Machine-annotated Rationales: Faithfully Explaining Text Classification

Abstract

We propose an approach to faithfully explaining text classification models, using a specifically designed neural network to find explanations in the form of machine-annotated rationales during the prediction process. This results in faithful explanations that are similar to human-annotated rationales, while not requiring human explanation examples during training. The quality of found explanations is measured on faithfulness, quantitative similarity to human explanations, and through a user evaluation.