Elamur Research⌕
☰
home / papers / 2026 / Aladdin-FTI @ AMIYA Three Wishes for Arabic NLP: Fidelity, Diglossia, and Multidialectal Generation

Aladdin-FTI @ AMIYA Three Wishes for Arabic NLP: Fidelity, Diglossia, and Multidialectal Generation

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/2026.vardial-1.27.pdf

Summary

Dialectal Arabic is still poorly served by NLP. Models trained mostly on MSA tend to flatten dialect features or slip back into MSA. For the AMIYA shared task at VarDial 2026 (closed data track, official training data only), team Aladdin-FTI fine-tuned a single LLM on two instruction-formatted objectives:

  • Translation: MSA↔DA and DA↔English, which sharpens the MSA/dialect distinction while keeping meaning intact.
  • Sentence completion: finishing a sentence in a named dialect from a three-word prefix, which pushes the model toward dialectal output.

The two losses are mixed with a weight λ. Evaluation follows the AL-QASIDA protocol, using ChrF++ for diglossia and Macro ADI2 for dialectal fidelity. Translation-only training helps diglossia but not fidelity, and generation-only training does the reverse. The joint setup improves both. With LoRA, a jointly trained Llama-3.1-8B-Instruct reaches fidelity close to GPT-OSS-120B. This suggests that how training objectives are designed can partly replace sheer model size for dialect control. The code and models are publicly released.

Key points

  • Dialects covered: Moroccan, Egyptian, Palestinian, Syrian and Saudi Arabic, with generation and two-way translation involving MSA and English. The completion templates draw on MADAR training data.
  • Objective weighting: λ ∈ {0, 0.5, 1}, where 0 means generation only, 1 means MT only, and 0.5 is the joint MT+GEN setting. Only assistant tokens contribute to the causal LM loss.
  • SmolLM3-3B (contrastive run): comparing the base model with the joint model, ChrF++ goes from 22.23 to 33.65 and Macro ADI2 from 0.003 to 0.067.
  • Llama-3.1-8B-Instruct (primary run, LoRA): ChrF++ goes from 32.99 to 35.09 and Macro ADI2 from 0.065 to 0.233. For comparison, GPT-OSS-120B scores 47.82 / 0.237 and Command R Arabic scores 46.60 / 0.053.
  • Availability and limitations: code is on GitHub (drvenabili/mt-finetune_amiya) and models are in the Hugging Face collection unige-fti/aladdin-fti-amiya. The study uses one architecture per experiment, does not test few-shot or in-context learning, and includes no human evaluation yet.

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

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.vardial-1.27

@inproceedings{mutal2026aladdin,
  title = {{Aladdin-FTI @ AMIYA Three Wishes for Arabic NLP: Fidelity, Diglossia, and Multidialectal Generation}},
  author = {Mutal, Jonathan and Almaoui, Perla Al and Hengchen, Simon and Bouillon, Pierrette},
  booktitle = {Proceedings of the 13th Workshop on NLP for Similar Languages, Varieties and Dialects},
  year = {2026},
  doi = {10.18653/v1/2026.vardial-1.27},
  url = {https://doi.org/10.18653/v1/2026.vardial-1.27}
}

References

42 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 2023. ASREgArbCSC: An Egyptian Arabic conversa tional speech corpus. Opensource dataset consisting of 5.5 hours of transcribed Egyptian Arabic conversa tional speech across nine twospeaker conversations. the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). find
  2. Azzedine Aftiss, Salima Lamsiyah, Christoph Schom mer, and Said Ouatik El Alaoui. 2025. Empirical evaluation of pretrained language models for summa rizing moroccan darija news articles. In Proceedings of the 4th Workshop on Arabic Corpus Linguistics (WACL4), pages 77–85. find
  3. Perla Al Almaoui, Pierrette Bouillon, and Simon Hengchen. 2025. Arabizi vs LLMs: Can the genie understand the language of aladdin? In Proceedings of Machine Translation Summit XX: Volume 2, pages 28–41, Geneva, Switzerland. European Association for Machine Translation. ↗
  4. Mohammad AlFetyani, Muhammad AlBarham, Gheith Abandah, Adham Alsharkawi, and Maha Dawas. 2021. Masc: Massive arabic speech corpus. find
  5. Karim AlHaff, Mustafa Jarrar, Tymaa Hammouda, and Fadi Zaraket. 2022. Curras + baladi: Towards a Lev antine corpus. In Proceedings of the Thirteenth Lan guage Resources and Evaluation Conference, pages 769–778, Marseille, France. European Language Re sources Association. find
  6. Naif Alanazi, Mohammed AlBatineh, and Hussein Abu Rayyash. 2025. Saudial: The saudi arabic dialects game localization dataset. Data in Brief, 62:111906. find
  7. Meshrif Alruily. 2018. Saudi tweets dataset. find
  8. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. Unsupervised statistical machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3632–3642, Brussels, Belgium. Association for Com putational Linguistics. find our card
  9. Hanin Atwany, Nour Rabih, Ibrahim Mohammed, Abdul Waheed, and Bhiksha Raj. 2024. OSACT 2024 task 2: Arabic dialect to MSA translation. In Proceedings of the 6th Workshop on OpenSource Arabic Corpora and Processing Tools (OSACT) with Shared Tasks on Arabic LLMs Hallucination and Dialect to MSA Machine Translation @ LRECCOLING 2024, pages 98–103, Torino, Italia. ELRA and ICCL. find
  10. Gheith Abandah, Ashraf Suyyagh, Iyad Jafar, Mo hammad AbdelMajeed, Rabie Otoum, Shorouq AlAwawdeh, and Moath Khaleel. 2025. JODA: A dataset of jordanian dialect and erroneous modern arabic sentences coupled with proper MSA and full diacritics. find
  11. Elie Bakouch, Loubna Ben Allal, Anton Lozhkov, Noua mane Tazi, Lewis Tunstall, Carlos Miguel Patiño, Edward Beeching, Aymeric Roucher, Aksel Joonas Reedi, Quentin Gallouédec, Kashif Rasul, Nathan Habib, Clémentine Fourrier, Hynek Kydlicek, Guil herme Penedo, Hugo Larcher, Mathieu Morlon, Vaib hav Srivastav, Joshua Lochner, and 4 others. 2025. SmolLM3: smol, multilingual, longcontext reasoner. https://huggingface.co/blog/smollm3. ↗
  12. Kathrein Abu Kwaik, Motaz Saad, Stergios Chatzikyr iakidis, and Simon Dobnik. 2018. Shami: A cor pus of Levantine Arabic dialects. In Proceedings of Hassan Barmandah. 2025. Saudidialectallam: Lora finetuning for dialectal arabic generation. Preprint, arXiv:2508.13525. ↗
  13. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdul rahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and Kemal Oflazer. 2018. The MADAR Arabic dialect corpus and lexicon. In Pro ceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources As sociation (ELRA). find
  14. Abdelaziz Bounhar and Abdeljalil El Majjodi. 2025. At laset dataset for moroccan darija: From data collec tion, analysis, to model trainings. Hugging Face Blog. find
  15. Mahmoud ElHaj. 2020. Habibi: A multidialect arabic song lyrics corpus. In Proceedings of the Language Resources and Evaluation Conference (LREC). find
  16. Shereen ElSayed and Mona Farouk. 2020. Gender iden tification for egyptian arabic dialect in twitter using deep learning models. Egyptian Informatics Journal, 21. find
  17. Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Mak ing pretrained language models better fewshot learn ers. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan guage Processing (Volume 1: Long Papers), pages 3816–3830, Online. Association for Computational Linguistics. find
  18. Nizar Habash, Ryan Roth, Owen Rambow, Ramy Es kander, and Nadi Tomeh. 2013. Morphological anal ysis and disambiguation for dialectal Arabic. In Pro ceedings of the 2013 Conference of the North Amer ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 426–432, Atlanta, Georgia. Association for Computa tional Linguistics. find
  19. Mohamed Motasim Hamed, Muhammad Hreden, Khalil Hennara, Zeina Aldallal, Sara Chrouf, and Safwan AlModhayan. 2025. Lahjawi: Arabic crossdialect translator. In Proceedings of the 4th Workshop on Arabic Corpus Linguistics (WACL4), pages 12–24, Abu Dhabi, UAE. Association for Computational Lin guistics. find
  20. Salima Harrat, Karima Meftouh, and Kamel Smaili. 2019. Machine translation for Arabic dialects (survey). Information Processing & Management, 56(2):262–273. find
  21. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan AllenZhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Lowrank adaptation of large language models. Preprint, arXiv:2106.09685. ↗
  22. Shereen Hussein, Mona Farouk, and ElSayed Hemayed. 2019. Gender identification of egyptian di alect in twitter. Egyptian Informatics Journal, 20(2):109–116. find
  23. Amr Keleg, Sharon Goldwater, and Walid Magdy. 2025. Revisiting common assumptions about Arabic di alects in NLP. In Proceedings of the 63rd Annual Meeting of the Association for Computational Lin guistics (Volume 1: Long Papers), pages 3309–3327, Vienna, Austria. Association for Computational Lin guistics. find
  24. Abdullah Khered, Youcef Benkhedda, and Riza Batista Navarro. 2025. A multitask learning approach to dialectal Arabic identification and translation to Mod ern Standard Arabic. In Proceedings of the First Workshop on Advancing NLP for LowResource Lan guages, pages 21–31, Varna, Bulgaria. INCOMA Ltd., Shoumen, Bulgaria. find
  25. Mateusz Krubiński, Hashem Sellat, Shadi Saleh, Adam Pospíšil, Petr Zemánek, and Pavel Pecina. 2023. Multiparallel corpus of North Levantine Arabic. In Proceedings of ArabicNLP 2023, pages 411–417, Sin gapore (Hybrid). Association for Computational Lin guistics. find
  26. Viet Lai, Chien Nguyen, Nghia Ngo, Thuat Nguyen, Franck Dernoncourt, Ryan Rossi, and Thien Nguyen. 2023. Okapi: Instructiontuned large language mod els in multiple languages with reinforcement learning from human feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Lan guage Processing: System Demonstrations, pages 318–327, Singapore. Association for Computational Linguistics. find
  27. Xun Liang, Hanyu Wang, Yezhaohui Wang, Shichao Song, Jiawei Yang, Simin Niu, Jie Hu, Dan Liu, Shunyu Yao, Feiyu Xiong, and Zhiyu Li. 2024. Con trollable text generation for large language models: A survey. Preprint, arXiv:2408.12599. ↗
  28. Kelly Marchisio, WeiYin Ko, Alexandre Berard, Théo Dehaze, and Sebastian Ruder. 2024. Understand ing and mitigating language confusion in LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6653–6677, Miami, Florida, USA. Association for Computational Linguistics. find
  29. Imane Momayiz, Aissam Outchakoucht, Omar Choukrani, and Ali Nirheche. 2024. Terjamabench: A culturally specific dataset for evaluating translation models for moroccan darija. find
  30. Jonathan Mutal, Raphael Rubino, and Pierrette Bouil lon. 2025. Factors affecting translation quality in incontext learning for multilingual medical domain. In Proceedings of the Tenth Conference on Machine Translation, pages 161–179, Suzhou, China. Associa tion for Computational Linguistics. find
  31. NLLB Team. 2022. No language left behind: Scaling humancentered machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL). find
  32. Aissam Outchakoucht and Hamza EsSamaali. 2024. The evolution of darija open dataset: Introducing version 2. Preprint, arXiv:2405.13016. ↗
  33. Maja Popović. 2015. chrF: character ngram Fscore for automatic MT evaluation. In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics. find
  34. Nathaniel Robinson, Shahd Abdelmoneim, Anjali Kan tharuban, Otba Alsboul, Salima Lamsiyah, Kelly Marchisio, and Kenton Murray. 2026. AMIYA shared task: Arabic Modeling In Your Accent at VarDial 2026. In Proceedings of the 13th Workshop on NLP for Similar Languages, Varieties and Dialects, Rabat, Morocco. Association for Computational Linguistics. find our card
  35. Nathaniel Robinson, Perez Ogayo, David R. Mortensen, and Graham Neubig. 2023. ChatGPT MT: Competi tive for high (but not low) resource languages. In Proceedings of the Eighth Conference on Machine Translation, pages 392–418, Singapore. Association for Computational Linguistics. find
  36. Nathaniel Romney Robinson, Shahd Abdelmoneim, Kelly Marchisio, and Sebastian Ruder. 2025. AL QASIDA: Analyzing LLM quality and accuracy sys tematically in dialectal Arabic. In Findings of the As sociation for Computational Linguistics: ACL 2025, pages 22048–22065, Vienna, Austria. Association for Computational Linguistics. find
  37. Toshiyuki Takezawa, Genichiro Kikui, Masahide Mizushima, and Eiichiro Sumita. 2007. Multilingual spoken language corpus development for communi cation research. In International Journal of Compu tational Linguistics & Chinese Language Processing, Volume 12, Number 3, September 2007: Special Issue on Invited Papers from ISCSLP 2006, pages 303–324. find
  38. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar ElShangiti, Hiba Zayed, Mo hamedou Cheikh Tourad, Rahaf Alhamouri, Rwaa Assi, and 1 others. 2024. Casablanca: Data and models for multidialectal arabic speech recognition. In Proceedings of the 2024 Conference on Empiri cal Methods in Natural Language Processing, pages 21745–21758. find
  39. Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Alpaca: A strong, replicable instructionfollowing model. Stanford Cen ter for Research on Foundation Models. https://crfm.stanford. edu/2023/03/13/alpaca. html, 3(6):7. ↗
  40. Taghreed Tarmom, William Teahan, Eric Atwell, and Mohammad Alsalka. 2014. Compression vs tradi tional machine learning classifiers to detect code switching in varieties and dialects: Arabic as a case study. Journal of Natural Language Processing. find
  41. Nikolai Vladimirovich Yushmanov. 1961. The structure of the arabic language. 2012. Machine translation of Arabic dialects. In Pro ceedings of the 2012 Conference of the North Amer ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 49–59, Montréal, Canada. Association for Computa tional Linguistics. find
  42. Zhihan Zhang, Wenhao Yu, Mengxia Yu, Zhichun Guo, and Meng Jiang. 2023. A survey of multitask learn ing in natural language processing: Regarding task relatedness and training methods. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 943–956, Dubrovnik, Croatia. Association for Com putational Linguistics. find

↑ back to top