A Stochastic Decoder for Neural Machine Translation

Philip Schulz, Wilker Aziz, Trevor Cohn

ACL2018 | arxiv | code, Sockeye |

Account for variation in sequnce prediction model with a Variational/stochastic decoder for neural machine translation.

Contributions:

A decoder capable of capturing word-level variation in translation data.

A sequence (one per target position) of latent variables instead of the setence-level (single) gausian used in Zhang et al. 2016.

img

img

img

img

img

img

Experiments:

From English to:

#### Issues & comments: