Google recently announced that its new AI-based translation software, Neural Machine Translation (NMT), has developed an internal language of its own to facilitate its translation of certain languages — and Google can’t explain how it did it.
Neural Machine Translation (NMT) was developed by Google to allow a more naturalistic automatic translation between languages: traditional translation software usually takes the sentences to be converted and breaks them down into individual components, leaving the resulting translations prone to errors due to differences in syntax and grammar. NMT, being AI-based, is intended to look at the overall sample sentence, and put it into a more naturalistic context, to provide a less mechanistic translation. Instead of having all of its programmed skills provided by its initial programming, NMT was taught to learn languages through experience — hence the “neural” part of its name.
Initially, Google taught the program to translate between English and Korean, then taught it to do the same with English and Japanese. They then tried to see if it would translate between Japanese and Korean, without having to resort to using its previous experience with English to use as a go-between — and it worked perfectly.
What NMT’s programmers found was that the program used what they’re calling “interlingua,” an internal language that NMT used to go between Japanese and Korean, but that it devised this secret language all on its own — a language that Google’s programmers can’t understand.