HKUST statistical machine translation experiments for IWSLT 2007

HKUST Statistical Machine Translation Experiments for IWSLT 2007 Yihai SHEN Chi-kiu LO Marine CARPUAT Dekai WU Human Language Technology Center HKUST Department of Computer Science, University of Science and Technology, Hong Kong � shenyh,jackielo,marine,dekai � @cs.ust.hk Abstract This paper describes the HKUST experiments in the IWSLT 2007 evaluation campaign on spoken lan- guage translation. Our primary objective was to com- pare the open-source phrase-based statistical machine translation toolkit Moses against Pharaoh. We focused on Chinese to English translation, but we also report re- sults on the Arabic to English, Italian to English, and Japanese to English tasks. 1. Introduction We describe experiments conducted at HKUST during the IWSLT 2007 evaluation campaign on spoken lan- guage translation. For our second participation in the IWSLT evaluation, our focus was on experimenting with Moses (Koehn et al., 2007), the new open-source toolkit for phrase-based Statistical Machine Translation (SMT), and on comparing it against its closed-source prede- cessor Pharaoh (Koehn, 2004) which we used in our IWSLT 2006 submission (Carpuat et al., 2006). Our main focus was on the Chinese-English task, which, this year, used clean text as opposed to the other tasks where speech transcriptions were to be translated. We also report results on all the language pairs, although we did not do any tuning or any language-specific pro- cessing for the Arabic to English, Japanese to English and Italian to English tasks. � This material is based upon work supported in part by the Defense Advanced Research Projects Agency (DARPA) under GALE Contract No. HR0011-06-C-0023, and by the Hong Kong Research Grants Council (RGC) research grants RGC6083/99E, RGC6256/00E, and DAG03/04.EG09. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the Defense Advanced Research Projects Agency. 2. Machine translation engine We focus on phrase-based statistical machine transla- tion since this approach has been shown to achieve com- petitive translation quality in many state-of-the-art sys- tems (e.g., Zens et al. (2005), Shen et al. (2006), Eck et al. (2006)), and therefore is a good benchmark to eval- uate decoder architectures. 2.1. Decoder We use the Moses decoder (Koehn et al., 2007), which is an open source toolkit for statistical machine trans- lation. Just like the Pharaoh decoder, Moses uses a log-linear model, which combines several knowledge sources in translation decisions. Moses differs from Pharaoh by its represention of each input word as a factor as opposed to the word suface form only. Fac- tors can include additional information such as part-of- speech, class, morphology and allow the phrase-based model to incorporate richer linguistic information. How- ever, we do not use the factored representation in this first set of experiments, and use the surface form of words, just like in Pharaoh. 2.2. Phrasal bilexicon The core phrasal bilexicon is obtained by collecting phrase pairs that are consistent with the IBM model 4 align- ments obtained with GIZA++ (Och and Ney, 2002). During phrase extraction, we tried two different meth- ods …

Content truncated to 500 words