Chinese Medical Continual Named Entity Recognition Based on Continual Learning and Knowledge Distillation

Chinese Medical Named Entity Recognition (CMNER) is the extraction of entities from Chinese unstructured medical electronic medical record text. Traditional CMNER models cannot adapt to multi-task continuous entity recognition and suffer from catastrophic forgetting in incremental learning when new medical entities continue to arrive. To solve these problems, this paper proposes a Chinese medical continuous entity recognition model based on continuous learning and knowledge distillation. Independent modeling at the span and entity levels, combined with continuous learning and knowledge distillation to retain memory, aims to learn new knowledge and retain old knowledge at the same time. Further, by introducing dynamic adaptation mechanisms and advanced feature fusion strategies, the performance of the model is optimized so that it can maintain a high level of recognition accuracy and stability in the face of constantly changing medical entities. Experiments were conducted on the Chinese medical terminology dataset CCKS2020 and the self-built breast cancer examination report dataset, and the F1 values were improved by 1.64% and 0.37% respectively compared with the best baseline model.