Enhancing Document-Level Relation Extraction with Attention-Convolutional Hybrid Networks and Evidence Extraction

Document-level relation extraction aims at extracting relations between entities in a document. In contrast to sentence-level correspondences, document-level relation extraction requires reasoning over multiple sentences to extract complex relational triples. Recent work has found that adding additional evidence extraction tasks and using the extracted evidence sentences to help predict can improve the performance of document-level relation extraction tasks, however, these approaches still face the problem of inadequate modeling of the interactions between entity pairs. In this paper, based on the review of human cognitive processes, we propose a hybrid network HIMAC applied to the entity pair feature matrix, in which the multi-head attention sub-module can fuse global entity-pair information on a specific inference path, while the convolution sub-module is able to obtain local information of adjacent entity pairs. Then we incorporate the contextual interaction information learned by the entity pairs into the relation prediction and evidence extraction tasks. Finally, the extracted evidence sentences are used to further correct the relation extraction results. We conduct extensive experiments on two document-level relation extraction benchmark datasets (DocRED/Re-DocRED), and the experimental results demonstrate that our method achieves state-of-the-art performance (62.84/75.89 F1). Experiments demonstrate the effectiveness of the proposed method.