Google’s edge in machine-learning and artificial intelligence is no secret. Now Google’s Translate tool has taken the next step when it comes to machine learning by learning to translate between languages where it wasn’t explicitly taught to do so by creating its own technical language. Google is calling this is “Zero-Shot Translation” which is now possible with Google’s Neural Machine Translation system (GNMT).
For starters, do note that all form of artificial intelligence that we see today is just the beginning and it has a long way to do. For now, most of AI-based computing is driven by machine learning, which means a computer is specifically taught to do a particular task. But with ‘zero-shot translation’ Google’s GNMT can basically do a task that it wasn’t taught.
In September, Google had announced that it was switching Translate to GNMT system, which is an end-to-end learning framework that learns from millions of examples. Now they’ve expanded on the previous GNMT system to allow a single system to translate between languages.
According to Google Research blog, “Google Translate has grown from supporting just a few languages to 103, translating over 140 billion words every day,” and this needs a system which can handle different language systems. The blog also notes that “scaling up to all the 103 supported languages presented a significant challenge,” for them.
With the’Zero-Shot Translation’ Google’s system was able to translate between two sets of languages without being trained to do so. According to an example by Google, the system was taught to translate between Japanese and English as well as Korean and English. But with the new multilingual system, it was able to “transfer the ‘translation knowledge’ from one language pair to the others. So in essence the system managed to translate from Japanese to Korean without having previous knowledge of the same.
The ‘transferring’ of knowledge bit is key here, and Google calls it “zero-shot” translation. According to the company, “this is the first time this type of transfer learning has worked in Machine Translation.”
According to the Google Researchers, what they observed in the translations showed them that the network appears to have created its own language of sorts in the network in order to carry out these translations. The blog notes, “…This means the network must be encoding something about the semantics of the sentence rather than simply memorizing phrase-to-phrase translations. We interpret this as a sign of existence of an interlingua in the network.”
Google says that this Multilingual Google Neural Machine Translation system is running in production today for all Translate users, and is being used in 10 of the recently launched 16 language pairs.