A Short Text Classification Method Based on Cross-Source Heterogeneous Graphs and Adaptive Enhancement

Senye Chen

Abstract: Short text classification often suffers from semantic sparsity and limited contextual information. To address this challenge, this study proposes a short text classification model based on cross source heterogeneous graphs and adaptive augmentation. The method integrates short text data with external knowledge to construct a heterogeneous graph composed of words, entities, and part of speech information. A graph convolutional network is used to learn semantic representations of different node types. Text node association graphs are then enhanced through adaptive augmentation to generate complementary views. These views are fused to produce robust text representations, and contrastive learning is employed to improve semantic alignment for classification. Experimental results on MR, Twitter, and Snippets datasets demonstrate that the proposed model improves classification accuracy compared with several baseline methods. The results indicate that the proposed approach effectively enhances semantic representation and improves short text classification performance.

Keywords: Short text classification, Graph neural networks, Contrastive learning, Adaptive augmentation