Enhancing cross-lingual hate speech detection through contrastive and adversarial learning

Abstract

The rise of hate speech on social media platforms, particularly in low-resource languages, necessitates innovative solutions. In response, we introduce a zero and few-shot model combining supervised contrastive learning and adversarial training. To address the scarcity of labeled data in diverse languages, our approach adapts features from well-resourced languages to efficiently detect hate speech in low-resource contexts. The proposed framework first leverages supervised contrastive learning, maximizing the utility of limited labeled data by transferring knowledge from source languages. This adaptation empowers the accurate detection of hate speech in underrepresented languages, optimizing available resources. We then introduce contrastive adversarial training, refining hate speech representations in low-resource languages. This approach ensures a nuanced understanding of hate speech across linguistic boundaries, significantly enhancing the model’s adaptability and accuracy. To validate our approach, we conducted zero-shot and few-shot cross-lingual evaluations in three languages. Our results demonstrate the superiority of the proposed contrastive learning-based models. To ensure reproducibility, the code associated with this paper is available on GitHub (Almahdi, 2024). .