A desirable property of learning systems is to be both effective and
interpretable. Towards this goal, recent models have been proposed that first
generate an extractive explanation from the input text and then generate a
prediction on just the explanation called explain-then-predict models