Scholarlink

<p>Fact verification involves verifying the claims’ truthfulness based on evidence. Multi-hop fact verification, which requires multiple evidence retrieval and reasoning steps, is challenging due to dependencies between intermediate steps. These dependencies lead to error propagation: errors in early steps can propagate through subsequent steps, ultimately distorting the conclusion. We identify that error propagation often stems from the evidence retrieval process, where models generate queries lacking essential contextual information, resulting in inaccurate evidence retrieval. To address this, we propose a propositional logic-based reasoning framework with two stages: (1) formalizing multi-hop verification as a propositional logic reasoning process to preserve critical contextual information and (2) using these propositions to guide the generation of more precise queries. Additionally, to overcome the limitations of smaller models, we employ knowledge distillation, transferring knowledge from state-of-the-art models to smaller ones, enhancing performance while reducing inference overhead. We evaluate our framework on the HoVer benchmark. The results show that our framework outperforms baselines and effectively mitigates specific error types.</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

Mitigating error propagation in multi-hop fact verification with logic reasoning

  • 期刊论文

  • 发表日期：

    2025年5月26日

查看全文
  • Xiucheng Lyu,
  • Chengyu Cao,
  • Mingwei Sun,
  • Ruifeng Xu

摘要

Fact verification involves verifying the claims’ truthfulness based on evidence. Multi-hop fact verification, which requires multiple evidence retrieval and reasoning steps, is challenging due to dependencies between intermediate steps. These dependencies lead to error propagation: errors in early steps can propagate through subsequent steps, ultimately distorting the conclusion. We identify that error propagation often stems from the evidence retrieval process, where models generate queries lacking essential contextual information, resulting in inaccurate evidence retrieval. To address this, we propose a propositional logic-based reasoning framework with two stages: (1) formalizing multi-hop verification as a propositional logic reasoning process to preserve critical contextual information and (2) using these propositions to guide the generation of more precise queries. Additionally, to overcome the limitations of smaller models, we employ knowledge distillation, transferring knowledge from state-of-the-art models to smaller ones, enhancing performance while reducing inference overhead. We evaluate our framework on the HoVer benchmark. The results show that our framework outperforms baselines and effectively mitigates specific error types.