A tree-to-tree model for statistical machine translation

Abstract

In this thesis, we take a statistical tree-to-tree approach to solving the problem of machine translation (MT). In a statistical tree-to-tree approach, first the source-language input is parsed into a syntactic tree structure; then the source-language tree is mapped to a target-language tree. This kind of approach has several advantages. For one, parsing the input generates valuable information about its meaning. In addition, the mapping from a source-language tree to a target-language tree offers a mechanism for preserving the meaning of the input. Finally, producing a target-language tree helps to ensure the grammaticality of the output. A main focus of this thesis is to develop a statistical tree-to-tree mapping algorithm. Our solution involves a novel representation called an aligned extended projection, or AEP. The AEP, inspired by ideas in linguistic theory related to tree-adjoining grammars, is a parse-tree like structure that models clause-level phenomena such as verbal argument structure and lexical word-order. The AEP also contains alignment information that links the source-language input to the target-language output. Instead of learning a mapping from a source-language tree to a target-language tree, the AEP-based approach learns a mapping from a source-language tree to a target-language AEP. The AEP is a complex structure, and learning a mapping from parse trees to AEPs presents a challenging machine learning problem. In this thesis, we use a linear structured prediction model to solve this learning problem. A human evaluation of the AEP-based translation approach in a German-to-English task shows significant improvements in the grammaticality of translations. This thesis also presents a statistical parser for Spanish that could be used as part of a Spanish/English translation system.