基于路径推理图的文档级关系抽取模型研究OA

Relation extraction(RE)has recently shifted from the sentence level to the document level,which requires aggregating document information and using entities and mentions for reasoning.Existing research ignores local contextual information around the target entity pairs.In addition,existing work focuses only on entity-level inference paths without considering the complex interactions between long-distance en-tities across multiple sentences in a document.To this end,a new document-level relation extraction model with information aggregation and long-distance cross-sentence reasoning is proposed.A document-level graph is first constructed to model the global information in a docu-ment,and a new node is added to aggregate the local contextual information of target entity pairs.In addition,various paths between target en-tity pairs are integrated into a simple inference graph structure for long-distance cross-sentence entity pairs and perform relation inference.The experimental results on the three public datasets of DocRED,CDR and GDA show that the path reasoning model outperforms the compari-son models on F1,verifying the validity of the model.