Small in Size, Big in Precision: A Case for Using Language-Specific Lexical Resources for Word Sense Disambiguation

Linked open data (LOD) presents an ideal platform for connecting the multilingual lexical resources used in natural language processing (NLP) tasks, but the use of machine translation to fill in gaps in lexical coverage for resource-poor languages means that large amounts of data are potentially unverified. For graph-based word sense disambiguation (WSD), one approach has been to first translate terms into English in order to disambiguate using richer, fuller lexical knowledge bases (LKBs) such as WordNet. In this paper, we show that this approach actually creates more ambiguity and is far less accurate than using languagespecific resources, which, regardless of their smaller size, can provide results comparable in accuracy to the state-of-theart reported for graph-based WSD in English. For LOD, this demonstrates the importance of continuing to grow and extend language-specific resources in order to continually verify and reintegrate them as accurate resources.