ArabFake: A Multitask Deep Learning Framework for Arabic Fake News Detection, Categorization, and Risk Prediction

Abstract

The spread of fake news among Arabic media including social media represents a great challenge to the integrity of information and the trust of the public in it. In this paper, we introduce a comprehensive deep-learning framework, named ArabFake, that goes beyond the binary classification on Arabic fake news detection. ArabFake, built over MARBERTv2 (a state-of-the-art model for multi-dialectal Arabic tweets), proficiently address the complexity of the Arabic language while performing three unified tasks which are fake news detection, content categorization and its risk assessment. The framework promotes efficiency and performance both by enabling multi-task learning through shared knowledge representation across tasks. In order to facilitate development and evaluation, we present the ArabFake Dataset consisting of 2,495 manually labelled news items with labels that are verified by experts regarding fake news categories and risk levels. ArabFake demonstrates robust performance, achieving an F1 score of 94.12% for fake news detection, 84.92% for categorization, and 88.91% for risk zone assessment, highlighting its reliability and effectiveness across multiple tasks. We improve interpretability and extract insight into manipulative techniques by integrating valence scoring as part of the framework that emphasizes misleading linguistic cues used to disseminate fake news within the produced image. The results show that ArabFake is a holistic Arabic fake news detection framework that has practical implications on news organizations and fact checking projects.

Keywords

• Fake news detection
• misinformation detection
• Arabic language
• OSNs