<p>Although hate speech shares common semantic patterns, the diversity of text sources and differences in language habits lead to significant cross-domain distribution shifts. Current deep learning-based detection methods, while significantly improving hate speech detection within a single dataset, face severe degradation in generalization ability across datasets due to insufficient modeling of domain-invariant features. To address this, this paper proposes an Unsupervised Contrastive Adversarial Domain Adaptation (UCADA) framework for implicit hate speech cross-domain detection. This framework integrates adversarial domain adaptation and contrastive learning to minimize domain differences and maximize class differences, effectively transferring knowledge between the source and target domains. Furthermore, since implicit hate speech exhibits similar semantic characteristics, this paper employs a multi-level data augmentation strategy to generate a pseudo-target domain. This approach bridges the domain gap between the source and target domains, forcing the model to learn more domain-invariant features. Extensive experiments demonstrate that UCADA significantly outperforms mainstream models in cross-dataset implicit hate speech detection, enhancing the model’s generalization capability across multiple datasets while exhibiting stronger robustness and accuracy.</p>