Related Experiment Video
Updated: May 24, 2026
14:34
A Bilingual Computational Workflow for Identifying Potential PLK1 Inhibitors in American Sign Language and English
Published on: April 3, 2026
GLiNER-BioMed: a suite of efficient models for open biomedical named entity recognition
Anthony Yazdani1, Ihor Stepanov2, Douglas Teodoro1
1Department of Radiology and Medical Informatics, Faculty of Medicine, University of Geneva, Geneva, 1202, Switzerland.
Bioinformatics (Oxford, England)
|May 22, 2026
Summary
We developed GLiNER-BioMed, a novel approach for biomedical named entity recognition (NER) that overcomes limitations of traditional models. GLiNER-BioMed achieves state-of-the-art performance in zero-shot and few-shot learning scenarios, offering efficient and accurate entity extraction for biomedical text.
Area of Science:
- Computational Biology
- Bioinformatics
- Natural Language Processing
Background:
- Biomedical named entity recognition (NER) faces challenges with specialized vocabularies and novel entities.
- Traditional NER models struggle with generalization due to fixed taxonomies and human annotations.
Purpose of the Study:
- To introduce GLiNER-BioMed, a domain-adapted suite of GLiNER models for efficient and accurate biomedical NER.
- To improve zero-shot and few-shot learning capabilities in biomedical text analysis.
Main Methods:
- Distilled annotation capabilities of large language models (LLMs) into smaller, efficient GLiNER models.
- Trained uni- and bi-encoder GLiNER architectures at multiple scales for biomedical NER.
- Conducted experiments on eight biomedical datasets to evaluate performance and efficiency.
Main Results:
- GLiNER-BioMed achieved state-of-the-art zero-shot performance (micro-F1 59.77%), outperforming baselines by 5.96 points.
- The bi-encoder variant reached 70.39% in 10-shot learning, consistently outperforming baselines.
- Uni-encoder GLiNER-BioMed excels in zero-shot, while bi-encoder offers superior few-shot gains and higher inference throughput (+39-568%).
Conclusions:
- GLiNER-BioMed effectively addresses limitations in biomedical NER, providing state-of-the-art performance.
- The uni- and bi-encoder architectures offer distinct advantages for different biomedical NLP applications.
- Combining synthetic biomedical pre-training with general-domain post-training is crucial for optimal performance.
