Research On Medical Entity Recognition And Relation Recognition For Chinese EMRs

With the rapid construction of medical information systems,the number of electronic medical records has increased dramatically.The electronic medical record contains not only the patient's clinical information(such as examination results,clinical diagnosis,etc.),but also a wealth of medical entities,most of which have some relation with each other.Using natural language processing technology to extract medical entities and semantic relation between entities from electronic medical record texts and construct a medical knowledge base that can be used for clinical decision making is of great significance for promoting the application of electronic medical records in smart medical treatment.At present,the research of electronic medical record medical entity recognition and entity relation recognition is mainly oriented to English electronic medical records,and there are few evaluations and corpora published on Chinese electronic medical records.In addition,existing research is mainly based on traditional machine learning methods,which rely on many manual constructed features.In summary,the research content of this article is the recognition of Chinese electronic medical records medical entities recognition and entity relation recognition.The main research work includes:(1)In terms of labeling rules and corpora,the Chinese electronic medical record entity relation annotate rules were developed,and the entity relation corpus was constructed by annotate.On the basis of referring to the English electronic medical record annotate specifications published in the 2010 I2B2/VA Challenge Evaluation,and based on the language and structural characteristics of the Chinese electronic medical record,the entity relation annotate rule applicable to the Chinese electronic medical record were formulated.Under the premise that the annotate personnel are familiar with the labeling rules,the traditional annotate scheme is adopted to ensure the authenticity of the annotate results by means of sample inspection.A total of annotate 3500 electronic medical records.(2)In the aspect of medical entity recognition,proposed a based on medical knowledge attention enhancement method.The medical knowledge dictionary has a detailed description of entity definitions,which can provide auxiliary information for entity recognition of electronic medical records.Therefore,in this paper,the word-level vectors extracted by CNN and the pre-trained word-level vectors are first spliced;then the sentences are encoded using bidirectional LSTM to extract the contextual representation of each word;then the medical knowledge dictionary is introduced using the attention mechanism to learn Chinese electronic medical records The shared semantics of entities in this and medical knowledge dictionary;finally,CRF is used to predict entity tags in sentences.Experiments on a manually marked electronic medical record entity corpus prove that the introduction of a medical knowledge dictionary can effectively improve the performance of entity recognition.The F1 value is 92.03%.(3)In the aspect of entity relation recognition,proposed a combining bidirectional GRU and attention method.Traditional relation recognition mostly uses sentences as processing units,ignores the influence of incorrectly annotation sentences in the corpus,and does not take full advantage of the mutual enhancement of multiple sentences containing entity pairs in the classification.Therefore,this paper first use bidirectional GRU to learn the context information of the character,obtain more fine-grained features.Then use the character attention mechanism to increase the weight of the character that determines the relation recognition.Finally,passes the sentence level attention mechanism acquired the features of multiple sentences and reduce the influence of error annotation on classification.The experimental results show that the method in this paper has achieved good results on the manually annotation corpus,with the F1 value of 82.17%.(4)In terms of joint recognition of medical entities and entity relation,proposed a joint recognition method of medical entity and relation recognition.Existing joint recognition method perform entity recognition and then relation recognition,which results errors in entity recognition propagating to relation recognition,ignoring the mutual support between the two tasks.Therefore,the paper firstly is spliced the word-level vector extracted by CNN with the pre-trained word-level vector;then the bidirectional LSTM is used to extract context representation of each word;then the CRF is used to predict the entity in the sentence;finally,the graph convolution network is used to entity nodes and relation nodes joint train,and output node-level vectors for relation classification.Experiments on a hand-annotated corpus show that the results of entity recognition and relation recognition are superior to other models,with F1 values of 88.05% and 84.81%.The experiment proves that our model effective entity recognition and relation recognition combined fully consider the mutual support between the two tasks.