String-to-Tree Multi Bottom-up Tree Transducers

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 815–824, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics String-to-Tree Multi Bottom-up Tree Transducers Nina Seemann and Fabienne Braune and Andreas Maletti Institute for Natural Language Processing, University of Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {seemanna,braunefe,maletti}@ims.uni-stuttgart.de Abstract We achieve significant improvements in several syntax-based machine translation experiments using a string-to-tree vari- ant of multi bottom-up tree transducers. Our new parameterized rule extraction al- gorithm extracts string-to-tree rules that can be discontiguous and non-minimal in contrast to existing algorithms for the tree-to-tree setting. The obtained models significantly outperform the string-to-tree component of the Moses framework in a large-scale empirical evaluation on several known translation tasks. Our linguistic analysis reveals the remarkable benefits of discontiguous and non-minimal rules. 1 Introduction We present an application of a variant of local multi bottom-up tree transducers (ℓMBOTs) as proposed in Maletti (2011) to statistical machine translation. ℓMBOTs allow discontinuities on the target language side since they have a sequence of target tree fragments instead of a single tree fragment in their rules. The original approach makes use of syntactic information on both the source and the target side (tree-to-tree) and a cor- responding minimal rule extraction is presented in (Maletti, 2011). Braune et al. (2013) imple- mented it as well as a decoder inside the Moses framework (Koehn et al., 2007) and demonstrated that the resulting tree-to-tree ℓMBOT system sig- nificantly improved over its tree-to-tree baseline using minimal rules. We can see at least two draw- backs in this approach. First, experiments investi- gating the integration of syntactic information on both sides generally report quality deterioration. For example, Lavie et al. (2008), Liu et al. (2009), and Chiang (2010) noted that translation quality tends to decrease in tree-to-tree systems because the rules become too restrictive. Second, minimal rules (i.e., rules that cannot be obtained from other extracted rules) typically consist of a few lexi- cal items only and are thus not the most suitable to translate idiomatic expressions and other fixed phrases. To overcome these drawbacks, we abol- ish the syntactic information for the source side and develop a string-to-tree variant of ℓMBOTs. In addition, we develop a new rule extraction algo- rithm that can also extract non-minimal rules. In general, the number of extractable rules explodes, so our rule extraction places parameterized restric- tions on the extracted rules in the same spirit as in (Chiang, 2007). In this manner, we combine the advantages of the hierarchical phrase-based ap- proach on the source side and the tree-based ap- proach with discontinuiety on the target side. We evaluate our new system in 3 large-scale ex- periments using translation tasks, in which we ex- pect discontinuiety on the target. MBOTs are pow- erful but asymmetric models since discontinuiety is available only on the target. We chose to trans- late from English to German, Arabic, and Chi- nese. In all experiments our new …

Content truncated to 500 words