Graph Convolutional Networks for Fast Text Classification

Recently, lots of studies have explored text classification methods based on graph convolutional neural network (GCN) technology. Compared with traditional deep learning methods, graph convolutional neural networks can capture global information while processing complex graph-structured data. However, when the previous GCN method deals with text classification problems, the entire corpus is often constructed as a complex heterogeneous graph. Such a graph structure faces the problems of the huge amount of calculation and long network training time when learning text representation. To solve the above problems, we simplified the construction of the adjacency matrix of the text heterogeneous graph based on Text GCN to reduce the amount of calculation during training. In addition, by constructing a special feature matrix, the graph convolutional neural network can extract a better text representation during training, while reducing the dimension of the feature matrix. The model performs text classification tasks on three data sets of R8, R52, and Ohsumed. The results show that the training speed of the proposed model on the three data sets of R8, R52, and Ohsumed is improved compared with the benchmark method (Text GCN) 71.5%, 72.6%, 78.6%. At the same time, the proposed model achieves an accuracy comparable to Text GCN on three data sets.