A Technical Review of Fake News Detection: Progression from Textual Analysis to Multimodal and Knowledge-Aware Learning

Abstract

The process of fake news detection has evolved from classification-based text processing to multimodal, propagation-based, knowledge-based, and heterogeneous graph learning for addressing the increasing complexity of disinformation on social media platforms. This paper presents an analysis of methodological advances made in this domain through the examination of the use of textual, visual, emotional, propagation, relational, and external knowledge data in fake news detection frameworks. The recent studies related to fake news detection have been categorized on the basis of fuzzy and multitask learning, attention and emotion modeling, multimodal fusion and consistency learning, propagation early detection, hyperbolic representation learning, knowledge-based vision-language models, heterogeneous graph neural networks, multilingual detection, and real-time analysis. This research study compares existing major datasets and evaluation practices to emphasize the impacts of modality, language, platform, temporal information, and labeling characteristics on the reported performance. The key issues related to cross-modality inconsistency, dataset biases, distribution shifts, interpretability, knowledge reliability, temporality, and computational scalability are identified and the literature-driven research directions towards robust, interpretable, multilingual, and evidence-based fake news detection are experimented.