Research on Optimization and Application of Retrieval-LLM-NER Method

Named Entity Recognition (NER) is a crucial task in Natural Language Processing (NLP); however, Large Language Models (LLMs) still underperform compared to traditional supervised learning methods in NER tasks. To bridge this gap, this paper proposes an LLM-based NER method (LLM-NER) that reformulates the sequence labeling task into a text generation task better suited for LLMs. The proposed approach consists of key steps including prompt construction, entity annotation generation, and output parsing to enhance the model's entity recognition capability. Furthermore, this study explores different retrieval methodsdense retrieval, multi-vector retrieval, and sparse/lexical retrieval-to optimize prompt construction. Additionally, a selfknowledge distillation technique is applied to integrate multiple retrieval strategies, further improving overall model performance. The experiments are conducted on a custom dataset, utilizing two versions of GPT-3.5 as the base models for entity recognition, with precision, recall, and $\mathbf{F 1}$-score as evaluation metrics. The results demonstrate that the proposed method effectively enhances model performance in entity recognition tasks, providing new insights for applying LLMs to NER.