Data Issues of the Multilingual Translation Matrix

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 395–400, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Data Issues of the Multilingual Translation Matrix Daniel Zeman Charles University in Prague, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics Malostranské náměstí 25, CZ-11800 Praha, Czechia zeman@ufal.mff.cuni.cz Abstract We describe our experiments with phrase- based machine translation for the WMT 2012 Shared Task. We trained one sys- tem for 14 translation directions between English or Czech on one side and English, Czech, German, Spanish or French on the other side. We describe a set of results with different training data sizes and sub- sets. 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 generated by the European Union and its institutions. At the same time, the EU also provides motivation and boosts potential market for machine translation outcomes. Most of the major European languages belong to one of three branches of the Indo-European language family: Germanic, Romance or Slavic. Such relatedness is responsible for many struc- tural similarities in European languages, al- though significant differences still exist. Within the language portfolio 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 or- der rules, shifting verbs from one end of the sen- tence to the other, easily create long-distance de- pendencies. Long German compound words are notorious for increasing out-of-vocabulary rate, which has led many researchers to devising unsu- pervised compound-splitting techniques. Also, uppercase/lowercase distinction is more impor- tant because all German nouns start with an uppercase 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 inflec- tion (especially the system of 7 grammatical cases) in Czech. While the free word order of Czech is a problem when translating to English (the text should be parsed first in order to de- termine the syntactic functions and the English word order), generating correct inflectional af- fixes 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 sparse- ness problem really severe, hindering both direc- tions. Our goal is to run one system under as similar conditions as possible to all fourteen translation directions, to compare their translation accura- cies and see why some directions are easier than others. Future work will benefit from knowing what are the special processing needs for a given language pair. The current version of the system does not include really language-specific tech- niques: we neither split German compounds, nor do we address the peculiarities of Czech mentioned above. 395 2 The …

Content truncated to 500 words