The Facebook usecase is machine translation. Instead of learning translations between each language pair (which scales horribly), just learn to and from a commo

The Facebook usecase is machine translation. Instead of learning translations between each language pair (which scales horribly), just learn to and from a common representation. I expect you could also use this for classification that works over many languages.