I'm working on a word prediction application and we have some common problems. I don't think that looking at you application as one of translation is a good approach. Translation is made with losses. English is a low-inflected language and that is a big advantage for you. I'm thinking at your application as a semantic-syntactic prediction problem. Read Text prediction systems: a survey[^] and see the biography. . You can find the pdf version(16 pages) here[^].