Prompt robust large language model for Chinese medical named entity recognition

Medical Named Entity Recognition (NER) is crucial for constructing healthcare knowledge graphs and enhancing intelligent medical systems, yet it faces three challenges: data scarcity, low recall in nested entities annotation and high prompt sensitivity of generative NER model. In this paper, we aim to address the three challenges simultaneously. First, we construct a Multi-Scenario Medical NER dataset which is the largest medical NER dataset, including over 40,000 samples and over 3400 entity types with eight major scenarios: medical web data, online consultation, medical book, etc. Second, we propose a decomposed question answering based data annotation and selection method, which improved F1 score by 6% compared to direct annotation. Third, to enhance the robustness of large models to diverse prompts in real-world scenarios, we construct diverse prompt templates and implements dynamic prompt strategy during the training phase. Finally, we conducted a comprehensive set of experiments, and the results demonstrate the effectiveness of our annotation method and robustness training approach. Notably, the proposed framework achieves a 5% performance improvement on the test set compared to conventional methods. Moreover, our method enables a 7B parameter model to surpass a 32B parameter model, highlighting its superior efficiency and capability.