Densely Connected Graph Attention Network Based on Iterative Path Reasoning for Document-Level Relation Extraction

Hongya Zhang, Zhen Huang, Zhenzhen Li, Dongsheng Li, Feng Liu

Abstract: Document-level relation extraction is a challenging task in Natural Language Processing, which extracts relations expressed with one or multiple sentences. It plays an important role in data mining and information retrieval. The key challenge comes from the indirect relations expressed across sentences. Graph-based neural networks have been proved effective for modeling structural information among the document. Existing methods enhance the graph models by using either the attention mechanism or the iterative path reasoning, which is not enough to capture all the effective structural information. In this paper, we propose a densely connected graph attention network based on iterative path reasoning  (IPR-DCGAT) for document-level relation extraction. Our approach uses densely connected graph attention network to model the local and global information among the document. In addition, we propose to learn dynamic path weights for reasoning relations across sentences. Extensive experiments on three datasets demonstrate the effectiveness of our approach. Our model achieves 84% F1 score on CDR, which is about 16.3%–22.5% higher than previous models with a significant margin. Meanwhile, the results of our approach are also comparably superior to the state-of-the-art results on the GDA and DocRED dataset.