A Lemma-Based Approach to a Maximum Entropy Word Sense Disambiguation System for Dutch

A Lemma-Based Approach to a Maximum Entropy Word Sense Disambiguation System for Dutch Tanja Gaustad Alfa-Informatica, University of Groningen Postbus 716 9700 AS Groningen, The Netherlands, T.Gaustad@let.rug.nl Abstract In this paper, we present a corpus-based super- vised word sense disambiguation (WSD) sys- tem for Dutch which combines statistical classi- fication (maximum entropy) with linguistic in- formation. Instead of building individual clas- sifiers per ambiguous wordform, we introduce a lemma-based approach. The advantage of this novel method is that it clusters all inflec- ted forms of an ambiguous word in one classi- fier, therefore augmenting the training material available to the algorithm. Testing the lemma- based model on the Dutch SENSEVAL-2 test data, we achieve a significant increase in accur- acy over the wordform model. Also, the WSD system based on lemmas is smaller and more robust. 1 Introduction: WSD for Dutch A major problem in natural language processing (NLP) for which no satisfactory solution has been found to date is word sense disambiguation (WSD). WSD refers to the resolution of lexical semantic ambiguity and its goal is to attribute the correct sense(s) to words in a certain context. For instance machine translation, information retrieval or docu- ment extraction could all benefi t from the accurate disambiguation of word senses. The WSD system for Dutch1 presented here is a corpus-based supervised algorithm combining stat- istical classifi cation with various kinds of linguistic information. The intuition behind the system is that linguistic information is benefi cial for WSD which means that it will improve results over purely stat- istical approaches. The linguistic information in- cludes lemmas, part-of-speech (PoS), and the con- text around the ambiguous word. In this paper, we focus on a lemma-based ap- proach to WSD for Dutch. So far, systems built individual classifi ers for each ambiguous wordform 1The interest in Dutch lies grounded in the fact that we are working in the context of a project con- cerned with developing NLP tools for Dutch (see http://www.let.rug.nl/˜vannoord/alp). (Hendrickx et al., 2002; Hoste et al., 2002). In the system presented here, the classifi ers built for each ambiguous word are based on its lemma in- stead. Lemmatization allows for more compact and generalizable data by clustering all inflected forms of an ambiguous word together, an effect already commented on by Yarowsky (1994). The more in- flection in a language, the more lemmatization will help to compress and generalize the data. In the case of our WSD system this means that less classifi ers have to be built therefore adding up the training ma- terial available to the algorithm for each ambiguous wordform. Accuracy is expected to increase for the lemma-based model in comparison to the wordform model. The paper is structured as follows: First, we will present the dictionary-based lemmatizer for Dutch which was used to lemmatize the data, followed by a detailed explanation of the lemma-based approach adopted in our WSD system. Next, the statistical classifi cation algorithm, namely maximum entropy, and Gaussian priors (used for smoothing purposes) are …

Content truncated to 500 words