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. In DocRE, we observe that (1) a subset of the sentences in a document, noted as the evidence sentences, are often sufficient for predicting the relation between a specific entity pair; (2) these evidence sentences can be extracted in an effective and lightweight manner: by multi-task learning along with the RE model or by heuristic rules. In this paper, we propose a novel DocRE framework called Eider that automatically extracts and makes use of evidence. Eider enhances a DocRE model by combining the inference results from the evidence sentences and the original document through a blending layer. The performance can be further improved by jointly training an RE model with an evidence extraction model via multi-task learning. If human-annotated evidence is not available, we can use the evidence extracted by this joint model or by several heuristic rules. Extensive experiments show that Eider achieves state-of-the-art performance on the DocRED, CDR, and GDA datasets. Remarkably, Eider outperforms the runner-up by 1.37/1.26 Ign F1/F1 on DocRED. In particular, Eider-RoBERTa$_\text{large}$ significantly improves the performance on entity pairs requiring co-reference/multi-hop reasoning by 1.98/2.08 F1, respectively.