One-shot Biomedical Named Entity Recognition via Knowledge-Inspired Large Language Model
Jnuyi Bian, Jiaxuan Zheng, Yuyi Zhang, Hong Jun Zhou, Shanfeng Zhu · 2024
Large Language Models (LLMs) have demonstrated exceptional performance in numerous natural language processing tasks, particularly in generative tasks. Nevertheless, their performance in non-generative tasks, such as information extraction, especially within specialized domain-specific extraction tasks like Biomedical Named Entity Recognition (NER), has been less successful when applied in an unsupervised manner. To address this challenge, we draw inspiration from the chain-of-thought concept and adopt a two-step approach for NER using LLMs: entity span extraction and entity type determination. Additionally, we introduce a framework for incorporating domain-specific entity knowledge to mitigate the LLM's inherent lack of domain expertise during entity category determination. Experimental results from four biomedical NER datasets illustrate a significant improvement in our approach when compared to prior LLM-based methods. Furthermore, our approach achieves results on par with other few-shot methods using just one shot, in contrast to their requirement of 50 shots.