Related Experiment Video
Updated: May 26, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
DA-BioNER: data augmentation based on few-shot learning and distant supervision for biomedical named entity
Yesol Park1, Gyujin Son2, Taeuk Kim1,2
1Department of Computer Science, Hanyang University, Seoul, Republic of Korea.
Bioinformatics (Oxford, England)
|May 24, 2026
Summary
This study introduces DA-BioNER, a novel framework for biomedical named entity recognition (NER). DA-BioNER enhances low-resource NER by refining existing annotations with large language models, improving accuracy and entity diversity.
Area of Science:
- Biomedical informatics
- Natural Language Processing
- Machine Learning
Background:
- Named Entity Recognition (NER) is crucial for structured knowledge extraction in biomedicine.
- Scarcity of high-quality annotated data hinders NER effectiveness in emerging biomedical domains.
- Existing data augmentation methods often suffer from limited entity diversity, noisy labels, and disrupted contextual integrity.
Purpose of the Study:
- To develop a context-preserving data expansion framework for biomedical NER.
- To address the challenges of low-resource settings and domain adaptation in biomedical NER.
- To improve the generalization ability of NER models in data-scarce environments.
Main Methods:
- DA-BioNER framework combines multiple base NER models for coarse annotations.
- A large language model (LLM) refines annotations using global biomedical knowledge.
- Annotation refinement occurs within existing sentences, preserving syntactic and semantic context.
- LLM's role is constrained to refinement, reducing hallucination and improving precision.
Main Results:
- DA-BioNER achieves high F1-scores on benchmark datasets (NCBI-Disease, BC5CDR, BioRED) in low-resource settings (e.g., 0.799 on BioRED in 40-shot).
- Outperforms state-of-the-art methods by up to 0.32 in 40-shot settings.
- Improves F1-scores by up to 0.08 in extreme few-shot settings.
- Generates an average of 1,391 additional unique entities, enhancing training diversity.
Conclusions:
- DA-BioNER offers a scalable and adaptable solution for robust biomedical NER.
- The framework is particularly effective for domain adaptation and low-resource scenarios.
- DA-BioNER preserves contextual integrity while significantly improving NER performance and data diversity.