Research And Application Of Document Level Relation Extraction Based On Heterogeneous Graph Neural Network

In the background of big data boom,huge amount of complex text data is gener-ated in various industries,which makes it difficult for people to filter key information.Extracting relational triplets from unstructured text is a key step in information filtering.Documents contain both intra-sentence and inter-sentence triplets,and previous studies have focused on the task of intra-sentence triplets extraction,which is difficult to cap-ture inter-sentence triplets with complex inter-sentence relational information.There-fore,the task of document-level relation extraction has received academic attention in recent years.However,the current document-level relation extraction tasks still suffer from the problems of difficult prediction inter-sentence triplets and correlation between relation labels.To address these problems,this thesis optimizes the model in terms of evidence guidance and label relevance.The main work of the thesis includes:(1)Proposing an Adaptive Evidence guidance Heterogenous Graph Neural Net-work model(Adapt Evi HGNN).The document evidence confidence is calculated using the evidence sentence prediction module,which guides the Heterogeneous Graph Neu-ral Network model to perform convolutional calculations on the nodes,which can better aggregate the local information of the nodes.The F1 scores of the Adapt Evi HGNN model are 61.12 and 69.84 by conducting comparison experiments on the Doc RED and CDR datasets,respectively.(2)Introducing a relation label sequence generation module based on the Adapt E-vi HGNN model.Proposing an Adaptive Evidence guidance Heterogenous Graph Neu-ral Network model(LSGDRE).The sequence generation module based on the Long and Short-Term Memory network completes the relation label prediction based on the histor-ical relation label classification results and the current document relation representation to further improve the performance of the Adapt Evi HGNN model.The final F1 score of the LSGDRE model is 61.41 in the experiments conducted on the Doc RED dataset.(3)Based on the LSGDRE model,we implement the knowledge-driven K-based search engine.First,crawling technology is used to crawl news documents,and then the LSGDRE model is used to extract the relational triplets in the documents and con-struct the knowledge graph.The K-based search engine makes full use of the structured information in the knowledge graph to improved the efficiency of knowledge search.