Using Word-Dependent Transition Models in HMM-Based Word Alignment for Statistical Machine Translation