Research On Document-level Relation Extraction Based On Graph Neural Networks And Relation Inference

With the advent of the era of big data and the continuous improvement in the abil-ity to collect and store data,the variety and quantity of long texts are increasing rapidly.However,these unstructured long texts usually cannot be used directly.Extracting struc-tured data from unstructured text and then enhancing the value of long texts has become an urgent problem.Document-level relation extraction is a powerful method to solve this problem,which can identify relations between entities from long texts.It provides sup-port for downstream natural language processing applications such as knowledge graphs,intelligent question and answer,etc.This thesis studies document-level relation extrac-tion based on graph neural networks and fused with relation inference.The main work is summarized as follows:（1）In order to alleviate the problem of long-tail dependency in long text relation ex-traction,a document-level relation extraction method based on heterogeneous graph neural networks and incorporating local semantic features is proposed,which can capture the in-teractions between entities by making full use of the type information of nodes and edges in the document graph.Firstly,a mention-sentence group graph is built.Relational graph convolutional neural network and heterogeneous attention neural network are used to per-form information flow and feature aggregation in the graph respectively.Based on both graph neural networks,56.88%and 57.24%Ign₁are achieved on the Doc RED dataset,showing an improvement of 9.52%and 7.65%over the baseline model（GAT,GCN）.These results illustrate the effectiveness of document-level relations extraction based on heterogeneous graph neural networks.Besides,the impact of information at different granularities（sentence granularity,sentence group granularity and global document gran-ularity）on relation extraction is investigated.The experimental results demonstrate that the proposed sentence cluster node extraction is the best,which shows that the sentence group node is more suitable than other granularities.This supports the effectiveness of the improved algorithm proposed in this thesis.（2）To further explore the implicit representations of relations and improve the re-lation extraction results,relation inference is incorporated into the heterogeneous graph neural network based model.Since inference is performed in the entity-relation graph,multi-hop path inference is utilized.As an important part of the representation,path infer-ence not only contains rich node information but also provides inference guidance.Three inference methods are used:shortest path inference,long-short path inference,and limited path inference.Experiments prove that the limited path reasoning achieves the best results.The model achieves 59.36%and 62.63%Ign₁on the Doc RED and DWIE datasets.It improves by 2.12%over the model that do not use relation inference,illustrating the im-portance of relation inference.Compared with recent representative models,the results are improved to some extent on both datasets.In summary,it shows that the document-level relation extraction method is effective combining graph neural networks,sentence group nodes and path inference.