Study exposes generalization crisis in Bangla hate speech detection

A recent study diagnoses why benchmark-trained models fail to identify implicit, context-dependent hate speech in under-resourced languages like Bangla. Researchers evaluated six architectures on benchmark and multi-source datasets, then validated them on an external set of social media posts containing explicit and implicit hate speech.

• BanglaBERT achieved an F1-score of 91.4% on benchmarks but declined to 75.3% on the external set and 63.4% for implicit hate speech involving sarcasm and emojis.
• FastText + CNN accuracy dropped from 78.0% to 51.2% under similar external validation conditions.
• Emoji-aware preprocessing improved implicit hate speech detection by up to 12%, while emoji removal caused a notable performance decline (F1: 0.75 to 0.63).
• Frequent misclassifications in politically charged or satirical comments revealed risks of over-policing.

The findings highlight the need for adaptive, emoji-aware, and culturally grounded frameworks to ensure ethical moderation while preserving freedom of expression.