--- title: "SHAMI-MT: A Syrian Arabic Dialect to Modern — Sibaee 2025" description: "Arabic has a deep split between Modern Standard Arabic (MSA), used in writing and formal settings, and the regional dialects people actually speak." authors: ["Serry Sibaee", "Omer Nacar", "Yasser Al-Habashi", "Adel Ammar", "Wadii Boulila"] year: 2025 venue: "arXiv:2508.02268" source: "https://arxiv.org/abs/2508.02268" doi: "10.48550/arxiv.2508.02268" arxiv: "2508.02268" licence: "CC BY-SA 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by-sa/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/sibaee-2025-shami-mt-syrian-arabic-dialect/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/sibaee-2025-shami-mt-syrian-arabic-dialect/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/) / SHAMI-MT: A Syrian Arabic Dialect to Modern Standard Arabic Bidirectional Machine Translation System # SHAMI-MT: A Syrian Arabic Dialect to Modern Standard Arabic Bidirectional Machine Translation System **Authors:** [Serry Sibaee](https://md.research.elamur.ai/free/en/authors/serry-sibaee/), [Omer Nacar](https://md.research.elamur.ai/free/en/authors/omer-nacar/), [Yasser Al-Habashi](https://md.research.elamur.ai/free/en/authors/yasser-al-habashi/), [Adel Ammar](https://md.research.elamur.ai/free/en/authors/adel-ammar/), [Wadii Boulila](https://md.research.elamur.ai/free/en/authors/wadii-boulila/) · **Year:** 2025 · **Venue:** arXiv:2508.02268 · **Licence:** [CC BY-SA 4.0 (arXiv)](http://creativecommons.org/licenses/by-sa/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2508.02268) · [PDF](https://arxiv.org/pdf/2508.02268) · [DOI 10.48550/arxiv.2508.02268](https://doi.org/10.48550/arxiv.2508.02268) · [arXiv 2508.02268](https://arxiv.org/abs/2508.02268) · [OpenAlex](https://openalex.org/W4417103126) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) · 2 in our library, of the 140884 papers whose bibliographies we read ## Summary Arabic has a deep split between Modern Standard Arabic (MSA), used in writing and formal settings, and the regional dialects people actually speak. Most Arabic NLP tools are trained on MSA, so they handle colloquial text poorly. The Syrian (Shami) dialect has had especially little attention. SHAMI-MT addresses this with two separate translation models, one for MSA→Shami and one for Shami→MSA. Both models fine-tune the Arabic-only encoder-decoder AraT5v2-base-1024 on the Nâbra corpus. Nâbra is a varied collection of Syrian Arabic drawn from social media, film and TV scripts, songs and proverbs, and it covers several cities. For testing, the authors held out 1,500 Damascus sentence pairs from MADAR. Instead of BLEU, they had GPT-4.1 grade each output on a 0–5 scale against both the source and the reference. The MSA→Shami model averaged 4.01. The models are released openly. They give an underserved language pair a usable tool for localization, research on dialect text, and education. The authors also show that a modest, specialised model can be competitive. ## Key points - The base model is UBC-NLP/AraT5v2-base-1024, a T5-style text-to-text model pretrained only on Arabic, with sequences up to 1024 tokens. - Training ran for 22 epochs (10,384 steps) with batch size 256 and a cosine schedule starting at 5e-5. Final training loss was 1.396 and evaluation loss was 0.771. - The blind test set is 1,500 Damascus-subset pairs from MADAR. GPT-4.1 scored each output for semantic accuracy, dialectal authenticity and fluency. MSA→Shami reached an average of 4.01 out of 5. - Both models are on Hugging Face as Shami-MT and SHAMI-MT-2MSA, part of the SHAMIYAT collection for Syrian-dialect NLP. - The error analysis found two causes of low scores. Some came from mismatched or reinterpreted references. Others came from genuine model trouble with very short idiomatic inputs, where the model chose a literal or formal phrasing. - Planned next steps are other Levantine dialects, LoRA fine-tuning, and a speech-to-speech pipeline. ## Interesting ideas - The LLM judge sees the source, the system output and the reference together, and scores against named criteria rather than surface overlap, so valid paraphrases are not penalised. — For Mova (learn): A rubric-driven teacher judge could grade generated text or frame extractions on separate axes, such as meaning kept versus naturalness, instead of a single string-match score. - Low-scoring cases are sorted into two kinds: reference faults (the gold is inconsistent or reinterprets the source) and true model faults. — For Mova (learn): When silver parses disagree with treebank gold, a triage step could label each case as "gold suspect" or "parser wrong" before training or writing repair rules. - The reverse model is meant to standardise dialect text into the formal variety so that existing tools can analyse it. — For Mova (tokens): A normalisation pass could map colloquial or social-media English toward standard forms, while keeping the original tokens alongside, before the tagger and lemmatizer run. - Very short, highly idiomatic inputs are flagged as a distinct failure class, where literal rendering misses the conventional meaning. — For Mova (frames): One-word replies and idioms could be routed to a lookup of conventional speech acts rather than compositional frame building. - The training corpus deliberately mixes genres (scripts, lyrics, proverbs, posts) and sub-regions to expose the model to real usage variation. — For Mova (learn): Silver data could be balanced across genres such as fables, dialogue and news, so that witness parsers are not tuned to one register. *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](https://md.research.elamur.ai/free/en/shelves/linguistics/) Topics: [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{sibaee2025shami, title = {{SHAMI-MT: A Syrian Arabic Dialect to Modern Standard Arabic Bidirectional Machine Translation System}}, author = {Sibaee, Serry and Nacar, Omer and Al-Habashi, Yasser and Ammar, Adel and Boulila, Wadii}, year = {2025}, doi = {10.48550/arxiv.2508.02268}, eprint = {2508.02268}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2508.02268} } ``` - [← previous SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training](https://md.research.elamur.ai/free/en/papers/chu-2025-sft-memorizes-rl-generalizes-comparative/) - [next → SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild](https://md.research.elamur.ai/free/en/papers/zeng-2025-simplerl-zoo-investigating-taming-zero/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/) ## References 14 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. M. Enis and M. Hopkins, “From LLM to NMT: Advancing low-resource machine translation with claude,” 2024, arXiv:2404.13813. [↗](https://arxiv.org/abs/2404.13813) 2. Saeed, A. M. A. (2025). Machine Translation Evaluation between Arabic and English during 2020 to 2024: A Review Study. Arts for Linguistic & Literary Studies, 7(2), 665-678. [find](https://scholar.google.com/scholar?q=Saeed%2C+A%2E+M%2E+A%2E+%282025%29%2E+Machine+Translation+Evaluation+between+Arabic+and+English+during+2020+to+2024%3A+A+Review+Study%2E+Arts+for+Linguistic+%26+Literary+Studies%2C+7%282%29%2C+665%2D678%2E) 3. Fatiha Sadat, Farnazeh Kazemi, and Atefeh Farzindar. 2014. Automatic identification of arabic dialects in social media. SoMeRA ’14, page 35–40, New York, NY, USA. Association for Computing Machinery. [find](https://scholar.google.com/scholar?q=Fatiha+Sadat%2C+Farnazeh+Kazemi%2C+and+Atefeh+Farzindar%2E+2014%2E+Automatic+identification+of+arabic+dialects+in+social+media%2E+SoMeRA+%E2%80%9914%2C+page+35%E2%80%9340%2C+New+York%2C+NY%2C+USA%2E+Association+for+Computing+Machinery%2E) 4. Libovický, J., Rosa, R., & Fraser, A. (2019). How language-neutral is multilingual BERT?. arXiv preprint arXiv:1911.03310. [↗](https://arxiv.org/abs/1911.03310) 5. Xue, L., Constant, N., Roberts, A., Kale, M., Al-Rfou, R., Siddhant, A., ... & Raffel, C. (2020). mT5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934. [↗](https://arxiv.org/abs/2010.11934) 6. Nacar, O., Sibaee, S., Alharbi, A., Ghouti, L., & Koubaa, A. (2024, August). ASOS at NADI 2024 shared task: Bridging Dialectness Estimation and MSA Machine Translation for Arabic Language Enhancement. In Proceedings of The Second Arabic Natural Language Processing Conference (pp. 748-753). [find](https://scholar.google.com/scholar?q=Nacar%2C+O%2E%2C+Sibaee%2C+S%2E%2C+Alharbi%2C+A%2E%2C+Ghouti%2C+L%2E%2C+%26+Koubaa%2C+A%2E+%282024%2C+August%29%2E+ASOS+at+NADI+2024+shared+task%3A+Bridging+Dialectness+Estimation+and+MSA+Machine+Translation+for+Arabic+Language+Enhancement%2E) [our card](https://md.research.elamur.ai/free/en/papers/nacar-2024-asos-nadi-2024-shared-task/) 7. El Moatez Billah Nagoudi, AbdelRahim Elmadany, and Muhammad Abdul-Mageed. AraT5: Text-to-Text Transformers for Arabic Language Generation. arXiv preprint arXiv:2109.12068v4, 2022. [↗](https://arxiv.org/abs/2109.12068) [our card](https://md.research.elamur.ai/free/en/papers/nagoudi-2022-arat5-text-text-transformers-arabic/) 8. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., ... & Liu, P. J. (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140), 1-67. [find](https://scholar.google.com/scholar?q=Raffel%2C+C%2E%2C+Shazeer%2C+N%2E%2C+Roberts%2C+A%2E%2C+Lee%2C+K%2E%2C+Narang%2C+S%2E%2C+Matena%2C+M%2E%2C+%2E%2E%2E+%26+Liu%2C+P%2E+J%2E+%282020%29%2E+Exploring+the+limits+of+transfer+learning+with+a+unified+text%2Dto%2Dtext+transformer%2E+Journal+of+machine+le) 9. Amal Nayouf, Tymaa Hasanain Hammouda, Mustafa Jarrar, Fadi A Zaraket, and Mohamad-Bassam Kurdy. Nâbra: Syrian Arabic dialects with morphological annotations. arXiv preprint arXiv:2310.17315, 2023. [↗](https://arxiv.org/abs/2310.17315) [our card](https://md.research.elamur.ai/free/en/papers/aclbulk-2023-nayouf-nabra-syrian-arabic-dialects-morphological/) 10. Bouamor, H., Habash, N., Salameh, M., Zaghouani, W., Rambow, O., Abdulrahim, D., ... & Oflazer, K. (2018, May). The MADAR Arabic dialect corpus and lexicon. In Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018). [find](https://scholar.google.com/scholar?q=Bouamor%2C+H%2E%2C+Habash%2C+N%2E%2C+Salameh%2C+M%2E%2C+Zaghouani%2C+W%2E%2C+Rambow%2C+O%2E%2C+Abdulrahim%2C+D%2E%2C+%2E%2E%2E+%26+Oflazer%2C+K%2E+%282018%2C+May%29%2E+The+MADAR+Arabic+dialect+corpus+and+lexicon%2E+In+Proceedings+of+the+eleventh+internationa) 11. Salam Khalifa, Nizar Habash, and Houda Bouamor. A Morphologically-Aware Fine-Grained Arabic to English SMT System. In Proceedings of the Second Workshop on Arabic Corpora and Processing Tools, pages 58–65, 2016. [find](https://scholar.google.com/scholar?q=Salam+Khalifa%2C+Nizar+Habash%2C+and+Houda+Bouamor%2E+A+Morphologically%2DAware+Fine%2DGrained+Arabic+to+English+SMT+System%2E+In+Proceedings+of+the+Second+Workshop+on+Arabic+Corpora+and+Processing+Tools%2C+pages+5) 12. Ahmed Abdelali, Sabit Hassan, Hamdy Mubarak, Kareem Darwish, and Younes Samih. Pre-training Bert on Arabic Tweets: Practical Considerations. arXiv preprint arXiv:2102.10684, 2021. [↗](https://arxiv.org/abs/2102.10684) [our card](https://md.research.elamur.ai/free/en/papers/abdelali-2021-pre-training-bert-arabic-tweets/) 13. Marwah Alian, Arafat Awajan, Ahmad Al-Hasan, and Raeda Akuzhia. Towards building arabic paraphrasing benchmark. In Proceedings of the Second International conference on Data Science E-learning and Information Systems, pages 1–5, 2019. [find](https://scholar.google.com/scholar?q=Marwah+Alian%2C+Arafat+Awajan%2C+Ahmad+Al%2DHasan%2C+and+Raeda+Akuzhia%2E+Towards+building+arabic+paraphrasing+benchmark%2E+In+Proceedings+of+the+Second+International+conference+on+Data+Science+E%2Dlearning+and+Inf) 14. Serry Sibaee and Omer Nacar. Shami-MT : A Machine Translation from MSA to Syrian Dialect. 2025. [find](https://scholar.google.com/scholar?q=Serry+Sibaee+and+Omer+Nacar%2E+Shami%2DMT+%3A+A+Machine+Translation+from+MSA+to+Syrian+Dialect%2E+2025%2E)