Scholarlink

In today’s business landscape, deploying Large Language Models (LLMs) for company tasks has become prevalent. However, leveraging LLMs for enterprise tasks poses significant challenges, primarily due to hallucinations. These hallucinations can lead to misinformation and poor decision-making. Additionally, internal data that cannot be directly memorized by public LLMs further complicates the utilization process. To address these issues, we propose a novel framework CoKGLM for hallucination detection that combines Knowledge Graphs (KGs) and Language Models (LMs). This approach enhances the use of both textual and structural information from KGs and thoroughly examines desirable KG information for hallucination detection. Our experimental results demonstrate a 38% improvement in hallucination detection and an 84% increase in robustness compared to using LMs alone. This approach enhances transparency and control over content verification while fully utilizing KG data.

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

CoKGLM: Detecting Hallucinations Generated by Large Language Models via Knowledge Graph Verification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Rie Hasegawa,
  • Ryutaro Ichise

摘要

In today’s business landscape, deploying Large Language Models (LLMs) for company tasks has become prevalent. However, leveraging LLMs for enterprise tasks poses significant challenges, primarily due to hallucinations. These hallucinations can lead to misinformation and poor decision-making. Additionally, internal data that cannot be directly memorized by public LLMs further complicates the utilization process. To address these issues, we propose a novel framework CoKGLM for hallucination detection that combines Knowledge Graphs (KGs) and Language Models (LMs). This approach enhances the use of both textual and structural information from KGs and thoroughly examines desirable KG information for hallucination detection. Our experimental results demonstrate a 38% improvement in hallucination detection and an 84% increase in robustness compared to using LMs alone. This approach enhances transparency and control over content verification while fully utilizing KG data.