Statistical multi-source translation

Statistical Multi-Source Translation Franz Josef Och and Hermann Ney Lehrstuhl f¨ur Informatik VI, Computer Science Department RWTH Aachen - University of Technology 52056 Aachen, Germany �och,ney �@informatik.rwth-aachen.de Abstract We describe methods for translating a text given in multiple source languages into a single target language. The goal is to improve translation quality in applications where the ultimate goal is to translate the same document into many languages. We describe a statistical approach and two specific statistical models to deal with this problem. Our method is generally applicable as it is independent of specific models, languages or application domains. We evaluate the approach on a multilingual corpus covering all eleven official European Union languages that was collected automatically from the Internet. In various tests we show that these methods can significantly improve translation quality. As a side effect, we also compare the quality of statistical machine translation systems for many European languages in the same domain. Keywords statistical machine translation, multi-source translation, automated acquisition of parallel corpora 1 Introduction In many applications for machine translation, it is nec- essary to translate a document into multiple languages. For example, in international organizations such as the European Union or the United Nations, all rele- vant documents must be translated into all official lan- guages. Very often, the document is originally written in one language, and then translated into the other lan- guages. If, for example, first is produced an English translation of a French document, then this translation should be used as additional knowledge source when producing a German translation. So far, existing ma- chine translation technology is not able to make use of these additional knowledge sources. From performing multi-source translation, we ex- pect a better machine translation quality due to the fol- lowing reasons: � Better word sense disambiguation: Often ambi- guities that need to be resolved between two lan- guages do not exist between other languages. � Better word reordering: A significant source of errors in statistical machine translation is the word reordering problem (Och et al., 1999). The word order between related languages is often very similar while the word order between dis- tant languages might differ significantly. By us- ing more source languages, we can expect that among the source languages there is one with a similar word order. � Reduction of the need for explicit anaphora reso- lution: By having various translations of a pro- noun in different languages the probability in- creases that it can be translated correctly without performing a full anaphora resolution. Our method for performing multi-source translation fits nicely into the statistical approach and is relatively easy to implement. We are able to deduce a general statistical approach to multi-source translation. Our method is very general as it is independent of specific models, languages or application domains. Ultimately, the approach boils down to a multiplicative combina- tion of various statistical translation models. In principle, multi-source translation is not re- stricted to a statistical approach and it would be pos- sible to pursue it …

Content truncated to 500 words