AAA at MAHED Text-based Hate and Hope Speech Classification: A Systematic Encoder Evaluation for Arabic Hope and Hate Speech Classification
Year: 2025
Type: conference-paper
Abstract: Arabic hate speech detection presents unique challenges due to the language's morphological complexity, dialectal diversity, and the subtle nature of emotional expressions in social media.In this paper, we present our submission to the MAHED shared task for Arabic hate speech classification, which aims to classify Arabic text into three categories: hope, hate, and not_applicable.This task is crucial for building safer online communities and has applications in content moderation, social media analysis, and digital wellbeing initiatives.We systematically evaluate six transformer-based encoders, comparing Arabic-specific models (MARBERT, AraBERT, ALCALM) against multilingual alternatives (XLM-RoBERTa, LaBSE, BGE).Our approach demonstrates that specialized Arabic models specially encoders trained on more than one dialect like marber significantly outperform their multilingual counterparts, with MAR-BERT achieving the best overall performance.Using our proposed methodology, we achieved competitive results on the MAHED shared task with a macro-F1 score of 0.707 on the test split, securing a strong position in the final competition rankings.
Authors: Ahmed Khalil Elzainy, Mohamed Amin, Ahmed Samir, Hazem Abdelsalam
Institution: Alexandria University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037639