Learning Thematic Role Relations for Wordnets

In this paper, I present a method for learning thematic role relations (selectional preferences) for wordnets by means of statistical corpus analysis. An evaluation on a gold standard, which I extracted from EuroWordNet, shows that this method achieves a learning accuracy of up to 77%. I also propose a preprocessing step for a partial lexical disambiguation of the input data. This disambiguation step reduces noise in the output of the learning algorithm, accompanied by a comparably moderate loss in performance.