Abstract
Automated fact checking has gained immense interest to tackle the growingmisinformation in the digital era. Existing systems primarily focus onsynthetic claims on Wikipedia, and noteworthy progress has also been made onreal-world claims. In this work, we release QuanTemp, a diverse, multi-domaindataset focused exclusively on numerical claims, encompassing temporal,statistical and diverse aspects with fine-grained metadata and an evidencecollection without leakage. This addresses the challenge of verifyingreal-world numerical claims, which are complex and often lack preciseinformation, not addressed by existing works that mainly focus on syntheticclaims. We evaluate and quantify the limitations of existing solutions for thetask of verifying numerical claims. We also evaluate claim decomposition basedmethods, numerical understanding based models and our best baselines achieves amacro-F1 of 58.32. This demonstrates that QuanTemp serves as a challengingevaluation set for numerical claim verification.