Distribution Matching for Rationalization

The task of rationalization aims to extract pieces input text as rationales justify neural network predictions on classification tasks. By definition, represent key used for prediction and thus should have similar feature distribution compared the original text. However, previous methods mainly focused maximizing mutual information between labels while neglecting relationship To address this issue, we propose a novel method that matches distributions in both space output space. Empirically, proposed matching approach consistently outperforms by large margin. Our data code are available.