Short Text Classification with Vertex Entanglement Enhanced Graph Contrastive Learning

Short text classification is one of the core tasks in the field of natural language processing, which has attracted extensive attention due to problems such as semantic sparsity and insufficient labeled data. Some recent researches have combined graph neural networks with contrastive learning to mitigate the aforementioned problems in short text classification. However, existing methods also have certain limitations in that they generate augmented views by randomly perturbing short texts or the nodes and edges of constructed text graphs, which results in the loss of important information. To this end, we propose a novel Vertex Entanglement Enhanced Graph Contrastive Learning Short Text Classification (VGSTC) method. Specifically, we construct a word-level component graph consisting of a word graph, an entity graph, and a part-of-speech tag graph. In addition, we design an enhancement strategy based on vertex entanglement, which preserves relatively important edges by perturbing unimportant connections to generate augmented views for contrastive learning. To validate the effectiveness of proposed method, we conducted extensive experiments on four benchmark datasets including MR, Snippets, StackOverflow and Ohsumed. The experimental results show that the VGSTC improves the short text classification performance by 1.89 \% 17.26 \%, and 0.04 \% 17.3 \%, in terms of accuracy and F1 score compared with the baseline methods.