We present SZED's solution for the ArGuard 2026 Track A, which focuses on Arabic hateful meme detection. The competition consists of two subtasks: binary classification (A1) and fine‑grained multi‑label tagging (A2).
For A1, we fuse textual embeddings from MARBERTv2 with visual embeddings from CLIP ViT‑L/14. Three fusion strategies are evaluated: simple concatenation, a self‑attention based fusion module, and an ensemble of multiple heads. The self‑attention approach achieves the best macro‑F1 while keeping the model compact.
For A2, we adopt a multi‑task label‑attention architecture. Each label is assigned a query vector that extracts label‑specific information from the fused representation. An auxiliary head performs cascade calibration to mitigate the long‑tail distribution of labels. This design yields an average macro‑F1 of 0.3142 across the 10 active classes.
We report reproducibility across different random seeds, observing a variance of ±0.002 for A1 and ±0.015 for A2. Error analysis per class shows that rare labels suffer from data scarcity, but cascade calibration notably improves their recall.
Overall, the system achieves macro‑F1 = 0.7597 on A1 and macro‑F1 = 0.3142 on A2.
Review