Abstract
Unsupervised neural machine translation (NMT) has attracted a lot ofattention recently. While state-of-the-art methods for unsupervised translationusually perform well between similar languages (e.g., English-Germantranslation), they perform poorly between distant languages, becauseunsupervised alignment does not work well for distant languages. In this work,we introduce unsupervised pivot translation for distant languages, whichtranslates a language to a distant language through multiple hops, and theunsupervised translation on each hop is relatively easier than the originaldirect translation. We propose a learning to route (LTR) method to choose thetranslation path between the source and target languages. LTR is trained onlanguage pairs whose best translation path is available and is applied on theunseen language pairs for path selection. Experiments on 20 languages and 294distant language pairs demonstrate the advantages of the unsupervised pivottranslation for distant languages, as well as the effectiveness of the proposedLTR for path selection. Specifically, in the best case, LTR achieves animprovement of 5.58 BLEU points over the conventional direct unsupervisedmethod.