Goto

Collaborating Authors

 language processing










Non-Monotonic Latent Alignmentsfor CTC-Based Non-Autoregressive Machine Translation

Neural Information Processing Systems

Follo14,36], we alignments, outputand Ta, (y) whichreturnsY including T 1 : Y 7! Yisthecollapsing aandthen alignmenta=( , A, A,)iscollapsedy =( withamonotonic alignment logp(y|x, ) = log X Model BLEUMETEOR Speed Base Transformer 27.54 54.38 1.0 Vanilla-NAT 19.32 45.79 15.5