Explanation Faithfulness under Intra-Language Distribution Shift in Low-Resource Bangla Sentiment Classification

1. First evaluation of explanation faithfulness under natural intra-language distribution shift in a low-resource South Asian language (Bangla).

2. All attribution methods show significant raw comprehensiveness drop (32.5–38.7%, p GradientSHAP > Attention) confirms Attention as least robust, with its margin over random shrinking >4× faster than Integrated Gradients.