Don't Take It Literally: An Edit-Invariant Sequence Loss for Text Generation