--- title: "Arabic Dialect Translation with Small LLMs: Enhancing through" description: "Most machine translation for Arabic is built on Modern Standard Arabic. 1 reference. Among them: 0." authors: ["Sohaila Abdulsattar", "Keith Ross"] year: 2026 venue: "Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script" source: "https://aclanthology.org/2026.abjadnlp-1.11/" doi: "10.18653/v1/2026.abjadnlp-1.11" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/abdulsattar-2026-arabic-dialect-translation-small-llms/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/abdulsattar-2026-arabic-dialect-translation-small-llms/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/22/) / Arabic Dialect Translation with Small LLMs: Enhancing through Reasoning-Oriented Reinforcement Learning # Arabic Dialect Translation with Small LLMs: Enhancing through Reasoning-Oriented Reinforcement Learning **Authors:** Sohaila Abdulsattar and Keith Ross · **Year:** 2026 · **Venue:** [Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script](https://md.research.elamur.ai/free/en/venues/abjadnlp/) ([Workshop on NLP for Languages Using Arabic Script 2026](https://md.research.elamur.ai/free/en/venues/abjadnlp/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.abjadnlp-1.11/) · [PDF](https://aclanthology.org/2026.abjadnlp-1.11.pdf) · [DOI 10.18653/v1/2026.abjadnlp-1.11](https://doi.org/10.18653/v1/2026.abjadnlp-1.11) · [OpenAlex](https://openalex.org/W7140141731) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary Most machine translation for Arabic is built on Modern Standard Arabic. Everyday speech and much online text, however, use regional dialects that vary widely, have no fixed spelling, and lack large parallel corpora. The strongest systems are large proprietary models, which communities with limited compute or offline needs cannot easily use. The authors ask whether compact open models can close this gap. They take Qwen-2.5-base models with 1.5B and 3B parameters and train them on the MADAR corpus, which covers 25 city dialects plus MSA aligned with English. Training uses Group-Relative Policy Optimization with a SacreBLEU reward and an output format that asks for a thinking segment before the translation. They compare these models with Qwen-2.5-7B, both pretrained and supervised fine-tuned. The small RL-trained models beat the 7B baselines. Dialect→English BLEU roughly doubles, and English→dialect gains appear mainly in COMET. The "thinking" segments contain no real reasoning, so the improvement comes from optimizing directly against the reward. The work shows that offline-deployable models can handle a hard, low-resource translation task. ## Key points - Data: MADAR-26 configuration, with 2,000 aligned sentence pairs for each of 25 city dialects and MSA, 52,000 pairs in total. The pairs derive from English BTEC travel sentences and needed no extra cleaning beyond tokenization. - Method: GRPO applied to Qwen-2.5 1.5B/3B with a `……` template and a SacreBLEU reward, plus an SFT warmup where needed. Baselines are Qwen-2.5-7B, pretrained and fine-tuned with next-token prediction on MADAR-26. - Dialect→English: BLEU rises from 17.4 (7B SFT) to 34.9 with the small RL models. - English→dialect: BLEU stays close to SFT (10.70 for 1.5B, 11.31 for 3B), but COMET climbs from 0.57 to 0.71–0.73. For example, RIY moves from 13.19→13.69 BLEU and from 0.53→0.79 COMET. - Reasoning traces degenerate into placeholders, echoes of the prompt, or copies of the final translation. The gains are therefore attributed to fitting the reward, not to emergent reasoning. ## Interesting ideas - Using a bigger model of the same family as an "upper bound" turns a scale comparison into a clean test of whether a training method adds something that size alone does not. - The authors read the intermediate "thinking" text and found it empty, even though accuracy improved, which shows that a structured trace is not proof of a reasoning process. - Reporting a surface-overlap metric next to a meaning-level metric exposed cases where surface scores barely moved while adequacy rose sharply. - Group-relative optimization rewards each sampled output by how it compares with its siblings for the same input, which avoids training a separate value model. - Prior work and this paper both find that models understand non-standard varieties better than they produce them, so the two directions should be evaluated and budgeted separately. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (learn): Compare each new training regime of a small parser against a larger same-family parser, to show whether a gain comes from the method rather than from capacity. - **I2** (learn): Add a check that an explanation actually caused the result, for example by perturbing or removing the trace as a negative control. This guards against explanations that only decorate the output. - **I3** (back): Score the text → graph → text compressor and generation with both n-gram overlap and a semantic measure, so that faithful paraphrases are not counted as failures. - **I4** (learn): Sample several candidate parses or frame analyses per sentence and prefer those that beat the group average under witness agreement, giving a cheap relative signal for self-play. - **I5** (tree): When converting or parsing dialectal or informal English treebanks, report per-variety results and treat analysis and generation of the variety as distinct difficulties. *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 Dhabi](https://md.research.elamur.ai/free/en/universities/new-york-university-abu-dhabi/) — University, Abu Dhabi, United Arab Emirates · 122 works · 4,602 citations 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works 🧭 Topics (2): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works ## Cite Cite as: Sohaila Abdulsattar and Keith Ross. 2026. Arabic Dialect Translation with Small LLMs: Enhancing through Reasoning-Oriented Reinforcement Learning. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{abdulsattar2026arabic, title = {{Arabic Dialect Translation with Small LLMs: Enhancing through Reasoning-Oriented Reinforcement Learning}}, author = {Sohaila Abdulsattar and Keith Ross}, booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script}, year = {2026}, doi = {10.18653/v1/2026.abjadnlp-1.11}, url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.11} } ``` - [← previous Arabic Citation Parsing using Part of Speech and Named Entity Recognition](https://md.research.elamur.ai/free/en/papers/karout-2026-arabic-citation-parsing-part-speech/) - [next → Arabic Morphosyntactic Tagging and Dependency Parsing with Large Language Models](https://md.research.elamur.ai/free/en/papers/adel-2026-arabic-morphosyntactic-tagging-dependency-parsing/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/22/) ## References 1 references, 1 of them with a card of ours. ↗ leads to the work. 1. Muhammad Abdul-Mageed, Amr Keleg, AbdelRahim Elmadany, Chiyu Zhang, Injy Hamed, Walid Magdy, Houda Bouamor, and Nizar Habash. 2024. [NADI 2024: The fifth nuanced Arabic dialect identification shared task](https://md.research.elamur.ai/free/en/papers/mageed-2024-nadi-2024-fifth-nuanced-arabic/). In Proceedings of the Second Arabic Natural Language Processing Conference, pages 709–728, Bangkok, Thailand. Association for Computational Linguistics. *Updated 1 November 2026*