Research On Named Entity Recognition Of Chinese Electronic Medical Records Based On Deep Learning And Attention Mechanism

Named Entity Recognition（NER）is a natural language processing technique used to identify entity names in text with specific meanings.The development of NER technology has promoted the construction and development of informationization in the medical field.In clinical medicine,electronic medical record NER can automatically extract key entity information from cases,improving the efficiency and accuracy of medical diagnosis,treatment,and research.However,the current research on Chinese electronic medical record NER is still in its early stages,with unclear entity boundary information,potential semantic conflicts between words,and a lack of comprehensive architecture.This dissertation focuses on the Chinese electronic medical record NER task,pre-processing two medical datasets from CCKS2017 and CCKS2021,and proposes two NER models.The details of the study can be concluded as following:（1）A Chinese electronic medical records NER method called IFLAT-Bi GRU-CRF is proposed,which combines Inter-Attention mechanism and Bidirectional Gated Recurrent Unit（Bi GRU）.The method addresses the issue of entity boundary conflicts in Chinese electronic medical records and the increased computational and memory costs of using the FLAT（Flat-Lattice Transformer）for processing long sequential text.The joint model uses a combination of the fused inter-attention FLAT,bidirectional gated recurrent units,and conditional random fields（CRF）to extract key medical information.Firstly,the characters and word encodings in the medical text are organized into a flat lattice,and an Inter Former module is used to enhance vocabulary and process long sequential text.Secondly,the enhanced word vectors by using Inter Former method are fused with FLAT to form IFLAT method,which effectively reduces computational and memory costs.Thirdly,the output of IFLAT is fed into the bidirectional gated recurrent unit to obtain richer contextual features and prevent the problems of vanishing and exploding gradients.Finally,the predicted label sequence is decoded using a conditional random field.Experimental results show that this method improves the accuracy of named entity recognition,reduces computational and memory costs,and increases the F₁ score by 5.06%and 1.54%on the CCKS2017 and CCKS2021 datasets,respectively.（2）A Chinese electronic medical records NER method that combines Radical features and BERT-Transformer-CRF（BTC）is proposed.This method addresses the problem of decreased long-term dependency ability of deep models when processing long sequential text by using a named entity recognition method based on pre-trained language models.Firstly,the fine-tuned BERT（Bidirectional Encoder Representation from Transformer）is pre-trained to extract features from medical texts.Then,Transformer is used to capture the dependency relationship between characters,a process that does not require consideration of distance between characters.In addition,as terminology dictionary information and Radical information of Chinese characters contain deeper semantic information,the features of the terminology dictionary and Radicals are incorporated into the model to improve its performance.Finally,the best label sequence is outputted through decoding with the Conditional Random Field module,which constrains the context labeling.Experimental results show that the proposed model achieved F₁ values of 96.22%and 84.65%on the CCKS2017 and CCKS2021 datasets respectively,indicating good recognition performance.This dissertation proposes a deep learning based named-entity recognition（NER）method by using mutual attention mechanism and multi-head attention mechanism,which combines the semantic information of electronic medical records such as Radical features and boundary information.This method improves the accuracy of Chinese electronic medical record entity recognition and promotes the development of medical intelligence.