TLDR: A new research paper introduces a unified framework for Biomedical Named Entity Recognition (BioNER) using Large Language Models (LLMs). It reformulates BioNER as a text generation task with a symbolic tagging strategy for both flat and nested entities, employs bilingual joint fine-tuning across multiple datasets for generalization, and includes a contrastive entity selector to filter out incorrect predictions. The framework achieves state-of-the-art performance and robust zero-shot generalization across Chinese and English biomedical texts.
A new research paper introduces a groundbreaking approach to Biomedical Named Entity Recognition (BioNER), a crucial task for extracting vital information from medical texts. Authored by Tengxiao Lv, Ling Luo, Juntao Li, Yanhua Wang, Yuchen Pan, Chao Liu, Yanan Wang, Yan Jiang, Huiyi Lv, Yuanyuan Sun, Jian Wang, and Hongfei Lin, this work addresses long-standing challenges in the field, such as identifying complex nested entities, resolving ambiguous entity boundaries, and ensuring the system works across different languages.
Biomedical texts, including electronic medical records and textbooks, are rich with information essential for clinical decisions and research. However, their unstructured nature makes direct use difficult. BioNER aims to extract structured information by identifying and classifying specific entities like genes, proteins, diseases, and chemicals. Traditional methods have struggled with the complexity of these texts, especially when entities are nested within each other (e.g., “IL-5” and “luciferase” both being proteins within a larger DNA entity) or when terms are highly specialized and ambiguous.
The researchers propose a unified BioNER framework that leverages the power of Large Language Models (LLMs). Instead of treating BioNER as a traditional sequence labeling problem, they reformulate it as a text generation task. This innovative approach allows the model to generate entities and their exact positions within the text, making it more practical for real-world applications.
Also Read:
- Unlocking Entity Understanding in Large Language Models
- GraphMERT: Building Factual and Scalable Knowledge Graphs for Domain-Specific AI
Key Innovations of the Framework:
Symbolic Tagging Strategy: The team designed and compared several ways to tag entities within the text. They found that a “symbolic tagging” strategy, which uses simple symbols like square brackets to mark the start and end of an entity and explicitly states its type, performed best. This method simplifies the annotation process for LLMs and helps them accurately identify both flat (single) and nested entities, providing precise boundary information.
Multi-dataset Joint Fine-tuning: To make the model more versatile and capable of handling different tasks and languages, the framework undergoes bilingual joint fine-tuning. This means it’s trained simultaneously on multiple Chinese and English BioNER datasets, including those with nested and flat entities. This diverse training helps the model learn more generalized patterns and improve its performance in scenarios where it encounters new, unseen data or different languages.
Contrastive Entity Selector: A common issue with entity recognition is errors in identifying the exact boundaries or types of entities. To combat this, the researchers introduced a contrastive learning-based entity selector. This selector is trained using both correct (positive) and intentionally incorrect (negative) examples of entities. By learning to distinguish between these, the selector can effectively filter out incorrect or ambiguous predictions made by the LLM, significantly enhancing the precision of the recognition process.
The experimental results are highly promising. The proposed method achieved state-of-the-art performance on four benchmark datasets and demonstrated robust generalization capabilities on two unseen corpora. This means the framework not only performs exceptionally well on data it was trained on but also adapts effectively to new, previously unencountered biomedical texts and languages without further training.
This unified framework represents a significant step forward in BioNER, offering a comprehensive solution that can jointly support flat and nested entities across both English and Chinese within a single, adaptable system. For more details, you can read the full research paper available here.
