Google's Translation Tool Has Taught Itself Its Own Language

Google's Neural Machine Translation system has taught itself a new language. Building on its original Google Translate tool to provide accurate and quick translation, researchers were recently surprised to find, after upgrades, the system had learned to translate between languages without bridging back to English, something referred to as "zero-shot translation."

Now, Google Translate can convert from Spanish to French without having any training in that medium, in turn making the output more human, more organic and reasonably accurate. Google switched to the deep-learning GNMT in late September, "an end-to-end learning framework that learns from millions of examples" which "provided significant improvements in translation quality."

The system works with 103 languages, though, which proved to be difficult in that thousands of language pairs needed to be translated. But researchers noticed the system was automatically grouping sentences with the same meanings from three different languages, or "encoding something about the semantics of the sentence rather than simply memorizing phrase-to-phrase translations."

"This means the network must be encoding something about the semantics of the sentence rather than simply memorizing phrase-to-phrase translations," the researchers wrote. "We interpret this as a sign of existence of an interlingua in the network."

• Trump: Google Hides Bad Hillary Stories – and It Actually Does
• Investors See Artificial Intelligence Destroying Millions of Jobs