mdp0999@CC-MMD 2026: Gated Cross-Attention Fusion of Qwen and CLIP for Cross-Cultural Misogynistic Meme Detection

Abstract

We present our system for Task B of the Cross-Cultural Misogynistic Meme Detection Challenge, addressing misogyny detection across Tamil, Malayalam, Chinese, and English memes. Meme comprehension inherently depends on the complex interplay between textual and visual cues, further complicated by culturally specific contexts that conventional fusion strategies struggle to model effectively due to their inability to account for asymmetric modality contributions. To address this limitation, we propose a Gated Cross-Attention Fusion (GCAF) architecture that dynamically weights image and text representations on a per-instance basis. Our framework integrates multilingual textual embeddings derived from Alibaba-Cloud Qwen/Qwen3.5-2B with visual embeddings extracted via OpenAI CLIP ViT-L/14, coupled through cross-attention mechanisms and adaptive sigmoid gating. To further mitigate the adverse effects of class imbalance, we incorporate Focal Loss, label smoothing, Weighted Random Sampling, and validation-driven threshold optimization into the training pipeline. Experimental results demonstrate that the proposed framework achieves competitive performance, underscoring the efficacy of adaptive multimodal fusion for culturally grounded misogynistic meme detection in multilingual, low-resource settings.