Google’s translation AI may have developed its own internal language in order to translate two languages without using English as a reference for definitions. Image INQUIRER.net Stock Photo
In September Google launched its new AI-powered language translation tool–the Neural Machine Translation system. After a little over a month of learning more languages to translate beyond Spanish, it appears to have created its own secret internal language.
This image illustrates how the zero-shot translation process happens. The AI first learns how to translate Japanese to English, as well as Korean to English. Then it is made to translate Japanese to Korean and vice versa. Image Google
TechCrunch reports that the discovery came about when Google researchers wondered if the AI could translate Japanese to Korean and vice versa without using the English language as a bridge. They call this “zero-shot translation.”
To their surprise, the AI was able to produce reasonable translations for the two languages.
The question that comes to the researchers’ mind is, “Has the computer formed its own concept of shared meaning for words found in the two languages?” Google’s researchers believe that it has.
Here is a visual representation of the system’s memory when translating a single sentence into different languages. Image Google
It’s difficult to determine exactly since the inner processes of neural networks are very hard to describe.
Just as it is difficult to know exactly what is going on inside a neural network, so too is the implication of this discovery difficult to ascertain. However, the fact remains that this is conceptually very powerful. Alfred Bayle