Yarowsky (1993) states that most clues for the purpose of disambiguation are present in a micro-context of 5 or 4 words. SCT seems to be an adequate approach to handle short contexts. Moreover, SCT, which are binary decision trees, permit a simple interpretation of the results, by recovering the successive questions asked along each path from the root to a leaf. Kuhn and

De Mori (1995) have shown that these extracted rules correspond to regular expressions. However, this approach requires a certain amount of data in order for the trees to be grown with reliable questions in its nodes. Relying on previous work in this field (Loupy etal, 2000), the training corpus was used to build one tree for each word to be disambiguated. While growing the trees, the list of possible questions is built at each node, taking into consideration the position of an element of the context (lemma in this case). The Gini impurity G(X) (Breiman et al, 1984) is then computed (formula 1) for each question in the list, in order to extract the one which generates the highest decrease in impurity AGq (formula 2).

Where P(s/X) is the probability of sense s given population X,

Here Yesq and Noq correspond respectively to the population answering yes or no to the question pYes (respectively )isthe proportion of population T answering yes

(respectively no) to question q. A more detailed description of our approach to SCT can be found in Crestan and El-Bèze (2001).

The data had to be pre-processed before they could be used. Motivated by conclusions drawn from recent work (see for example Loupy and El-Bèze (2000)), the context was lemmatized, except for the word to be disambiguated. The determiners, possessive pronouns, adverbs and adjectives were removed, because they bring more noise to the tree growing process than they help capture relevant clues. However, some adjectives were preserved, when they were part of a compound noun, as in "short circuit'. For the part-of-speech (POS) tagging process and lemmatization process, the English Tree-Tagger (Schmid, 1994) was used.

Regarding previous work using SCT, the novelty of our approach consists in the introduction of rough semantic features into the context in order to increase the coverage of the trees. The process of tree growing can quickly suffer from lack of data. The ability of our system to view the context, not only as a succession of lemmas, but also as a multi-level view makes it more robust and reliable. We used the Semantic Classes (SC) proposed in Wordnet in order to improve the coverage of the trees. There are 26 SC associated with nouns (e.g. <noun.body> for body related nouns) and 15 SC associated with verbs (e.g. <verb.motion> for motion related verbs). Because most of the adjectives and adverbs were removed during the pre-processing phase and because they have only one or two possible SC, their respective SC are not employed.

During the SCT building process, there is now not just one question to ask at a given position in a training sentence, but n+J (where n is the number of possible SC associated with a lemma). For example, the sentence sample in figure I leads to 16 possible questions if considering SC, and only 7 questions if considering only lemmas.

Figure 1: Example of SC usage

SC are added regardless of the POS. In the example above, the term offer can only be a verb, but we still associate with it the classes _04 (noun.act) and 10 (noun.communication), which are associated with the noun-senses of offer. There are two reasons for this choice: First, in the case of erroneous POS tagging, we would not be able to characterize a sense using the adequate SC. Second, tests have shown that results obtained using POS related SC or all the SC are comparable. This last point could be explained by the aptitude of SCT to select the best questions. Therefore, SCT are able to partially disambiguate the local context at a coarse-grained sense level when enough data are available. Consequently, it seems useless to make assumptions about POS. Experiments carried on the SENSEVAL-1 data, has shown an improvement of about 2.5% on nouns and about 3% on verbs when using the Semantic Classes.

Experience has shown us that a window size of WS=3 is enough for disambiguation in many cases, but there are still numerous cases for which a larger window size is required. However, if a larger window size can provide more information for sense detection, it may also add more noise. In order to cope with this drawback, a similarity measure is employed (a technique usually applied in the field of document retrieval), as a ruler to decide which sense seems the more likely, considering the whole sentence (Figure 2). Firstly, three different Window Sizes (WS) are considered and run through the appropriate SCT process (trained on the same WS). Secondly, for each sense proposed by the SCT systems (Eh E2, and E3), a pseudo-document is built with the corresponding sentences from the training corpus. Then, a similarity measure as those used in document retrieval is computed between the test sentence (WS=\S\) and each of the pseudo-documents (i.e. senses). Finally, only the sense having the best score is kept. The similarity measure used here is the Cosine measure (Salton andMcGill, 1983).

The analysis of the results has shown that monitoring several SCT based views of the context by using the here above described technique leads to an average precision improvement of about 2%.

In a previous experiment (Loupy et al, 1998), HMM were applied directly to disambiguate senses at fine-grained level using a unisem-bisem model, after training on the SemCor (Miller et al, 1993). However, even if this method achieves correct results (72 % of correct assignation), it does not really improve

Mr Portillo, however, keeps his cabinet post.

Similarity Measure

over the unisem model. Therefore, it is recognized that there are not enough data to correctly learn the transitions between senses. On the other hand, an HMM unisem-bisem model brings a slight improvement as compared to unisem alone when applied to a coarser semantic level, that is SC (Loupy et al, 1998). We adopt the following two-step strategy: • Firstly, determine the SC associated with each word in the text (formula 4) where G is the set of possible coarse-grained semantic classes associated to the lemma L. Secondly, assign the most probable fine­grained sense according to the word and the previously retrieved SC (formula 5).

where S is the set of possible senses associated with the lemma L and its possible semantic classes G.

In view of our experience with the lexical sample task, we decided to take advantage of it. The most frequent words among those to be disambiguated in the all-words task and which were also present in the SENSEVAL-2 lexical sample task were extracted. For those words, the technique presented in Section 1 was applied. In this way, 4 verbs (call, develop, find and use) and 2 nouns (child and church) were disambiguated by the SCT-Cosine method, as described in Section 1.3.

Results and Conclusion

As mentioned in section 1, the scores for the second edition of the lexical sample task are much lower than for the first edition (about 20%). However, our system achieved satisfactory results comparing to other participants (see table 1) and even accessed the top-5 systems. The use of SC as a multi-level view of the context has generated significant improvements in the results. As well as, the combination of different window sizes using similarity measure on a larger context as a judge has shown noticeable improvements.

For the all-words task, our system has proven to be one of the bests, achieving an average precision/recall of 61.8%, and this, despite the absence of mapping between Wordnet 1.6 senses used for training purpose (SemCor) and Wordnet 1.7 senses used as test references.