Google's interlingua

Incase you missed this: earlier this year, Google's translation services was changed to a machine learning system, and invented its very own language in the process.

Google has done something quite impressive behind the scenes - it's switched its translation system from "phrase based translation" to "Google Neural Machine Translation" (GNMT).

"Phrase based translation" is essentially the equivalent of translating word-by-word - like you thumbing through a Lonely planet dictionary, but super fast. It's limited by its vocabulary, and has no understanding of linguistic structures. It maps words and phrases from one language to another.

GNMT, on the other hand, got smart. It is able to learn from the people that use it. It can transfer the "translation knowledge" from language to language, and so able to translate phrases and languages it may never have seen before.

What's more, when Google engineers looked under the 'bonnet' of the system, they found something pretty amazing: it had created its own language termed "interlingua" - as the most efficient way to solve the problem of translation.

The system encodes the patterning behind the semantics of the language, instead of memorising phrase-to-phrase translations.

Tuesday morning #TechForHumanity geek-out. What does this say about the way our own brains codify and translate languages...