This is a brief summary of paper for me to study and organize it, Improving Neural Machine Translation Models with Monolingual Data (Sennrich et al., ACL 2016) that I read and studied.

This paper propose new data augmentation they call back-tranlsation.

They focused that NMT has obtained state-of-the art performance for several language pair and that target-side monolingual data plays an important role in boosting fluency for phrased-based statistical mahicne translation.

So, they investigated the use of monolingual data for NMT.

They experiemt two different methods to fill source side of monolingual training instance.

One is to use a dummy source sentence, i.e. They pair monlingual setenences on target side with a single-word dummy source side **** to allow processing of both parallel and monolingual trainig examples with the same network graph.

The other is to use a source sentence obtained via back-translation, which they call synthetic source sentence taht is obtained from automatically translating the target sentence into the source language.

They find that the latter is more effecitve.

Reference