Google has launched a Google Translate upgrade utilizing enhanced deep-learning techniques to produce more accurate translations.
The neural machine translation system considers the entire sentence as one unit to be translated. The system relies on a recurrent neural network algorithm consisting of layered nodes, and a network of eight layers acts as the encoder and transforms the input into a list of vectors representing all possible meanings of each word. The second eight-layer network acts as the decoder and generates the translation one word at a time. Meanwhile, an attention network connects the encoder and decoder by directing the decoder to refer back to certain weighted vectors. Rare words are broken into a set of smaller, common subunits the system can handle more easily.
Google researchers say this approach has reduced the number of translation errors by at least 60% versus the phrase-based method.
The researchers used a chip designed for deep learning to make translations three times faster than ordinary chips without sacrificing accuracy.
Google Translate already has begun using neural machine translation for its 18 million daily translations between English and Chinese, with improved translations for many more language pairs rolling out in the coming months.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found