When Humor Hides Harm: A Fusion-Based Approach to Decoding Sentiment in Bengali Memes

Novel Dataset: A manually curated and annotated dataset of 3,000 Bengali social media memes, labeled for both sentiment (Offensive/Non-Offensive) and thematic category (Entertainment, Sports, Religion, Politics, Others) — a valuable resource for a low-resource language where such multimodal datasets are scarce.
Multimodal Fusion Framework: An OR-operation-based decision-level fusion approach that combines textual predictions (SVM, LR, RF, BERT) with visual predictions (ResNet50, VGG19) to jointly determine offensive content, demonstrating that multimodal integration outperforms unimodal approaches.
Comparative Model Evaluation: Systematic benchmarking of eight text–image fusion combinations, establishing that BERT+VGG19 achieves the highest performance (77% accuracy), and showing VGG19 consistently outperforms ResNet50 as the visual backbone across all pairings.
Thematic Categorization: Introduction of a five-category content classification scheme (beyond simple polarity labels), enabling more nuanced, culturally grounded analysis of Bengali meme content — an angle largely unexplored in prior Bengali sentiment analysis work.
Contribution to Low-Resource NLP: Advances content moderation tooling for Bengali, a linguistically and culturally underrepresented language in multimodal sentiment analysis research, addressing a real gap relative to well-resourced languages like English.