SC-DGCN: Sentiment Classification Based on Densely Connected Graph Convolutional Network

Abstract

Recently, various neural network frameworks have achieved good results in sentiment classification task, such as Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN). However, these methods only take into account semantic information in local contexts and ignore the global syntactic structure information due to the network structure. To solve this problem, we propose a novel neural architecture called SC-DGCN that combines Graph Convolutional Network (GCN) and Bi-LSTM. In SC-DGCN model, we utilize a GCN over the dependency tree of a sentence to exploit syntactical information and words dependencies. In addition, we further introduce dense connection strategy into GCN blocks to aggregate more syntactic information from neighbors and multi-hops in the dependency tree, and employ attention mechanism to generate the final representation of text. Our proposed SC-DGCN model can automatically extract semantic feature in local contexts and the global syntactic structure feature. A series of experiments on MR and SST datasets also indicate that our model is effective for sentiment classification task.