Google Translate 'now almost as good as a human'

Google has developed a new version of its Translate tool and according to the company, it's almost as good as human translation.

The app, like many other computer-powered translation services, lets tourists or people abroad for business speak in their own language and then translates it into that of the country they're in.

However, comically bad mistranslations are common and are seen as inevitably associated with automated translation. The new software Google is rolling out, which starts with a Mandarin to English version today, should change that.

The Week

Escape your echo chamber. Get the facts behind the news, plus analysis from multiple perspectives.

SUBSCRIBE & SAVE

Sign up for The Week's Free Newsletters

From our morning news briefing to a weekly Good News Newsletter, get the best of The Week delivered directly to your inbox.

From our morning news briefing to a weekly Good News Newsletter, get the best of The Week delivered directly to your inbox.

Sign up

Google calls the new method Neural Machine Translation, says Quartz, and it is "radical" change from the previous system.

At the moment, the app breaks sentences down and tries to find words and phrases in a specially-enlarged dictionary. The new system uses two neural networks. One breaks down the sentence to work out what it means and the other generates text in the destination language, says Quartz.

Thanks to artificial intelligence, the networks learn dynamically, ignoring rules about how to translate and instead focusing purely on how good the result is. Google believes the new system should make 80 per cent fewer errors.

Compared side-to-side with translations provided by humans, the new system scores highly. Human translators average 5.1 out of 6 on a scale of accuracy, while the old Google Translate managed just 3.6. The new version averages 5.0.

The system is "yet another example of the success of deep learning", says Nature, combining artificial neural networks -"layers of computational units that mimic the way neurons connect in the brain" - on huge sets of data.

It's an approach that has had notable successes with game playing – a computer beat a human champion of the complex board game Go - and image recognition, among other applications.

That Google has started with Mandarin Chinese is perhaps indicative of its "larger ambitions", says Wired, but it plans to replace the existing system for all 10,000 language pairs it currently translates.

"All the big internet giants" are "training deep neural nets using translations gathered from across the internet", says Wired. At least for now, Google seems to have got there first.