Shallow Local Multi-Bottom-up Tree Transducers in Statistical Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 811–821, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Shallow Local Multi Bottom-up Tree Transducers in Statistical Machine Translation Fabienne Braune and Nina Seemann and Daniel Quernheim and Andreas Maletti Institute for Natural Language Processing, University of Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {braunefe,seemanna,daniel,maletti}@ims.uni-stuttgart.de Abstract We present a new translation model in- tegrating the shallow local multi bottom- up tree transducer. We perform a large- scale empirical evaluation of our obtained system, which demonstrates that we sig- nificantly beat a realistic tree-to-tree base- line on the WMT 2009 English → German translation task. As an additional contribu- tion we make the developed software and complete tool-chain publicly available for further experimentation. 1 Introduction Besides phrase-based machine translation sys- tems (Koehn et al., 2003), syntax-based systems have become widely used because of their abil- ity to handle non-local reordering. Those systems use synchronous context-free grammars (Chi- ang, 2007), synchronous tree substitution gram- mars (Eisner, 2003) or even more powerful for- malisms like synchronous tree-sequence substitu- tion grammars (Sun et al., 2009). However, those systems use linguistic syntactic annotation at dif- ferent levels. For example, the systems proposed by Wu (1997) and Chiang (2007) use no linguis- tic information and are syntactic in a structural sense only. Huang et al. (2006) and Liu et al. (2006) use syntactic annotations on the source lan- guage side and show significant improvements in translation quality. Using syntax exclusively on the target language side has also been success- fully tried by Galley et al. (2004) and Galley et al. (2006). Nowadays, open-source toolkits such as Moses (Koehn et al., 2007) offer syntax-based components (Hoang et al., 2009), which allow experiments without expert knowledge. The im- provements observed for systems using syntactic annotation on either the source or the target lan- guage side naturally led to experiments with mod- els that use syntactic annotations on both sides. However, as noted by Lavie et al. (2008), Liu et al. (2009), and Chiang (2010), the integration of syntactic information on both sides tends to de- crease translation quality because the systems be- come too restrictive. Several strategies such as (i) using parse forests instead of single parses (Liu et al., 2009) or (ii) soft syntactic constraints (Chi- ang, 2010) have been developed to alleviate this problem. Another successful approach has been to switch to more powerful formalisms, which al- low the extraction of more general rules. A par- ticularly powerful model is the non-contiguous version of synchronous tree-sequence substitu- tion grammars (STSSG) of Zhang et al. (2008a), Zhang et al. (2008b), and Sun et al. (2009), which allows sequences of trees on both sides of the rules [see also (Raoult, 1997)]. The multi bottom-up tree transducer (MBOT) of Arnold and Dauchet (1982) and Lilin (1978) offers a middle ground between traditional syntax-based models and STSSG. Roughly speaking, an MBOT is an STSSG, in which all the discontinuities must oc- cur on the target language …

Content truncated to 500 words