Static analysis of lots of human translators' output. Given the input and output of a strange human filter, they're trying to approximate the filter using machine learning. When their algorithm's output approximates the best translations, they're done.