Advancements in Reordering Models for Statistical Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 322–332, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Advancements in Reordering Models for Statistical Machine Translation Minwei Feng and Jan-Thorsten Peter and Hermann Ney Human Language Technology and Pattern Recognition Computer Science Department, RWTH Aachen University, Aachen, Germany <surname>@cs.rwth-aachen.de Abstract In this paper, we propose a novel re- ordering model based on sequence label- ing techniques. Our model converts the reordering problem into a sequence label- ing problem, i.e. a tagging task. Results on five Chinese-English NIST tasks show that our model improves the baseline sys- tem by 1.32 BLEU and 1.53 TER on av- erage. Results of comparative study with other seven widely used reordering mod- els will also be reported. 1 Introduction The systematic word order difference between two languages poses a challenge for current statistical machine translation (SMT) systems. The system has to decide in which order to translate the given source words. This problem is known as the re- ordering problem. As shown in (Knight, 1999), if arbitrary reordering is allowed, the search problem is NP-hard. Many ideas have been proposed to address the reordering problem. Within the phrase-based SMT framework there are mainly three stages where improved reordering could be integrated: In the preprocessing: the source sentence is re- ordered by heuristics, so that the word order of source and target sentences is similar. (Wang et al., 2007) use manually designed rules to reorder parse trees of the source sentences. Based on shal- low syntax, (Zhang et al., 2007) use rules to re- order the source sentences on the chunk level and provide a source-reordering lattice instead of a sin- gle reordered source sentence as input to the SMT system. Designing rules to reorder the source sen- tence is conceptually clear and usually easy to im- plement. In this way, syntax information can be in- corporated into phrase-based SMT systems. How- ever, one disadvantage is that the reliability of the rules is often language pair dependent. In the decoder: we can add constraints or mod- els into the decoder to reward good reordering op- tions or penalize bad ones. For reordering con- straints, early work includes ITG constraints (Wu, 1995) and IBM constraints (Berger et al., 1996). (Zens and Ney, 2003) did comparative study over different reordering constraints. This paper fo- cuses on reordering models. For reordering mod- els, we can further roughly divide the existing methods into three genres: • The reordering is a classification problem. The classifier will make decision on next phrase’s relative position with current phrase. The classifier can be trained with maximum likelihood like Moses lexicalized reordering (Koehn et al., 2007) and hierarchical lexical- ized reordering model (Galley and Manning, 2008) or be trained under maximum entropy framework (Zens and Ney, 2006). • The reordering is a decoding order problem. (Mari˜no et al., 2006) present a translation model that constitutes a language model of a sort of bilanguage composed of bilingual units. From the …

Content truncated to 500 words