Rapid development of RBMT systems for related languages

摘要

A research of already available and accessible language processing tools and materials, mostly corpora, revealed that there is a reasonably big amount of work already done for Slovenian language, less for Serbian. The tools for Slovene language are (reasonable or even good quality): part of speech tagger (Erjavec et al., 2000), lemmatizer (Erjavec et al., 2004), stemmer, none of these tools exists for Serbian language. Both languages have solid monolingual reference corpora (going into hundreds of millions) and a small bilingual corpus that was used mostly for evaluation purposes. Evaluation was conducted on the functional machine translation system and the results presenting coverage using referential corpus and selected evaluation metrics are shown. Objective and subjective evaluation methods were used as only a correct mixture of methods minimizes evaluation bias. Translation quality evaluation was conducted using subjective evaluation methods where a set of native speakers scored translations. Automatic objective measures NIST and BLEU (Papineni et al., 2001) were used to ensure wider coverage. Bilingual corpus was used in both automatic evaluations. Conclusions present strong and weak points of this approach and explore grounds for further work.