Research On Document Level Relation Extraction Method Based On Graph Convolutional Neural Network

With the rapid development of social economy and the wide application of Internet technology,how to accurately obtain users target information from the vast amount of information on the Internet has become an urgent problem to be solved.In order to solve this problem,relation extraction technology emerges at the right moment.It aims to identify the semantic relationship between entity pairs from a given natural language text.It is widely used in knowledge graph,automatic question answering and other fields.Through the research on the existing relationship extraction methods,on the one hand,it can be found that the document-level relationship extraction has the problem of insufficient capture of entity information and long-distance dependence;on the other hand,it lacks effective fusion with other information(such as co-referential relationship),and insufficient analysis of global dependency information among document-level information.Based on the above problems,dependency are used to capture syntactic information,attention mechanism and graph convolution to capture topological relationships in the graph,and the contextual semantic features contained in relationship instances are combined to achieve effective relationship extraction.Firstly,to solve the problem of insufficient acquisition of entity information and longdistance dependence,this thesis proposes an entity relation extraction model(Bi LSTMGCN)based on bidirectional long and short term memory network and graph convolutional neural network.In the Graph Convolutional module,a Context Graph Convolutional Network(C-GCN)is used to encode the Graph information and construct a topology diagram to solve the problem of insufficient long-distance dependence and loss of entity information.Secondly,we propose a Graph Convolution Network Model Based on Bert and Attention Mechanism(BGA)to solve the problem of insufficient analysis of global dependence information,such as co-denotation.Will refer to the model as the main characteristics of application level in the document entity relationship extraction task,and use the BERT blend in syntax,refers to the characteristics of graph information,then input to the is based on the analysis of the interdependence syntactic convolution code in the module,and through the concentration mechanism to distinguish the characteristic of different entities,analysis refers to the relationship between total as dependent on global information,in the document class entity relation extraction task.Finally,this thesis takes the current mainstream relational extraction methods as the benchmark,conducts comparative experiments and results analysis on the proposed two models on the open data set of Doc RED,which verifies the feasibility and effectiveness of the proposed method.