Memes, combining images and text, are a popular social media medium that can spread humor or harmful content, including misogyny—hatred or discrimination against women. Detecting misogynistic memes in Malayalam is challenging due to their multimodal nature, requiring analysis of both visual and textual elements. A Shared Task on Misogyny Meme Detection, organized as part of DravidianLangTech@NAACL 2025, aimed to address this issue by promoting the advancement of multimodal machine learning models for classifying Malayalam memes as misogynistic or non-misogynistic. In this work, we explored visual, textual, and multimodal approaches for meme classification. CNN, ResNet50, Vision Transformer (ViT), and Swin Transformer were used for visual feature extraction, while mBERT, IndicBERT, and MalayalamBERT were employed for textual analysis. Additionally, we experimented with multimodal fusion models, including IndicBERT+ViT, MalayalamBERT+ViT, and MalayalamBERT+Swin. Among these, our MalayalamBERT+Swin Transformer model performed best, achieving the highest weighted F1-score of 0.87631, securing 1st place in the competition. Our results highlight the effectiveness of multimodal learning in detecting misogynistic Malayalam memes and the need for robust AI models in low-resource languages.

Social media has become a widely used platform for communication and entertainment, but it has also become a space where abuseand harassment can thrive. Women, in particular, face hateful and abusive comments that reflect gender inequality. This paper discussesour participation in the Abusive Text Targeting Women in Dravidian Languages shared task at DravidianLangTech@NAACL 2025, whichfocuses on detecting abusive text targeting women in Malayalam social media comments. The shared task provided a dataset of YouTubecomments in Tamil and Malayalam, focusing on sensitive and controversial topics where abusive behavior is prevalent. Our participationfocused on the Malayalam dataset, where the goal was to classify comments into these categories accurately. Malayalam-BERT achievedthe best performance on the subtask, securing 3rd place with a macro f1-score of 0.7083, highlighting the effectiveness of transformer models for low-resource languages. These results contribute to tackling gender-based abuse and improving online content moderation for underrepresented languages.

Detecting misogyny in multimodal content remains a notable challenge, particularly in culturally conservative and low-resource contexts like Bangladesh. While existing research has explored hate speech and general meme classification, the nuanced identification of misogyny in Bangla memes, rich in metaphor, humor, and visual-textual interplay, remains severely underexplored. To address this gap, we introduce BanMiMe, the first comprehensive Bangla misogynistic meme dataset comprising 2,000 culturally grounded samples where each meme includes misogyny labels, humor categories, metaphor localization, and detailed human-written explanations. We benchmark the various performance of open and closed-source vision-language models (VLMs) under zero-shot and prompt-based settings and evaluate their capacity for both classification and explanation generation. Furthermore, we systematically explore multiple fine-tuning strategies, including standard, data-augmented, and Chain-of-Thought (CoT) supervision. Our results demonstrate that CoT-based fine-tuning consistently enhances model performance, both in terms of accuracy and in generating meaningful explanations. We envision BanMiMe as a foundational resource for advancing explainable multimodal moderation systems in low-resource and culturally sensitive settings.