Applying Morphological Decompositions to Statistical Machine Translation

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 195–200, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Applying morphological decomposition to statistical machine translation Sami Virpioja and Jaakko V¨ayrynen and Andr´e Mansikkaniemi and Mikko Kurimo Aalto University School of Science and Technology Department of Information and Computer Science PO BOX 15400, 00076 Aalto, Finland {svirpioj,jjvayryn,ammansik,mikkok}@cis.hut.fi Abstract This paper describes the Aalto submission for the German-to-English and the Czech- to-English translation tasks of the ACL 2010 Joint Fifth Workshop on Statistical Machine Translation and MetricsMATR. Statistical machine translation has focused on using words, and longer phrases con- structed from words, as tokens in the sys- tem. In contrast, we apply different mor- phological decompositions of words using the unsupervised Morfessor algorithms. While translation models trained using the morphological decompositions did not im- prove the BLEU scores, we show that the Minimum Bayes Risk combination with a word-based translation model produces significant improvements for the German- to-English translation. However, we did not see improvements for the Czech-to- English translations. 1 Introduction The effect of morphological variation in languages can be alleviated by using word analysis schemes, which may include morpheme discovery, part-of- speech tagging, or other linguistic information. Words are very convenient and even efficient rep- resentation in statistical natural language process- ing, especially with English, but morphologically rich languages can benefit from more fine-grained information. For instance, statistical morphs dis- covered with unsupervised methods result in bet- ter performance in automatic speech recognition for highly-inflecting and agglutinative languages (Hirsim¨aki et al., 2006; Kurimo et al., 2006). Virpioja et al. (2007) applied morph-based models in statistical machine translation (SMT) between several language pairs without gaining improvement in BLEU score, but obtaining re- ductions in out-of-vocabulary rates. They uti- lized morphs both in the source and in the tar- get language. Later, de Gispert et al. (2009) showed that Minimum Bayes Risk (MBR) com- bination of word-based and morph-based trans- lation models improves translation with Arabic- to-English and Finnish-to-English language pairs, where only the source language utilized morph- based models. Similar results have been shown for Finnish-to-English and Finnish-to-German in per- formance evaluation of various unsupervised mor- pheme analysis algorithms in Morpho Challenge 2009 competition (Kurimo et al., 2009). We continue the research described above and examine how the level of decomposition affects both the individual morph-based systems and MBR combinations with the baseline word-based model. Experiments are conducted with the WMT10 shared task data for German-to-English and Czech-to-English language pairs. 2 Methods In this work, morphological analyses are con- ducted on the source language data, and each dif- ferent analysis is applied to create a unique seg- mentation of words into morphemes. Translation systems are trained with the Moses toolkit (Koehn et al., 2007) from each differently segmented ver- sion of the same source language to the target lan- guage. Evaluation with BLEU is performed on both the individual systems and system combina- tions, using different levels of decomposition. 2.1 Morphological models for words Morfessor …

Content truncated to 500 words