Machine translation has made a lot of advances in using syntactic properties of languages, but the question of whether or not semantic information can also be effective in improving the quality of translations has only been narrowly investigated. In this proposal, we examine this question by incorporating the semantic structure of a language, in the form of predicate and argument roles, into the grammar of a machine translation system. We also propose some additional methods such as features for modeling the transformations of semantic roles between the two languages, and building different kinds of features derived from different semantic roles.
We also explore the topic of parameter tuning, which relates to the ways of optimizing the weights of the features which are used in decoding. As we increase the number of the features in our translation system, we raise the need for better methods for tuning the weights of those features. In this proposal, we introduce a new effective method for parameter tuning in machine translation