Research On Medical Named Entity Recognition Method Based On Deep Learning

Named entity recognition(NER)is a very fundamental task in the field of natural language processing(NLP).It is widely applied in many NLP tasks,such as machine translation and question answering system.And,the performance of named entity recognition directly determines the effectiveness of downstream tasks..Medical NER refers to the recognizing of specific medical named entities from medical texts,which is the fundamental work of intelligent medical service,medical knowledge mining,medical clinical decision support system and other high-level applications.Because the medical NER technology started late in China,the existing work on Chinese clinical text has many limitations.Therefore,this paper attempts to integrate and improve the current deep learning model to carry out the research on Chinese medical NER task.The main work and contributions are reflected in the following three aspects:(1)This dissertation proposes a medical NER method based on stacked neural network.Because the text in the medical field is highly professional,the traditional methods are not good at recognizing the entity with fuzzy boundary.In addition,as medical texts are usually long and information is dependent on a long distance,single layer neural network is difficult to capture these long-distance features.To solve this problem,this dissertation proposes a method of medical NER based on stacked neural network,which encodes medical texts by using multi-layer stacked bidirectional recurrent neural network,and make use of conditional random field(CRF)to select the optimal sequence annotation path,so as to improve the performance of NER task.(2)This dissertation proposes a medical NER method based on iterated dilated convolutional neural network.Considering that the application of convolutional neural network in NER task will ignore the previous input,if the convolutional neural network is superimposed simply,more local information will be extracted and the features of long sentences can’t be acquired.In view of this problem,this paper proposes a medical NER method based on the iterated dilated convolutional neural network,which can solve the problem of incomplete feature capture by utilizing the iterated dilated convolutional neural network and expanding specific feature words to the data sets,so as to improve the entity recognition performance in the medical field.(3)This dissertation proposes a medical NER method based on BERT pre-training.Considering to make further optimization in the direction of word representation,this dissertation designs a medical NER method based on BERT pre-training.Because the static word for traditional word representation method are not flexibile enough,the poor selection of feature extractor and unidirection language model affect the performance,this dissertation attempts to adopt BERT pre-training character embedings to replace the traditional words embeddings for CRF network to provide enough prior knowledge(information)such as grammar,meaning,so as to improve the performance of medical NER task.