Google Translate has in all of its ten years of existence been a quite useful tool – appreciated by some (especially those who know about the difficulty of machine based translation), ridiculed by others, used by most. It has though never really lived up to professional standards. And, to take that upfront, that hasn’t changed since you last used it half an hour ago. But Google is about to use its main advantage. Its data. Its 140 billion translated words daily.
In order to do that, the ‘machine’ had to be made adaptive. One main obstacle down that road was that, considering 103 supported languages, the idea of creating comprehensive language pairs (meaning direct translation from each one language to possibly every other supported one), was close to impossible. The translation would usually go via English. And thereby any real progress was hampered because even if you managed to improve the result, what you would have improved would not have been the link between the two target languages but between either or both of those and English. As soon as you took English out of the equation, the progress would probably be lost. Translation is a delicate field, if you aim for good results.
What you want to do instead, according to American and German researchers, is creating a so-called neural network. A new common language that may serve as a link between languages that the machine has not been trained to directly translate. If it was able to translate between Hindi and Hebrew and Hebrew and French before, then it can now also translate between Hindi and French without any middle step. Another reason why neural network systems have hailed as the new star in the machine translation universe is that they take into account contextual meaning. Before, and still to an extent because all machine based translation by definition has to be statistical if doesn’t want to be random, a sentence was translated by translating the single words or idioms individually and then putting them together. This is how things like this happen:
This brings us to the main point: Google’s new common language (or any neural network language for that matter), other than English, can be changed, improved, adjusted with the help of the users’ requests.
Which tempts some to the conclusion that for one of the first times, AI, Artificial Intelligence, is actually at play in a user-related field. But that’s a topic for another day …