According to Och, a good base for developing a usable statistical device translation system for your new pair of languages from scratch would encompass a bilingual text corpus (or parallel collection) of much more than one hundred fifty–two hundred million text, and two monolingual corpora Every single of much more than a billion words and phrase