Google had announced that the Google Neural Machine Translation System (GNMTS) was live and functional this Past September. The system translates entire sentences instead of small pieces. Later, they conducted an experiment where they first trained GNMTS to translate English into Korean and Japanese, and further tasked it with translating Korean into Japanese, without using English. They found that the system could make reasonably correct translations between the two languages, marking a noteworthy achievement for deep neural networks.
Its achievement is a bit of a mystery in itself. It appears that the GNMT has created its own internal language which exists at a deeper level where the system understands the similarities in representation of a sentence or word in the three languages. The internal processes of such difficult networks are still uncertain. However, the independent creation of such an “interlingua” by the system is a milestone in itself.