On Word Alignment Models for Statistical Machine Translation