Title: Comparative analysis of Arabic transformers using an Arabic fake news dataset

Authors: Amine Mammasse; Khaled Bedjou; Faical Azouaou

Addresses: LITAN Laboratory, École Supérieure en Sciences et Technologies de l'Informatique et du Numérique (ESTIN), Algeria ' Laboratoire LIMED, Faculté des Sciences Exactes, Université de Bejaia, 06000 Bejaia, Algeria ' LITAN Laboratory, École Supérieure en Sciences et Technologies de l'Informatique et du Numérique (ESTIN), Algeria

Abstract: This study presents a comparative analysis of 12 state-of-the-art Arabic transformer models, addressing the challenges of fake news detection in Arabic natural language processing (NLP). We evaluate models including AraT5v2-base-1024, MARBERT, AceGPT, and ArabianGPT using a unified corpus integrating three publicly available datasets (AFND, ANS, and AraFacts). Our novelty lies in systematically analysing how model architecture, pre-training data quality, and fine-tuning strategies influence performance across Arabic dialectal variations and linguistic complexities. Results demonstrate ArabianGPT achieving superior performance with an F1-score of 82.4%, followed by PhoenixMultiple-Langs-v1 and BERT-base-QARIB. The findings reveal decisive correlations between detection accuracy and factors such as architectural design, training data diversity, and dialectal adaptability. This pioneering analysis uncovers critical insights into the importance of morphological richness, script-specific pre-processing, and comprehensive dialect coverage in Arabic NLP. Our research underscores the need for more diverse training data and specialised architectures to advance technologies combating misinformation across Arabic-speaking regions.

Keywords: Arabic natural language processing; NLP; transformer models; fake news detection; FND; Arabic datasets; neural networks.

DOI: 10.1504/IJAISC.2025.149626

International Journal of Artificial Intelligence and Soft Computing, 2025 Vol.9 No.2, pp.106 - 133

Received: 16 Nov 2024
Accepted: 21 Jul 2025
Published online: 07 Nov 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article