To demystify the "black box" property of deep neural networks for natural
language processing (NLP), several methods have been proposed to interpret
their predictions by measuring the change in prediction probability after
erasing each token of an input. Since existing methods replace