikitas Theodoropoulos, Giorgos Filandrianos, Vassilis Lyberatos, Maria Lymperaiou, and Giorgos Stamou. 2024. BERTtime stories: Investigating the role of synthetic story data in language pre-training. In The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning, page… find
  23. Yulia Tsvetkov, Manaal Faruqui, Wang Ling, Brian MacWhinney, and Chris Dyer. 2016. Learning the curriculum with Bayesian optimization for task-specific word representation learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa… find
  24. Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R. Bowman. 2020. BLiMP: The benchmark of linguistic minimal pairs for English. Transactions of the Association for Computational Linguistics, 8:377– 392. find
  25. Wikimedia. 2023. Simple english wikipedia dump. https://dumps.wikimedia.org/simplewiki/ 20230301/. Accessed: 2023-07-31. ↗
  26. Mira L Nencheva and Casey Lew-Williams. 2022. Understanding why infant-directed speech supports learning: A dynamic attention perspective. Developmental Review, 66:101047. find
  27. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and 3 others. 2020. Huggin… our card
  28. Elissa L Newport, Henry Gleitman, and Lila R Gleitman. 2020. Mother, i’d rather do it myself. Sentence first, arguments afterward: Essays in language and learning, 141. find
  29. Guanghao You, Balthasar Bickel, Moritz M Daum, and Sabine Stoll. 2021. Child-directed speech is optimized for syntax-free semantic inference. Scientific Reports, 11(1):16527. find
  30. OpenAI. 2025. Gpt-5 mini. https://openai.com. Lightweight variant of GPT-5 large language model. ↗
  31. Yue Yu, Elizabeth Bonawitz, and Patrick Shafto. 2019. Pedagogical questions in parent–child conversations. Child development, 90(1):147–161. find
  32. Meredith L Rowe. 2008. Child-directed speech: Relation to socioeconomic status, knowledge of child development and child vocabulary skill. Journal of child language, 35(1):185–205. find
  33. Jessica F Schwab and Casey Lew-Williams. 2016. Repetition across successive sentences facilitates young children’s word learning. Developmental psychology, 52(6):879. find
  34. Catherine E. Snow and Charles A. Ferguson, editors. 1977. Talking to Children: Language Input and Acquisition. Cambridge University Press. find
  35. Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteer. 2000. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational linguistics, 26(3):339–373. our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/poh-2025-what-did-you-say-generating/