MultiMinds at MAHED 2025: Multimodal and Multitask Approaches for Detecting Emotional, Hate, and Offensive Speech in Arabic Content

Le résumé fourni par la source

This paper describes the MultiMinds team's participation in the MAHED 2025 shared task at ArabicNLP 2025, which targets the detection of hate speech, hope speech, and emotional expression in Arabic content.We addressed two subtasks.For the text-based subtask (Task 2), we experimented with multiple models, including Support Vector Machines with TF-IDF and AraBERT embeddings, XGBoost with fused AraBERT and XLM-RoBERTa embeddings optimized via Optuna, and a fine-tuned AraBERT model and GPT-5 (gpt-oss-20b).The fine-tuned AraBERT achieved the best performance with an F1 score of 0.68.For the multimodal subtask (Task 3), we proposed an architecture combining DistilBERT for text representation with a lightweight ELU-Net enhanced by a cross-attention mechanism, reaching 75% accuracy.Major challenges included dataset imbalance and noisy text, which we mitigated through preprocessing, class-weighted optimization, and feature fusion.Our results demonstrate the benefits of combining multiple embedding layers for text classification and leveraging lightweight multimodal architectures for robust hate speech detection in Arabic.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.