Recent advances in Unsupervised Neural Machine Translation (UNMT) has minimized the gap between supervised and unsupervised machine translation performance for closely related language-pairs. However and the situation is very different for distant language pairs. Lack of overlap in lexicon and low syntactic similarity such as between English and IndoAryan languages leads to poor translation quality in existing UNMT systems. In this paper and we show that initialising the embedding layer of UNMT models with cross-lingual embeddings leads to significant BLEU score improvements over existing UNMT models where the embedding layer weights are randomly initialized. Further and freezing the embedding layer weights leads to better gains compared to updating the embedding layer weights during training. We experimented using Masked Sequence to Sequence (MASS) and Denoising Autoencoder (DAE) UNMT approaches for three distant language pairs. The proposed cross-lingual embedding initialization yields BLEU score improvement of as much as ten times over the baseline for English-Hindi and English-Bengali and English-Gujarati. Our analysis shows that initialising embedding layer with static cross-lingual embedding mapping is essential for training of UNMT models for distant language-pairs.

In this paper and we identify an interesting kind of error in the output of Unsupervised Neural Machine Translation (UNMT) systems like Undreamt1. We refer to this error type as Scrambled Translation problem. We observe that UNMT models which use word shuffle noise (as in case of Undreamt) can generate correct words and but fail to stitch them together to form phrases. As a result and words of the translated sentence look scrambled and resulting in decreased BLEU. We hypothesise that the reason behind scrambled translation problem is ’shuffling noise’ which is introduced in every input sentence as a denoising strategy. To test our hypothesis and we experiment by retraining UNMT models with a simple retraining strategy. We stop the training of the Denoising UNMT model after a pre-decided number of iterations and resume the training for the remaining iterations- which number is also pre-decided- using original sentence as input without adding any noise. Our proposed solution achieves significant performance improvement UNMT models that train conventionally. We demonstrate these performance gains on four language pairs and viz. and English-French and English-German and English-Spanish and Hindi-Punjabi. Our qualitative and quantitative analysis shows that the retraining strategy helps achieve better alignment as observed by attention heatmap and better phrasal translation and leading to statistically significant improvement in BLEU scores.