Detecting Bias in Arabic Political News: A Transformer-Based Ensemble Stacking Approach
A. Abdelhameed, Ensaf Hussein Mohamed, Walaa Medhat · 2025
Bias detection in Arabic text is challenging due to complex bias representation, class imbalances, and limited labeled datasets. This study proposes an ensemble stacking model integrating pre-trained transformers (AraBERT, MARBERT, QARiB) with a meta-classifier combining gradient boosting, random forest, and logistic regression for improved accuracy and robustness. Evaluated on four datasets (FigNews, UsVsThem, SemEval, ThatiAR), the model addresses class imbalances using GPT-3.5 Turbo for data augmentation, advanced preprocessing, SMOTE, and hyperparameter optimization. Results show the ensemble model consistently outperforms traditional and standalone transformer models, achieving higher F1 scores, highlighting its effectiveness for low-resource linguistic challenges.