Data Augmentation for Learning Bilingual Word Embeddings with Unsupervised Machine Translation

Unsupervised bilingual word embedding (BWE) methods learn a linear transformation matrix that maps two monolingual embedding spaces that are separately trained with monolingual corpora. This method assumes that the two embedding spaces are structurally similar, which does not necessarily hold true in general. In this paper, we propose using a pseudo-parallel corpus generated by an unsupervised machine translation model to facilitate structural similarity of the two embedding spaces and improve the quality of BWEs in the mapping method. We show that our approach substantially outperforms baselines and other alternative approaches given the same amount of data, and, through detailed analysis, we argue that data augmentation with the pseudo data from unsupervised machine translation is especially effective for BWEs because (1) the pseudo data makes the source and target corpora (partially) parallel; (2) the pseudo data reflects some nature of the original language that helps learning similar embedding spaces between the source and target languages.