Table-Text Alignment: Explaining Claim Verification Against Tables in Scientific Papers

Xanh Ho, Sunisth Kumar, Yun-Ang Wu, Florian Boudin, Atsuhiro Takasu, Akiko Aizawa · 2025

Read Table-Text Alignment: Explaining Claim Verification Against Tables in Scientific Papers by Xanh Ho, Sunisth Kumar, Yun-Ang Wu, Florian Boudin, Atsuhiro Takasu, Akiko Aizawa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-10486.txt.

Description

Six authors argue that checking a claim against a table should come with an explanation, not just a verdict. Standard systems predict whether a table supports or refutes a claim, but a bare label hides the reasoning and gives users little to trust. The team reframes the task: a model must also point to the exact table cells that decide the claim. They extended the SciTab benchmark with human-marked cell sets, where annotators confirmed each label and marked the smallest set of cells behind their call.

Experiments gave two findings. Feeding alignment information into the check improved the verdict accuracy. Yet most large language models, while often landing on the right label, failed to recover the cells the humans had marked. The authors read that gap as a sign the models' answers do not rest on faithful reasoning.

Catalog record

CREATORXanh Ho, Sunisth Kumar, Yun-Ang Wu, Florian Boudin, Atsuhiro Takasu, Akiko Aizawa
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-10486

Browse the full library · About the Tower of Babel