Design of an Iterative Cross-Modal and Context-Aware Deep Analytical Framework for Hate Speech and Fake Post Detection on Social Media Sets
Rakesh Bharati, Jyoti Bharti, Vasudev Dehalwar, Jaydeep Kishore · Applied Sciences · 2026
Abstract The exponential growth of user-generated content on social media platforms has increased the diffusion of hate speech and fake information considerably, which poses serious risks to the integrity of societies and the public trust people vest in them. Existing approaches tend to be critically limited by core restrictions such as unimodal processing, absence of user-context integration, poor cross-modal synchronizations, and weak robustness against adversarial manipulations. Such restrictions lead compounded losses in generalizations, misclassifications, and unreliability of deployments in real-life scenarios. For these challenges, this work offers a complete, modular, and analytically validated pipeline around five newly developed methods for multimodal hate speech and fake post detection from a Twitter corpus. The first component is Context-Driven Social Vector Extraction, which produces enriched contextual embeddings by capturing text, image, temporal, and behavioural metadata. These embeddings are processed by the Cross-Modal Label Fusion via Mutual Co-Attention (CMF-MCA) module, which utilizes a dual-transformer with co-attention to jointly annotate textual and visual information. The third module brought into play is Semantic Propagation Graph for Hate & Fake Correlation (SPG-HFC), which implements a multi-relational graph attention mechanism that captures semantic influence and community-level propagation patterns. The next step is located in the Adaptive Modality Routing via Reinforcement (AMR-R) module which selects, depending on the input complexity, between machine learning and deep learning classifiers using a reinforcement policy network. Finally, the Counterfactual Consistency Validation Engine (CCVE) performs post-prediction verification by evaluating counterfactual variants to ensure robustness and semantic consistency. Thus, detection accuracy, scalability, and inference reliability significantly increase in this integrated framework. It transcends the realm by making permissible cross-modal, context-aware, and behaviour-driven classification, which in turn greatly supports the development of large-scale secure and trustworthy content moderation systems in process.