The CNGL-DCU-Prompsit Translation Systems for WMT13

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 213–218, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics The CNGL-DCU-Prompsit Translation Systems for WMT13 Raphael Rubino†, Antonio Toral†, Santiago Cortés Vaíllo⋆, Jun Xie§, Xiaofeng Wu‡, Stephen Doherty♭, Qun Liu‡ †NCLT, Dublin City University, Ireland ⋆Prompsit Language Engineering, Spain §ICT, Chinese Academy of Sciences, China ‡,♭CNGL, Dublin City University, Ireland †,‡{rrubino, atoral, xfwu, qliu}@computing.dcu.ie ⋆santiago@prompsit.com §junxie@ict.ac.cn ♭stephen.doherty@dcu.ie Abstract This paper presents the experiments con- ducted by the Machine Translation group at DCU and Prompsit Language Engineer- ing for the WMT13 translation task. Three language pairs are considered: Spanish- English and French-English in both direc- tions and German-English in that direc- tion. For the Spanish-English pair, the use of linguistic information to select paral- lel data is investigated. For the French- English pair, the usefulness of the small in- domain parallel corpus is evaluated, com- pared to an out-of-domain parallel data sub-sampling method. Finally, for the German-English system, we describe our work in addressing the long distance re- ordering problem and a system combina- tion strategy. 1 Introduction This paper presents the experiments conducted by the Machine Translation group at DCU1 and Prompsit Language Engineering2 for the WMT13 translation task on three language pairs: Spanish- English, French-English and German-English. For these language pairs, the language and trans- lation models are built using different approaches and datasets, thus presented in this paper in sepa- rate sections. In Section 2, the systems built for the Spanish- English pair in both directions are described. We investigate the use of linguistic information to se- lect parallel data. In Section 3, we present the sys- tems built for the French-English pair in both di- 1http://www.nclt.dcu.ie/mt/ 2http://www.prompsit.com/ rections. The usefulness of the small in-domain parallel corpus is evaluated, compared to an out- of-domain parallel data sub-sampling method. In Section 4, for the German-English system, aiming at exploring the long distance reordering problem, we first describe our efforts in a dependency tree- to-string approach, before combining different hi- erarchical systems with a phrase-based system and show a significant improvement over three base- line systems. 2 Spanish-English This section describes the experimental setup for the Spanish-English language pair. 2.1 Setting Our setup uses the MOSES toolkit, version 1.0 (Koehn et al., 2007). We use a pipeline with the phrase-based decoder with standard pa- rameters, unless noted otherwise. The decoder uses cube pruning (-cube-pruning-pop-limit 2000 -s 2000), MBR (-mbr-size 800 -mbr-scale 1) and monotone at punctuation reordering. Individual language models (LMs), 5-gram and smoothed using a simplified version of the im- proved Kneser-Ney method (Chen and Goodman, 1996), are built for each monolingual corpus using IRSTLM 5.80.01 (Federico et al., 2008). These LMs are then interpolated with IRSTLM using the test set of WMT11 as the development set. Fi- nally, the interpolated LMs are merged into one LM preserving the weights using SRILM (Stol- cke, 2002). We use all the parallel corpora available for this language pair: Europarl (EU), News Com- mentary (NC), United Nations (UN) and Common …

Content truncated to 500 words