Abstract

Princeton WordNet is one of the most important resources for natural languageprocessing, but is only available for English. While it has been translatedusing the expand approach to many other languages, this is an expensive manualprocess. Therefore it would be beneficial to have a high-quality automatictranslation approach that would support NLP techniques, which rely on WordNetin new languages. The translation of wordnets is fundamentally complex becauseof the need to translate all senses of a word including low frequency senses,which is very challenging for current machine translation approaches. For thisreason we leverage existing translations of WordNet in other languages toidentify contextual information for wordnet senses from a large set of genericparallel corpora. We evaluate our approach using 10 translated wordnets forEuropean languages. Our experiment shows a significant improvement overtranslation without any contextual information. Furthermore, we evaluate howthe choice of pivot languages affects performance of multilingual word sensedisambiguation.

 

Introduction (beta)

None

 

Conclusion (beta)

None