Large language model augmented framework with domain-specific knowledge integration for medical named entity recognition.

Abstract

Medical named entity recognition aims to recognize medical named entities and assign entity types from medical texts, which serves as the foundation for domain-specific applications. Although large language models (LLMs) and named entity recognition frameworks have presented impressive capabilities, they face challenges including domain specialization limitations, limited accessibility of medical knowledge resources, and lack of standardization. We propose a novel LLM augmented framework with domain-specific knowledge integration. It decomposes complex medical named entity recognition into recognition, refinement, and entity type identification. A medical LLM and a task-adaptive recognizer are employed to perform systematic medical named entity recognition. Subsequently, iterative refinement and standardization are conducted through the knowledge augmentation method to incorporate domain-specific knowledge and ensure consistency. We introduce an entity type identifier to effectively determine the entity type utilizing the unified contextual information. Extensive experiments are conducted on over 26,000 diverse samples from three benchmark datasets in the medical domain, which cover 16 entity types from medical encyclopedias, electronic health records, and question answering texts. Experimental results validate the effectiveness of the framework. It achieves average evaluation metrics of 78.24%, 95.89%, and 96.08% on the 3 benchmark datasets, respectively. The framework outperforms 12 representative baseline methods from general and medical domains by approximately 3%, which demonstrates enhanced performance in medical named entity recognition. As the initial effort to integrate domain-specific knowledge with the knowledge augmentation method for progressive medical named entity recognition, the framework provides an innovative approach to knowledge-intensive tasks.