If you don't mind it would be useful to have word* encoders instead of making transformations before encoding.