NUS at WMT09: Domain Adaptation Experiments for English-Spanish Machine Translation of News Commentary Text

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 75–79, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics NUS at WMT09: Domain Adaptation Experiments for English-Spanish Machine Translation of News Commentary Text Preslav Nakov Department of Computer Science National University of Singapore 13 Computing Drive Singapore 117417 nakov@comp.nus.edu.sg Hwee Tou Ng Department of Computer Science National University of Singapore 13 Computing Drive Singapore 117417 nght@comp.nus.edu.sg Abstract We describe the system developed by the team of the National University of Singa- pore for English to Spanish machine trans- lation of News Commentary text for the WMT09 Shared Translation Task. Our approach is based on domain adaptation, combining a small in-domain News Com- mentary bi-text and a large out-of-domain one from the Europarl corpus, from which we built and combined two separate phrase tables. We further combined two language models (in-domain and out-of-domain), and we experimented with cognates, im- proved tokenization and recasing, achiev- ing the highest lowercased NIST score of 6.963 and the second best lowercased Bleu score of 24.91% for training without us- ing additional external data for English-to- Spanish translation at the shared task. 1 Introduction Modern Statistical Machine Translation (SMT) systems are typically trained on sentence-aligned parallel texts (bi-texts) from a particular domain. When tested on text from that domain, they demonstrate state-of-the art performance, but on out-of-domain test data the results can deteriorate significantly. For example, on the WMT06 Shared Translation Task, the scores for French-to-English translation dropped from about 30 to about 20 Bleu points for nearly all systems when tested on News Commentary instead of the Europarl1 text, which was used for training (Koehn and Monz, 2006). 1See (Koehn, 2005) for details about the Europarl corpus. Subsequently, in 2007 and 2008, the WMT Shared Translation Task organizers provided a limited amount of bilingual News Commentary training data (1-1.3M words) in addition to the large amount of Europarl data (30-32M words), and set up separate evaluations on News Commen- tary and on Europarl data, thus inviting interest in domain adaptation experiments for the News do- main (Callison-Burch et al., 2007; Callison-Burch et al., 2008). This year, the evaluation is on News Commentary only, which makes domain adapta- tion the central focus of the Shared Translation Task. The team of the National University of Singa- pore (NUS) participated in the WMT09 Shared Translation Task with an English-to-Spanish sys- tem.2 Our approach is based on domain adapta- tion, combining the small in-domain News Com- mentary bi-text (1.8M words) and the large out- of-domain one from the Europarl corpus (40M words), from which we built and combined two separate phrase tables. We further used two language models (in-domain and out-of-domain), cognates, improved tokenization, and additional smart recasing as a post-processing step. 2 The NUS System Below we describe separately the standard and the nonstandard settings of our system. 2.1 Standard Settings In our baseline experiments, we used the follow- ing general setup: First, we tokenized the par- 2The task organizers invited submissions …

Content truncated to 500 words