Elamur Research⌕
☰
home / papers / 2026 / REIGNITE at AbjadMed: Imbalance-Aware Fine-Tuning of Pretrained Arabic Transformers for Arabic Medical Text Classification Task

REIGNITE at AbjadMed: Imbalance-Aware Fine-Tuning of Pretrained Arabic Transformers for Arabic Medical Text Classification Task

Paper · 2026 · Nahid Montasir Rifat and Foyez Ahmed Dewan

Summary

This paper describes the REIGNITE team's entry to AbjadNLP Shared Task 4, which assigns Arabic medical question–answer pairs to one of 82 categories. The main difficulty is extreme label skew: the 27,951 training examples range from 7 to over 600 per class.

The authors combine three steps:

  • Preprocessing: Arabic orthographic normalization.
  • Augmentation: synthetic data generated only for rare classes.
  • Training: fine-tuning several Arabic and multilingual transformers with an identical setup and a class-weighted focal loss.

The three strongest models, CAMeLBERT, AraBERT-base v02 and AraBERT-large v02, are combined by hard majority voting. The ensemble reaches a macro F1 of 0.4052 on the private leaderboard, placing 15th. This is above every single model, though the margin is small. Augmentation helped on the public split, but private-split results were mixed. No ablation was run, so the contribution of each component remains unclear.

The work is a practical, reproducible baseline for imbalanced Arabic domain classification. Code is on GitHub.

Key points

  • Data: the dataset (Nagarajan et al., 2025) has 27,951 training and 18,634 test samples across 82 categories. Roughly a third of the classes have under 100 training examples.
  • Augmentation: classes with fewer than 50 samples were expanded to 150 using three methods:
    • back-translation with Helsinki-NLP OPUS-MT Arabic↔English (~60%);
    • substitution from a hand-built Arabic medical synonym lexicon (~30%);
    • random word swaps (~10%).
  • Loss and hyperparameters:
    • focal loss with γ = 2.5;
    • inverse-frequency class weights clipped to [0.5, 10.0];
    • shared settings for all models: max length 384, batch size 16, learning rate 2e-5, up to 15 epochs, early-stopping patience 3.
  • Baselines without augmentation (macro F1, public/private):
    • AraBERT-base v02: 0.3718 / 0.3836
    • AraBERT-large v02: 0.3653 / 0.3974
    • CAMeLBERT: 0.3701 / 0.3856
    • MARBERTv2, XLM-RoBERTa and AraBERT-base v2 scored lower.
  • Augmentation and ensemble results:
    • Augmentation raised public scores by about 0.6–1.3 points.
    • On the private split, it helped AraBERT-base v02 (0.3925) but lowered CAMeLBERT and AraBERT-large v02.
    • The ensemble scored 0.4048 (public) and 0.4052 (private).
    • Due to GPU limits, only the last 18 layers of AraBERT-large v02 were fine-tuned.

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

🎓Rajshahi University of Engineering and TechnologyRajshahi · Bangladesh3 works

 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 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Nahid Montasir Rifat and Foyez Ahmed Dewan. 2026. REIGNITE at AbjadMed: Imbalance-Aware Fine-Tuning of Pretrained Arabic Transformers for Arabic Medical Text Classification Task. 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. DOI 10.18653/v1/2026.abjadnlp-1.19

@inproceedings{rifat2026reignite,
  title = {{REIGNITE at AbjadMed: Imbalance-Aware Fine-Tuning of Pretrained Arabic Transformers for Arabic Medical Text Classification Task}},
  author = {Nahid Montasir Rifat and Foyez Ahmed Dewan},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.19},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.19}
}

References

9 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sulaiman Aftan and Habib Shah. 2023. Using the arabert model for customer satisfaction classification of telecom sectors in saudi arabia. Brain Sciences, 13(1). ↗
  2. Salman Almutrash and Shadi Abudalfa. 2024. Comparative study on the efficiency of using palm and camelbert for arabic entity sentiment classification. Publisher Copyright: © 2024 1st Saudi Conference on Information Systems, SaudiCIS 2024. All rights reserved.; 1st Saudi Conference on Information Systems, SaudiCIS 2024 ; Conference date: 19-11-2024 Through 21-11-2024. find
  3. Abrar Alotaibi, Atta ur Rahman, Raheel Alhaza, Wala Alkhalifa, Narjes Alhajjaj, Atheer Alharthi, Dhai Abushoumi, Maryam Alqahtani, and Dania Alkhulaifi. 2022. Spam and sentiment detection in arabic tweets using marbert model. Mathematical Modelling of Engineering Problems. find
  4. 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, with a Shared Task on Offensive Language Detection, pages 9–15, Marseille, France. European Language Resource Association. ↗
  5. Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for NLP. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 968–988, Online. Association for Computational Linguistics.
  6. Pranav Gupta, Niranjan Kumar M, Balaji Nagarajan, Imed Zitouni, and Mo El-Haj. 2026. Abjadmed: Arabic medical text classification at abjadnlp 2026. In Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script (AbjadNLP 2026), co-located with the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), Rabat, Morocco.
  7. Balaji Nagarajan, Niranjan kumar M, and Pranav Gupta. 2025. Eacl 2026 abjad nlp shared task 4. https://kaggle.com/competitions/eacl-2026-abjad-nlp-shared-task-medical-text-classification-in-arabic. Kaggle. ↗
  8. Sayar Ghosh Roy, Ujwal Narayan, Tathagata Raha, Zubair Abid, and Vasudeva Varma. 2021. Leveraging multilingual transformers for hate speech detection. CoRR, abs/2101.03207. ↗
  9. Hongzhi Zhang and M Omair Shafiq. 2024. Survey of transformers and towards ensemble learning using transformers for natural language processing. Journal of big Data, 11(1):25. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rifat-2026-reignite-abjadmed-imbalance-aware-fine/