Elamur Research⌕
☰
home / papers / 2026 / L3IA-Subtask 1 at AraSentEval Shared Task: Multi-Dialect Arabic Sentiment Classification via a Transformer-Based Approach

L3IA-Subtask 1 at AraSentEval Shared Task: Multi-Dialect Arabic Sentiment Classification via a Transformer-Based Approach

Paper · 2026 · Mohamed M'haouach, Kaouthar Elyoussoufi, Hamza Alami et al.

Summary

The authors take on three-way sentiment classification (positive, negative, neutral) of Arabic dialect text for AraSentEval 2026 Subtask 1. The task is hard because dialects have no standard spelling and vary widely in vocabulary and morphology, and most earlier work targets Modern Standard Arabic. The data is the Multi-Dialect-Sent (MDS-3) corpus: 3,000 hotel-review sentences, translated into several dialects and labelled by native speakers. The pipeline is simple. Spelling variants are normalized, labels are mapped to integers, and the text is tokenized with each model's own tokenizer. The [CLS] representation then feeds a classifier. Five encoders are fine-tuned under one shared configuration: BERT, AraBERT, MARBERT, CAMeLBERT and DarijaBERT. MARBERT, which was pre-trained on both MSA and dialect data, comes out on top on the internal validation split. The official test score is much lower (micro-F1 75.96%), which suggests the validation split overestimates real performance. The work is useful mainly as a controlled baseline comparison of Arabic encoders on dialectal sentiment.

Key points

  • Dataset: MDS-3, with 3,000 hotel-review sentences carrying three sentiment labels. The texts were translated into multiple dialects and annotated by native speakers. The authors split it 85% train and 15% validation, stratified by label.
  • Shared fine-tuning setup for all models: AdamW, learning rate 2×10⁻⁵, weight decay 0.01, 5 epochs, batch size 16, cross-entropy loss. Training ran on Hugging Face Transformers with PyTorch in Google Colab.
  • Validation F1 by model:
    • MARBERT: 94.92 (accuracy 95.10%)
    • AraBERT: 92.54
    • DarijaBERT: 90.45
    • CAMeLBERT: 90.16
    • Multilingual/general BERT: 88.50
  • The best system scored a micro-F1 of 75.96% on the official test set, roughly 19 points below its validation F1.
  • Preprocessing is deliberately light: orthographic normalization, plus the repetition reduction mentioned in the abstract. Sentence classification uses the final [CLS] hidden state.

Interesting ideas

  • I1. Unify spelling variants and squash repeated characters before subword tokenization, which cuts sparsity in noisy, non-standardized text without discarding linguistic content.
  • I2. Compare several encoders under a single fixed hyperparameter configuration, so that differences reflect pre-training data rather than tuning effort.
  • I3. The large gap between the in-house validation split and the official test set shows that a random stratified split of a small, translation-derived corpus can badly overstate performance.
  • I4. Pre-training that matches the target variety (MARBERT's dialect coverage) beat both general and narrowly single-dialect models. This suggests that broad in-variety coverage matters more than either generality or a narrow focus.
  • I5. Building a multi-dialect dataset by translating one source domain into several varieties gives parallel, label-preserving data across dialects.

For Mova

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

  • I1 (tokens): A similar pre-tokenizer pass for informal English (e.g. "sooooo", variant spellings) could map tokens to canonical forms for the tagger and lemmatizer, while keeping the original surface form for explainability.
  • I2 (learn): When choosing a backbone for gparse or the witness parsers, freeze one training recipe across candidates so the comparison is attributable to the encoder itself.
  • I3 (learn): Evaluation should include genuinely held-out sources (different domains or authors) alongside random splits, and should report the gap as a sanity check next to negative controls.
  • I4 (tree): For UD parsing of non-standard English (social media, dialectal or spoken treebanks), prefer an encoder or silver data that covers many informal varieties over a single-dialect specialization.
  • I5 (back): Parallel paraphrases of the same content in different registers could test whether the text → graph → text compressor yields the same meaning graph regardless of surface variety.

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

🎓Sidi Mohamed Ben Abdellah UniversityFes · Morocco8 works · 156 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 works

Cite

Cite as Mohamed M'haouach, Kaouthar Elyoussoufi, Hamza Alami and Abdessamad Benlahbib. 2026. L3IA-Subtask 1 at AraSentEval Shared Task: Multi-Dialect Arabic Sentiment Classification via a Transformer-Based Approach. The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4tdk6bsmy4dj

@inproceedings{mhaouach2026l3ia,
  title = {{L3IA-Subtask 1 at AraSentEval Shared Task: Multi-Dialect Arabic Sentiment Classification via a Transformer-Based Approach}},
  author = {Mohamed M'haouach, Kaouthar Elyoussoufi, Hamza Alami and Abdessamad Benlahbib},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/4tdk6bsmy4dj},
  url = {https://doi.org/10.63317/4tdk6bsmy4dj}
}

References

12 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi. 2021. ARBERT & MARBERT: Deep bidirectional transformers for arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021), Volume 1: Long Papers, pages 7088–7105. Association for Computational Linguistics. ↗
  2. Omar Al-Harbi. 2019. Classifying sentiment of dialectal arabic reviews: A semi-supervised approach. International Arab Journal of Information Technology. find
  3. Maram Alharbi, Saad Ezzini, Hansi Hettiarachchi, Tharindu Ranasinghe, and Ruslan Mitkov. 2025a. Evaluating large language models on arabic dialect sentiment analysis. In Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing (RANLP). find
  4. Maram I Alharbi, Salmane Chafik, Saad Ezzini, Ruslan Mitkov, Tharindu Ranasinghe, and Hansi Hettiarachchi. 2025b. Ahasis: Shared task on sentiment analysis for arabic dialects. In Proceedings of the Shared Task on Sentiment Analysis for Arabic Dialects, pages 1–6.
  5. Maram I Alharbi, Saad Ezzini, Hansi Hettiarachchi, Tharindu Ranasinghe, and Ruslan Mitkov. 2025c. Ador: Dataset for arabic dialects in hotel reviews: A human benchmark for sentiment analysis. In Proceedings of the First Workshop on Advancing NLP for Low-Resource Languages, pages 187– 191.
  6. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based model for arabic language understanding. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT 2020), with a Shared Task on Offensive Language Detection, pages 9–15. European Language Resources Association (ELRA). ↗
  7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2019), pages 4171–4186, Minneapolis, Minnesota, USA. Association for Computational Linguistics.
  8. Saad Ezzini, Shadi Abudalfa, Maram Alharbi, Salmane Chafik, Samaher Alghamdi, Reem Alotaibi, Hamzah Luqman, Mo El-Haj, and Paul Rayson. 2026. AraSentEval: A shared task on sentiment analysis and swapping in arabic. In Proceedings of the 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7), co-located with the Language Resources and Evaluation Conference (LREC 2026), Palma, Mallorca, Spain. find
  9. Kamel Gaanoun, Abdou Mohamed Naira, Anass Allak, and Imade Benelallam. 2025. Darijabert: A step forward in nlp for the written moroccan dialect. International Journal of Data Science and Analytics, 20(2):917–929. ↗
  10. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The interplay of variant, size, and task type in arabic pre-trained language models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop (WANLP 2021), pages 92–104, Kyiv, Ukraine (Online). Association for Computational Linguistics.
  11. Youssef Matrane et al. 2023. A systematic literature review of arabic dialect sentiment analysis. Journal of King Saud University – Computer and Information Sciences. ↗
  12. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Jamie Brew, Canwen Huang, HuggingFace Inc., et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/haouach-2026-l3ia-subtask-1-arasenteval-shared/