SMT Experiments for Romanian and German Using JRC-ACQUIS

One of the LT-applications that ensures the access to the information, in the user’s mother tongue, is machine translation (MT). Unfortunately less spoken languages a category in which the Balkan and Slavic languages can be included have to overcome a major gap in language resources, reference-systems and tools. In its simplest form, statistical machine translation (SMT) is based only on the existence of a big parallel corpus and therefore it seems to be a solution for these languages. In this paper the performance of a Moses-based SMT system, for Romanian and German, is investigated using test data from two different domains legislation (JRCACQUIS) and a manual of an electronic device. The obtained results are compared with the ones given by the Google on-line translation tool. An analysis of the obtained translation results gives an overview of the main challenges and sources of errors in translation, in these experimental settings.