Document-Level Entity Relation Extraction Based on Evidence Graph Reasoning
[Research purpose]To alleviate the sentence noise problem in the document-level entity relation extraction task and improve the extraction performance,we propose a document-level entity relation extraction method based on evidence graph reasoning,which of-fers valuable insights for research on document-level entity relation extraction and knowledge discovery.[Research method]Firstly,we capture the evidence sentence path information required by reasoning of relations between entity pairs through heuristic rules.Secondly,we incorporate the evidence sentence path information into the heterogeneous document graph based on graph structure learning ideas.Next,relational reasoning is performed using relational graph convolutional networks to enhance the document graph's ability to aggregate evi-dence sentence information.Finally,relation extraction results are predicted with the help of feed-forward neural networks.[Research conclusion]The model proposed in this paper achieves F1 values of 71.4％and 85.4％on the international public document-level evalu-ation datasets CDR and GDA,respectively,which are 1.2％and 1.1％higher than the benchmark model EIDER.Experimental results demonstrate that the approach presented in this paper can effectively select evidence paths required for entity relation reasoning,thereby en-hancing the performance of document-level relation extraction.
万方数据
维普
