Edinburgh’s Submission to all Tracks of the WMT 2009 Shared Task with Reordering and Speed Improvements to Moses

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 160–164, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Edinburgh’s Submission to all Tracks of the WMT2009 Shared Task with Reordering and Speed Improvements to Moses Philipp Koehn and Barry Haddow School of Informatics University of Edinburgh pkoehn@inf.ed.ac.uk bhaddow@inf.ed.ac.uk Abstract Edinburgh University participated in the WMT 2009 shared task using the Moses phrase-based statistical machine transla- tion decoder, building systems for all lan- guage pairs. The system configuration was identical for all language pairs (with a few additional components for the German- English language pairs). This paper de- scribes the configuration of the systems, plus novel contributions to Moses includ- ing truecasing, more efficient decoding methods, and a framework to specify re- ordering constraints. 1 Introduction The commitment of the University of Edinburgh to the WMT shared tasks is to provide a strong sta- tistical machine translation baseline with our open source tools for all language pairs. We are again the only institution that participated in all tracks. The shared task is also an opportunity to incor- porate novel contributions and test them against the best machine translation systems for these lan- guage pairs. In this paper we describe the speed improvements to the Moses decoder (Koehn et al., 2007), as well as a novel framework to specify re- ordering constraints with XML markup, which we tested with punctuation-based constraints. 2 System Configuration We trained a default Moses system with the fol- lowing non-default settings: • maximum sentence length 80 • grow-diag-final-and symmetrization of GIZA++ alignments • interpolated Kneser-Ney discounted 5-gram language model • msd-bidrectional-fe lexicalized reordering Language ep nc news intpl. English 449 486 216 192 French 264 311 147 131 German 785 821 449 402 Spanish 341 392 219 190 Czech *:1475 1615 752 690 Hungarian hung:2148 815 786 Table 1: Perplexity (ppl) of the domain-trained (ep = Europarl (CzEng for Czech), nc = News Com- mentary, news = News) and interpolated language models. 2.1 Domain Adaptation In contrast to last year’s task, where news transla- tion was presented as a true out-of-domain prob- lem, this year large monolingual news corpora and a tuning set (last year’s test set) were pro- vided. While still no in-domain news parallel cor- pora were made available, the monolingual cor- pora could be exploited for domain adaption. For all language pairs, we built a 5-gram lan- guage model, by first training separate language models for the different training corpora (the par- allel Europarl and News Commentary and new monolingual news), and then interpolated them by optimizing perplexity on the provided tuning set. Perplexity numbers are shown in Table 1. 2.2 Truecasing Our traditional method to handle case is to low- ercase all training data, and then have a separate recasing (or recapitalization) step. Last year, we used truecasing: all words are normalized to their natural case, e.g. the, John, eBay, meaning that only sentence-leading words may be changed to their most frequent form. To refine last …

Content truncated to 500 words