MSAD4GL: Low-Resource Text Embedding and Classification with Multi-scale Attention Diffusion for Graph Learning

This paper presents the Multi-Scale Attention Diffusion Graph Learning (MSAD4GL) model, a groundbreaking solution for low-resource text embedding and classification. By addressing the constraints of traditional transformer-based models, which demand extensive data and computational power, MSAD4GL employs graph neural networks (GNNs) to represent textual data as graphs, effectively preserving long-range word relationships and global structural information. The model introduces a deep multi-scale attention diffusion mechanism that captures both local and global text graph representations, leading to significant enhancements in text classification performance, especially in low-resource and language-independent scenarios. MSAD4GL achieves remarkable results, outperforming traditional methods (NB, SVM, DT, word2vec, fastText) by up to 85.62% in F1 score, and surpassing contemporary models like PhoBERT and DADGNN by 6.69% and 3.96%, respectively, on real-world datasets.