Modelling the Lexicon in Unsupervised Part of Speech Induction