Research On Medical Relation Extraction And Distributed Representation For Knowledge Graph

With the continuous development of society and economic,people’s living standard have been greatl y improved,more and more attention has been paid to personal health and life problems.At the same time,the development of society also brings techno logical innovation and progress,In order to solve the contradiction between the increase of medical expenses and demands and the shortage of high-qualit y medical resources,smart medical technology emerges at the right moment,which will be a technology that can make full use of medical text knowledge to simulate doctors’ learning and diagnosis and treatment.The development of smart medical is inseparable from the stud y of medical knowled ge,on the whole,the sources of medical knowledge can be summarized as medical books,clinical data and network data.Based on the data from medical books and network corpus,this paper studied the relationship extraction technolog y and the distributed re presentation of knowledge graph in the process of knowledge graph construction.(1)Annotate medical data and train supervised relationship extraction model.A piecewise pooling convolutional neural network classification model based on self-attention mech anism is proposed,and features such as entit y category are introduced to make F1 reach 87.2%.(2)The distant supervision relationship extraction without manual annotation was explored,and attempts were made on how to reduce the noise of remote annotatio n.Rule-based method,attention-based method and reinforcement learning method were successivel y adopted for noise reduction,and the final AUC reached 0.489.(3)For the basic knowledge graph constructed,it is transformed from symbolic representation to dense low-dimensional real value vector representation which is easier to provide information input for other tasks.Based on the trained vector representation model,the relationship in the knowledge graph is predicted to achieve the goal of completing th e knowledge graph.After the above research,a basic version of medical knowledge graph was constructed from various medical books and websites.In addition,the distant supervision relationship extraction framework constructed in this paper can be applied to most medical corpus,even clinical texts,without manual annotation,which means that almost all medical knowledge can be included in our existing knowledge graph s ystem in the future.Further,through researching the distributed representation of the knowledge graph,the vector representation information of the graph can be directl y provided to tasks such as dialogue question answeri ng,diagnostic reasoning,etc.which can better play the role of smart medical treatment,reduce the number of doctors’ diagnosis and diagnosis time,promote high-qualit y medical treatment,accelerate the sinking of high-qualit y medical resources and make up for the shortage of medical resources.