Abstract
Whereas existing literature on unsupervised machine translation (MT) focuses on exploiting unsupervised techniques for low-resource language pairs where bilingual training data is scare or unavailable, we investigate whether unsupervised MT can also improve translation quality of high-resource language pairs where sufficient bitext does exist. We compare the style of correct translations generated by either supervised or unsupervised MT and find that the unsupervised output is less monotonic and more natural than supervised output. We demonstrate a way to combine the benefits of unsupervised and supervised MT into a single system, resulting in better human evaluation of quality and fluency. Our results open the door to discussions about the potential contributions of unsupervised MT in high-resource settings, and how supervised and unsupervised systems might be mutually-beneficial.
Abstract (translated)
URL
https://arxiv.org/abs/2106.15818
