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

In 2010,my country issued a notice on the pilot work of electronic medical records,which promoted the rapid development of electronic medical records.Now electronic medical records have become the core of hospital information systems.The use of electronic medical record named entity recognition technology can efficiently use a large amount of electronic medical record data,promote the accelerated development of medical information,and is the basis for tasks such as subsequent relationship extraction and auxiliary diagnosis.Electronic medical record named entity recognition is different from general field named entity recognition,and Chinese electronic medical record named entity recognition is different from English electronic medical record named entity recognition.In Chinese electronic medical records,entity boundaries are difficult to determine and there are polysemous words and nested entities,which makes the recognition of Chinese electronic medical records more difficult.Therefore,the recognition of named entities in Chinese electronic medical records requires in-depth research.Aiming at the phenomenon of polysemous words and nested entities in electronic medical records,the paper proposes an innovative model BERT-SPAN based on the pre-training model BERT.In different contexts,the BERT model will output different word vectors through the 12-layer Transformer,which can solve some polysemous problems.The BERT-SPAN model uses double sequences to label the head and tail of the entity,uses two Linear layers for decoding,and adds the information extracted from the initial sequence to the decoding process of the end sequence to extract entities,which can obtain more semantic information in the text.In addition,using 20 pieces of electronic medical record data from the Ai Ai medical website,they were first manually labeled,and then experimented with the BERT-SPAN model.From the results,it can be seen that the model obtains an F1 value of 82.02%.Optimized on the basis of the BERT-SPAN model,and proposed the BERT-BiLSTM-SPAN model.Comparing this model with other researchers’ models,it is concluded that the BERT-BiLSTM-SPAN model does not require additional medical data to obtain an F1 value of 85.15%.At the same time,the BERT-BiLSTM-SPAN model was tested using the data from the Ai Ai medical website,and it was concluded that the model obtained an F1 value of 83.03%.It shows that the BERT-BiLSTM-SPAN model has high application value and can lay the foundation for subsequent electronic medical record text analysis.