In its early days, Google Translate wasn’t exactly perfect, but over the years, it has gotten smarter and smarter at translating text. Thanks to the recent boom in machine learning research, Google Inc.’s translation service could soon become the universal translator we’ve always wanted.

Now, according to a new post on the Google Research Blog, Google Translate has become so smart that it is starting to come up with its own translations without being taught.

Google revealed the details of its progress with Google Translate in a blog post written by Mike Schuster and Nikhil Thorat of the Google Brain Team, as well as Melvin Johnson of Google Translate.

“In the last 10 years, Google Translate has grown from supporting just a few languages to 103, translating over 140 billion words every day,” the Google team said. “To make this possible, we needed to build and maintain many different systems in order to translate between any two languages, incurring significant computational cost. With neural networks reforming many fields, we were convinced we could raise the translation quality further, but doing so would mean rethinking the technology behind Google Translate.”

The biggest change to Google Translate came back in September, when Google announced that it would be shifting to a new system called Google Neural Machine Translation, which takes advantage of neural networks to correlate words and phrases of one language with another.

Initially, the team explained, GNMT worked well with the languages on which is was first tested, but scaling it to the full 103 languages supported by Google Translate proved more difficult. This is partly from the fact that the demand for translations differs for the different languages, so while the system might be frequently tasked with translating English to Spanish, it likely does not receive the same number of requests for Maori to Latin. Because of this, Google Translate is naturally trained to understand translations between more common languages, leaving others in the dust.

Google’s system solves this problem in a way that is surprisingly intuitive for a machine. Rather than having to learn every possible translation between every supported language, GNMT can infer meaning based on word pairs it already knows.

For example, if Google Translate knows the Spanish translation of an English word and it also knows the Korean translation for that same English word, then it can translate that word between Spanish and Korean using its existing knowledge. Effectively, this means that GNMT’s algorithms work as a sort of universal internal language that can translate new language combinations based on what it already knows.

If you want a more detailed explanation of how Google Translate’s new system works, you can read the research team’s paper published by Cornell University.

Image courtesy of Google