Gender-based hate speech research in Bengali has largely centered on misogyny, leaving anti-male hate, or misandry, unexamined. According to a new arXiv preprint, this creates a systematic blind spot in moderation systems trained on existing data, which may fail to recognize or classify this form of abuse.
The paper introduces a Bengali dataset built for detecting misandry, along with a late-fusion transformer framework designed to identify such content. The approach combines multiple transformer representations, though the abstract does not specify which features are fused or how the dataset was constructed.
By targeting an understudied category, the work aims to broaden the scope of Bengali hate-speech detection. The authors argue that moderation systems need balanced coverage of both misogyny and misandry to be effective, rather than focusing on one direction of gender-based hate.