A Robust Graph Convolutional Network for Relation Extraction by Combining Edge Information

Graph convolutional network was widely used in natural language processing this year and achieved great success. However, existed Graph-based methods do not account for the edge information for node representation. In this paper, we propose a new graph convolutional network for relation extraction, which can embed the edge information for node representation. Specifically, we apply an edge embedding model to embed the edge label into vectors. Then we construct a new adjacent matrix which has multi-channel and can represent the edge information. During the computation of the hidden representation of the sentence, we split the adjacent matrix into several different 2D matrices where each one represents one channel matrix. Then the 2D matrices and tokens' embedding of the sentence are fed into the Attention Guided Graph Convolutional Networks (AGGCN) model to get the hidden representation of each one channel. Finally, the representations of different channels are concatenated together and then make a relation prediction through a feed-forward neural network (FFNN) and logistic regression classifier. The experiment results show that edge information contributes to the model's performance for relation extraction.