joy_2004114 at MAHED Shared Task : Filtering Hate Speech from Memes using A Multimodal Fusion-based Approach
Year: 2025
Type: conference-paper
Abstract: Social media platforms have become major spaces for sharing opinions, humor, and information through memes that blend images with text.While many memes are harmless, some promote hate speech against individuals or communities based on cultural, religious, gender, or national identity.Detecting such content in Arabic is particularly challenging due to linguistic complexity, cultural context, and limited annotated data.In this study, we present an effective approach for detecting hateful content in Arabic memes using the QCRI Prop2Hate-Meme dataset, which contains image-text pairs labeled for hatefulness.We experimented with several multimodal configurations, and the best performance was achieved using a combination of InceptionNet for visual features and multilingual BERT for text.These representations were fused after applying normalization and augmentation to enhance robustness.Our InceptionNet with mBERT configuration achieved a macro F1-score of 63 percent and secured the sixth position on the official Cod-aBench leaderboard.These findings highlight the strength of our multimodal model and support its potential for detecting harmful Arabic content in low-resource settings.
Authors: Joy Das, Alamgir Hossain, Mohammed Moshiul Hoque
Institution: Chittagong University of Engineering & Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037632