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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, short

Summary

Discovery@FI developed a system for multilingual clinical Named Entity Recognition (NER), submitted to the MultiClinNER subtask of MultiClinAI 2026. This system addresses seven languages and three critical entity classes: disease, symptom, and procedure. The approach involves training one binary token classifier ensemble for each entity class, utilizing cross-lingual fine-tuning of XLM-RoBERTa-large, with all languages processed jointly. A key component is character-level ensembling, applied across six models derived from two encoder variants and three cross-validation folds. This ensembling technique enhances the granularity of probability estimates compared to single-model classifiers, enabling more flexible tuning of the precision-recall trade-off. The system demonstrated strong performance, achieving character-level F1 scores ranging from 0.70 to 0.88 on the official test set.

Key takeaway

For NLP Engineers or Research Scientists developing multilingual clinical Named Entity Recognition systems, you should consider implementing character-level ensembling. This method, combined with cross-lingual fine-tuning of models like XLM-RoBERTa-large, provides more granular probability estimates. This allows you to achieve a more flexible precision-recall trade-off, crucial for adapting your models to specific clinical requirements across seven or more languages, ultimately improving system robustness and accuracy.

Key insights

Character-level ensembling on cross-lingually fine-tuned XLM-RoBERTa-large improves multilingual clinical NER precision-recall flexibility.

Principles

• Joint language training improves multilingual NER.
• Character-level ensembling refines probability estimates.
• Ensembling allows flexible precision-recall tuning.

Method

Train one binary token classifier ensemble per entity class using cross-lingual fine-tuning of XLM-RoBERTa-large, then apply character-level ensembling over six models.

In practice

• Apply character-level ensembling for granular NER probabilities.
• Use XLM-RoBERTa-large for cross-lingual fine-tuning.
• Jointly train across multiple languages for efficiency.

Topics

• Multilingual NER
• Clinical NLP
• XLM-RoBERTa-large
• Ensemble Learning
• Cross-lingual Fine-tuning
• Precision-Recall Tuning

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Ensembles of Large Language Models for Identifying EQ-5D Studies in PubMed Based on Their Abstracts — Ensemble LLM frameworks reliably automate biomedical text classification, balancing performance and interpretability.
• Predicting relations between SOAP note sections: The value of incorporating a clinical information model — Fine-tuning a RoBERTa-base model on manually annotated clinical data improves NER for SOAP note sections.
• MMed-Bench-IR: A Heterogeneous Benchmark for Multilingual Medical Information Retrieval — MMed-Bench-IR reveals severe cross-lingual performance gaps in medical information retrieval, highlighting the need for benchmarks that assess multilingual and biomedical expertise jointly.
• Beyond Bilingual Transfer: Multilingual Code-Switching in Instruction Tuning — Multilingual code-switching instruction tuning improves LLM performance beyond bilingual settings.
• EDEN: A Large-Scale Corpus of Clinical Notes for Italian — EDEN provides the largest Italian clinical note corpus and a CRF-filling benchmark for LLM-based information extraction.
• When Multiple Scripts Matter: Evaluating ASR in Clinical Settings — Multiscript variability in non-English clinical ASR requires specialized evaluation and script unification during training for accurate performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.