Abstract:
Humor and sarcasm detection is essential for intelligent agents to revolutionize customer
service, healthcare, and other industries by enabling truly natural and empathetic in
teractions. However, detecting sarcasm and humor in Bengali videos, one of the most
spoken languages in the world, is significantly challenged by the language’s intricate
linguistic nuances and cultural specificity. As per our knowledge, this work introduces
the first publicly available multimodal dataset for Bengali humor detection, consisting of
600 annotated video clips categorized into humor, sarcasm, and normal. Context-Aware
Multi-Modal Fusion Framework (CAMFusion) is proposed, integrating visual and textual
features. Visual features are extracted from a Time-Distributed MobileNetV2 with 0.985
million parameters, enhanced with reverse fusion mechanism and squeeze and excitation
blocks. In contrast, textual features are processed through Bi-LSTM and Bi-GRU networks
with group-wise enhancement. A new Bengali Text Extraction Algorithm ensures robust
text retrieval from complex video frames. CAMFusion achieves an accuracy of 90%
through dynamic attention-based fusion with 13.88 million parameters, significantly out
performing state-of-the-art baselines, including MobileNetV3 with XLM-R (81.10%) and
TimeSformer-L with DeBERTaV3 (84.45%). This work establishes a new benchmark for
Bengali humor detection, addressing a critical gap in low-resource language research and
advancing multimodal computational linguistics for culturally sensitive AI applications.
