Research On Named Entity Recognition Of Chinese Image Reports Based On Recurrent Neural Networks

The rapid development of medical information system has produced a large number of clinical texts,which record a large amount of health information in the form of text.These clinical texts are valuable to improve clinical nursing and support research.At the same time,it is a great challenge to automatically manage and effectively utilize these massive unstructured data.Image report is a kind of typical clinical text.Extracting medical concepts from unstructured image reports can provide a good support for medical record classification,fine patient classification and automatic question answering system.As a kind of professional clinical text,the Chinese image report has distinct language features and more complex language forms.At the same time,we lack public corpus of the Chinese clinical texts.Thus,it would be faced more difficulties to analyze and apply the Chinese image report,compared with general domain texts.Aiming at the Chinese image reports,this thesis studies the methods of named entity recognition.The main work includes the following aspects:Firstly,with reference to the I2B2 experience,this thesis defines the annotation of named entity recognition of image reports combined with the characteristics of Chinese image reports and the guidance of medical staff.In this paper,two image report corpus with a scale of 39954 characters are established according to word segmentation and character segmentation respectively.Secondly,according to the image report corpus,taking the bidirectional recurrent neural network(RNN)as the basic framework,this thesis constructs three Chinese image report named entity recognition models based on Long Short-Term Memory(LSTM)unit,gated recurrent unit(GRU)and conditional random field(CRF).Thirdly,the Chinese image report named entity recognition models are trained through the image report corpus.Several experiments are conducted to evaluated the performance of models;the experimental results show that compared with the traditional CRF model,the bi-directional recurrent neural network has better adaptability in Chinese image reporting entity recognition;in addition,compared with the word segmentation tagging method,the better performance could be obtained by character segmentation and annotation.Finally,this thesis designs and implements a prototype system of Chinese image report named entity recognition,which directly shows the process and results of Chinese image report named entity recognition.