Heterogeneous Graph Neural Network for Short Text Classification

Abstract

Aiming at the sparsity of short text features, lack of context, and the inability of word embedding and external knowledge bases to supplement short text information, this paper proposes a text, word and POS tag-based graph convolutional network (TWPGCN) performs short text classification. This paper builds a T-W graph of text and words, a W-W graph of words and words, and a W-P graph of words and POS tags, and uses Graph Convolutional Network (GCN) to learn its feature and performs feature fusion. TWPGCN only focuses on the structural information of text graph, and does not require pre-training word embedding as initial node features, which improves classification accuracy, increases computational efficiency, and reduces computational difficulty. Experimental results show that TWPGCN outperforms state-of-the-art models on five publicly available benchmark datasets. The TWPGCN model is suitable for short text or ultra-short text, and the composition method in the model can also be extended to more fields.