The paper introduces the problem of minimal evidence group (MEG) identification for claim verification. In real-world settings, claim verification often requires aggregating a complete set of evidence pieces that collectively provide full support to the claim. However, the problem becomes particularly challenging when there exist distinct sets of evidence that could be used to verify the claim from different perspectives.
The paper formally defines the MEG identification problem and shows that it can be reduced from the Set Cover problem. The key aspects are:
The paper proposes a practical approach that decomposes the problem into two steps: (1) predicting whether a candidate set of evidence pieces fully supports, partially supports, or does not support the claim, and (2) bottom-up merging of partially supporting groups to search for a fully supporting group. The approach leverages the properties of MEGs to prune the search space efficiently.
Experiments on the WiCE-MEG and SciFact-MEG datasets show that the proposed approach significantly outperforms direct LLM prompting and classic claim verification baselines, achieving 18.4% and 34.8% absolute improvements in precision, respectively. The paper also demonstrates the benefits of MEGs in downstream applications such as claim generation, where MEGs provide more compact and sufficient evidence compared to classic claim verification approaches.
翻譯成其他語言
從原文內容
arxiv.org
深入探究