Scholarlink

<p>In recent years, fake news detection has received widespread attention from researchers, technology companies, and relevant policymakers. Detecting fake news requires both efficient processing of key information from long news texts and a deep understanding of real-world background knowledge, which remains a significant challenge for current fake news detection methods. Inspired by the human process of reading news, summarizing news, retrieving information, and inferring news authenticity, this paper proposes a framework based on the Fact-Augmented LLM Generation with co-attention (FALG). This framework extracts key elements through semantic summarization using large language models, obtains external facts through knowledge-enhanced retrieval, and implements alignment and fusion of text and facts using Co-Attention mechanisms. Compared to the state-of-the-art baseline, our FALG model achieves 2.5% and 7.9% higher accuracy on the English GossipCop and Chinese Weibo21 datasets, respectively, demonstrating its strong generalization ability across different languages. Our intelligent summarization-knowledge retrieval-adaptive fusion process not only enhances the information richness of news content but also improves the reliability and accuracy of the detection system.</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

Enhancing fake news detection through fact-augmented LLM generation with Co-Attention

  • 期刊论文

  • 发表日期：

    2025年11月15日

查看全文
  • Kun Huang,
  • Xiaoming Li,
  • Salah Uddin

摘要

In recent years, fake news detection has received widespread attention from researchers, technology companies, and relevant policymakers. Detecting fake news requires both efficient processing of key information from long news texts and a deep understanding of real-world background knowledge, which remains a significant challenge for current fake news detection methods. Inspired by the human process of reading news, summarizing news, retrieving information, and inferring news authenticity, this paper proposes a framework based on the Fact-Augmented LLM Generation with co-attention (FALG). This framework extracts key elements through semantic summarization using large language models, obtains external facts through knowledge-enhanced retrieval, and implements alignment and fusion of text and facts using Co-Attention mechanisms. Compared to the state-of-the-art baseline, our FALG model achieves 2.5% and 7.9% higher accuracy on the English GossipCop and Chinese Weibo21 datasets, respectively, demonstrating its strong generalization ability across different languages. Our intelligent summarization-knowledge retrieval-adaptive fusion process not only enhances the information richness of news content but also improves the reliability and accuracy of the detection system.