Joint Extraction Of Enhanced Entity And Relation Based On Graph Neural Network

Relationship extraction aims to identify the target relationships of text entities from such unstructured text information,which is an important technical link for building knowledge graphs and can also provide support for downstream tasks such as search engines and QA system,and has important research significance.The current deep learning-based relationship extraction techniques fail to fuse entity and relationship information well before extracting entities,and have the problem of ambiguous semantic representation of entities.To this end,this paper conducts a study on enhanced joint entity and relationship extraction based on graph neural networks,and the main research contents are as follows:(1)To address the ambiguity problem of previous relationship extraction models in expressing entity semantics,which fail to highlight the semantic information of entities and the contextual information between entities well,this thesis constructs a novel model structure based on Bert and Bi LSTM as the encoder model,and enhances the entity semantic information by using entity tagging and span interception methods.Firstly,the use of entity tagging makes the model efficient in distinguishing entity positions and improving the focus of the model on entities,while enabling the model to reduce the interference of irrelevant words.Second,the semantic vectors of entities are obtained using the span interception approach and combined with encoder head nodes to jointly enrich the contextual semantic information,and finally the classification results are obtained through the pooling and classification layers.The proposed model is experimented on Sem Eval 2010 Task8 dataset and obtains an F1 value of 89.41% and,which is a significant improvement over the traditional model,and the experiment proves that the method is effective and can improve the accuracy of relationship extraction.(2)For the current existing entity-relationship joint extraction model,the entity and relationship information are not well integrated before extracting entities,and both entity and relationship models exist in separate forms and are input into the model,ignoring the hidden connections between entities and relationships.In this thesis,we use Bert and Bi LSTM as the base model,and the entity model and relationship model are input to GAT by Bert encoder and parameter sharing,and use entities and relationships as nodes of the graph,so that all entity nodes are integrated into relationships and all relationships are integrated into entities,and enhance semantic information mutually by this iterative fusion.Sequence tagging method is used to mark the start position and end position of entities,and finally the correlation between two entities and relations is calculated to determine the triad.The experiments were conducted under three datasets,Sem Eval 2010 Task8,NYT,and Web NLG and obtained F1 values of 90.20%,91.92%,and 92.27%,which have significantly improved the accuracy compared with other improved models and verified that the model has better performance.