demo

This demos is built upon the Flickr data set. You can download pre-processed files by running

make flickr/data

Pre-trained models can also be downloaded as:

Conditional:

make flickr/models/conditional

AEVNMT:

make flickr/models/aevnmt/supervised

Conditional NMT

In conditional neural machine translation, we learn a conditional distribution $P(y|x)$ over sentence pairs on bilingual parallel data.

Training you can train a German-English model by running:

./conditional-training.sh

Prediction you can translate the dev set by running:

./conditional-translate.sh

Interactive demo you can play with it yourself by entering raw (no preprocessing required) German sentences in the terminal:

./conditional-interactive-translate.sh

AEVNMT

Auto-Encoding Variational Neural Machine Translation, AEVNMT for short, is a joint generative model of translation data (sentence pairs). It learns a joint distribution $P(x, y)$ over sentence pairs as a marginal from a deep generative model $P(x, y, z) = p(z)P(x|z)P(y|z,x)$, where a latent sentence-pair embedding $z$ is assumed to be normally distributed.

In AEVNMT training is performed via variational inference with a posterior approximation $q(z|x)$. We limit the posterior approximation to condition only on the source side of the data and by doing so we can use it both during training (when both $x$ and $y$ are known) and for predictions (when only $x$ is known).

Training you can train a German-English model by running:

./aevnmt-training.sh

Prediction you can translate the dev set by running:

./aevnmt-translate.sh

Interactive demo you can play with it yourself by entering raw (no preprocessing required) German sentences in the terminal:

./aevnmt-interactive-translate.sh