Prediction without justification has limited applicability. As a remedy, we
learn to extract pieces of input text as justifications -- rationales -- that
are tailored to be short and coherent, yet sufficient for making the same
prediction. Our approach combines two modular components,