Elamur Research⌕
☰
home / papers / 2026 / Computational Benchmarks for Egyptian Arabic Child Directed Speech

Computational Benchmarks for Egyptian Arabic Child Directed Speech

Paper · 2026 · Salam Khalifa, Abed Qaddoumi, Nizar Habash et al.

Summary

ARABABYTALK-EGY makes Arabic child-adult interaction data usable for NLP. The Egyptian Arabic CHILDES corpus (EGYCHILDES) existed only as near-phonemic IPA transcriptions. Arabic NLP tools need orthographic input, so the corpus could not be used even for basic processing. The authors enriched it with three layers: fully diacritized Arabic script following CODA conventions, lemmas, and core POS tags based on the Buckwalter stem tagset. All layers stay aligned with the original IPA lines and audio timestamps.

The workflow ran in four steps:

  • build a lexicon of word types from the IPA transcripts,
  • have GPT-4o draft diacritized spellings for each type,
  • correct the drafts by hand and add a lemma and POS to each entry,
  • project the lexicon back onto every session and validate it in context, checking the audio when needed.

The result is roughly 26K annotated tokens. The paper also reports two benchmarks: morphological disambiguation and ASR with GPT-4o-transcribe. ASR error rates are much higher than those reported for other dialectal Arabic speech corpora, which shows how hard this genre is. The resource fills a gap for Arabic acquisition research and speech technology.

Key points

  • Source data. EGYCHILDES contains 10 monolingual children from Alexandria (5 girls, 5 boys), aged 1.6 to 3.7 years. Each child has one 30-minute session with an investigator and/or a parent.
  • Lexicon. The type lexicon has 4,170 entries with a Zipf-like frequency distribution.
  • LLM drafting setup. GPT-4o spellings were requested in batches of 20 words with schema-enforced structured output. Pilot runs showed that larger batches tended to lose diacritics.
  • Disambiguation benchmark. It uses the system of Inoue et al. (2022) and scores diacritized form (Diac), lemma (Lex), POS, and all three jointly (DLP).
  • ASR benchmark.
    • GPT-4o-transcribe was tested with a minimal Arabic prompt and temperature 0.
    • Audio chunks of 30 s worked better than 60 s or 1,000 s.
    • Scoring used WER/CER via Jiwer, after removing punctuation and normalizing hamzated Alif.
    • The noisiest recordings (FR, ZY) produced fluent transcripts unrelated to the reference.

Interesting ideas

  • I1. Annotators labeled each unique word type once, then pushed the labels onto all tokens and checked them in context; this cuts manual effort while still catching context-dependent errors.
  • I2. Related word forms were grouped by a crude vowel-stripped skeleton so a reviewer could fix whole families at once, even though the key is not a true root.
  • I3. LLM drafts came from small batches under an enforced output schema, and batch size was chosen by a pilot because large batches silently degraded output detail.
  • I4. Child mispronunciations were kept next to the adult target form, and the cleaned text stays line-aligned with the raw annotated source, so nothing is lost.
  • I5. On very noisy audio, the ASR model wrote fluent text with no relation to the reference, a reminder that fluency says nothing about faithfulness.

For Mova

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

  • I1 (tokens): when adapting the en tagger and lemmatizer to a new genre, build a reviewed type-level lexicon first, then use token-level checks only where context changes the analysis.
  • I2 (learn): self-induced repair rules or silver-data disagreements could be clustered by a cheap stem key, so each family is reviewed and fixed in one pass.
  • I3 (learn): LLM teacher tiers could use strict schemas and pilot-tuned batch sizes, with checks for silently dropped fields rather than only malformed output.
  • I4 (tree): treebank and UD dialect conversion could store original and normalized tokens side by side, so trees can be traced back to the exact surface input.
  • I5 (learn): evaluation with negative controls could feed deliberately degraded input and require the system to abstain or flag low support instead of producing a confident answer.

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

🎓New York University Abu DhabiAbu Dhabi · United Arab Emirates122 works · 4,602 citations🎓Stony Brook UniversityStony Brook · United States43 works · 2,571 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksPsychology · 1,552 worksDevelopmental and Educational Psychology · 499 works

🧭 Topics 3Research subjects and languages the work is about.Semitic languages · 2,908 worksEvaluation & benchmarks · 1,220 worksPsycholinguistics & language acquisition · 102 works

Cite

Cite as Salam Khalifa, Abed Qaddoumi, Nizar Habash and Owen Rambow. 2026. Computational Benchmarks for Egyptian Arabic Child Directed Speech. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-long.102

@inproceedings{khalifa2026computational,
  title = {{Computational Benchmarks for Egyptian Arabic Child Directed Speech}},
  author = {Salam Khalifa, Abed Qaddoumi, Nizar Habash and Owen Rambow},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.102},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.102}
}

References

38 references, 15 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Omri Abend, Tom Kwiatkowski, Nathaniel J Smith, Sharon Goldwater, and Mark Steedman. 2017. Bootstrapping language acquisition. Cognition, 164:116– 143. ↗
  2. Caleb Belth. 2021. The Greedy and Recursive Search for Morphological Productivity. ↗
  3. Rui Huang. 2016. An evaluation of pos taggers for the childes corpus. Master’s thesis, City University of New York (CUNY). find
  4. 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.
  5. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The interplay of variant, size, and task type in Arabic pre-trained language models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 92– 104, Kyiv, Ukraine (Virtual). Association for Computational Linguistics.
  6. Tim Buckwalter. 2004. Buckwalter Arabic Morphological Analyzer Version 2.0. LDC catalog number LDC2004L02, ISBN 1-58563-324-0. ↗
  7. Go Inoue, Salam Khalifa, and Nizar Habash. 2022. Morphosyntactic tagging with pre-trained language models for Arabic and its dialects. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1708–1719, Dublin, Ireland. Association for Computational Linguistics.
  8. Steven Y. Feng, Noah D. Goodman, and Michael C. Frank. 2024. Is child-directed speech effective training data for language models? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 22055–22071, Miami, Florida, USA. Association for Computational Linguistics.
  9. Salam Khalifa, Abdelrahim Qaddoumi, Jordan Kodner, and Owen Rambow. 2025. Learning cross-dialectal morphophonology with syllable structure constraints. In Proceedings of the 12th Workshop on NLP for Similar Languages, Varieties and Dialects, pages 157– 167, Abu Dhabi, UAE. Association for Computational Linguistics. find
  10. Hassan Gadalla, Hanaa Kilany, Howaida Arram, Ashraf Yacoub, Alaa El-Habashi, Amr Shalaby, Krisjanis Karins, Everett Rowson, Robert MacIntyre, Paul Kingsbury, David Graff, and Cynthia McLemore. 1997. CALLHOME Egyptian Arabic transcripts LDC97T19. Web Download. Philadelphia: Linguistic Data Consortium. find
  11. H. Kilany, H. Gadalla, H. Arram, A. Yacoub, A. El-Habashi, and C. McLemore. 2002. Egyptian Colloquial Arabic Lexicon. LDC catalog number LDC99L22. find
  12. Shai Gretz, Alon Itai, Brian MacWhinney, Bracha Nir, and Shuly Wintner. 2015. Parsing hebrew childes transcripts. Language Resources and Evaluation, 49(1):107–145. find
  13. Christo Kirov and Ryan Cotterell. 2018. Recurrent neural networks in linguistic theory: Revisiting pinker and prince (1988) and the past tense debate. Transactions of the Association for Computational Linguistics, 6:651–665. ↗
  14. Nizar Habash, Ramy Eskander, and Abdelati Hawwari. 2012. A Morphological Analyzer for Egyptian Arabic. In Proceedings of the Workshop of the Special Interest Group on Computational Morphology and Phonology (SIGMORPHON), pages 1–9, Montréal, Canada.
  15. Jordan Kodner and Salam Khalifa. 2022. SIGMORPHON–UniMorph 2022 shared task 0: Modeling inflection in language acquisition. In Proceedings of the 19th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 157–175, Seattle, Washington. Association for Computational Linguistics. Nizar Habash, Salam Khalifa, Fadhl Eryani, Owen Rambow, Dana Abdulrahim, Alexander Erdmann, Reem Faraj, Wajdi Zaghouani, Houda Bouamor, Nasser Zalmout, Sara Hassan, Faisal Al shargi, Zoey Liu and Emily Prud’hommeaux. 2023. Data-driven parsing evaluation for child-parent interactions. Transactions of the Association for Computational Linguistics, 11:1734–1753.
  16. Mohamed Maamouri, Ann Bies, Seth Kulick, Michael Ciul, Nizar Habash, and Ramy Eskander. 2014. Developing an egyptian arabic treebank: Impact of dialectal morphology on annotation and tool development. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC-2014). European Language Resources Association (ELRA). find
  17. Brian MacWhinney. 2000. The CHILDES Project: Tools for Analyzing Talk, 3rd edition. Lawrence Erlbaum Associates, Mahwah, NJ. Supported by NICHD Grant HD082736. find
  18. Brian MacWhinney. 2025. Understanding language through talkbank. Current Directions in Psychological Science, 34(2):75–81. ↗
  19. Reham Marzouk, Sondos Krouna, and Nizar Habash. 2025. A derivational ChainBank for Modern Standard Arabic. In Proceedings of the 1st Workshop on NLP for Languages Using Arabic Script, pages 78– 87, Abu Dhabi, UAE. Association for Computational Linguistics. ↗
  20. Kate McCurdy, Sharon Goldwater, and Adam Lopez. 2020. Inflecting When There’s No Majority: Limitations of Encoder-Decoder Neural Networks as Cognitive Models for German Plurals. Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 1745–1756. find
  21. Tala Nazzal. 2021. CHILDES Palestinian Arabic Nazzal Corpus. https://doi.org/10.21415/ VJYY-KA80. DOI:10.21415/VJYY-KA80. ↗
  22. Dimitrios Ntelitheos and Ali Idrissi. 2017. Language Growth in Child Emirati Arabic. Perspectives on Arabic Linguistics, 29:229–248. ↗
  23. Ossama Obeid, Nasser Zalmout, Salam Khalifa, Dima Taji, Mai Oudah, Bashar Alhafni, Go Inoue, Fadhl Eryani, Alexander Erdmann, and Nizar Habash. 2020. CAMeL tools: An open source python toolkit for Arabic natural language processing. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 7022–7032, Marseille, France. European Language Resources Association.
  24. Jan Odijk, Alexis Dimitriadis, Martijn van der Klis, Marjo van Koppen, Meie Otten, and Remco van der Veen. 2018. The AnnCor CHILDES treebank. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). find
  25. OpenAI. 2024a. GPT-4o. Large language model accessed via OpenAI API (model name: gpt-4o, snapshot 2025-07-06). find
  26. OpenAI. 2024b. Gpt-4o system card. Technical Report arXiv:2410.21276, OpenAI. ↗
  27. OpenAI. 2025. GPT/̄4o/̄transcribe. https://platform.openai.com/docs/models/ gpt-4o-transcribe (accessed July 25, 2025). Large language model, API documentation. ↗
  28. Arfath Pasha, Mohamed Al-Badrashiny, Mona Diab, Ahmed El Kholy, Ramy Eskander, Nizar Habash, Manoj Pooleery, Owen Rambow, and Ryan Roth. 2014. Madamira: A fast, comprehensive tool for morphological analysis and disambiguation of Arabic. In Proceedings of the Language Resources and Evaluation Conference (LREC), pages 1094–1101, Reykjavik, Iceland.
  29. Kenji Sagae, Eric Davis, Alon Lavie, Brian MacWhinney, and Shuly Wintner. 2007. High-accuracy annotation and parsing of CHILDES transcripts. In Proceedings of the Workshop on Cognitive Aspects of Computational Language Acquisition, pages 25–32, Prague, Czech Republic. Association for Computational Linguistics. ↗
  30. Kenji Sagae, Eric Davis, Alon Lavie, Brian Macwhinney, and Shuly Wintner. 2010. Morphosyntactic annotation of childes transcripts. Journal of Child Language, 37(3):705–29. ↗
  31. Heba Salama. 2015. CHILDES Egyptian Arabic Salama Corpus. https://doi.org/10.21415/ 78CE-VW65. DOI:10.21415/78CE-VW65. ↗
  32. Mahmoud Salhab, Marwan Elghitany, Shameed Sait, Syed Sibghat Ullah, Mohammad Abusheikh, and Hasan Abusheikh. 2025. Advancing arabic speech recognition through large-scale weakly supervised learning. Preprint, arXiv:2504.12254. ↗
  33. Alessandro Sanchez, Stephan C Meylan, Mika Braginsky, Kyle E MacDonald, Daniel Yurovsky, and Michael C Frank. 2019. childes-db: A flexible and reproducible interface to the child language data exchange system. Behavior research methods, 51(4):1928–1941. ↗
  34. Ida Szubert, Omri Abend, Nathan Schneider, Samuel Gibbon, Louis Mahon, Sharon Goldwater, and Mark Steedman. 2024. Cross-linguistically consistent semantic and syntactic annotation of child-directed speech. Language Resources and Evaluation, pages 1–50. ↗
  35. Nik Vaessen. 2025. Jiwer: Similarity measures for automatic speech recognition evaluation. find
  36. Alex Warstadt, Aaron Mueller, Leshem Choshen, Ethan Wilcox, Chengxu Zhuang, Juan Ciro, Rafael Mosquera, Bhargavi Paranjabe, Adina Williams, Tal Linzen, and Ryan Cotterell. 2023. Findings of the BabyLM challenge: Sample-efficient pretraining on developmentally plausible corpora. In Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning, pages 1–34, Singapore. Association for Computational Linguistics. find
  37. Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. 2021. Raise a child in large language model: Towards effective and generalizable fine-tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9514– 9528, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. ↗
  38. Xiulin Yang, Zhuoxuan Ju, Lanni Bu, Zoey Liu, and Nathan Schneider. 2025. Ud-english-childes: A collected resource of gold and silver universal dependencies trees for child language interactions. Preprint, arXiv:2504.20304. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/khalifa-2026-computational-benchmarks-egyptian-arabic-child/