To better learn node dependence and make use of structural information

this paper proposes a new method that takes the tree of complete dependence as the direct input. The method uses the graph convolutional network and combines two parallel attention modules to learn how to select the useful information. The method represents the samples as nodes on the graph. One module is used to compute the influence between positions of node features

which allows the feature vector to contain a wider range of semantic information. The other one is used to compute the relational features of node dependence

so as to enhance the global dependence between nodes. The two modules promote each other in parallel to obtain complete feature representation. The experimental results on the TACRED and SemEval datasets show that the method can obtain more useful information for relation extraction

thus achieve better performances on various evaluation metrics.