Statistical Machine Reordering

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 70–76, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Statistical Machine Reordering Marta R. Costa-juss`a and Jos´e A. R. Fonollosa Department of Signal Theory and Communications TALP Research Center (UPC) Barcelona 08034, Spain (mruiz,adrian)@gps.tsc.upc.edu Abstract Reordering is currently one of the most important problems in statistical machine translation systems. This paper presents a novel strategy for dealing with it: sta- tistical machine reordering (SMR). It con- sists in using the powerful techniques de- veloped for statistical machine translation (SMT) to translate the source language (S) into a reordered source language (S’), which allows for an improved translation into the target language (T). The SMT task changes from S2T to S’2T which leads to a monotonized word alignment and shorter translation units. In addition, the use of classes in SMR helps to infer new word reorderings. Experiments are reported in the EsEn WMT06 tasks and the ZhEn IWSLT05 task and show significant im- provement in translation quality. 1 Introduction During the last few years, SMT systems have evolved from the original word-based ap- proach (Brown et al., 1993) to phrase-based trans- lation systems (Koehn et al., 2003). In parallel to the phrase-based approach, the use of bilin- gual n-grams gives comparable results, as shown by Crego et al. (2005a). Two basic issues differ- entiate the n-gram-based system from the phrase- based: training data are monotonously segmented into bilingual units; and, the model considers n- gram probabilities rather than relative frequencies. This translation approach is described in detail by Mari˜no et al. (2005). The n-gram-based system follows a maximum entropy approach, in which a log-linear combination of multiple models is im- plemented (Och and Ney, 2002), as an alternative to the source-channel approach. In both systems, introducing reordering capabil- ities is of crucial importance for certain language pairs. Recently, new reordering strategies have been proposed in the literature on SMT such as the reordering of each source sentence to match the word order in the corresponding target sentence, see Kanthak et al. (2005) and Crego et al. (2005b). Similarly, Matusov et al. (2006) describe a method for simultaneously aligning and monotonizing the training corpus. The main problems of these ap- proaches are: (1) the fact that the proposed mono- tonization is based on the alignment and cannot be applied to the test sets, and (2) the lack of reorder- ing generalization. This paper presents a reordering approach called statistical machine reordering (SMR) which improves the reordering capabilities of SMT sys- tems without incurring any of the problems men- tioned above. SMR is a first-pass translation performed on the source corpus, which converts it into an intermediate representation, in which source-language words are presented in an order that more closely matches that of the target lan- guage. SMR and SMT are performed using the same modeling tools as n-gram-based systems but using different statistical log-linear models. In order to be able to infer new reorderings we use word …

Content truncated to 500 words