English-to-Czech Factored Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 232–239, Prague, June 2007. c⃝2007 Association for Computational Linguistics English-to-Czech Factored Machine Translation Ondˇrej Bojar Institute of Formal and Applied Linguistics ´UFAL MFF UK, Malostransk´e n´amˇest´ı 25 CZ-11800 Praha, Czech Republic bojar@ufal.mff.cuni.cz Abstract This paper describes experiments with English-to-Czech phrase-based machine translation. Additional annotation of input and output tokens (multiple factors) is used to explicitly model morphology. We vary the translation scenario (the setup of multi- ple factors) and the amount of information in the morphological tags. Experimental results demonstrate significant improvement of translation quality in terms of BLEU. 1 Introduction Statistical phrase-based machine translation (SMT) systems currently achieve top performing results.1 Known limitations of phrase-based SMT include worse quality when translating to morphologically rich languages as opposed to translating from them (Koehn, 2005). One of the teams at the 2006 sum- mer engineering workshop at Johns Hopkins Uni- versity2 attempted to tackle these problems by in- troducing separate FACTORS in SMT input and/or output to allow explicit modelling of the underlying language structure. The support for factored transla- tion models was incorporated into the Moses open- source SMT system3. In this paper, we report on experiments with English-to-Czech multi-factor translation. After a brief overview of factored SMT and our data (Sec- tions 2 and 3), we summarize some possible trans- lating scenarios in Section 4. Section 5 studies the 1http://www.nist.gov/speech/tests/mt/ 2http://www.clsp.jhu.edu/ws2006/ 3http://www.statmt.org/moses/ level of detail useful for morphological representa- tion and Section 6 compares the results to a setting with more data available, albeit out of domain. The second part (Section 7) is devoted to a brief analysis of MT output errors. 1.1 Motivation for Improving Morphology Czech is a Slavic language with very rich morphol- ogy and relatively free word order. The Czech mor- phological system (Hajiˇc, 2004) defines 4,000 tags in theory and 2,000 were actually seen in a big tagged corpus. (For comparison, the English Penn Treebank tagset contains just about 50 tags.) In our parallel corpus (see Section 3 below), the English vocabulary size is 35k distinct token types but more than twice as big in Czech, 83k distinct token types. To further emphasize the importance of morphol- ogy in MT to Czech, we compare the standard BLEU (Papineni et al., 2002) of a baseline phrase- based translation with BLEU which disregards word forms (lemmatized MT output is compared to lem- matized reference translation). The theoretical mar- gin for improving MT quality is about 9 BLEU points: the same MT output scores 12 points in stan- dard BLEU and 21 points in lemmatized BLEU. 2 Overview of Factored SMT In statistical MT, the goal is to translate a source (foreign) language sentence f J 1 = f1 . . . fj . . . fJ into a target language (Czech) sentence cI 1 c1 . . . cj . . . cI. In phrase-based SMT, the assump-= tion is made that the target sentence can be con- structed by segmenting source sentence into phrases, translating …

Content truncated to 500 words