Tag-less back-translation

An effective method to generate a large number of parallel sentences for training improved neural machine translation (NMT) systems is the use back-translations target-side monolingual data. The standard back-translation has been shown be unable efficiently utilize available huge amount existing data because inability models differentiate between authentic and synthetic during training. Tagging, or using gates, used enable distinguish data, improving also enabling iterative on language pairs that underperformed back-translation. In this work, we approach as domain adaptation problem, eliminating need explicit tagging. -- \emph{tag-less back-translation} are treated out-of-domain in-domain respectively and, through pre-training fine-tuning, model able learn more from them Experimental results have outperforms tagged approaches low resource English-Vietnamese English-German translation.