Medical named entity recognition based on domain knowledge and position encoding

A model for recognizing named entities in Chinese electronic medical records is proposed, focusing on accurate boundary detection, by leveraging medical domain knowledge and positional encoding. Medical domain-specific terms are integrated into a BERT module by a lexical adapter firstly. After pre-training, the model captures the dynamic character feature representation containing lexical information and boundary information. In the feature encoding module, Star-Transformer and BiLSTM are employed to extract local features and long-distance features respectively in order to generate the sequence's feature representation. Additionally, considering the influence of the relative position information between characters in the text on recognition results, Rotary Position Embedding (RoPE) is incorporated to improve Star-Transformer to enhance the ability of extracting semantic features. Experimental results on the CCKS2020 dataset show an improvement in the F1-score, reaching 85.78%. Compared to the baseline model, the F1-score increases by 2.96%. For the self-build breast cancer ultrasound report dataset, improvement is also observed, which proves the effectiveness and applicability of the model in medical field.