Fact Checking with Insufficient Evidence

Abstract Automating the fact checking (FC) process relies on information obtained from external sources. In this work, we posit that it is crucial for FC models to make veracity predictions only when there sufficient evidence and otherwise indicate not enough. To end, are first study what consider by introducing a novel task advancing with three main contributions. First, conduct an in-depth empirical analysis of new fluency-preserving method omitting at constituent sentence level. We identify remaining (in)sufficient FC, based trained different Transformer architectures datasets. Second, ask annotators whether omitted was important resulting in diagnostic dataset, SufficientFacts1, evidence. find least successful detecting missing adverbial modifiers (21% accuracy), whereas easiest date (63% accuracy). Finally, propose data augmentation strategy contrastive self-learning employing proposed omission combined tri-training. It improves performance Evidence Sufficiency Prediction up 17.8 F1 score, which turn 2.6 score.