A Multiagent-Based Document-Level Relation Extraction System with Entity Pair Awareness and Sentence Significance

Document-level relation extraction extracts relations between entity pairs from multiple sentences in a document by reasoning about semantic relationships between entities and entity mentions. In this article, we propose a multiagent system that extracts relations between entity pairs by capturing entity semantics, mentions, and pairs. First, we employ a pretrained language model to capture entities, entity pairs, and sentence context. Then, we design a multisemantic relation graph between entities and entity pairs, based on which a graph convolutional network is constructed for information reasoning and relation extraction in a relation extraction agent. Next, we introduce an evidence sentence extraction agent to focus on a relevant subset of inputs. Finally, we implement collaboration between two agents in a multiagent collaborative optimization module. Experimental results on three public datasets show that our method outperforms the existing methods in terms of F1 (harmonic mean of precision and recall) and IgnF1 (F1 on ignored relations). Specifically, our method outperforms the state-of-the-art method by 0.91 F1 and 0.65 IgnF1 on DocRED, 1.2 F1 on CDR, and 1.2 F1 on GDA.