What Can Syntax-Based MT Learn from Phrase-Based MT?

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 755–763, Prague, June 2007. c⃝2007 Association for Computational Linguistics What Can Syntax-based MT Learn from Phrase-based MT? Steve DeNeefe and Kevin Knight Information Sciences Institute The Viterbi School of Engineering University of Southern California 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292 {sdeneefe,knight}@isi.edu Wei Wang and Daniel Marcu Language Weaver, Inc. 4640 Admiralty Way, Suite 1210 Marina del Rey, CA 90292 {wwang,dmarcu}@languageweaver.com Abstract We compare and contrast the strengths and weaknesses of a syntax-based machine translation model with a phrase-based ma- chine translation model on several levels. We briefly describe each model, highlight- ing points where they differ. We include a quantitative comparison of the phrase pairs that each model has to work with, as well as the reasons why some phrase pairs are not learned by the syntax-based model. We then evaluate proposed improvements to the syntax-based extraction techniques in light of phrase pairs captured. We also compare the translation accuracy for all variations. 1 Introduction String models are popular in statistical machine translation. Approaches include word substitution systems (Brown et al., 1993), phrase substitution systems (Koehn et al., 2003; Och and Ney, 2004), and synchronous context-free grammar systems (Wu and Wong, 1998; Chiang, 2005), all of which train on string pairs and seek to establish connections be- tween source and target strings. By contrast, ex- plicit syntax approaches seek to directly model the relations learned from parsed data, including models between source trees and target trees (Gildea, 2003; Eisner, 2003; Melamed, 2004; Cowan et al., 2006), source trees and target strings (Quirk et al., 2005; Huang et al., 2006), or source strings and target trees (Yamada and Knight, 2001; Galley et al., 2004). It is unclear which of these important pursuits will best explain human translation data, as each has ad- vantages and disadvantages. A strength of phrase models is that they can acquire all phrase pairs con- sistent with computed word alignments, snap those phrases together easily by concatenation, and re- order them under several cost models. An advan- tage of syntax-based models is that outputs tend to be syntactically well-formed, with re-ordering influ- enced by syntactic context and function words intro- duced to serve specific syntactic purposes. A great number of MT models have been re- cently proposed, and other papers have gone over the expressive advantages of syntax-based approaches. But it is rare to see an in-depth, quantitative study of strengths and weaknesses of particular models with respect to each other. This is important for a scientific understanding of how these models work in practice. Our main novel contribution is a com- parison of phrase-based and syntax-based extraction methods and phrase pair coverage. We also add to the literature a new method of improving that cover- age. Additionally, we do a careful study of several syntax-based extraction techniques, testing whether (and how much) they affect phrase pair coverage, and whether (and how much) they …

Content truncated to 500 words