Year: 2025
Type: conference-paper
Abstract: We present our system for the MAHED 2025 Shared Task on Arabic Hate Meme Detection (subtask 3), a binary classification task to determine whether a multimodal meme containing Arabic text and an image conveys a hateful message.Our approach uses multimodal fusion combining a visual encoder and an Arabic text encoder.We explored four fusion strategies-transformer fusion, early fusion, cross-attention, and bilinear fusion-and found transformer fusion offered the best singlemodel trade-off, while an ensemble of all four achieved the highest score.To address the severe class imbalance (90.05% not-hate vs. 9.95% hate), we applied class-weighted loss, focal loss, strong regularization, and light augmentation.Our best submission reached a macro-F1 score of 0.75 on the gold test set.
Author: Yassir El Attar
Institution: University of Stuttgart
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037752