Work

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

Year: 2026

Type: conference-paper

Abstract: We present a system for multilingual clinical named entity recognition (NER) submitted to the MultiClinNER subtask of MultiClinAI 2026, covering all seven languages and three entity classes (disease, symptom, procedure).Our approach trains one binary token classifier ensemble per entity class using cross-lingual fine-tuning of XLM-RoBERTa-large, with all languages handled jointly.We apply characterlevel ensembling over six models (two encoder variants × three cross-validation folds).This ensembling method provides more granular probability estimates than single-model classifiers, allowing for more flexible precisionrecall trade-off tuning.The system achieves character-level F1 scores of 0.70-0.88 on the official test set.

Source: Social Media Mining for Health Research and Applications Workshop and Shared Tasks (SMM4H)

DOI: 10.18653/v1/2026.smm4h-1.28

Authors: Petr Zelina, Vit Novacek

Institution: Masaryk University

Language: English

Topic: Biomedical Text Mining and Ontologies

Subfield: Molecular Biology

Field: Biochemistry, Genetics and Molecular Biology

Domain: Life Sciences

SDG: Good health and well-being

Open Access status: hybrid

Funders: European Commission, Masarykova Univerzita