| 相甍甍,纪泽旭,郭佳鸣,王秀琨.证据链增强的多智能体分组讨论事实核查框架研究[J].数字图书馆论坛,2026,22(5):20~33 |
| 证据链增强的多智能体分组讨论事实核查框架研究 |
| Research on an Evidence Chain Enhanced Multi-Agent Group Discussion Fact-Checking Framework |
| 投稿时间：2026-04-08 |
| DOI：10.3772/j.issn.1673–2286.2026.05.003 |
| 中文关键词: 虚假信息；事实核查；多智能体；证据链；大语言模型 |
| 英文关键词: Misinformation; Fact-checking; Multi-agent Systems; Evidence Chain; Large Language Models |
| 基金项目:本研究得到国家社会科学基金一般项目“认知和心理账户视角下社交媒体用户隐私悖论行为及干预机制研究”（编号：21BTQ057）资助。 |
| 作者 | 单位 | | 相甍甍 | 吉林财经大学人工智能与计算机学院，吉林财经大学大数据与交叉科学研究院 | | 纪泽旭 | 吉林财经大学人工智能与计算机学院 | | 郭佳鸣 | 吉林财经大学人工智能与计算机学院 | | 王秀琨 | 吉林财经大学人工智能与计算机学院 |
|
| 摘要点击次数: 38 |
| 全文下载次数: 35 |
| 中文摘要: |
| 事实核查是甄别虚假信息的重要手段，而单一大语言模型在事实核查中存在“幻觉”问题、可解释性不足、准确率低等局限，本研究旨在提出一种证据链增强的多智能体分组讨论事实核查（Evidence Chain Enhanced Multi-Agent Group Discussion for Fact-Checking，ECEMA-GD-FC）框架，以期有效提升事实核查任务的准确性、鲁棒性与可解释性。ECEMA-GD-FC框架由主持人智能体负责任务调度、结果整合与共识裁决，下设科学证据、媒体证据与事实证据三个核查智能体小组，各小组内部遵循证据检索、证据链构建、真伪判定、报告生成的标准化流程，实现“统筹—分工—聚合”闭环核查机制。基于事实核查数据集的实验结果表明，本研究所提ECEMA-GD-FC框架在真伪判定准确率与核查报告生成质量上展现出良好的性能，能显著提升事实核查结果的可信度。 |
| 英文摘要: |
| Fact-checking serves as a critical approach for identifying and mitigating misinformation. However, individual large language models (LLMs) suffer from inherent limitations in fact-checking tasks, including hallucination issues, insufficient interpretability, and limited accuracy. This study proposes an Evidence Chain Enhanced Multi-Agent Group Discussion Fact-Checking Framework to improve the accuracy, robustness, and explainability of fact-checking systems. The proposed framework consists of a moderator agent responsible for task orchestration, result aggregation, and consensus adjudication, along with three specialized fact-checking agent groups focusing on scientific evidence, media evidence, and factual evidence, respectively. Within each group, agents follow a standardized fact-checking workflow involving evidence retrieval, evidence chain construction, truthfulness assessment, and report generation, thereby establishing a closed-loop mechanism of “Coordination–Specialization–Integration” Experimental results on benchmark fact-checking datasets demonstrate that the proposed framework achieves competitive performance in both veracity classification accuracy and fact-checking report generation quality, substantially enhancing the credibility of fact-checking results. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
