Zero-Shot Multi-Modal Learning for Cross-Lingual Visual Hate Speech Detection

This paper attempts to deal with the practical challenge associated with hate speech detection in multiple languages and content formats across online environments. The approach is a novel zero-shot, multimodal, cross-lingual framework with the aid of pretrained models such as CLIP for visual feature extraction and XLM-RoBERTa for multilingual text processing. These two pretrained models help eliminate dependency on labeled datasets and make it robust in complex multimodal scenarios. Evaluations on the MMHS150K show high accuracy and generalization capabilities, which underline its practical applicability in fostering safer digital spaces and lay a foundation for future multimodal cross-lingual applications.