Research On Entity Relationship Extraction Technology Based On Graph Neural Network

Relation extraction is an important research direction in the field of natural language processing.How to effectively extract relation facts from a large number of texts has become a research hotspot in recent years.The mainstream methods at this stage usually use deep learning techniques for relation extraction,and most of these methods focus on relation extraction in a single sentence.However,most relational facts can only be extracted from a paragraph or a document.Due to the long input of paragraphs and documents,the existing deep learning methods cannot accurately locate the position of the relationship facts in the text,and cannot reason about the relationship across sentences.In response to the above problems,this thesis combines graph neural network technology to construct entity graphs to reason about the cross-sentence relationship,and combines attention distribution prediction to locate the relationship fact position.The main contributions of this thesis include the following three parts:1.This thesis proposes a document-level heterogeneous graph neural network algorithm(DH-GNN),whose purpose is to infer cross-sentence relationships.DH-GNN first uses deep learning related technology to obtain the high-level semantic feature representation of the entity.Secondly,it constructs the heterogeneous graph of the entity,and combines the graph neural network to extract the structural information of the entity heterogeneous graph.Finally,it uses the entity vector splicing method to present and predict relationship facts.The experimental results show that this method has a significant effect on the accuracy of document-level relation extraction compared with the existing deep learning methods.2.This thesis proposes a graph neural network relation extraction algorithm(JAHGNN)for joint attention distribution prediction,whose purpose is to effectively locate the position of the relation fact to improve the accuracy of relation prediction.JA-HGNN first obtains the local representation of the entity by predicting the attention distribution of the relationship facts,and combines the DH-GNN algorithm to obtain the global representation of the entity.Secondly,after aggregating the information of the entity’s global and local representations,the relationship facts are through the bilinear layer to make predictions.Finally,JA-HGNN conducts joint training on attention distribution prediction and relationship prediction to reduce the propagation error caused by pipeline training.Experimental results show that this method can effectively locate the relationship fact position,reduce the interference of irrelevant information,and improve the effect of document-level relationship extraction.3.In order to solve the problem that the existing relation extraction system has a large error in document-level relation extraction,this thesis combines the DH-GNN and JAHGNN algorithms to design and implement the graph neural network relation extraction prototype system.The graph neural network relation extraction system is based on the B/S architecture design,including three modules: model training,model prediction,and report recording.The browser side(B)is mainly responsible for the visualization of report records,and the server side(S)is mainly responsible for model training and model prediction.The system can effectively extract complex relationships among multiple entities in a document,and these extracted relationship facts can be used in the construction of knowledge graphs,intelligent question answering and other downstream tasks.