After the preprocessing stage, WordNetl.7 was used to guide the sense disambiguation pro­cess. For every combination of a word form and a POS, WordNet was consulted to determine whether this combination had one or more pos­sible senses. In case of only one possible sense (about 20% of the test words), the appropriate WordNet sense was assigned. In case of more possible senses, a threshold of 11 occurrences in the Semcor training data was determined. For all words below this threshold, the most frequent sense according to WordNet was as­signed as sense-tag. For the other words, which represent more than 60% of the word forms to be sense-tagged, word experts were built for each word form-POS combination, leading to 568 word experts for the SENSEVAL2 test data.

These word experts consist of different trained subcomponents (see Figure 1) which make use of different knowledge.

The first subcomponent is trained us­ing TiMBL, a package containing several memory-based learning algorithms and metrics (Daelemans et al., 2000). It takes as input a vector representing the local context of the focus word in a window of three word forms to the left and three to the right. For the focus word, also the lemma and POS are provided. For the context word forms, POS information is given. E.g., the following is a training instance: many JJ times MS , , yet yet RB on IN each JJ occasion NN yet%4:02:02 : :. During training, those instances are stored in memory and during sense-tagging, the instance most similar to that of the ambiguous word and its context is selected and the associated class is returned as sense-tag.

A second subcomponent of each word ex­pert trained with TiMBL is trained with in­formation about possible disambiguating con­tent keywords in a context of three sentences. The method used to extract these keywords for each sense is based on the work of (Ng and Lee, 1996). They determine the probability of a sense s of a focus word / given keyword k by dividing Ns^ioc (the number of occurrences of a possible local context keyword k with a par­ticular focus word-POS combination w with a particular sense s) by N^ioc (the number of oc­currences of a possible local context keyword kloc with a particular focus word-POS combination w ignoring its sense). In addition, we also took into account the frequency of a pos­sible keyword in the complete training corpus

A word is a keyword for a given sense if (i) the word occurs more than Mi times in that sense s, where Mi is a predefined minimum number of times and if (ii) p(s\k) > M2 for that sense s, where M2 is some predefined minimum proba­bility. Due to time restrictions Mi was not op­timized by cross-validation, but arbitrarily set to 3 and M2 to 0.001.

In addition to the keyword information ex­tracted from the local context of the focus word, possible disambiguating content words were also extracted from the examples that accompany the different sense definitions for a given focus word in WordNet. For each combination of a word form, POS and sense, all content words were extracted and added to the input vector of the memory-based learner. Both the contex­tual keywords and the example keywords were represented as binary features, with a value of 1 when the keyword was present in the example and 0 if not.

The third subcomponent of each word expert was trained with Ripper (Cohen, 1995), a rule learning algorithm, allowing both single-valued and set-valued attributes. In our disambigua­tion task, the ripper input vector contained lo­cal context feature values (as the first TiMBL), and a set-valued feature with all content words in a context of three sentences.

For the first TiMBL memory-based learner, backward sequential selection (BSS) (Aha and

Bankert, 1994) was performed for each word form-POS combination. BSS starts from the complete feature set and generates in each iter­ation new subsets by discarding a feature. The feature string with the best performance is re­tained. Furthermore, the use of different fea­ture weighting possibilities was explored, viz. gain ratio weighting, information gain weight­ing, chi-squared weighting and shared variance weighting. For each feature weighting possi­bility, the k value, representing the number of nearest neighbours used for extrapolation, was varied between 1 and 19. Leave-one-out was used as testing method: testing was done on each instance of the training file, while the re­mainder of the training file functioned as train­ing material.

Due to the size of the feature vectors for the second memory-based learner, which takes content words from the surrounding sentences and from the example sentences in the Word-Net definitions as input, no feature selection was performed. For the same reasons, 10-fold cross-validation was used as testing method: the training data was split into 10 different parts and in each iteration, one part served as test set, while the remainder was used to train the classifier. The k value was varied (1-19), dif­ferent weighting techniques (gain ratio weight­ing, chi-squared weighting and log likelihood weighting) and different distance metrics (num­ber of mismatches, number of matches, number of matches minus number of mismatches) were explored.

For Ripper, the default parameter settings were used, due to time constraints and the slow­ness of the cross-validation process. 10-fold-cross-validation was used as testing method.

On the output of these three (optimized) classi­fiers and the default Word .Not 1.7. most frequent sense, both majority voting and weighted vot­ing was performed. In case of majority voting, each sense-tagger is given one vote and the tag with most votes is selected. In weighted vot­ing, more weight is given to the taggers with a higher overall accuracy. In case of ties when voting over the output of 4 classifiers, the first decision (TiMBL) was taken as output class. Voting was also performed on the output of the three learning classifiers without taking into account the WordNet class. Table 1 shows the best performing classifiers per word form-POS combination of the Semcor train set: both op­timized memory-based learners outperform the other classifiers.