Year: 2025
Type: conference-paper
Abstract: The detection of harmful online content, including hate speech and propaganda, is particularly challenging in multimodal and multilingual contexts such as Arabic social media.This work addresses Sub-task 1: Text-based Hate and Hope Speech Classification in the MA-HED2025 (Zaghouani et al., 2025) challenge, where the goal is to classify Arabic text into hate, hope, or not_applicable.We develop a system based on pre-trained Arabic BERT models with three fine-tuning strategies, combined with a custom preprocessing pipeline for noise removal, normalization, and diacritic stripping.To address class imbalance and lexical sparsity, we augment the training data with synthetically generated paraphrases via the OpenAI API.Experimental results on the official test set demonstrate that our best configuration, BERT-base-AraBERTv02 + NN with cleaning and generated data, achieves a macro-F1 score of 0.6747 F1.Error analysis reveals that mislabeled training instances significantly limit model performance, suggesting that future improvements may be achieved through systematic dataset refinement.Our approach highlights the importance of preprocessing, augmentation, and careful architectural choices for robust Arabic text classification.
Institutions: Prince Sultan University, Tuwaiq Academy
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: hybrid
Funder: Prince Sultan University
OpenAlex ID: W4416037789