Off-the-Shelf Unsupervised NMT
This work addresses the problem of machine translation without parallel data for researchers and practitioners, though it is incremental as it builds on existing unsupervised MT and multi-task learning approaches.
The authors tackled unsupervised machine translation by framing it as multi-task learning using off-the-shelf neural MT architectures, achieving performance competitive with purpose-built models for unsupervised MT and extending applicability to low-resource language pairs like English-Turkish.
We frame unsupervised machine translation (MT) in the context of multi-task learning (MTL), combining insights from both directions. We leverage off-the-shelf neural MT architectures to train unsupervised MT models with no parallel data and show that such models can achieve reasonably good performance, competitive with models purpose-built for unsupervised MT. Finally, we propose improvements that allow us to apply our models to English-Turkish, a truly low-resource language pair.