Delemmatization strategies for Dutch

Abstract

In this paper we investigate whether, for Dutch open class words, it is possible to generate the surface form on the basis of the lemma and the POS tag, using a lexicon and a machine learning system. When testing on an annotated corpus, we are able to generate more than 97% of the gold standard types correctly and over 99% of the gold standard tokens. The most efficient strategy appears to be the pure machine learning one, even for those words that do occur in the lexicon. Apart from overall statistics, we look at specific machine learner settings in more detail and also investigate the errors made by the best scoring strategy.