Google is super-charging its translation software with artificial intelligence (AI).
Simply put, neural networks are AI modeled on connections in the human brain, capable of learning and improving over time.
Google has been aggressively introducing the tech into its core products in attempts to improve efficiency. It has been used for everything from reducing the power bill of the company's enormous data centers, to defeating the human world champion of ancient Chinese board game Go in a highly publicized bout.
It's now being used in Google Translate, starting with Chinese-to-English translations, with plans for other language pairs in the works.
The Google Neural Machine Translation system (or GMNT, as Google is catchily calling it) is reducing translation errors by between 55 and 85%, the company says.
And according to its research, its translations score only just below those of human translations. Take a look at the chart below: A theoretical perfect translation scores a 6, but these basically never happen, even with human translators. But humans still translate much more effectively than traditional phrase-based translation software. GMNT manages to close the gap considerably.
That said, it's definitely not perfect, and Google admits as much. "GNMT can still make significant errors that a human translator would never make," two research scientists for the Google Brain Team wrote in a blog post, "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 GMNT software is live now on all Chinese-to-English translations -- of which Google says it processes 18 million a day.