Study On Named Entity Recognition Of Chinese Electronic Medical Record Based On Deep Learning

Medical Name Entity Recognition(MNER)refers to the recognition of medical entities with special meanings in medical texts.It is an important component of medical information retrieval,relationship extraction,knowledge mapping,machine translation and other tasks.With the rapid development of intelligent medical information technology,entity recognition in medical field has gradually become the focus of interdisciplinary research.How to accurately and quickly identify medical entities with research significance from massive medical records is of great significance in the current development of domestic medical informatization.This project aims to make full use of massive Chinese Electronic Medical Records(CEMR)resources and obtain highly valuable medical information,so as to promote the research,production and practice in the field of clinical medicine.The specific research contents include the following parts:(1)Aiming at the problem of low accuracy caused by the polysemy of Chinese characters in CEMR,a named entity recognition model of CEMR based on BERT-Bi LSTM-CRF is proposed.This paper introduces a Chinese character level pre-training model named BERT,which can effectively solve the problem of "polysemy" of Chinese characters in CEMR.At the same time,due to the phenomenon of entity nesting and complex entity types in CEMR,Bi LSTM-CRF is introduced to extract the context information in semantic structure and obtain the relationship between entity labels,so as to achieve accurate recognition.Experiments show that the proposed model has a good recognition effect on named entities in CEMR.(2)Aiming at the difficulty of word boundary in CEMR,this paper proposes a model based on two granularity fusion coding of Chinese character and vocabulary and Attention mechanism.Based on the model of bet Bi LSTM-CRF,the word vector obtained by FMM and the word vector generated by BERT are fused to solve the noise problem caused by using word segmentation tools.At the same time,the word level semantic information is input into the model to obtain the word information.In addition,due to the limitation of computer hardware and software,the BERT used in this paper is trained by fixed parameters.In order to make up for the lack of recognition accuracy caused by fixed parameters,the Multi-head Attention mechanism is introduced into the model to analyze the structure between entities.Experimental results show that the performance of the proposed model is better than that of single word coding.(3)Aiming at the problems of numerous parameters and low computational efficiency of the pre-training model and Bi LSTM neural network structure model,an accelerated method of named entity recognition for CEMR based on SGRU is proposed.The ALBERT-SGRUCRF model is built,and the parallel computing of feature extraction is realized by using SGRU.In addition,the lightweight BERT and GRU network are used in the model,which greatly reduces the total parameters of the model.The experimental results show that the performance of the proposed model is significantly improved compared with the traditional model,and the computational efficiency of the model can be effectively improved.