Back in September, Google released its Neural Machine Translation system. The search giant uses deep learning to produce better and more natural translations between languages. Google Translate in last 10 years has grown from supporting just a few languages to 103.
Google Neural Machine Translation (GNMT) team is looking forward to adding a new feature to the translation system named as “Zero-shot translation”. Instructing the machine to translate English to Korean and vice versa, and also to translate English to Japanese and vice versa, the challenge arises when it is given the translation of Korean to Japanese without taking help of English translation in between. To make it simple to understand, Google has made a gif to show the idea of Zero-shot translation.
The gif shows how the translation system works when this multilingual system is given the translation of English to Japanese and vice versa, also the translation of English to Korean to vice versa. Google’s multilingual system works as a single GNMT system, sharing its parameters in order to translate these four different language pairs. Sharing the parameters enables the translation system to transfer the “translation knowledge” from one language pair to the others.
The visualization shows how the translation system works while translating a single sentence in different directions. Using a 3D representation of internal network data, the system translates a set of sentences including all possible pairs of the Japanese, Korean, and English languages.
This “interlingua” shows a great similarity between a sentence or word in all the three languages. The interlingua is constructed with the complex neural network, therefore it is difficult to describe.
Part (a) shows an overall geometry of these translations. The colors shown in the view means that the translation of a sentence from English to Korean is same as the sentence translation from Japanese to English. Part (b) zooms into one of the groups. Part (c) shows a sentence with the same meaning but from three different languages.