[Purpose/significance] In the context of sudden emergency events, content on social media poses severe challenges to online environments and emergency management due to the amplification effect of public opinion dissemi⁃nation. Traditional detection methods suffer from issues such as imbalanced classification data and low detection accu⁃racy, necessitating efficient solutions. [Method/process] This study constructs a dedicated dataset of content from Wei⁃bo comments in sudden emergency events. Generative Artificial Intelligence (GenAI) technology is employed to gener⁃ate semantically equivalent pseudo-toxic content through few-shot prompt learning, balancing the sample distribution.Furthermore, the MACBert-Att model is proposed by integrating the MACBert pre-trained model with a global atten⁃tion mechanism, enhancing the semantic capture capability for domain-specific terms and emotional expressions. [Re⁃sult/conclusion] Experiments demonstrate that with GenAI data augmentation, the MACBert-Att model achieves an F1 score of 0.95, representing a 15% improvement over the baseline Bert model and significantly outperforming tradi⁃tional augmentation methods like SMOTE. This validates the collaborative effectiveness of GenAI-based semantic-lev⁃el data augmentation and the model architecture.