BIT.UA at #SMM4H–HeaRD 2026: Towards Multi-Class Multilingual Clinical Entity Recognition with Multi-Head CRF Ensembles

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research · Depth: Expert, medium

Summary

The BIT.UA system, developed for the MultiClinNER shared task at #SMM4H–HeaRD 2026, addresses multilingual clinical named entity recognition (NER) across seven languages for Disease, Procedure, and Symptom entity types. This system extends the Multi-Head CRF architecture, originally for Spanish clinical text, to a multilingual context. It incorporates an adaptive text consolidation pipeline, which preserves over 94% of annotations, enabling joint multi-entity training despite dataset variations. A key finding is that a single "xlm-roberta-large" model, jointly trained on all seven languages and three entity types, achieved competition rank 2 for five of seven languages. This model surpassed dedicated monolingual models by up to +6.94 F1 points while using only one set of weights. Further, ensembling multiple seeds of this model secured rank 1 for those five languages, and its combination with monolingual models achieved rank 1 for the remaining two. The code and models are publicly available.

Key takeaway

For NLP Engineers developing clinical NER systems, consider adopting a multilingual, multi-head CRF approach. Your team can achieve superior performance across diverse languages and entity types, potentially reducing model complexity by using a single "xlm-roberta-large" backbone. Explore ensembling techniques and adaptive text consolidation to maximize F1 scores, as demonstrated by the BIT.UA system's rank 1 results. This strategy offers a path to efficient, high-accuracy clinical entity extraction.

Key insights

A single multilingual "xlm-roberta-large" model with Multi-Head CRF ensembles excels in clinical NER across seven languages.

Principles

• Multilingual models can outperform monolingual ones significantly.
• Ensembling model seeds boosts performance and robustness.
• Adaptive data consolidation supports joint multi-entity training.

Method

The system extends Multi-Head CRF to multilingual NER, using an adaptive text consolidation pipeline for joint training. It employs a single "xlm-roberta-large" model, then ensembles its seeds and combines with monolingual models for optimal performance.

In practice

• Use "xlm-roberta-large" for multilingual clinical NER.
• Implement Multi-Head CRF for multi-class entity recognition.
• Explore model ensembling for F1 score improvements.

Topics

• Multi-Head CRF
• Clinical Named Entity Recognition
• Multilingual NLP
• xlm-roberta-large
• Model Ensembling
• MultiClinNER Shared Task

Code references

• ieeta-pt/Multi-Head-CRF

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• 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.
• Multilingual Coreference Resolution via Cycle-Consistent Machine Translation — Coreference resolution in low-resource languages can be significantly improved by cycle-consistent machine translation data augmentation.
• Analysing Lightweight Large Language Models for Biomedical Named Entity Recognition on Diverse Ouput Formats — Lightweight LLMs offer competitive biomedical NER performance, with specific output formats proving more effective than multi-format tuning.
• Noise reduction in BERT NER models for clinical entity extraction — A post-processing Noise Removal model significantly boosts clinical NER precision by analyzing probability distributions and contextual features.
• End-to-End Evaluation of a RAG System for Hospital Documents in Portuguese — Optimizing RAG components individually enhances performance for querying specialized documents in resource-limited settings.
• sebis at CRF Filling 2026: A Two-Stage Local LLM Pipeline for Medical CRF Filling — A two-stage local LLM pipeline using MedGemma-27B achieves competitive, privacy-preserving medical CRF filling with strict evidence adherence.

Open in AIssential →

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