Propositional claim detection: a task and dataset for the classification of claims to truth

This paper introduces Propositional Claim Detection (PCD), an NLP task for classifying claims to truth, and presents a publicly available dataset for it. PCD is applicable in practical scenarios, for instance, for the support of fact-checkers, as well as in many areas of communication research. By leveraging insights from philosophy and linguistics, PCD is a more systematic and transparent version of claim detection than previous approaches. This paper presents the theoretical background for PCD and discusses its advantages over alternative approaches to claim detection. Extensive experiments on models trained on the dataset are conducted and result in an $$\hbox {F}_{1}$$ F 1 -score of up to 0.91. Moreover, PCD’s generalization across domains is tested. Models trained on the dataset show stable performance for text from previously unseen domains such as different topical domains or writing styles. PCD is a basic task that finds application in various fields and can be integrated with many other computational tools.