Authors: Vu Thanh Nhan, Luong Truong An
ABSTRACT: The exponential proliferation of digital misinformation across Vietnamese online media platforms poses substantial challenges to information integrity, public safety, and institutional trust. Traditional text classification paradigms utilizing Pretrained Language Models (PLMs) conventionally process input by concatenating headlines and body texts into a single monolithic sequence. This design fails to capture the subtle inter-textual semantic dissonance characteristic of sensational clickbait and deceptive realism, where fabricated headlines deliberately diverge from or contradict body contexts. Furthermore, conventional static deployment settings remain incapable of handling continuous distribution shifts as deception tactics dynamically evolve.
To address these limitations, we propose an end-to-end framework integrating a Dual-Branch Cross-Attention PhoBERT architecture, a Faithfulness-Verified Explainable AI (XAI) mechanism via Local Interpretable Model-agnostic Explanations (LIME), and an Active Continual Learning pipeline with experience replay. Rather than treating input as an aggregated flat sequence, our dual-branch encoder extracts contextual representations for headlines (
) and body texts () asynchronously, explicitly modeling cross-modal semantic discrepancies via Multi-Head Cross-Attention.
Extensive experiments on a standardized Vietnamese benchmark dataset demonstrate that our proposed model achieves an Accuracy of 94.41%, a Precision of 89.74%, an outstanding Recall of 98.59%, and an F1-Score of 93.96%, suppressing false positives by 0.99% over traditional baselines while effectively mitigating false-negative omissions on malicious content and maintaining competitive performance against standard single-branch models.
Furthermore, quantitative deletion faithfulness evaluation (
) validates that model decisions are strictly governed by genuine linguistic markers, exhibiting a monotonic confidence degradation on the misinformation class from 
(
) down to (), successfully triggering a decision-flip at . Coupled with an uncertainty-driven active replay buffer and a tri-state client-side verification engine, the proposed framework provides an accurate, transparent, and continuously adaptive paradigm for digital misinformation mitigation
Keywords: Fake News Detection, PhoBERT, Multi-Head Cross-Attention, Explainable AI, Deletion Faithfulness, Active Continual Learning, Vietnamese NLP