Elamur Research⌕
☰
home / papers / 2026 / Context-Aware Dialectal Arabic Machine Translation with Interactive Region and Register Selection

Context-Aware Dialectal Arabic Machine Translation with Interactive Region and Register Selection

Summary

Arabic MT systems tend to flatten dialectal input into Modern Standard Arabic (MSA), and users can rarely choose which vernacular the output should use. The authors call this "Dialect Erasure" and build a steerable translator to address it. Using a Rule-Based Data Augmentation (RBDA) pipeline, they grow an initial set of 3,000 sentences into a parallel set of roughly 57,000 sentences, evenly balanced across eight regional varieties, including Egyptian, Levantine, Gulf and Maghrebi. They then fine-tune a single mT5-base model with short tags prepended to the input. The tags name the target region and the context or register, so the model learns to produce region-appropriate wording on request. NLLB scores higher on BLEU (13.75 vs. 8.19), but mostly because it falls back to MSA. An LLM-assisted judgment of cultural authenticity favours the new model strongly (4.80/5 vs. 1.0/5). The authors name this gap the "Accuracy Paradox": n-gram metrics computed against MSA-heavy references penalise faithful dialectal output. The work argues for dialect-aware evaluation and for treating variety and register as explicit, user-controlled variables.

Key points

  • RBDA turns a 3,000-sentence seed into a balanced corpus of about 57,000 sentences (57,600 in the limitations section) across eight Arabic varieties. It works mainly through lexical substitution.
  • Translation is modelled as P(Y | X, C), where C is a control vector of dialect, domain and register. C is realised as prepended tokens, e.g. [EGYPTIAN][MEDICAL].
  • One multilingual mT5-base model serves all varieties at once, sharing parameters across dialects with different roots.
  • Results:
    • BLEU: NLLB 13.75, the proposed model 8.19.
    • LLM-rated authenticity: 4.80/5 for the proposed model vs. 1.0/5 for the baseline.
    • Per-region view: baselines score near zero on dialects such as Moroccan, while the fine-tuned model stays in a steady band.
  • Additional deliverables and caveats:
    • A Gradio web interface lets users pick region and register live.
    • Acknowledged limits are shallow morphological and syntactic coverage, artifacts from synthetic data, non-standard orthography and Arabizi, and an LLM judge that cannot replace human evaluation.

Interesting ideas

  • I1. The target variety becomes an explicit conditioning signal, a few tags in front of the input, rather than a separate model per variety, so one network can be steered at inference time. — For Mova (back): Generation from the meaning graph could take prepended register/style tags (formal, child-level, dialect) so the same graph yields controllable surface text, and the chosen tag stays visible as part of the explanation.
  • I2. A small hand-made seed is multiplied into a large balanced corpus by rule-based substitution, giving every variety equal weight instead of letting the dominant one swamp training. — For Mova (learn): Silver data for rare constructions could be balanced the same way, by applying controlled lexical rewrites to verified seed sentences so that rare classes are not drowned out by frequent ones.
  • I3. The "Accuracy Paradox" shows that an overlap metric scored against a biased reference set can reward the system that ignores the requested condition. — For Mova (learn): Evaluation should include a check of whether the output actually honours the requested condition, beyond overlap with references; a negative control (the same input with a different tag) would reveal systems that score well while ignoring the tag.
  • I4. Breaking results down per variety (radar-style per-region scores) exposes collapse that a single aggregate number hides. — For Mova (answer): Test scores for readers should be reported per text genre or problem type, so that a high average cannot mask near-zero performance on, say, fables or multi-step word problems.
  • I5. Substituting dialect-specific lexical items onto a shared source is a cheap first approximation to dialect conversion, with known limits on morphology and syntax. — For Mova (tree): When converting treebanks between UD dialects or annotation conventions, a lexical/label substitution pass could serve as a transparent first stage, with structural rewrites added only where substitution demonstrably fails.

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: Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2604.06456 · arXiv 2604.06456

@misc{nowshin2026context,
  title = {{Context-Aware Dialectal Arabic Machine Translation with Interactive Region and Register Selection}},
  author = {Nowshin, Afroza and Porag, Prithweeraj Acharjee and Jeelani, Haziq and Syed, Fayeq Jeelani},
  year = {2026},
  doi = {10.48550/arxiv.2604.06456},
  eprint = {2604.06456},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2604.06456}
}

References

30 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed Abdelali, Hamdy Mubarak, Shammur Chowdhury, et al. LAraBench: Benchmarking Arabic AI with large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, pages 487–520, 2024. doi: 10.18653/v1/2024.eacl-long.30. URL https://aclanthology.org/2024.eacl-long.30. ↗ our card
  2. Muhammad Abdul-Mageed, Amr Keleg, et al. NADI 2024: The fifth nuanced Arabic dialect identification shared task. In Proceedings of the Second Arabic Natural Language Processing Conference, pages 709–728, 2024. doi: 10.18653/v1/2024.arabicnlp-1.79. URL https://aclanthology.org/ 2024.arabicnlp-1.79. ↗
  3. Kathrein Abu Kwaik, Motaz Saad, et al. Shami: A corpus of Levantine Arabic dialects. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, 2018. URL https://aclanthology.org/L18-1576. ↗
  4. Kylian AI. Arabic dialects: A guide to regional arabic varieties, 2025. URL https://www.kylian.ai/blog/en/arabic-dialects. Accessed: 13 Dec 2025. ↗
  5. Rawan Al-Matham, Kareem Darwish, Raghad Al-Rasheed, et al. BALSAM: A platform for benchmarking arabic large language models. arXiv preprint arXiv:2507.22603, 2025. URL https://arxiv.org/abs/2507.22603. ↗
  6. Laila Abdullah Al Suwaiyan. Diglossia in the arabic language. International Journal of Language and Linguistics, 5(3):228–238, 2018. doi: 10.30845/ijll.v5n3p22. URL https://doi.org/10. 30845/ijll.v5n3p22. ↗
  7. Abdullah Alabdullah, Lifeng Han, and Chenghua Lin. Advancing dialectal arabic to modern standard arabic machine translation. arXiv preprint arXiv:2507.20301, 2025. URL https://arxiv.org/ abs/2507.20301. ↗ our card
  8. Yazeed Alnumay, Alexandre Barbet, Anna Bialas, et al. Command r7b arabic: A small, enterprise focused, multilingual, and culturally aware arabic llm. arXiv preprint arXiv:2503.14603, 2025. URL https://arxiv.org/abs/2503.14603. ↗
  9. Arwa Alsindy. Language Variation: Arabic Dialects in Madinah, Saudi Arabia. PhD thesis, The University of Mississippi, 2024. URL https://egrove.olemiss.edu/etd/3001. ↗
  10. Fakhraddin Alwajih, Abdellah El Mekki, et al. Palm: A culturally inclusive and linguistically diverse dataset for arabic llms. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, pages 32871–32894, 2025. URL https://github.com/UBC-NLP/palm. ↗
  11. Shengnan An, Zexiong Yang, Luoru Zheng, Dongdong Zhao, Rui Song, Bo Hu, Zuhao Chen, and Ming Zhou. Make your LLM fully utilize the context, 2024. URL https://arxiv.org/abs/2404. 16811. arXiv preprint arXiv:2404.16811. ↗
  12. Julia Baronova, Catherine Stevens, Luke Tennant, and Alistair MacPhee. Dynamic context-aware representation for semantic alignment in large language models, 2024. URL https://osf.io/ preprints/osf/svcn3. OSF Preprints. ↗
  13. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, et al. The MADAR arabic dialect corpus and lexicon. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, 2018. URL https://aclanthology.org/L18-1535. ↗
  14. Houda Bouamor, Sabit Hassan, and Nizar Habash. The MADAR shared task on arabic fine-grained dialect identification. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 199–207, 2019. URL https://aclanthology.org/W19-4622. ↗ our card
  15. Junxuan Chen, Xiang Li, Jiarui Zhang, Chulun Zhou, Jianwei Cui, Bin Wang, and Jinsong Su. Modeling discourse structure for document-level neural machine translation. In Proceedings of the 1st Workshop on Automatic Simultaneous Translation, pages 30–36, Seattle, WA, USA, 2020. doi: 10.18653/v1/2020. autosimultrans-1.4. URL https://aclanthology.org/2020.autosimultrans-1.4. ↗
  16. Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672, 2022. URL https://arxiv.org/abs/2207.04672. ↗
  17. William M. Cotter and Rudolf de Jong. Regional variation. In The Routledge Handbook of Arabic Sociolinguistics, chapter 1. Routledge, 2019. doi: 10.4324/9781315722450-4. URL https://doi.org/10.4324/9781315722450-4. ↗
  18. Discover Discomfort. Arabic dialects compared, 2019. URL https://discoverdiscomfort.com/arabic-dialects-maghrebi-egyptian-levantine-gulf-hejazi-msa/. ↗
  19. Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, et al. Fanar: An arabic-centric multimodal generative ai platform. arXiv preprint arXiv:2501.13944, 2025. URL https://arxiv.org/abs/ 2501.13944. ↗ our card
  20. Strommen Inc. Arabic dialects explained: A complete guide for language learners, 2025. URL https://strommeninc.com/ arabic-dialects-explained-a-complete-guide-for-language-learners/. Accessed: 13 Dec 2025. ↗
  21. Amr Keleg, Sharon Goldwater, and Walid Magdy. ALDi: Quantifying the arabic level of dialectness of text. arXiv preprint arXiv:2310.13747, 2023. URL https://arxiv.org/abs/2310.13747. ↗
  22. Fajri Koto, Haonan Li, Sara Shatnawi, et al. ArabicMMLU: Assessing massive multitask language understanding in Arabic. In Findings of the Association for Computational Linguistics: ACL 2024, pages 5622–5640, 2024. doi: 10.18653/v1/2024.findings-acl.334. URL https://aclanthology.org/2024.findings-acl.334. ↗ our card
  23. Sangmin-Michelle Lee. A systematic review of context-aware technology use in foreign language learning. Computer Assisted Language Learning, 35(3):294–318, 2022. doi: 10.1080/09588221.2019.1688836. URL https://doi.org/10.1080/09588221.2019.1688836. ↗
  24. Shafqat Maqsood, Abdul Shahid, Fatima Nazar, Muhammad Asif, Muhammad Ahmad, Manuel Mazzara, and Salvatore Distefano. C-POS: A context-aware adaptive part-of-speech language learning framework. IEEE Access, 8:58320–58334, 2020. doi: 10.1109/ACCESS.2020.2982425. URL https://doi.org/10.1109/ACCESS.2020.2982425. ↗
  25. Nina Markl. Language variation and algorithmic bias: understanding algorithmic bias in british english automatic speech recognition. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 521–534, 2022. doi: 10.1145/3531146.3533117. URL https://doi.org/10.1145/3531146.3533117. ↗
  26. Sameen Maruf, Fahimeh Saleh, and Gholamreza Haffari. A survey on document-level neural machine translation: Methods and evaluation. ACM Computing Surveys, 54(2):1–36, 2021. doi: 10.1145/ 3412844. URL https://doi.org/10.1145/3412844. ↗
  27. Omar Nassra. Arabic dialects: Different types of arabic language, 2025. URL https://www. omarnassramethod.com/blog/Learn%20Arabic:%20Mastering%20’I%20Want’. ↗
  28. Tatoeba Project. Tatoeba project english to arabic datasets, 2025. URL https://tatoeba.org/ en/. ↗
  29. Yu-Hsiang Tseng, Cing-Feng Shih, Pin-Er Chen, Hsin-Yu Chou, Mei-Chen Ku, and Sheng-Kai Hsieh. CxLM: A construction and context-aware language model. In Proceedings of the 13th Language Resources and Evaluation Conference, pages 6361–6368, 2022. URL https://aclanthology.org/2022.lrec-1.683. ↗
  30. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Abheek Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483–498, 2021. doi: 10.18653/v1/2021.naacl-main.41. URL https://aclanthology.org/2021.naacl-main.41. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nowshin-2026-context-aware-dialectal-arabic-machine/