Named Entity Recognition In Medical Field

With the development of electronic information technology,electronic medical records are more and more popular in many hospitals.Different from the traditional paper medical records,electronic medical records are easier for storage management.Electronic medical records generally store many important information of patients with symptoms of the disease,clinical diagnosis and treatment.It has a close connection with patients' health.So the mining and analysis of the electronic medical data has received extensive attention in recent years.And the entity recognition regarded as an important basic task of Natural Language Processing is of great significance.Deep learning has been developing rapidly and popularized in recent years.Recurrent neural network(RNN),especially the long and short term neural network(LSTM),has achieved very good results in many fields of natural language processing.Meanwhile,convolution neural network achieves excellent performance in extracting character level features of words.Firstly the paper uses LSTM to model the named entity recognition task in medical field,introduces the model structure in detail,and adds character vectors in the model.The experiment shows that adding character vectors can improve the performance of the model.Then it introduces how to use CNN to extract the expression information based on characters,and put it into the LSTM model.In the experimental part,the effect of word tagging and character tagging on data is compared and analyzed.Finally,through combining the bidirectional long and short term memory network with conditional random fields and the convolution neural network,the BLSTM-CNN-CRF model is constructed.This model realizes the end to end prediction without any manual selection of features and preprocessed data.Therefore,the applicability of the model is very wide and can be applied to various fields.In the experiment part,a number of comparative experiments are conducted.The F1 value of the BLSTM-CNN-CRF model on test data hasreached 90.97%,which is 1.29% higher than the BLSTM model and is 1.64%higher than the BLSTM-CNN model.