Scholarlink

With the rapid growth of social media and online platforms, detecting misinformation has become a critical task for ensuring information credibility, particularly in the health domain, where misinformation poses significant risks to public health and safety. Although Large Language Models (LLMs) excel at semantic understanding, their performance in processing long texts is often limited due to dispersed information, intertwined arguments, and implicit logical inconsistencies, which reduce detection accuracy and efficiency. To address these challenges, we propose HierDaC, a novel LLM-based detection framework that utilizes a divide-and-conquer strategy. This strategy decomposes health-related misinformation detection into independent, semantically complete paragraphs, focusing on locally suspicious segments to reduce information overload. Furthermore, to address the lack of dedicated datasets in the Chinese health domain, we introduce the CHMDD dataset, which contains both fake and authentic samples across various health topics. Experimental results demonstrate that our divide-and-conquer strategy significantly enhances LLM performance in misinformation detection. On the CHMDD dataset, detection performance improves from 75.2% at baseline to 86.2% (+11.0%). While on the CoAID dataset with Open-ended generation in LLMFake, detection performance increases from 30.6% at baseline to 62.9% (+32.3%).

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

HierDaC: Detecting Long-Text Misinformation via Hierarchical Divide-and-Conquer

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Liyang Zhang,
  • Zhiteng Song,
  • Tongxuan Zhang,
  • Guiyun Zhang

摘要

With the rapid growth of social media and online platforms, detecting misinformation has become a critical task for ensuring information credibility, particularly in the health domain, where misinformation poses significant risks to public health and safety. Although Large Language Models (LLMs) excel at semantic understanding, their performance in processing long texts is often limited due to dispersed information, intertwined arguments, and implicit logical inconsistencies, which reduce detection accuracy and efficiency. To address these challenges, we propose HierDaC, a novel LLM-based detection framework that utilizes a divide-and-conquer strategy. This strategy decomposes health-related misinformation detection into independent, semantically complete paragraphs, focusing on locally suspicious segments to reduce information overload. Furthermore, to address the lack of dedicated datasets in the Chinese health domain, we introduce the CHMDD dataset, which contains both fake and authentic samples across various health topics. Experimental results demonstrate that our divide-and-conquer strategy significantly enhances LLM performance in misinformation detection. On the CHMDD dataset, detection performance improves from 75.2% at baseline to 86.2% (+11.0%). While on the CoAID dataset with Open-ended generation in LLMFake, detection performance increases from 30.6% at baseline to 62.9% (+32.3%).