SagDRE
Ying Wei, Qi Li
Abstract
Relation extraction (RE) is an important task for many natural language processing applications. Document-level relation extraction task aims to extract the relations within a document and poses many challenges to the RE tasks as it requires reasoning across sentences and handling multiple relations expressed in the same document. Existing state-of-the-art document-level RE models use the graph structure to better connect long-distance correlations. In this work, we propose SagDRE model, which further considers and captures the original sequential information from the text. The proposed model learns sentence-level directional edges to capture the information flow in the document and uses the token-level sequential information to encode the shortest paths from one entity to the other. In addition, we propose an adaptive margin loss to address the long-tailed multi-label problem of document-level RE tasks, where multiple relations can be expressed in a document for an entity pair and there are a few popular relations. The loss function aims to encourage separations between positive and negative classes. The experimental results on datasets from various domains demonstrate the effectiveness of the proposed methods.