Left to right: Agreement between datasets’ original labels and the majority label according to our (discretized) re-annotation; accuracy of BERT NLI model against original labels; accuracy of BERT against re-annotation labels; number of p/h pairs (out of 100) on which all three label sources (original, re-annotation, model prediction) agree on the most likely label. Our analysis in §5.3 is performed only over pairs in ∩.