HateRaid: Multimodal Hate Speech Detection

Fine-tuning cross-modal transformers for contextual meme classification.

The Challenge: Detecting hate speech in memes requires understanding the non-linear relationship between text and visual context (e.g., benign text paired with a hostile image).
Technical Approach: Deployed and fine-tuned a Multimodal Bi-transformer (MMBT) on the Meta Hateful Memes Dataset. The model fuses ResNet visual features with BERT text tokens inside a unified cross-attention space.
Results & Validation: Reached a validation accuracy of 92.6%, utilizing structured data augmentation (image masking, text synonym replacement) to counter label imbalance.
Links: GitHub Codebase

Our Approach Diagram A visual explanation showing how text and image combine to create hateful content.