SmolLab_SEU at MAHED Shared Task: Do Arabic-Native Encoders Surpass Multilingual Models in Detecting the Nuances of Hope, Hate, and Emotion?
Year: 2025
Type: conference-paper
Abstract: The dynamic interplay of hope and hate speech on Arabic social media presents a critical challenge for content moderation and digital discourse analysis.This paper presents our systems for the MAHED 2025 shared task on Multimodal Detection of Hope and Hate Emotions in Arabic Content, addressing the two text-based subtasks.Our approach centers on a systematic, empirical comparison of Arabicnative versus large-scale multilingual Transformer encoders to determine the optimal pretraining strategy for this nuanced domain.Comprehensive evaluations demonstrate the clear superiority of Arabic-native models, with our ARBERTv2-based system achieving the highest performance.We secured 11 th place in Subtask 1 with a macro F1-score of 0.682 and 5 th place in Subtask 2 with a macro F1-score of 0.514.Error analysis reveals persistent challenges in interpreting implicit language and overcoming severe class imbalance, particularly in distinguishing targeted hate from general offensiveness.This work contributes a robust benchmark for this comparison and underscores the importance of language-specific pre-training for nuanced affective computing in Arabic.
Authors: Md Abdur Rahman, Md Sabbir Dewan, Md. Tofael Ahmed Bhuiyan, Md Ashiqur Rahman
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Funder: Southeast University
OpenAlex ID: W4416037630