Research On Named Entity Recognition Of Chinese Electronic Medical Records Based On BERT

With the rapid development of information technology,many medical institutions use electronic medical records instead of thesis medical records to record patients’ treatment information,resulting in a large number of electronic medical records.Using named entity recognition technology to process and analyze electronic medical record text,high-value information can be extracted from it,which provides support for subsequent professional medical knowledge base,drug identification and clinical decision-making,so it has been concerned by many scholars.At present,there is often a problem in the Chinese electronic medical record named entity recognition task that ignores the single feature of short distance semantic information and word vector between entities.To solve the problem of ignoring short-distance semantic information between entities,this thesis first presents a named entity recognition model for Chinese electronic medical records based on SCNN-BERT.On the basis of this model,considering the single feature of word vectors during training,a named entity recognition model for Chinese electronic medical records based on DCNN-BERT-Fusion is proposed.This study mainly focuses on the following two aspects:(1)Named entity recognition of Chinese electronic medical records based on SCNN-BERT.Most studies only consider contextual semantic information and ignore the impact of short-distance semantic information between entities on entity recognition.Firstly,the semantics of the electronic medical record text is extracted by introducing the BERT pre-training model with two-way encoding,and the dynamic word vector of the text is obtained according to the context.Then,a single channel neural network(SCNN)is designed by concatenating two-way long-term and short-term memory network and convolution neural network.It not only extracts the context semantic information features,but also extracts the short-distance semantic information features between entities,so as to make good use of the local and global feature information.Finally,the extracted results are input into a conditional random field to automatically learn the constraints of the hidden layer,decode the sequence labels,and get the named entity of the electronic medical record.The experiment obtained 84.89% F1 value,and the result shows that the model achieves good results in the electronic medical record named entity recognition.(2)Named entity recognition of Chinese electronic medical records based on DCNN-BERT-Fusion.On the basis of model(1),considering the single feature of word vector in training process,a named entity recognition model for Chinese electronic medical records based on DCNN-BERT-Fusion is proposed.First,the text of the electronic medical record is input into the bidirectional encoding BERT model for character level encoding,and then the dynamic word vector containing contextual information is trained.Secondly,the word vectors are input into the convolution neural network and the dual-channel neural network(DCNN)which is composed of two-way gated loop unit in parallel.Two internal features of the medical record text are extracted by DCNN,which are local and global.Then,it constructs four external features,namely,side feature,stroke feature,word feature and dictionary feature,and uses the difference fusion method to fuse the internal and external features with word vector,which can enrich the feature information of word vector and enhance the expression ability of words.Finally,a conditional random field is input for sequential labeling,which reduces the output probability of error labels and yields five medical record entities.The experiment obtained 89.45% F1 value.The result shows that the model can achieve a good effect in the electronic medical record named entity recognition.Figure [22] Table [13] Reference [67]...