Text domain generalization via domain fuzzification and fuzzy relation-aware contrastive learning

Domain generalization aims to train models on multiple source domains to generalize to unseen target domains. However, in natural language processing, domain boundaries are often ambiguous and overlapping, and target domain data are unavailable during training, making conventional methods, such as adversarial training or extensive data augmentation, inefficient and unstable. To address these challenges, we propose Fuzzy Domain Generalization (FDG), a framework that explicitly models domain uncertainty using fuzzy logic. FDG comprises two key modules: (i) a Domain Fuzzification module that assigns probabilistic domain memberships to samples, facilitating lightweight learning of domain-invariant features; and (ii) a Fuzzy Relation-Aware Contrastive Learning module that constructs fuzzy relation matrices to capture semantic similarities without costly augmentation. Convergence analysis shows that the fuzzification objective exhibits local convexity near convergence, leading to more stable optimization compared to adversarial alternatives. Extensive experiments on seven benchmark datasets demonstrate that FDG consistently outperforms 13 strong baselines, improving Macro-F1 by up to 0.87%, while reducing runtime by over 67.51% and introducing only 0.0008M additional parameters. Our implementation is publicly available at https://github.com/Balding-Lee/FDG.