Research on Semi-Supervised Text Classification Based on Graph Attention Network

Abstract

Text classification is a basic task in natural language processing. Due to the sparsity of text data and the finiteness of tags, the feature construction method and the neural network used to learn features in text affect the subsequent classification effect. In order to make full use of text label data and improve the accuracy of text classification, we propose a semi-supervised text classification method combining topic information and graph attention network. Firstly, we construct text heterogeneous information graph by combining text self-information and additional information to better mine text semantics. Then, the Heterogeneous Graph Convolutional Network(HGCN) is used to learn the text graph. In particular, a two-layer attention mechanism is added to our method, including type-level attention and node-level attention. Through the attention mechanism, the model can understand the importance of different adjacent nodes and the importance of various types of information in the graph to the current node. A large number of experimental results show that our proposed model is significantly superior to other methods in three benchmark data sets.