Quasi-Synchronous Phrase Dependency Grammars for Machine Translation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 474–485, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Quasi-Synchronous Phrase Dependency Grammars for Machine Translation Kevin Gimpel Noah A. Smith Language Technologies Institute Carnegie Mellon Univeristy Pittsburgh, PA 15213, USA {kgimpel,nasmith}@cs.cmu.edu Abstract We present a quasi-synchronous dependency grammar (Smith and Eisner, 2006) for ma- chine translation in which the leaves of the tree are phrases rather than words as in pre- vious work (Gimpel and Smith, 2009). This formulation allows us to combine structural components of phrase-based and syntax-based MT in a single model. We describe a method of extracting phrase dependencies from paral- lel text using a target-side dependency parser. For decoding, we describe a coarse-to-fine ap- proach based on lattice dependency parsing of phrase lattices. We demonstrate performance improvements for Chinese-English and Urdu- English translation over a phrase-based base- line. We also investigate the use of unsuper- vised dependency parsers, reporting encourag- ing preliminary results. 1 Introduction Two approaches currently dominate statistical ma- chine translation (MT) research. Phrase-based mod- els (Koehn et al., 2003) excel at capturing local reordering phenomena and memorizing multi-word translations. Models that employ syntax or syntax- like representations (Chiang, 2005; Galley et al., 2006; Zollmann and Venugopal, 2006; Huang et al., 2006) handle long-distance reordering better than phrase-based systems (Auli et al., 2009) but often re- quire constraints on the formalism or rule extraction method in order to achieve computational tractabil- ity. As a result, certain instances of syntactic diver- gence are more naturally handled by phrase-based systems (DeNeefe et al., 2007). In this paper we present a new way of combin- ing the advantages of phrase-based and syntax-based MT. We propose a model in which phrases are orga- nized into a tree structure inspired by dependency syntax. Instead of standard dependency trees in which words are vertices, our trees have phrases as vertices. We describe a simple heuristic to extract phrase dependencies from an aligned parallel cor- pus parsed on the target side, and use them to com- pute target-side tree features. We define additional string-to-tree features and, if a source-side depen- dency parser is available, tree-to-tree features to cap- ture properties of how phrase dependencies interact with reordering. To leverage standard phrase-based features along- side our novel features, we require a formalism that supports flexible feature combination and effi- cient decoding. Quasi-synchronous grammar (QG) provides this backbone (Smith and Eisner, 2006); we describe a coarse-to-fine approach for decod- ing within this framework, advancing substantially over earlier QG machine translation systems (Gim- pel and Smith, 2009). The decoder involves generat- ing a phrase lattice (Ueffing et al., 2002) in a coarse pass using a phrase-based model, followed by lat- tice dependency parsing of the phrase lattice. This approach allows us to feasibly explore the combined search space of segmentations, phrase alignments, and target phrase dependency trees. Our experiments demonstrate an average im- provement of +0.65 BLEU in Chinese-English translation across three test sets and an improvement …

Content truncated to 500 words