Statistical Machine Translation of Serbian-English

In this work we present the first results of statistical approach to the machine translation of Serbian language into English and vice versa. The experiments are performed on the Assimil language course, bilingual parallel corpus which consists of about 3k sentences and 20k running words from unrestricted domain. The error rates for the translation of Serbian into English are about 35-45% and for the other direction about 45-55%. The results are comparable with those for the other language pairs having been translated using statistical approach. Reducing Serbian words into stems has decreased error rates for the translation into English for about 8% relative.