DRGIA: A document-level relation extraction method integrating graph Inference and attention mechanism

Abstract

Document-level relation extraction aims to extract relation facts between entities from unstructured documents. Compared to sentence-level relation extraction, document-level contains a large number of cross-sentence relations that requiring complex relational reasoning across multiple sentences. Existing methods face challenges including inadequate modeling of semantic information in documents, noise from irrelevant sentences in the document, and insufficient interaction among entity mentions. Therefore, we propose a document-level relation extraction model integrating graph aggregation and attention mechanism (DRAIN). The structural relations between entities, mentions and sentences are leveraged to construct the document structure diagram, and the paper combines with the Relational-Graph Convolutional Networks (R-GCN) to realize the aggregation of entity information. In addition, the key sentence extraction method and attention-based path coding mechanism are proposed to improve the relational reasoning ability and enhance the accuracy of relation extraction. Experimental results show that the proposed model outperforms well than existing models on two publicly available datasets, DocRED and CDR, demonstrating the effectiveness of the proposed method.