ProoFVer: Natural Logic Theorem Proving for Fact Verification

Abstract Fact verification systems typically rely on neural network classifiers for veracity prediction, which lack explainability. This paper proposes ProoFVer, uses a seq2seq model to generate natural logic-based inferences as proofs. These proofs consist of lexical mutations between spans in the claim and evidence retrieved, each marked with logic operator. Claim is determined solely based sequence these operators. Hence, are faithful explanations, this makes ProoFVer by construction. Currently, has highest label accuracy second best score FEVER leaderboard. Furthermore, it improves 13.21% points over next dataset counterfactual instances, demonstrating its robustness. As show better overlap human rationales than attention-based highlights help humans predict decisions correctly more often using directly.1