Abstract Natural Language Inference (NLI) models have shown impressive performance on benchmark datasets, but recent studies suggest they may perpetuate and amplify societal biases present in training data. We propose…