Incorporating Evidence Retrieval into Document-Level Relation Extraction by Guiding Attention

Abstract

Document-level relation extraction (DocRE) is the task of identifying all relations between each entity pair in a document. Evidence, a set of sentences containing enough clues for deciding the relation between an entity pair, has been shown to benefit relation extraction. Previous works tackle Evidence Retrieval (ER) and DocRE as separate tasks, while this work propose to incorporate ER directly into the DocRE model. Specifically, we guide the self attention mechanism to assign higher weights on evidence when encoding entity pairs. In this way we obtain contextualized representations focused on evidence. We further propose to learn ER on massive data without evidence annotations from automatically-generated evidence. Experimental results show that our approach exhibits state-of-the-art performance on DocRED and Re-DocRED, two popular benchmarks for DocRE, in both DocRE and ER.