Elamur Research⌕
☰
home / papers / 2026 / A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation

A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation

Summary

The paper tackles Arabic medical answer generation. In this setting, ordinary fine-tuning treats every training pair as equally important, even though the cases range from trivial complaints to emergencies. The authors take a 32K-pair subset of the MAQA (Medical Arabic Question Answering) dataset and give each question one of three urgency labels: Mild, Moderate or Critical. They assign these labels with a rule-based tagger that normalizes the Arabic text and then searches it for symptom keywords. Training then follows a curriculum: the model is fine-tuned in successive stages, starting with mild cases and moving to moderate and then critical ones, with cross-entropy loss at each stage. The authors compare ten small Arabic-capable language models under three conditions: no tuning, standard fine-tuning and the curriculum. Every model scores higher with the curriculum, and the best result, 68.00%, comes from LFM2-700M. The authors argue that the gain comes from how the data is ordered rather than from any particular architecture. This makes the approach a cheap, model-agnostic way to improve medical generation in a low-resource language.

Key points

  • Data: a 32K question–answer subset of MAQA. Each question is labeled Mild, Moderate or Critical by keyword rules.
  • Labeling rules:
    • Preprocessing removes diacritics and punctuation and normalizes letter variants.
    • Emergency terms such as chest pain, fainting, heavy bleeding or breathing trouble mark a question as Critical.
    • Fever, vomiting or persistent pain mark it as Moderate.
    • When several cues fire, the most severe one wins.
    • Questions with no matching keyword default to Mild.
  • Models tested: AraGPT2-Base, AraGPT2-Medium, Bloomz-560M, Al-Atlas-0.5B, GPT2-Small-Arabic, LFM2-350M, LFM2-700M, Dialect-ar-gpt-2021, Qwen2.5-0.5B and Qwen3-0.6B.
  • Reported gains: the curriculum beats the baseline by about +4 to +7 points and standard fine-tuning by about +3 to +6, according to the authors. The table's own per-model differences are less tidy:
    • Bloomz-560M gains only about +2 over baseline.
    • AraGPT2-Base gains about +9 over baseline.
    • Standard fine-tuning alone adds about +5.7 for AraGPT2-Base, more than the stated ~2% ceiling.
  • Smaller models benefit most from the curriculum. Larger ones gain less but reach the top scores, with LFM2-700M at 68.00%.

Interesting ideas

  • I1. A crude rule-based difficulty label (a keyword lexicon, highest-severity-wins, a safe default) is enough to impose a curriculum without any human annotation. — For Mova (learn): order silver treebank sentences for the perceptron parser and gparse by a transparent difficulty proxy, such as length, non-projective arcs or witness-parser disagreement, and train in stages from easy to hard.
  • I2. The labeling resolves conflicts with an explicit priority order, so every label can be traced to the cue that triggered it. — For Mova (frames): in crate prag, speech-act or irony cue lexicons could use the same ranked "strongest cue wins" rule, so each tagged frame carries an explainable reason.
  • I3. The evaluation includes both an untuned baseline and a standard fine-tune on the same data, which isolates the effect of ordering alone. — For Mova (learn): add a negative control in which the same silver data is shuffled instead of staged, to show whether any curriculum gain is real.
  • I4. The observation that low-capacity models profit most from staged exposure suggests curricula matter most where model capacity is the bottleneck. — For Mova (tree): test the curriculum first on the small perceptron parser, where the payoff may be larger than for the GPU transformer gparse.
  • I5. The paper's stated gain ranges do not fully match its own table, which shows why deltas should be recomputed from the raw scores. — For Mova (learn): have evaluation reports compute per-model differences automatically from stored scores instead of quoting summary ranges by hand.

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 10.48550/arxiv.2604.06365 · arXiv 2604.06365

@misc{alansary2026severity,
  title = {{A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation}},
  author = {Alansary, Ahmed and Mohamed, Molham and Hamdi, Ali},
  year = {2026},
  doi = {10.48550/arxiv.2604.06365},
  eprint = {2604.06365},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2604.06365}
}

References

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

  1. 1. Abdelhay, M., Mohammed, A., Hefny, H.A.: Deep learning for arabic healthcare: Medicalbot. Social Network Analysis and Mining 13(1), 71 (2023) 2. Ahmed, S., Allam, A., Hamdi, A., Mohammed, A.: Arabic symptom classification and diagnosis using transformer models and llm-based augmentation. find
  2. IEEE, Egypt (July 2025) 3. Alkhurayyif, Y.: Enhancing medical knowledge access: Development and assessment of an open-domain medical question answering system with deep learning. In: 2025 4th International Conference on Computing and Information Technology (ICCIT). pp. 231–237. find
  3. IEEE (2025) 4. Allam, A., Ahmed, S., Hamdi, A., Mohammed, A.: Arabic large language models for medical text generation. find our card
  4. IEEE, Malaysia (February 2025) 5. Chang, E., Yeh, H.S., Demberg, V.: Does the order of training samples matter? improving neural data-to-text generation with curriculum learning. Association for Computational Linguistics (Apr 2021) 6. Garg, R., Gupta, A.: A systematic review of nlp applications in clinical healthcare: Advancement and challenges. In: International Conference on Advances in Data-driven Computing and Intelligent Systems. pp. 31–44. find
  5. Springer (2023) 7. Guo, J., Tan, X., Xu, L., Qin, T., Chen, E., Liu, T.Y.: Fine-tuning by curriculum learning for non-autoregressive neural machine translation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 7839–7846 (2020) 8. Hamdi, A., Mohamed, M., Emad, R., Shaban, K.: An ensemble classification approach in a multi-layered large language model framework for disease prediction. find our card
  6. IEEE, Qatar (October 2025) 9. Klila, J., Souihi, S.B., Boujelben, R., Semmar, N., Belguith, L.H.: Adapting large language models to biomedical domain: A survey of techniques and approaches. find
  7. Springer Nature Switzerland, Cham (June 2024) 10. Kumichev, G., Blinov, P., Kuzkina, Y., Goncharov, V., Zubkova, G., Zenovkin, N., Goncharov, A., Savchenko, A.: Medsyn: Llm-based synthetic medical text generation framework. find
  8. Springer (2024) 11. Liu, F., Ge, S., Wu, X.: Competence-based multimodal curriculum learning for medical report generation. Association for Computational Linguistics (Aug 2021) 12. Mohamed, M., Emad, R., Hamdi, A.: Ensemble transformers for arabic medical text classification. In: 2025 Intelligent Methods, Systems, and Applications (IMSA). pp. 1–6. find
  9. IEEE, Egypt (2025) 13. Mohamed, M., Emad, R., Hamdi, A.: A multi-layered large language model framework for disease prediction. In: International Congress on Information and Communication Technology. pp. 259–270. find
  10. Springer Nature Singapore, Singapore (2025) 14. Nazi, Z.A., Peng, W.: Large language models in healthcare and medical domain: A review. Informatics 11(3), 57 (August 2024) 15. Ren, M., Yan, Y., Chen, H., Hu, D., Xu, J., Zeng, X.: Tacl: Threshold-adaptive curriculum learning strategy for enhancing medical text understanding. arXiv preprint arXiv:2510.15269 (2025) 16. Sarella, P.N.K., Mangam, V.T.: Ai-driven natural language processing in healthcare: transforming patient-provider communication. Indian Journal of Pharmacy Practice 17(1), 21–26 (2024) 17. Shen, L., Feng, Y.: CDL: Curriculum dual learning for emotion-controllable response generation. Association for Computational Linguistics (Jul 2020) 18. Zhou, B., Yang, G., Shi, Z., Ma, S.: Natural language processing for smart healthcare. IEEE Reviews in Biomedical Engineering 17, 4–18 (2022) ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alansary-2026-severity-based-curriculum-learning-strategy/