Off-the-Shelf Unsupervised NMT

TL;DR

This paper demonstrates that off-the-shelf neural MT architectures can be effectively adapted for unsupervised translation without parallel data, achieving competitive results and extending to low-resource language pairs like English-Turkish.

Contribution

It introduces a novel approach of using off-the-shelf neural MT models for unsupervised translation, combining multi-task learning insights and enabling application to low-resource languages.

Findings

Unsupervised models achieve competitive performance with purpose-built models.

The approach extends to low-resource language pairs like English-Turkish.

Proposed improvements enhance applicability to truly low-resource settings.

Abstract

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.