Named Entity Recognition in Chinese Electronic Medical Records Based on Global Pointer and Medical BERT

Chinese electronic medical record named entity recognition suffers from low accuracy and high complexity. In this paper, we propose a MBBLEGP model based on medical pre-training, which firstly masks the medical entities in the input medical record text data by MC-BERT to obtain the medical text word vector representations; then the word vectors are inputted into BiLSTM to extract the forward and backward local feature information to obtain the feature vectors; finally, a self-attention layer for GlobalPointer is used to Finally, rotational positional coding is introduced to GlobalPointer's self-attention layer to improve its sensitivity to the sequence structure, and the loss function is optimized using the sharing matrix to reduce the number of parameters when increasing the entity types to complete the construction of the model. The experimental results show that the effect is significantly improved by using MBBLEGP hybrid model, and compared with the mainstream BERT-BiLSTM-CRF, the F1 value is improved by 2.21% and 1.21%, respectively, and the speed of convergence on F1 is improved, which verifies the effectiveness of the present model in the recognition of medical named entities.