It uses a deep sequence to sequence model, and the drop-down thesaurus are phrase embeddings trained on large corpuses like Reddit and Wikipedia.