XML files made available by the task organizers have been processed with an XML parser. As for lemmatization and part-of-speech tagging the Tree Tagger, developed at the University of Stuttgart (Schmid, 1994) has been used, both for English and Italian. The WordNet mor­phological analyser has also been used in order to resolve ambiguities and lemmatization mis­takes. After this process texts are represented as vectors of triples: word lemma, WordNet part of speech and position in the text.

The basic procedure in domain driven disam­biguation is a function that, given a lemma L, associates a score to each domain defined for that lemma in Wordnet Domains. Such a score is the relative frequency of the domain in L, computed on the basis of the occurrences of the synsets of L in Semcor. Semcor occurrences for synsets with multiple domain annotations are repeated for each domain (e.g. if a synset has 2 occurrences and 2 labels it is counted as having 4 occurrences), while synsets with 0 occurrences are counted as 0.5. As an example, consider the lemma "bank" in Table 1. According to our scoring method, it has 57 total occurrences in Semcor. The Geology domain collects contri­butions from senses 2 and 7, for a total of 16 occurrences in Semcor, which corresponds to a frequency .28 (i.e. fq[DGeoiogy](bank) = 0.28).

The data structure that collects domain infor­mation is called a Domain Vector (DV). Intu­itively a DV represents the domains that are relevant for a certain lemma (or word sense) in a certain context. We have considered three kinds of DV's: a DV for a lemma L within a context C (DV£), for the case of test data; a DV for a synset S of a lemma L within a context C (DV$), for the case of training data; and a DV for a synset S of a lemma L in WordNet (DVs), which is used when no training data are available.

DV for a lemma in context (DV£). Given a set of domains D\ ... Z?n, a DV for a lemma L in a position K within a text represents the rele­vance of those domains for that lemma, i.e. each component DVi[i] gives the degree of relevance of the domain D{ for the lemma L. Given a con­text of ±C words before and after the lemma L in the position if, each component of the do­main vector is defined with the following for­mula:

where gauss is the normal distribution cen­tered on the position K. In the current al­gorithms C is set to 50 because our experi­ments with Semcor showed that the precision decreases below that thresold.

Intuitively, the above formula takes into ac­count the contribution of the lemmas in the con­text C to the sense of the target lemma L. In addition a DV actually selects a set of relevant domains rather than just one domain.

DV for a synset in context (DVg) In case a training corpus is available where lemmas are annotated with the correct sense, Domain Vec­tors are computed with the formula above. In­stead of considering a lemma in a position K within a text, we have a sense for that lemma (i.e. a synset). DV$ represents a "typical" vec­tor for a sense S of a lemma L.

DV for a synset without context (DVs) When a training corpus is not available (as for the 'all words' task), a simpler way to build a DV for a certain synset is to compute it with respect to WordNet Domains. Given a synset S in WordNet Domains, the domain vector DVs is a vector that has l's in the position of its domain(s) and 0's otherwise. A more accurate DV could be obtained by considering contextual information such as the synset gloss.

To disambiguate a lemma L (i.e. the target lemma) in a text, first its DV£ is computed. The next step consists of comparing the DV of the target lemma L with the domain vec­tors for each sense of L derived either from the training set, when available, or from WordNet Domains, when training data are not available. The sense vector DVs which maximizes the sim­ilarity is selected as the appropriate sense of L in that text. The similarity between two DV's is calculated with the standard scalar product: DVx • DV2 = Et DVX[%\ * DVx[i\.

From the point of view of domain driven dis­ambiguation, the 'lexical sample' task was in­herently more difficult than the 'all words' task for two reasons. First the context provided for disambiguation was generally shorter than the 100 words we used to build a semantic vector. Second, the high number of "factotum" words to be disambiguated resulted in a recall even lower (i.e. about 0.24) than for the 'all words' task. The improvement of performance from the fine grained to the coarse grained evalua­tion seems to confirm that, at least to some de­gree, domain clustering corresponds to the sense grouping created by the task organizers.

The low results obtained for the Italian 'lexical sample' task may have several causes. First of all, the absence of a training set and the ab­sence of any tagged text for Italian forced us to use a similarity function (see 3.4) trained to an English corpus. This was possible be­cause we maintained the mappings between the English and the Italian wordnets. However, these multiple mappings (i.e. from Word-NetI.6 to WordNet1.5 and then to the Ital­ian synsets through the equivalence links) are another source of possible errors, especially con­cerning the domain information associated with Italian synsets.

We have described an approach to word sense disambiguation based on domain information. The underlying assumption is that domains con­stitute a fundamental feature of text coherence. As a consequence, word senses occurring in a co­herent portion of text tend to maximize domain similarity. Three systems have been imple­mented, integrating knowledge-based and sta­tistical techniques, for the three task we partic­ipated in. As for lexical resources, the systems make use of WordNet Domains, an extension of English Wordnet 1.6, where synsets have been annotated with domain information. The dis­ambiguation algorithm is based on domain vec­tors that collect contextual information with re­spect to the target word. At this moment only domain information is used in our system. A promising research direction is the use of local information (e.g. syntax) to capture word be­haviors that lie outside the domain approach.