Researchers at the University of the Basque Country (UPV) in Spain and Facebook have separately developed unsupervised machine-learning techniques for teaching neural networks to translate between languages with no parallel texts.
Each method employs as training strategies back translation and denoising; in the first process, a sentence in one language is approximately translated into the other, then translated back into the original language, with networks adjusted to make subsequent attempts closer to identical. Meanwhile, denoising adds noise to a sentence by rearranging or removing words, and attempts to translate that back into the original.
The UPV method translates more frequently during training, while the Facebook technique, in addition to encoding a sentence from one language into a more abstract representation before decoding it into the other language, also confirms the intermediate language is truly abstract.
When translating between English and French in a vast sentence database, both systems received a bilingual evaluation understudy score of about 15 in both directions, compared to Google Translate, which scores about 40, and humans, who can score more than 50.
View Full Article
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA