Google Translate Invents Own AI Language 'Translation' -- How Does It Work?
Google Translate has just become a ton smarter than how we used to be. It can be recently remembered that it just used a neural network to translate between some of its most popular languages. But now, it's smart enough to use language pairs on languages it's not trained for -- and it does it by using a new "AI language."
According to New Scientist, traditional machine-translation systems break sentences into words and phrases. They translate these words individually. However, in September, Google Translate unveiled a new "system" that uses a neural network to work on sentences themselves. This allows it to give more context to the sentences and figure out the best translation. The system is now in action for eight of the most common language pairs Google Translate is working on.
Though neural machine-translation systems are vastly becoming popular, most only work on a single pair of language. So different systems are needed to translate in between languages.
Apparently, Google has extended its system so it can handle multiple pairs -- to the point of translating between two languages it's not trained to do so.For instance, if the neural network has been taught to translate between English and Japanese, and English and Korean, it can translate from Japanese to Korean without even going through English.
New Scientist reports that according to Kyunghyun Cho of New York University, this is a big advance in the field of neural networks. Google now thinks their system can achieve this by finding a common ground where sentences with the same meaning are represented in similar ways regardless of language.
This is an example of "interlingua," in a sense, because it's a "new common language" that is specific for the task of translation and is not readable or usable for humans.This is also called zer-shot translation, and it still doesn't perform as well as the simpler approach of translating via a common language.