Evidence-driven Entity Relation Matching Enhances Document-level Relation Extraction

To address the issues of limited evidence and a single extraction method in Document-level Relation Extraction (DocRE), we propose an evidence self-training approach and a novel DocRE processing method. We introduce two methods to enhance the performance of existing models: (1) utilizing an evidence self-training approach to generate evidence sentences on distantly supervised data and assisting the model’s performance in DocRE through data augmentation. (2) Inspired by Triplet completion in the knowledge graph, we present an entity and predefined relation-matching method, guiding the model to focus on evidence sentences in the document via evidence distribution to enhance the model’s inference and prediction of entity relations. The EdEM model proposed in this paper achieves 63.43 in F1 and 60.31 in Ign F1 on the Dev set of the DocRED dataset, and 62.27 in F1 and 60.26 in Ign F1 on the Test set. Experimental results demonstrate that our improved model outperforms current baseline models. Evidence self-training effectively addresses the issue of limited evidence and guides the model’s focus on the location of evidence sentences through evidence distribution, enhancing the performance of DocRE tasks.