Research On Joint Entity Relation Extraction Based On Graph Neural Network

Entity relation extraction is a basic task of information extraction,which can provide basic support for many downstream tasks such as automatic question answering,semantic analysis,and knowledge base filling.Traditional entity relation extraction adopts pipeline method,which has significant shortcomings,namely error propagation,lack of interaction,information redundancy,and pipeline method cannot solve the problem of entity-relation overlap.Therefore,researchers propose a joint entity relation extraction method.In order to alleviate the shortcomings of the pipeline method and solve the problem of entity-relation overlap,this paper mainly uses the graph neural network to carry out the following three aspects:1.To solve the problem of entity-relation overlap,we proposed a joint entity relation extraction model BSGB（Bi LSTM+SDA-GAT+Bi GCN）combining semantic dependency and graph attention network.BSGB is divided into two stages.The first stage extends semantic dependency analysis to semantic dependency graph,proposes a graph attention network（SDA-GAT）fused with semantic dependency graph,extracts sentence sequence and local dependency features,and performs entity span detection and Preliminary relationship prediction;in the second stage,the relationship weighted GCN is constructed,and the interaction between entities and relationships is further modeled.A graph is constructed for each relationship in the sentence,and graph convolution operation is performed,which solves the overlapping of entity pairs to a certain extent（EPO）,single entity overlap（SEO）type sentence overlap problem,and finally use the Softmax function to complete entity relationship triple extraction.The experimental results on the NYT dataset show that the F1 value of the model reaches 67.1%,which is 5.2%higher than the baseline model Graph Rel in this dataset.2.Aiming to the problem of automatic syntactic dependency tree generation and noise in dependency tree may bring confusion to relation classification and affect the overall efficiency of model,we proposed A joint entity relation extraction model BCAG（BERT+CNN+A-GCN）,which combines potential relation prediction and global correspondence.,the model is divided into three parts:（1）The attention graph convolutional neural network model（A-GCN）is used to model the dependencies between words and the types of relationships.The force mechanism distinguishes the importance of different contextual features,identifies noise in syntactic knowledge,predicts latent relations in sentences,and generates a global correspondence matrix.（2）Through two sequence annotations,two entities in the sentence are respectively identified.（3）Use the global correspondence matrix to align the identified entities and relationships to solve the entity-relation overlap problem.The experimental results on the NYT dataset show that the overall F1 value of the model reaches89.8%,which is 0.2%higher than the baseline model Cas Rel _BERT.3.Through sorting out and integrating the above research contents,the proposed model is encapsulated,and a prototype system of joint entity relation extraction is designed and developed for practical application.