option 1: To model the generative likelihood of word sequences