EIDER: Evidence-enhanced Document-level Relation Extraction

Anonymous

Abstract: Document-level relation extraction (DocRE) aims to extract the semantic relations among entity pairs in a document. Typical DocRE methods blindly take the full document as input, while a subset of the sentences in a document, noted as the evidence, are often sufficient for humans to predict the relation of an entity pair. In this paper, we propose an evidence-enhanced DocRE framework called Eider that automatically extracts and leverages evidence. We first train an evidence extraction model together with relation extraction via multi-task learning, which allows the two tasks to benefit from shared representations and improve each other. Experiments show that even if human annotation of evidence is unavailable, using silver evidence labels extracted by heuristic rules still leads to better RE performance. We further design a simple yet effective evidence-enhanced inference process that makes RE predictions on both extracted evidence and the full document and fuses the predictions through a blending layer. This allows Eider to focus on the important context while still having access to all the information in the document. Extensive experiments show that \ours outperforms state-of-the-art methods on three benchmark datasets, e.g., by 1.37/1.26 Ign F1/F1 on DocRED.