Quasar at MAHED Shared Task : Decoding Emotions and Offense in Arabic Text using LLM and Transformer-Based Approaches
Year: 2025
Type: conference-paper
Abstract: The escalating presence of propaganda and hate speech on social media platforms underscores the need for robust automated detection systems to preserve the integrity of public discourse.Our team, participated in the MAHED 2025 Shared Task at the ArabicNLP 2025 conference, co-located with EMNLP 2025, focusing on Subtask 1 (Text-based Hate and Hope Speech Classification) and Subtask 2 (Emotion, Offensive, and Directed Hate Detection) in Arabic content.In Subtask 1, we experimented with models including XLM-RoBERTa-Large, Davlan/xlm-roberta-base-finetuned-arabic, asafaya/bert-base-arabic, aubmindlab/bertbase-arabertv2, Google Gemma-7B, and Qwen2.5-14B-Instruct,achieving the highest macro-f1 of 0.674 with Gemma-7B and ranking 12th on the leaderboard.In Subtask 2, using models such as aubmindlab/bert-base-arabertv2, Google Gemma-7B, Qwen2.5-14B-Instruct,asafaya/bert-base-arabic, and domain-specific hate-speech models, our best macro-f1 was 0.48 with both Gemma-7B and aubmindlab/bert-base-arabertv2, placing us 6th in the leaderboard.
Authors: Md Sagor Chowdhury, Adiba Fairooz Chowdhury
Institution: Chittagong University of Engineering & Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037625