Unified Metadata-Aware Transformer Framework for Misleading News Classification

Abstract

Misleading information appearing on various digital platforms poses serious risks to public trust. Recently developed approaches to misleading news detection rely primarily on binary classification; however, real-world fact-checking scenarios demand multiclass formulations that capture varying degrees of truthfulness. To address this gap, our research introduces MetaRoBERTa, a unified transformer-based representation learning framework, as a robust solution for multiclass misleading news classification. MetaRoBERTa integrates claims, contextual information, justification text, and speaker-related credibility metadata into a single textual representation, enabling end-to-end learning without explicit linguistic or behavioral feature engineering. The proposed approach achieves an accuracy of 80.44% on the semantically rich LIAR2 dataset, when justification text is available, surpassing state-of-the-art results. Ablation studies further show that the complete regularized training strategy yields an improvement over a cross-entropy baseline that is nominally significant at the 0.05 level before correction for multiple comparisons, with individual components contributing small, consistently positive gains, and per-class analysis reveals dataset-dependent behavior that provides practical guidance for model selection and deployment.