ILLC-UvA translation system for EMNLP-WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 413–419, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics ILLC-UvA translation system for EMNLP-WMT 2011 Maxim Khalilov and Khalil Sima’an Institute for Logic, Language and Computation University of Amsterdam P.O. Box 94242 1090 GE Amsterdam, The Netherlands {m.khalilov,k.simaan}@uva.nl Abstract In this paper we describe the Institute for Logic, Language and Computation (Uni- versity of Amsterdam) phrase-based statisti- cal machine translation system for English- to-German translation proposed within the EMNLP-WMT 2011 shared task. The main novelty of the submitted system is a syntax- driven pre-translation reordering algorithm implemented as source string permutation via transfer of the source-side syntax tree. 1 Introduction For the WMT 2011 shared task, ILLC-UvA submit- ted two translations (primary and secondary) for the English-to-German translation task. This year, we directed our research toward addressing the word order problem for statistical machine translation (SMT) and discover its impact on output translation quality. We reorder the words of a sentence of the source language with respect to the word order of the target language and a given source-side parse tree. The difference from the baseline Moses-based translation system lies in the pre-translation step, in which we introduce a discriminative source string permutation model based on probabilistic parse tree transduction. The idea here is to permute the order of the source words in such a way that the resulting permutation allows as monotone a translation process as possible is not new. This approach to enhance SMT by using a reordering step prior to translation has proved to be successful in improving translation quality for many translation tasks, see (Genzel, 2010; Costa-juss`a and Fonollosa, 2006; Collins et al., 2005), for example. The general problem of source-side reordering is that the number of permutations is factorial in n, and learning a sequence of transductions for explaining a source permutation can be computationally rather challenging. We propose to address this problem by defining the source-side permutation process as the learning problem of how to transfer a given source parse tree into a parse tree that minimizes the diver- gence from target word order. Our reordering system is inspired by the direction taken in (Tromble and Eisner, 2009), but differs in defining the space of permutations, using local prob- abilistic tree transductions, as well as in the learn- ing objective aiming at scoring permutations based on a log-linear interpolation of a local syntax-based model with a global string-based (language) model. The reordering (novel) and translation (standard) components are described in the following sections. The rest of this paper is structured as follows. After a brief description of the phrase-based translation sys- tem in Section 2, we present the architecture and de- tails of our reordering system (Section 3), Section 4 reviews related work, Section 5 reports the experi- mental setup, details the submissions and discusses the results, while Section 6 concludes the article. 2 Baseline system 2.1 Statistical machine translation In SMT the translation problem is formulated as se- lecting the …

Content truncated to 500 words