Elamur Research⌕
☰
home / papers / 2026 / CASPAR: A Context-Aware Span Refinement Approach for Decision Support

CASPAR: A Context-Aware Span Refinement Approach for Decision Support

Paper · 2026 · Jing Tao, Amir Eskandari and Farhana Zulkernine

Summary

The paper presents CASPAR, a system for the MedExACT shared task at BioNLP @ ACL 2026. The task asks systems to find clinical decision spans in ICU discharge summaries from the MedDec dataset and label each span with one of nine DICTUM categories. Results are also checked for robustness across eight demographic subgroups. The authors name three difficulties: documents are longer than encoder input limits, the categories overlap in meaning, and exact-match scoring counts any small boundary error as a full miss.

CASPAR works in two stages:

  • Stage 1 tags tokens with BIO labels using RoBERTa-base, a single-layer Bi-GRU and a CRF. Long notes are split into overlapping windows. Window logits are merged with triangular weights that give less weight to tokens near window edges, and a global CRF then decodes the whole document.
  • Stage 2 re-encodes each candidate span within a small local window. It then uses confidence-gated correction to fix the span's category and its left and right boundaries.

The system placed 10th of 37 teams with a Final Score of 0.5668, clearly ahead of the organizer baseline. This suggests that adding context before CRF decoding and refining spans afterwards both help strict span extraction.

Key points

  • Official test results: Final 0.5668, Worst-Group F1 0.5602, Span F1 0.5060 (baseline 0.348) and Token F1 0.6409. The 1st-place Final Score was 0.5965.

  • Validation ablation (Span F1):

    • RoBERTa only: 0.3939
    • adding CRF: 0.4065
    • adding Bi-GRU (full Stage 1): 0.4396
    • adding Stage 2: 0.4873

    Token F1 over the same steps rose from 0.5601 to 0.6249.

  • Main error types: spans that are too short (under-span, 24.21%) and too long (over-span, 19.88%). Label confusion is concentrated in Therapeutic procedure (cat 6, 26.1%).

  • Subgroup gaps: Hispanic patients had the highest label-confusion rate (31.69%) and the largest gap between Token F1 and Span F1 (0.519 vs. 0.368).

  • Limitation: Stage 2 is trained only on Stage 1 candidates that overlap a gold span. It therefore cannot remove false positives or recover spans Stage 1 missed. Code is available at github.com/schorm/caspar-medexact.

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

🎓Queen's UniversityKingston · Canada11 works · 1,637 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksBiochemistry, Genetics and Molecular Biology · 234 worksMolecular Biology · 231 works

Cite

Cite as Jing Tao, Amir Eskandari and Farhana Zulkernine. 2026. CASPAR: A Context-Aware Span Refinement Approach for Decision Support. Proceedings of the BioNLP 2026 (Shared Tasks).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-2.21

@inproceedings{tao2026caspar,
  title = {{CASPAR: A Context-Aware Span Refinement Approach for Decision Support}},
  author = {Jing Tao, Amir Eskandari and Farhana Zulkernine},
  booktitle = {Proceedings of the BioNLP 2026 (Shared Tasks)},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-2.21},
  url = {https://doi.org/10.18653/v1/2026.bionlp-2.21}
}

References

21 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. 2019. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32. find
  2. Alessandro Cocchieri, Viet Dac Lai, Sukannya Purkayastha Boro, and et al. 2025. Openbioner: Lightweight open-domain biomedical named entity recognition through entity type description. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 847–860. Association for Computatio… find
  3. Mohamed Elgaar, Jiali Cheng, Nidhi Vakil, Hadi Amiri, and Leo Anthony Celi. 2024. MedDec: A dataset for extracting medical decisions from discharge summaries. In Findings of the Association for Computational Linguistics: ACL 2024, pages 16442–16455, Bangkok, Thailand. Association for Computational L… find
  4. Mohamed Elgaar, Jiali Cheng, Nidhi Vakil, Mehrnaz Sadrolashrafi, Mitra Mohtarami, Adrian Wong, Hadi Amiri, and Leo A. Celi. 2026. Overview of medical decision extraction, analysis, and classification task (medexact) 2026. In The 25th Workshop on Biomedical Natural Language Processing and BioNLP Shar… find
  5. Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional LSTM-CRF models for sequence labeling. arXiv preprint arXiv:1508.01991.
  6. Richard A. A. Jonker, Tiago Almeida, Rui Antunes, João R. Almeida, and Sérgio Matos. 2024. Multi-head crf classifier for biomedical multi-class named entity recognition on spanish clinical notes. Database, 2024:baae068. find
  7. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692.
  8. Ling Luo, Chih-Hsuan Wei, Po-Ting Lai, Robert Leaman, Qingyu Chen, and Zhiyong Lu. 2023. Aioner: all-in-one scheme-based biomedical named entity recognition using deep learning. Bioinformatics, 39(5):btad310. find
  9. Eirik H. Ofstad, Jan C. Frich, Edvin Schei, Richard M. Frankel, and Pål Gulbrandsen. 2016. What is a medical decision? A taxonomy based on physician statements in hospital encounters: a qualitative study. BMJ Open, 6(2):e010098. find
  10. Lance Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In Third Workshop on Very Large Corpora.
  11. Lev Ratinov and Dan Roth. 2009. Design challenges and misconceptions in named entity recognition. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009), pages 147–155. Association for Computational Linguistics.
  12. João Ruano, Gonçalo Correia, Leonor Barreiros, and Afonso Mendes. 2025. Effective multi-task learning for biomedical named entity recognition. In Proceedings of the 24th Workshop on Biomedical Language Processing, pages 225–239, Viena, Austria. Association for Computational Linguistics. find
  13. Mario Sänger, Alan Akbik, Amir Pouran Ben Veyseh, and et al. 2024. Hunflair2 in a cross-corpus evaluation of biomedical named entity recognition and normalization tools. Bioinformatics, 40(10):btae564. find
  14. Jiuding Yang, Jinwen Luo, Weidong Guo, Di Niu, and Yu Xu. 2023. Exploiting hierarchically structured categories in fine-grained chinese named entity recognition. In Findings of the Association for Computational Linguistics: ACL 2023, pages 3407–3421, Toronto, Canada. Association for Computational Li… find
  15. Sheng Zhang, Hao Cheng, Jianfeng Gao, and Hoifung Poon. 2023. Optimizing bi-encoder for named entity recognition via contrastive learning. In The Eleventh International Conference on Learning Representations. find
  16. Yongliang Shen, Xinyin Ma, Zeqi Tan, Shuai Zhang, Wen Wang, and Weiming Lu. 2021. Locate and label: A two-stage identifier for nested named entity recognition. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Na…
  17. Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. Diffusionner: Boundary diffusion for named entity recognition. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7186–7201. Association for… find
  18. Yongliang Shen, Xiaobin Wang, Zeqi Tan, Guangwei Xu, Pengjun Xie, Fei Huang, Weiming Lu, and Yueting Zhuang. 2022. Parallel instance query network for named entity recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9… find
  19. Minghao Tang, Yongquan He, Yongxiu Xu, Hongbo Xu, Wenyuan Zhang, and Yang Lin. 2023. A boundary offset prediction network for named entity recognition. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14834– 14846, Singapore. Association for Computational Linguistics. find
  20. Harsh Verma, Sabine Bergler, and Narjesossadat Tahaei. 2023. Comparing and combining some popular NER approaches on biomedical tasks. In Proceedings of the 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks, pages 273– 279, Toronto, Canada. Association for Computational… find
  21. Ying Wei and Qi Li. 2023. Scdner: Span-based consistency-aware document-level named entity recognition. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15677–15685. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tao-2026-caspar-context-aware-span-refinement/