While Google Translate can be a lifesaver for any foreign traveler or lazy student, the tool still remains far from perfect. But with Google’s new system of “neural machine translation,” the service could start parsing entire sentences at a time, nailing the nuance in languages that word-by-word translation often misses.
“At a high level, the Neural system translates whole sentences at a time, rather than just piece by piece,” says Barak Turovsky, product lead at Google Translate. “It uses this broader context to help it figure out the most relevant translation, which it then rearranges and adjusts to be more like a human speaking with proper grammar.”
Prior to the addition of deep learning, Google Translate worked based on individual phrases, which often resulted in inaccuracies based on grammatical construction and context.
Google first announced it would be applying neural machine learning to translation in September. At the time researchers said that Google’s Neural Machine Translation system (GNMT) could teach itself how languages translate — to a point where even researchers were unclear on how the machine operated.
The initial tests only worked with English, Spanish, French, and Chinese. Now the application can also translate to and from German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish. According to Google, these eight languages make up over 35 percent of translation requests, though the company hopes to expand to all 103 languages in the system in the next few months.
The company also announced it would be opening up its Neural Machine Translation API to developers and businesses through the Google Cloud.
“Todays step towards Neural Machine Translation is a significant milestone for Google Translate, but theres always more work to do and well continue to learn over time,” says Turovsky.
Translate is just one of many Google products that has benefited from the company’s venture into neural networks, with other applications including geolocation detection and a futuristic image compression system.
Photos via Google, Flickr / akk_rus