Google's new A.I. translator comes close to human accuracy
Google says its new translation tool, which has been bolstered by artificial intelligence, can achieve results nearly on par with human translators.
In a paper published this week, the tech giant said that side-by-side comparisons of Google Translate’s new method, called Google Neural Machine Translation, “approaches the accuracy achieved by average bilingual human translators” in some tests.
On average, the new system also cuts down on roughly 60 per cent of translation errors that result from Google Translate’s current algorithm.
It’s a marked departure the company’s previous method, which was phrase-based.
In this method, the algorithm breaks down a sentence into words or phrases and attempts to match it in a massive dictionary.
In contrast, the GNMT system considers the original sentence as a whole.
It relies on two neural networks that have been trained using the same large dictionary: one to realize the meaning of the sentence and the other one to translate it into another language.
According to Quartz, because A.I. systems don’t have to rely on human logic, they can independently come up with their own solutions, rather than be handcuffed by the original code work.
“You don’t have to make design choices,” Mike Schuster, a Google engineer who worked on the project, told the online publication.
“The system can entirely focus on translation.”
Google says its new system reduces translation errors by between 55 per cent and 85 per cent on several major languages.
The biggest improvement is with English to Spanish and French to English, with the software’s results also approaching levels of human translation in several tests.
For example, Google Translate previously scored a 4.9, with a 6 being the perfect rendition, in English to Spanish, but that number was bumped 5.4 thanks to the GNMT systems.
Meanwhile, humans generally scored a 5.6.
Despite these promising results, Google admits that its new translation tool isn’t perfect.
“GNMT can still make significant errors that a human translator would never make, like dropping words and mistranslating proper names or rare terms, and translating sentences in isolation rather than considering the context of the paragraph or page,” the company said in a blog post.
“There is still a lot of work we can do to serve our users better. However, GNMT represents a significant milestone.”
Google is initially rolling out the technology for Mandarin, with new languages set to be incorporated in the coming months.