Scholarlink

<p>This research study explores how integrating a multi-agent system with a knowledge graph as shared memory enhances news bias analyses, detection, and fact-checking. Automated fact-checking and verification systems have evolved significantly but often struggle with complex contextual information and advanced narratives. This research examines the potential advantages of structured knowledge graphs, which offer significant advantages over both direct LLM prompting and unstructured RAG retrieval approaches in multi-agent systems for news evaluation. This system was designed and implemented as a framework showing multiple specialized agents using Large Language Models (LLMs) to collaborate while building and referencing a dynamic shared knowledge graph. Experiments conducted on diverse political news article datasets comparing three approaches-RAG baseline, LLM-only, and KG-augmented system-with statistical significance testing showing that integrating knowledge graphs improved performance over both RAG baseline and LLM-only. The result indicates that the multi-agent system with a knowledge graph outperforms baselines (<i>p</i> &lt; 0.01), achieving an F1 of 0.901 for bias detection (vs. 0.287 RAG, 0.713 LLM-only) and an F1 of 0.794 for fact-checking (vs. 0.661 RAG, 0.720 LLM-only). These results show that adding a structured knowledge graph representation to a multi-agent system is much better than unstructured retrieval, with statistically significant improvements in both the RAG baseline and LLM-only.</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

Improved multi-agent knowledge sharing system using knowledge graphs for news bias detection and fact-checking

  • 期刊论文

  • 发表日期：

    2026年3月14日

查看全文
  • Modupeola Fagbenro,
  • Christopher Washer,
  • Pavani Chella,
  • Amir Jafari

摘要

This research study explores how integrating a multi-agent system with a knowledge graph as shared memory enhances news bias analyses, detection, and fact-checking. Automated fact-checking and verification systems have evolved significantly but often struggle with complex contextual information and advanced narratives. This research examines the potential advantages of structured knowledge graphs, which offer significant advantages over both direct LLM prompting and unstructured RAG retrieval approaches in multi-agent systems for news evaluation. This system was designed and implemented as a framework showing multiple specialized agents using Large Language Models (LLMs) to collaborate while building and referencing a dynamic shared knowledge graph. Experiments conducted on diverse political news article datasets comparing three approaches-RAG baseline, LLM-only, and KG-augmented system-with statistical significance testing showing that integrating knowledge graphs improved performance over both RAG baseline and LLM-only. The result indicates that the multi-agent system with a knowledge graph outperforms baselines (p < 0.01), achieving an F1 of 0.901 for bias detection (vs. 0.287 RAG, 0.713 LLM-only) and an F1 of 0.794 for fact-checking (vs. 0.661 RAG, 0.720 LLM-only). These results show that adding a structured knowledge graph representation to a multi-agent system is much better than unstructured retrieval, with statistically significant improvements in both the RAG baseline and LLM-only.