A word-to-word model of translational equivalence