A Machine Learning Approach to Building Aligned Wordnets

WordNet is a lexical database describing English words and their senses. We propose a method for automatically producing similar resources for new languages by taking advantage of the original WordNet in conjunction with translation dictionaries. A small set of training mappings is used to learn a model for predicting associations between terms and senses. The associations are represented using a variety of scores that take into account structural properties as well as semantic relatedness and corpus frequency information. For evaluation, we created a German-language wordnet, and the data indicate a significantly better coverage and higher precision than previous heuristics. The resulting resources provide not only valuable information for monolingual NLP tasks but also enable a high degree of cross-lingual interoperability.