Note that this algorithm does not require translations as input. It takes large corpus of the languages of interest, models them and then find parallel between

Note that this algorithm does not require translations as input. It takes large corpus of the languages of interest, models them and then find parallel between both models which are used as a crude basis for a model. They use it to train a better model and iterate. This stems from earlier work on natural language translation (by the same team) and opens quitte a number of doors.