Using Grammatical Roles to Improve Statistical Machine Translation

Statistical machine translation systems often struggle to preserve predicateargument structure. We present a new hierarchical machine translation model that explicitly captures the grammatical roles taken on by the words and phrases being translated (e.g., subject, object, and indirect object). Although existing hierarchical and syntax-based grammars can capture how many arguments a predicate takes, they have limited awareness of what should fill the argument slots. This results in difficult to interpret translations with scrambled predicate argument relationships. Our model adds grammatical role typing to the translation rules that helps preserve and correctly order predicate-argument structures. We find that grammatical typing systematically outperforms both traditional hierarchical and syntax augmented models on Chineseto-English translation across a variety of NIST MT evaluation sets.