Google first unveiled on September that neural networks would start powering Google Translate. More than two weeks into November, Google finally pushed the technology out for a number of languages, improving translations for the better.
Named Neural Machine Translation, the technology employs cutting-edge training techniques to deliver more accurate and context-appropriate translations instead of flimsy, awkward and oftentimes broken translations.
Now, Google has a new trick under the Neural Machine Translation technology called "zero-shot" translations that can translate a pair of languages without prior learning.
A post published on Google Research Blog explained how zero-shot translation works:
"Our proposed architecture requires no change in the base [Google Neural Machine Translation] system, but instead uses an additional 'token' at the beginning of the input sentence to specify the required target language to translate to," wrote Mike Schuster, Nikhil Thorat from the Google Brain Team and Melvin Johnson from Google Translate.
The Google Neural Machine Translation System extracts the parameters of a multilingual system to form translations between different language pairs.
"This sharing enables the system to transfer the 'translation knowledge' from one language pair to the others," according to the post.
To simplify things, imagine teaching a translation system to translate Korean to Japanese and vice versa. Now, imagine teaching it English to Korean and vice versa. The new technology allows the system to translate Japanese to English without resorting first to its the original translated pair. Therefore, the system translates languages easily even when these languages weren't originally paired together.
Has Google Created Its Own Language?
"All right, don't panic, but computers have created their own secret language and are probably talking about us right now," TechCrunch's Devin Coldewey wrote, and he's not wrong. At least not entirely.
If the translation system has found a way to determine meaning from sentences and words that haven't been linked together, does this mean that the system has developed a shared meaning for the words, thereby suggesting that it has, more or less, achieved an understanding that surpasses the concept of simply matching a word to its equivalent word in another language?
To put it in simpler terms, has Google Translate invented its own set of rules for how it translates one language to another? According to artificial intelligence researchers, all signs point to yes.
A paper discussing multilanguage translations describes zero-shot translations as "[a] surprising benefit of modeling several language pairs in a single model ... " with the operative word in this case being "surprising benefit," as if zero-shot translations came as a natural corollary with the inroad of Neural Machine Translation.
The question is, might this be an example of where AI is headed? An AI capable of generating its own set of concepts and applying it to translate input into different language pairs independently? It's too early to determine such a conclusion, but ultimately, Google's zero-shot translation system is a step forward in breaking down language barriers more efficiently.