Scholarlink

<p>Multimodal fake news videos require comprehensive consideration of the correlations and consistency between different modalities, making their detection results often lack explainability. To fill this gap, this paper constructs the Multimodal Fake News Video Explanation (MFNVE) dataset, which systematically reveals the contradiction cues of fake videos through fine-grained manual annotation. Building upon this, we propose a Large Language Models (LLMs)-driven Fake News Video Contradiction Explanation (FNVCE) framework that can automatically identify the veracity of news videos and generate explainable evidence based on contradiction reasoning. FNVCE consists of four core modules: multimodal summary generation, multi-dimensional question construction and answer retrieval, contradictory answer filtering, and contradiction explanation generation. Experiments on the MFNVE dataset demonstrate that the proposed method achieves significant results in both veracity prediction and explanation generation tasks, not only enhancing the explainability of the detection process but also deepening the understanding mechanism of news video content. This study provides a new foundational paradigm for automated and explainable fake news detection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal fake news video contradiction explanations with large language models

  • 期刊论文

  • 发表日期：

    2026年5月8日

查看全文
  • Kuangda Hu,
  • Yan Liao,
  • Hao Liu,
  • Hengxuan Lin,
  • Zikun Hu

摘要

Multimodal fake news videos require comprehensive consideration of the correlations and consistency between different modalities, making their detection results often lack explainability. To fill this gap, this paper constructs the Multimodal Fake News Video Explanation (MFNVE) dataset, which systematically reveals the contradiction cues of fake videos through fine-grained manual annotation. Building upon this, we propose a Large Language Models (LLMs)-driven Fake News Video Contradiction Explanation (FNVCE) framework that can automatically identify the veracity of news videos and generate explainable evidence based on contradiction reasoning. FNVCE consists of four core modules: multimodal summary generation, multi-dimensional question construction and answer retrieval, contradictory answer filtering, and contradiction explanation generation. Experiments on the MFNVE dataset demonstrate that the proposed method achieves significant results in both veracity prediction and explanation generation tasks, not only enhancing the explainability of the detection process but also deepening the understanding mechanism of news video content. This study provides a new foundational paradigm for automated and explainable fake news detection.