The purpose of boosting-based algorithms is to find a highly accurate classification rule by combining many weak classifiers (or weak hypotheses), each of which may be only moderately accurate. The weak hypotheses are learned sequentially, one at a time, and, conceptually, at each iteration the weak hypothesis is biased to classify the examples which were most difficult to classify by the preceding weak hypotheses. The learned weak hypotheses are linearly combined into a single rule called the com­bined hypothesis.

The particular algorithm used in our system to perform the classification of senses is the gener­alized AdaBoost.MH with confidence-rated pre­dictions (Schapire and Singer, 1999). This algo­rithm is able to deal straightforwardly with mul­ticlass multi-label problems, and has been previ­ously applied, with significant success, to a num­ber of NLP disambiguation tasks, including, among others: Part-of-speech tagging and PP-attachment (Abney et al., 1999), text categorization (Schapire and Singer, 2000), and shallow parsing (Carreras and Màrquez, 2001). The weak hypotheses used in this work are decision stumps, which can be seen as extremely simple decision trees with one internal node testing the value of a single binary feature (e.g. "the word dark appears in the context of the word to be disambiguated?") and two leaves that give the prediction of the senses based on the feature value.

The "Lazy" Boosting, is a simple modification of the AdaBoost.MH algorithm, which consists of re­ducing the feature space that is explored when learn­ing each weak classifier. More specifically, a small proportion of attributes are randomly selected and the best weak rule is selected only among them. This modification significantly increases the effi­ciency of the learning process with no loss in ac­curacy (Escudero et al., 2000).

Three kinds of information have been used to de­scribe the examples and to train the classifiers. These features refer to local and topical contexts, and domain labels.

More particularly, let .. ws w-2 w~i w w+i w+2 ..." be the context of consecutive words around the word w to be disambiguated, and p±i (—3<i<3) be the part-of-speech tag of word w±il. Feature patterns referring to local context are the following 13:

where the last three correspond to collocations of two consecutive words.

The topical context is formed by ci,...,cm, which stand for the unordered set of open class words appearing in a medium-size 21-word win­dow centered around the target word.

The more innovative use of semantic domain in­formation is detailed in the next section,

We began this exercise by selecting a representa­tive sample, containing the most frequent words of the SENSEVAL-2 training data, and applying the LazyBoosting system straightforwardly on this sample. The results achieved after a 10-fold cross-validation procedure were very bad, mainly due to the fact that most of the words contain too many senses and too few examples per sense to induce reliable classifiers. With the aim of improving the performance of the learning algorithm, we have re­duced the number of senses by performing a hier­archical decomposition of the multiclass problem, following the idea of (Yarowsky, 2000).

Two different simplifications have been carried out. Firstly, multiword training examples have been processed separately. During training, multiwords have been saved into a separate file. At test time, all examples found in this multiword file are auto­matically tagged as multiwords. As an example, the word bar appears in the training set with 22 labels. But only the 10 senses showed in the left table of figure 1 are single words. The remaining 12 are multiwords which are considered unambigu­ous (Yarowsky, 1993).

*In this work, the English versions of MACO+ morphologi­cal analyzer and RELAX part-of-speech tagger have been used for tagging (Carmona et al., 1998).

Secondly, we have reduced the sense granularity, by hierarchically decomposing the learning process in two steps. In the first level, the learning algorithm is trained to classify between the labels correspond­ing to the WordNet semantic files, and, addition­ally the semantic-file labels with less than 10 train­ing examples are automatically discarded. If less than two senses remain, no training is performed and, simply, the Most-frequent-sense Classifier is applied.

As an example, for the word *bar\ in this first step the system is trained to classify between the labels of the top-right table of figure 1. Note that senses bar%l:04y bar%l:23 and bar%l:17 have been dropped out because there are not enough training examples.

In the second level, one classifier is trained for each of the resulting semantic-file labels of the first step in order to distinguish between their particular senses. Note that the same simplifying rules of the previous level are also applied. For instance, the bottom-right table of figure 1 shows the labels for bar%J:06, where 02:: has been rejected.

When classifying a new test example, the classi­fiers of the two levels are applied sequentially. That is, the semantic-file classifier is applied first. Then, depending on the semantic-file label output by this classifier, the appropriate 2nd level classifier is se­lected. The resulting label assigned to the test ex­ample is formed by the concatenation of the outputs of both previous levels.

In the official competition, labels 'IT and T' have been completely ignored. Thus, the examples labelled with these classes have not been considered during the training, and no test examples have been tagged with them.

Despite the simplifying assumptions and the loss of information, we have observed that all these changes together significantly improved the accu­racy on the training set. However, the components of the system were not tested separately due to the lack of time. Next section includes some evaluation about this issue.