CUni Multilingual Matrix in the WMT 2013 Shared Task

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 85–91, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics CUni Multilingual Matrix in the WMT 2013 Shared Task Karel Bílek Daniel Zeman Charles University in Prague, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics Malostranské náměstí 25, CZ-11800 Praha, Czechia kb@karelbilek.com, zeman@ufal.mff.cuni.cz Abstract We describe our experiments with phrase-based machine translation for the WMT 2013 Shared Task. We trained one system for 18 translation directions between English or Czech on one side and English, Czech, Ger- man, Spanish, French or Russian on the other side. We describe a set of re- sults with different training data sizes and subsets. For the pairs containing Russian, we describe a set of indepen- dent experiments with slightly different translation models. 1 Introduction With so many official languages, Europe is a paradise for machine translation research. One of the largest bodies of electronically available parallel texts is being nowadays gen- erated by the European Union and its insti- tutions. At the same time, the EU also pro- vides motivation and boosts potential market for machine translation outcomes. Most of the major European languages be- long to one of three branches of the Indo- European language family: Germanic, Ro- mance or Slavic. Such relatedness is respon- sible for many structural similarities in Eu- ropean languages, although significant differ- ences still exist. Within the language portfo- lio selected for the WMT shared task, English, French and Spanish seem to be closer to each other than to the rest. German, despite being genetically related to English, differs in many properties. Its word order rules, shifting verbs from one end of the sentence to the other, easily cre- ate long-distance dependencies. Long Ger- man compound words are notorious for in- creasing out-of-vocabulary rate, which has led many researchers to devising unsupervised compound-splitting techniques. Also, upper- case/lowercase distinction is more important because all German nouns start with an up- percase letter by the rule. Czech is a language with rich morphology (both inflectional and derivational) and rela- tively free word order. In fact, the predicate- argument structure, often encoded by fixed word order in English, is usually captured by inflection (especially the system of 7 grammat- ical cases) in Czech. While the free word order of Czech is a problem when translating to En- glish (the text should be parsed first in order to determine the syntactic functions and the English word order), generating correct inflec- tional affixes is indeed a challenge for English- to-Czech systems. Furthermore, the multitude of possible Czech word forms (at least order of magnitude higher than in English) makes the data sparseness problem really severe, hinder- ing both directions. Most of the above characteristics of Czech also apply to Russian, another Slavic language. Similar issues have to be expected when trans- lating between Russian and English. Still, there are also interesting divergences between Russian and Czech, especially on the syntactic level. Russian sentences typically omit cop- ula in the …

Content truncated to 500 words