Google is using Neural Networks for Chinese to English machine translation

Google has deployed the Google Neural Machine Translation (GNMT) system in Google Translate for Chinese to English translation. The system handles 18 million requests every day. Chinese to English is a particularly hard language pair for machines to handle, and the system is expected to be rolled out to more language pairs. Google Translate supports translation for over ten thousand language pairs. All the translations served by Google Translate, on both mobile and web, from Chinese to English are handled by the new GNMT system. The earlier system used Phrase-Based Machine Translation (PBMT). This approach broke down the sentences to be translated into phrases and words for translation. Neural Machine Translation (NMT) by contrast, considers the entire sentence as an input. The NMT approach has a key advantage of requiring fewer engineering decisions. However, when it was first tested, NMT results were as good as PBMT results. Google engineers improved the accuracy of the NMT system by making it identify and treat rare and obscure words separately. These included bilingual humans to continuously rate the results of the NMT. [caption id=“attachment_316058” align=“aligncenter” width=“640”]  The Tensor Processing Unit[/caption] Eventually, the Google engineers came up with a solution that was good enough to be deployed on Google Translate, to quickly and accurately handle the large volume of requests. The production deployment was possible because of the TensorFlow , an open source frame work for machine learning. The custom Tensor Processing Units , the secret to Google AI power, provide the computation necessary to deliver the translations rapidly. Parsey McParseface , one of the most powerful natural language parser, is also based on the TensorFlow framework.