The accuracy of Google Translate can be hit or miss. It’s often a reliable tool for navigating foreign language websites, but occasionally the tech slips up and confuses terms as different as clitoris and broccoli rabe. In an effort to improve the program, Google has unveiled a new version of Translate called Google Neural Machine Translation or GMNT, Fast Company reports.
The major difference between GMNT and the former phrase-based machine translation system (PBMT) is the way it tackles text. In the past, Translate worked with the individual components of sentences, words, and phrases to translate them separately. The new system looks at whole sentences at a time, improving the technology’s accuracy by roughly 60 percent. This means that even languages as distinct as English and Chinese can be translated to a more precise degree.
GMNT is able to achieve these results by simultaneously running data through multiple cores in computer graphic chips. Each processing layer is allowed limited room for error, which means more layers can be running at once to increase the chances of producing more accurate results (you can read Google’s full paper on the technology here).
Google believes neural networks like this one can be used to expand more than just their translation tool. Researchers at Google Brain have used 11,000 novels to improve the technology’s conversational style and help products like the Google App communicate more fluidly with users.
[h/t Fast Company]
Know of something you think we should cover? Email us at firstname.lastname@example.org.