Cross-Corpus Transfer and Label-Calibrated Learning for Imbalanced Multi-Label Bangla Hate Speech Detection

The research introduces a cross-corpus encoder transfer framework for mismatched Bangla hate-speech taxonomies, using BANHATE as the source corpus and BOISHOMMO as the imbalanced multi-label target task. It further contributes a leakage-aware evaluation protocol, imbalance-aware weighted learning, validation-only threshold calibration, and three-seed probability ensembling without using the test set for model selection. The final system is evaluated once on a frozen test set with bootstrap uncertainty and post-hoc explainability analysis using Integrated Gradients.