Elamur Research⌕
☰
home / papers / 2025 / AraSim: Optimizing Arabic Dialect Translation in Children's Literature with LLMs and Similarity Scores

AraSim: Optimizing Arabic Dialect Translation in Children's Literature with LLMs and Similarity Scores

Paper · 2025 · Alaa Bouomar and Noorhan Abbas

Summary

The paper addresses the lack of Egyptian Arabic versions of children's stories. Young readers often struggle with the formal register and grammar of Modern Standard Arabic (MSA), and dialectal parallel data is scarce. The authors built a pipeline in which Claude drafts MSA→Egyptian (Cairo) translations line by line. They fine-tuned an AraT5 model on the MADAR Cairo subset and used it for back-translation to MSA. Before fine-tuning, native Egyptian speakers found and corrected errors in the MADAR Cairo data, and retraining on the cleaned data lowered both training and validation loss. Quality was measured with embedding cosine similarity between source and output and with BLEU. Fine-tuning raised average similarity from 0.945 to 0.971 and BLEU slightly. Human experts checked only the lines that scored below a similarity threshold. The result is a parallel MSA–Egyptian corpus of 130 children's stories, to be released on GitHub. The work matters because it shows a cheap way to build dialect resources for an under-served audience by combining an LLM, a specialised model and targeted human review.

Key points

  • Source texts: 130 stories, selected for moral and educational value, taken from Al-Sulaiti et al.'s (2016) Arabic Children's Corpus. That corpus has 2,950 documents and roughly 2 million words.
  • Starting model: AraT5 base (PRAli22/arat5-base-arabic-dialects-translation), fine-tuned on MADAR (Cairo).
  • Data cleaning: two native reviewers corrected about 4,500 lines each. First-epoch training loss fell from 0.0892 to 0.0484, and validation loss from 0.05547 to 0.0360.
  • Results: mean semantic similarity rose from 0.945 to 0.971 after fine-tuning. BLEU (NLTK, smoothing method4, whitespace tokens) rose from 0.0828 to 0.0867.
  • Model choice: a native speaker preferred Claude over ChatGPT for dialect output. ChatGPT also drifted in coherence when given long paragraphs.

Interesting ideas

  • I1. Instead of reviewing every output, only lines whose source–output similarity falls under a fixed cutoff (96%) are sent to human experts, which focuses scarce reviewer time on likely failures.
  • I2. Native speakers audited the "gold" MADAR data before training, and the effect was measured by comparing loss curves before and after cleaning, which treats data quality as a controlled experimental variable.
  • I3. Translating strictly line by line, even for long stories, keeps sentence alignment intact and stops the generator from adding content or losing coherence across paragraphs.
  • I4. The authors pair a general LLM for drafting with a smaller specialised model for the reverse direction, so each component does what it is best at.
  • I5. Each corpus line is shipped with its similarity score, so users can filter or weight examples by estimated quality.

For Mova

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

  • I1 (learn): a confidence or agreement score from witness parsers could decide which silver sentences go to the LLM teacher tier, rather than sending everything.
  • I2 (learn): run training twice, on raw and on repaired treebank or silver data, to show that a repair rule actually helps rather than assuming it does.
  • I3 (back): generating or translating per sentence node, with the world of the text as shared context, would keep the text → graph → text output aligned and checkable unit by unit.
  • I4 (back): an LLM could propose surface realisations while Mova's own parser re-reads them back to a graph, accepting only drafts whose round trip matches the original meaning.
  • I5 (learn): storing a per-sentence agreement or confidence score with silver data would let training curricula and negative-control evaluations pick their own quality thresholds.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Alaa Bouomar and Noorhan Abbas. 2025. AraSim: Optimizing Arabic Dialect Translation in Children's Literature with LLMs and Similarity Scores. Proceedings of the 4th Workshop on Arabic Corpus Linguistics (WACL-4).

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

@inproceedings{bouomar2025arasim,
  title = {{AraSim: Optimizing Arabic Dialect Translation in Children's Literature with LLMs and Similarity Scores}},
  author = {Alaa Bouomar and Noorhan Abbas},
  booktitle = {Proceedings of the 4th Workshop on Arabic Corpus Linguistics (WACL-4)},
  year = {2025},
  url = {https://aclanthology.org/2025.wacl-1.11/}
}

References

12 references, 11 of them with a card of ours. ↗ leads to the work.

  1. Latifa Al-Sulaiti, Noorhan Abbas, Claire Brierley, Eric Atwell, and Abdulmohsen Alghamdi. 2016. Compilation of an Arabic children's corpus. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), pages 1808–1812, Portorož, Slovenia. European Language Resources Association (ELRA). https://aclanthology.org/L16-1285 ↗
  2. Abdulrahman Al-Thubaity, Abdullah Al-Khateeb, Bassam Al-Salhi, and Mohammed Al-Ghamdi. 2023. Evaluating ChatGPT and Bard AI on Arabic sentiment analysis. Computing Research Repository, arXiv:2305.14745. https://github.com/alaabouomar/Optimizing-Arabic-Dialect-Translation-for-Children-s-Literature-Using-Neural-Models.git https://www.semanticscholar.org/paper/d4c0ee9f7ea 7451216845c851d069dff95545faa Association for Computational Linguistics. https://aclanthology.org/2022.acl-long.46/ ↗
  3. Zaid Alyafeai, Anwar Al-Omari, Iman Al-Kindi, Samah Al-Riyami, and Alaeddin Al-Maqaleh. 2023. Taqyim: evaluating Arabic NLP tasks using ChatGPT models. Computing Research Repository, arXiv:2305.14849. https://www.semanticscholar.org/paper/d14aa448b1 7fdc8d4ea12b43ee1a2b1254c38703 ↗
  4. Ossama Obeid, Nasser Zalmout, Dima Taji, Salam Khalifa, Bashar Alhafni, Koichi Inoue, Fadhl Eryani, and Nizar Habash. 2020. CAMeL tools: an open source Python toolkit for Arabic natural language processing. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 7022–7032, Marseille, France. European Language Resources Association. https://www.semanticscholar.org/paper/995ec006ac9 8a697ea38bd4eea8c1f3170a8adb4 ↗
  5. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and Kemal Oflazer. 2018. The MADAR Arabic dialect corpus and lexicon. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). https://aclanthology.org/L18-1535 ↗
  6. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and Sandhini Agarwal. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020), pages 1877-1901. Curran Associates, Inc. https://proceedings.neurips.cc/paper/2020/file/1457c 0d6bfcb4967418bfb8ac142f64a-Paper.pdf Taoufik Kadaoui, Houda Bouamor, Fathi Badran, and Nizar Habash. 2023. TARJAMAT: evaluation of Bard and ChatGPT on machine translation of ten Arabic varieties. Computing Research Repository, arXiv:2305.14786. https://www.semanticscholar.org/paper/796b894c4e 1a3cb46715cc0b45a39e91ee5910e6 ↗
  7. Quentin Malartic, Hamza Alobeidli, Danilo Mazzotta, Gabriel Penedo, Giulia Campesan, Muhammad Farooq, Mansoor Alhammadi, Julien Launay, and Badr Noune. 2023. AlGhafa evaluation benchmark for Arabic language models. In Proceedings of ArabicNLP 2023. Association for Computational Linguistics. https://aclanthology.org/2023.arabicnlp-1.1 ↗
  8. Ritu Mullappilly, Mohammed Al-Awlaqi, Saleh Al-Yami, and Faisal Al-Dossary. 2023. Arabic Mini-ClimateGPT: a climate change and sustainability tailored Arabic LLM. Computing Research Repository, arXiv:2305.14824. https://www.semanticscholar.org/paper/6da8e97de0 981b867b1038e12e98608928ad4c0e El Moatez Billah Nagoudi, AbdelRahim Elmadany, and Muhammad Abdul-Mageed. 2022. AraT5: text-to-text transformers for Arabic language generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 628-647, Dublin, Ireland. ↗
  9. Chatrine Qwaider, Nizar Habash, Houda Bouamor, and Fathi Badran. 2018. Shami: a corpus of Levantine Arabic dialects. Computing Research Repository, arXiv:1805.05190. https://www.semanticscholar.org/paper/654af2f5747 126447e5d8fce220c6a1915761143 ↗
  10. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017), pages 5998-6008. Curran Associates, Inc. https://papers.nips.cc/paper/7181-attention-is-all-you-need.pdf ↗
  11. Ahmed Waheed, Muhammad Abdul-Mageed, El Moatez Billah Nagoudi, Badr Noune, Amir Hamdi, AbdelRahim Elmadany, and Massimo Poesio. 2023. GPTAraEval: a comprehensive evaluation of ChatGPT on Arabic NLP. Computing Research Repository, arXiv:2305.14976. https://arxiv.org/abs/2305.14976 ↗
  12. Ahmed Abdelali, Hamdy Mubarak, Shammur Chowdhury, Maram Hasanain, Bekhzod Mousi, Sofiane Boughorbel, Said Abdaljalil, Yasir El Kheir, Dema Izham, Fahim Dalvi, Mohamad Hawasly, Nada Nazar, Youssef Elshahawy, Ahmed Ali, Nadir Durrani, Natasa Milic-Frayling, and Firoj Alam. 2024. LAraBench: benchmarking Arabic AI with large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 487-520, St. Julian's, Malta. Association for Computational Linguistics. https://aclanthology.org/2024.eacl-long.30 ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bouomar-2025-arasim-optimizing-arabic-dialect-translation/