Research On Entity Relation Extraction In Medical Knowledge Graph

With the development of biomedicine and Internet technology,digital text information such as relevant materials,documents,and data in the field of biomedicine has shown an exponential growth trend in recent years.It is a very important task to transform a large amount of medical texts into medical resources through natural language processing technology and improve medical quality and health level.This paper conducts in-depth research on medical knowledge graph and entity relation extraction technology,studies data labeling issues and model interpretability issues,and finally constructs a medical knowledge graph.The main research of this paper is as follows:(1)Supervised algorithms have a strong dependence on labeled data,while manual labeling of data is time-consuming,laborious and expensive.Aiming at the time-consuming and labor-intensive problem of large-scale data labeling,this paper proposes a remote supervised relation extraction algorithm MILRE(Multiple Instance Learning Relation Extraction)based on MIL(Multiple Instance Learning).The model uses a paragraph-based encoding mechanism to embed contextual information,relaxes the traditional distant supervision assumption by using a self-attention mechanism,and uses knowledge distillation to reduce the error of machine annotation.On the data set NYT-10,the AUC score of MILRE reached 54.6,and the P@M score was 86.0,which can play an effective role in the research of automatic data labeling.(2)The entity relation in the medical field is complex,and the relation prediction work requires a high interpretability of the model.The existing models are difficult to meet the needs.Therefore,this paper proposes a medical entity relation extraction algorithm based on GCN(Graph Convolutional Network)MGCN(Medical Graph Convolutional Network).The model uses co-occurrence graph and graph convolutional network to build relation networks between entities,which can combine contextual information to provide global interpretability for relation predictions of medical entities;use Open World Assumption to construct potential relations between related entities,and leverage knowledge-aware attention mechanism gives the relation prediction of the entity pairs of attention,which can effectively solve the problem of relational extraction across sentence levels.On the data set CTF,the F1 score of MGCN reached 0.831,which proves its effectiveness in extracting medical entity relations and has important medical significance.(3)This paper uses MILRE to construct an electronic information health record corpus,further trains MGCN on it and performs relation extraction to obtain(entity,relation,entity)triples,and then stores the triples related information in the graph database Neo4 j.A medical knowledge graph was constructed and visualized.The research results of this paper provide convenience for the research work of scholars and doctors to a certain extent,and have very important significance and application value for the development of medical data mining and knowledge discovery research.