Detecting Hate and Inflammatory Content in Bengali Memes: A New Multimodal Dataset and Co-Attention Framework
Researchers have introduced Bn-HIB, a new dataset designed to detect hate and inflammatory content within Bengali internet memes. This dataset, containing 3,247 manually annotated memes, is notable for being the first to differentiate between inflammatory content and direct hate speech in Bengali. Alongside the dataset, a Multi-Modal Co-Attention Fusion Model (MCFM) was proposed, which analyzes both visual and textual elements of memes to improve classification accuracy. Experiments indicate that MCFM outperforms existing state-of-the-art models on the Bn-HIB dataset, and the dataset has been made publicly available. AI
IMPACT This research addresses a critical gap in low-resource language NLP, potentially improving content moderation and safety for Bengali-speaking online communities.