Thinking Nodes at MAHED: A Comparative Study of Multimodal Architectures for Arabic Hateful Meme Detection
Year: 2025
Type: conference-paper
Abstract: This paper describes our system for Task 3 of the Arabic NLP 2025 competition: detecting hateful content in Arabic memes.The task requires a robust understanding of both visual and textual information and their interplay.We developed and compared three distinct multimodal fusion architectures: a Cross-Attention model, a progressive CNN-based fusion model, and a two-stage model using custom-trained embeddings with a gated fusion classifier.All models leverage pre-trained CLIP and MAR-BERT encoders for image and text representation, respectively.We detail our approach to handling the significant class imbalance in the dataset through data re-splitting and the application of a weighted Focal Loss.Our postcompetition analysis, training on all available data, shows that the CNN-based fusion model achieved the highest macro F1-score of 0.779, demonstrating the effectiveness of its hierarchical feature extraction for this task.
Author: Itbaan Safwan
Institution: Institute of Business Administration Karachi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037626