Temporal-Aware Binary Claim–Tweet Alignment with Dynamic Calibration for Evolving Fact-Checking

Abstract

In automated fact-checking, the rapid transformation of information on social media networks is a significant threat, particularly in maintaining accuracy over time. To cope with the dynamism of fact-checking cases, this paper introduces a novel temporal-aware binary claim-tweet alignment framework with a dynamic calibration system. By taking advantage of temporal embeddings, adaptive thresholding, and real-time calibration, our approach enhances the accuracy of aligning claims that can evolve in validity. Comparison with a static baseline methodology reveals significant gains in experimental results on a dataset of 15,000 claim-tweet pairs over 18 months: a 7.5% F1-score gain, a 15.7% temporal consistency gain, and an 8.9% calibration score gain. The proposed framework is remarkably robust when dealing with breaking news scenarios and evolving scientific assertions.