System description of Supervised and Unsupervised Neural Machine Translation approaches from 'NL Processing' team at DeepHack.Babel task

摘要

A comparison ' of supervised and unsupervised Neural Machine Translation (NMT) models was done for the corpora provided by the DeepHack.Babel competition. It is shown that for even small parallel corpus, fully supervised NMT gives better results than fully unsupervised for the case of constrained domain of the corpus. We have also implemented a fully unsupervised and a semi-supervised NMT models which have not given positive results compared to fully supervised models. A blind set-up is described where participants know at no point what language pair is used for translation, so no extra data could be integrated in pre-submission phase or during training. Finally, future competition organizers should find ways to protect their competition set-ups against various attacks in order to prevent from revealing of language pairs. We have reported two possible types of attacks on the blind set-up.