Bayesian Models for Multilingual Word Alignment