A diffusion-inspired noise augmentation framework for robust multilingual text classification

To improve the robustness of text classifiers under spelling errors and semantic ambiguity, we propose NoiseDiffuser (ND), a task-oriented text augmentation framework inspired by diffusion-based perturbation and recovery principles. ND incorporates two key mechanisms. First, a length-aware dynamic noise schedule adjusts perturbation intensity according to text length, allowing stronger perturbations for redundant long documents and weaker perturbations for semantically sparse short texts. Second, a lightweight multilingual semantic recovery strategy uses HIT-CIR Tongyici Cilin for Chinese and WordNet for English to support synonym-based perturbation and lexical filtering. Evaluated on six corpora (e.g., THUCNews, 20Newsgroups), ND reduces the absolute magnitude of the generalization gap by 41.7%, increases feature cosine similarity by up to 33.33%, and reduces FGSM-induced accuracy degradation by approximately 40%. Furthermore, under TextFooler word-level adversarial attacks, ND-enhanced models achieve higher accuracy in most settings than baseline models. ND improves short-text F1 scores of baseline classifiers and BERT by approximately 38.63%. Overall, ND offers an efficient and compatible augmentation strategy for text classification in noisy environments while largely preserving the original semantics.