HTU at MAHED Shared Task: Ensemble-Based Classification of Arabic Hate and Hope Speech Using Pre-trained Dialectal Arabic Models
Year: 2025
Type: conference-paper
Abstract: Emotional contagion, the phenomenon where emotions spread between individuals, shows the importance of detecting both hope and hate speech in digital communications.This emotional transmission can amplify positive sentiments that foster community resilience or propagate harmful content that divides societies.While hate speech detection in Arabic has been extensively studied, hope speech detection has received comparatively limited attention, creating an imbalance in the understanding of emotional influence online.To address this gap, MAHED 2025 sub-task 1 introduced the task of detecting both hope and hate speech using a substantial dataset designed for developing robust classification models.This paper presents an ensemble approach combining three Transformer-based encoder models with soft voting and weighted loss functions to address class imbalance issues.Those models, ArabicDeBERTa-DA, BERT-DA, and MAR-BERTV2, have been continually pre-trained on different domains of Arabic, showing the benefits of continual pre-training both on downstream performance and computational efficiency.The proposed ensemble model achieved the highest performance in the competition with an F1 macro score of 72.3% using an ensemble voting of the best-performing variants.
Authors: Abdallah Saleh, Mariam M Biltawi
Institution: Al Hussein Technical University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037797