Researc On Event Extraction Method Based On Chinese Electronic Medical Record

In recent years,the research on information extraction has received more and more attention,especially in domain-specific event extraction.Event extraction,as a key sub-task in information extraction,has become the focus of related research.Numerous research results have emerged,and event extraction methods based on rules,traditional machine learning and deep learning have been proposed one after another.However,there are not many researches on event extraction in the field of biomedicine.The reasons include the scarcity of relevant corpus,more professional nouns in biomedical texts,and differences in writing styles between biomedical texts and general texts.4What’s more,two different models are designed for these two stages to jointly complete the event extraction of Chinese electronic medical record.The main work of this paper is as follows:(1)Based on the existing Chinese electronic medical record,a Chinese electronic medical record event representation model is designed.This presentation model has four event categories,including diagnostic events,sign events,examination events,and treatment events,as well as eight types of event arguments.Based on the event representation model,this paper sorts out and annotates the existing Chinese electronic medical record,and constructs a Chinese electronic medical record corpus.(2)Recognition and classification of trigger words in electronic medical records based on dynamic multi-pooling convolutional neural network.The main task of the event discovery stage is to identify the event trigger words in the text and classify the events corresponding to the trigger words.This model addresses the problem that it is difficult for the existing Chinese event extraction model to capture the dependency between the trigger word and other word which in the same sentence.On the basis of integrating the semantic and syntactic features of dependency syntactic tree,several candidate trigger words of related events in a sentence are proposed by dynamic multi-pooled convolutional neural network.At the same time,the word level feature and the weight of the word level feature in the event trigger word recognition and trigger word classification are considered,and the weight is trained as a parameter of the neural network to realize the automatic distribution of the weight.(3)Event element extraction of graph convolutional neural network based on mixed syntactic features.Through experiments,this paper found that syntactic features play an important role in achieving the accuracy of event element extraction.Previous studies generally used word embedding to obtain semantic information,but they could not make full use of syntactic features.In this paper,we design a novel model called Hybrid Syntactic Graph Convolutional Networks,which mixes the feature representations at character level,word level and sentence level.And the graph convolutional neural network is used to obtain syntactic features and achieve more accurate extraction of event elements.Finally,according to the characteristics of electronic medical record,the results of event elements are treated with certain rules.