Elamur Research⌕
☰
home / papers / 2026 / Discovery@FI at #SMM4H–HeaRD 2026: Ensemble Character Classifier for Multilingual Clinical NER

Discovery@FI at #SMM4H–HeaRD 2026: Ensemble Character Classifier for Multilingual Clinical NER

Paper · 2026 · Petr Zelina and Vit Novacek

Summary

The paper describes the Faculty of Informatics, Masaryk University submission to the MultiClinNER subtask of MultiClinAI 2026. The task is to find disease, symptom and procedure mentions in clinical texts in seven languages. The non-Spanish data were produced by translation, and those texts differ between entity classes. The authors therefore train a separate binary token classifier for each class, with all languages pooled in one model.

Each class uses an ensemble of six XLM-RoBERTa-large models: two encoder variants, each trained on three cross-validation folds. Token probabilities are converted to character-level probabilities and averaged across models. Characters are labelled with a 0.8 threshold, and the resulting spans are expanded to whole words. Because decoding happens at the character level, models with different tokenizers can be combined, and averaging gives smoother probabilities for tuning precision against recall.

On the official test set the system places roughly mid-field and does better on recall than on precision. The authors suggest it could serve as a fast candidate-retrieval stage before a more precise filter, such as a yes/no LLM check.

Key points

  • Ensemble composition: six models combine base FacebookAI/xlm-roberta-large (561M parameters) with a variant further pretrained for 10K MLM steps on ~200M words of private Czech clinical notes. The Czech fav-kky/FERNET-C5 was tested but reached only ~83% of XLM's validation score, so it was dropped.
  • Test results: character F1 ranges from 0.707 (Czech, symptom) to 0.837 (English, disease), and strict F1 ranges from 0.54 to 0.74.
  • Hardest cases: symptom is consistently the most difficult entity class, and Czech and Dutch score lowest.
  • Translation noise: training-set scores show a clear gap between Spanish and the other languages, which the authors attribute to translation noise. This gap does not appear on the test set.
  • Compute cost: all six models train in parallel on one H100 in about 16 minutes, and inference runs at roughly 21K tokens/s. Code is available at github.com/ZepZep/ner_character_ensemble.

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

🎓Masaryk UniversityBrno · Czechia11 works · 437 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksBiochemistry, Genetics and Molecular Biology · 263 worksMolecular Biology · 260 works

🧭 Topics 2Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 882 worksNamed entity recognition · 214 works

Cite

Cite as Petr Zelina and Vit Novacek. 2026. Discovery@FI at #SMM4H–HeaRD 2026: Ensemble Character Classifier for Multilingual Clinical NER. Proceedings of the 11th Social Media Mining for Health Research and Applications (SMM4H-HeaRD 2026) Workshop and Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.smm4h-1.28

@inproceedings{zelina2026discovery,
  title = {{Discovery@FI at \#SMM4H–HeaRD 2026: Ensemble Character Classifier for Multilingual Clinical NER}},
  author = {Petr Zelina and Vit Novacek},
  booktitle = {Proceedings of the 11th Social Media Mining for Health Research and Applications (SMM4H-HeaRD 2026) Workshop and Shared Tasks},
  year = {2026},
  doi = {10.18653/v1/2026.smm4h-1.28},
  url = {https://doi.org/10.18653/v1/2026.smm4h-1.28}
}

References

5 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. CoRR, abs/1911.02116.
  2. Fernando Gallego-Donoso, Salvador Lima-López, Judith Rosell, Eulàlia Farré-Maduell, and Martin Krallinger. 2026. The MultiClinAI Shared Task on Multilingual Clinical Corpus Construction and Concept Extraction: Systems, Evaluation, and Datasets. In Proceedings of the 11th Social Media Mining for Heal… find
  3. Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc. find
  4. Jan Lehečka and Jan Švec. 2021. Comparison of czech transformers on text classification tasks. In Statistical Language and Speech Processing, pages 27–37, Cham. Springer International Publishing. find
  5. Guillermo Lopez-Garcia, Jose Miguel Acitores Cortina, Jacob Berkowitz, Joey Chan, Ganesh Chandrasekar, Sumon Kanti Dey, Ivan Flores Amaro, Fernando Gallego, Lauren Gryboski, Ari Z Klein, Martin Krallinger, Salvador Lima-López, Tomohiro Nishiyama, Lisa Raithel, Ahmad Rezaie Mianroodi, Roland Roller,… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zelina-2026-discovery-fi-smm4h-heard-2026/