Improving English-Spanish Statistical Machine Translation: Experiments in Domain Adaptation, Sentence Paraphrasing, Tokenization, and Recasing

Proceedings of the Third Workshop on Statistical Machine Translation, pages 147–150, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Improving English-Spanish Statistical Machine Translation: Experiments in Domain Adaptation, Sentence Paraphrasing, Tokenization, and Recasing Preslav Nakov∗ EECS, CS division University of California at Berkeley Berkeley, CA 94720 nakov@cs.berkeley.edu Abstract We describe the experiments of the UC Berke- ley team on improving English-Spanish ma- chine translation of news text, as part of the WMT’08 Shared Translation Task. We ex- periment with domain adaptation, combin- ing a small in-domain news bi-text and a large out-of-domain one from the Europarl corpus, building two separate phrase transla- tion models and two separate language mod- els. We further add a third phrase transla- tion model trained on a version of the news bi-text augmented with monolingual sentence- level syntactic paraphrases on the source- language side, and we combine all models in a log-linear model using minimum error rate training. Finally, we experiment with differ- ent tokenization and recasing rules, achieving 35.09% Bleu score on the WMT’07 news test data when translating from English to Span- ish, which is a sizable improvement over the highest Bleu score achieved on that dataset at WMT’07: 33.10% (in fact, by our sys- tem). On the WMT’08 English to Spanish news translation, we achieve 21.92%, which makes our team the second best on Bleu score. 1 Introduction Modern Statistical Machine Translation (SMT) sys- tems are trained on sentence-aligned bilingual cor- pora, typically from a single domain. When tested on text from that same domain, they demonstrate ∗After January 2008 at the Linguistic Modeling Depart- ment, Institute for Parallel Processing, Bulgarian Academy of Sciences, nakov@lml.bas.bg state-of-the art performance, but on out-of-domain test data the results can get significantly worse. For example, on the WMT’06 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 rather than Europarl text, which was used on training (Koehn and Monz, 2006). Therefore, in 2007 the Shared Task organizers provided 1M words of bilingual News Commentary training data in addition to the 30M Europarl data, thus inviting interest in domain adaptation experi- ments. Given the success of the idea, the same task was offered this year with slightly larger training bi- texts: 1.3M and 32M words, respectively. 2 System Parameters The team of the University of California at Berkeley (ucb) participated in the WMT’08 Shared Transla- tion Task with two systems, English→Spanish and Spanish→English, applied to translating News Com- mentary text, for which a very limited amount of training data was provided. We experimented with domain adaptation, combining the provided small in-domain bi-text and the large out-of-domain one from the Europarl corpus, building two phrase trans- lation models and two language models. We further added a third phrase translation model trained on a version of the news bi-text augmented with mono- lingual sentence-level syntactic paraphrases on the source-language side, and we combined all models in one big …

Content truncated to 500 words