Enhanced graph convolutional network based on node importance for document-level relation extraction

Document-level relation extraction aims to reason complex semantic relations among entities expressed by multiple associated mentions in a document. Existing methods construct document-level graphs to model interactions between entities. However, these methods only pay attention to the connection relationship of nodes, yet ignore the importance of nodes decided by topological structure. In this paper, we propose a novel method, named Enhanced Graph Convolutional Network (EGCN), to extract document-level relations. Unlike previous methods that only model the connection relationship between two nodes, we further exploit the global topological structural information by measuring node importance. We merge these non-local relationship into a Graph Convolutional Network to aggregate relevant information. In addition, to model semantic and syntactic interactions in documents, we design a novel strategy to construct document-level heterogeneous graphs with different types of edges. Experimental results demonstrate that our EGCN outperforms the previous models by 5.54%, 1.7%, and 2.9% F1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_1$$\end{document} on three public document-level relation extraction datasets.