Anyone knows why translations are performed on language pairs? In my humble non linguistic common sense I think translations would be more optimal if word meanings independent of base language were the basis of translations between languages.
The neural network will come up with its own internal meanings for words after reading lots and lots of input data. There are papers on "zero-shot" tests, where a network trained to translate languages A to B and B to C can do A to C without additional data because it maps the meanings of similar words in the different languages the same way.