Few-shot Hate Speech Detection Based on the MindSpore Framework (arXiv)

Online communities are seriously threatened by the spread of hate speech on social media, necessitating efficient detection tools. Although deep learning models have demonstrated potential, their dependence on sizable annotated corpora sometimes causes them to perform worse in few-shot or low-resource scenarios. Using the MindSpore deep learning platform, the researchers provide MS-FSLHate, a prompt-enhanced neural framework for few-shot hate speech identification. To enhance generalization, the model incorporates synonym-based adversarial data augmentation, a CNN-BiLSTM backbone with attention pooling, and learnable prompt embeddings. Our method beats competitor baselines in accuracy, recall, and F1-score, according to results on two benchmark datasets: HateXplain and HSOL. The framework also exhibits great scalability and efficiency, indicating that it is appropriate for implementation in contexts with limited resources. These results demonstrate the possibility for strong and flexible hate speech identification in few-shot circumstances by integrating adversarial augmentation with prompt-based learning.

https://arxiv.org/abs/2504.15987