L0DROP

Sequence-to-sequence models usually transfer all encoder outputs to the decoder for genera- tion. In this work, by contrast, we hypothesize that these encoder outputs can be compressed to shorten the …...