Mapping Multilingual Hierarchies Using Relaxation Labeling

摘要

This paper explores the automatic construction of a multilingual LexicalKnowledge Base from pre-existing lexical resources. We present a new and robustapproach for linking already existing lexical/semantic hierarchies. We used aconstraint satisfaction algorithm (relaxation labeling) to select --among allthe candidate translations proposed by a bilingual dictionary-- the rightEnglish WordNet synset for each sense in a taxonomy automatically derived froma Spanish monolingual dictionary. Although on average, there are 15 possibleWordNet connections for each sense in the taxonomy, the method achieves anaccuracy over 80%. Finally, we also propose several ways in which thistechnique could be applied to enrich and improve existing lexical databases.