Improving Named Entity Recognition Performance for Large Language Models Using a Prompt Framework

Named entity recognition is the core task of information extraction, aiming to identify specific classes of named entities from unstructured text. However, existing methods have limitations in the determination of complex entity boundaries, which affects the accuracy and generalisation ability of information extraction. To this end, this paper proposes a NER prompt framework based on a large language model that fuses structured annotation specification and few-sample learning to improve the accuracy and robustness of complex entity boundary recognition. The framework includes: task-oriented baseline prompt, which constructs instruction templates based on predefined task goals and standard output formats to ensure output consistency; annotation guideline prompt, which combines the entity discrimination rules of the knowledge base with the boundary determination logic to enhance the recognition capability of complex entities; error analysis, which optimizes the iterative mechanism based on the attribution of historical errors to improve the generalisation capability of the model; role prompt, giving model expert roles to enhance domain knowledge and decision-making ability; and sample less learning templates, using high-quality labelled samples to improve learning ability in low-resource scenarios. The experimental results show that the F1 scores on public and self-built datasets are up to 0.858 as well as 0.831, which validates the effectiveness of the framework.