FakeNews-TransAug: an enhanced AraBERT-Based deep learning model with data augmentation for addressing class imbalance in Arabic fake news detection

Abstract

The rapid proliferation of fake news and fake news in digital media, particularly in the Arabic language, poses serious threats to public trust and societal stability. This article introduces FakeNews-TransAug, a novel deep learning model designed to enhance Arabic fake news detection through advanced language modelling and data augmentation. The model integrates Arabic Bidirectional Encoder Representations from Transformers (AraBERT), a transformer-based language model, fine-tuned with domain-specific features and augmented with synthetic samples generated via tailored prompt engineering. By addressing class imbalance and capturing the linguistic and contextual complexities of Arabic, the model accuracy is 86.5%, precision is 74%, recall is 62%, F1-score is 68%, and Area under the Receiver Operating Characteristic Curve (AUC-ROC) is 90%. Tests using the JoFakeNews dataset show that the suggested approach works, especially when it comes to finding categories that are not well represented. These results show that FakeNews-TransAug is a strong and flexible standard for finding fake news in Arabic, offering a practical solution for combating fake news in the evolving digital media landscape and demonstrating strong performance in imbalanced data scenarios. This is the first Arabic fake-news detection framework that integrates AraBERT with prompt-engineered data augmentation and parameter-efficient fine-tuning (Quantized Low-Rank Adaptation (QLoRA), BitFit, Prefix-Tuning) to directly mitigate class imbalance while maintaining scalability and high discriminative performance.