Word Sense Disambiguation (WSD) is a difficult Natural Language Processing task which requires that for every content word (noun, adjective, verb or adverb) the appropriate meaning is automatically selected from the available sense inventory. Tradi­tionally, the WSD algorithms are divided into two rough classes: supervised and unsupervised. The supervised paradigm relies on sense annotated cor­pora, with the assumption that neighbouring disam-biguate words provide a strongly discriminating and generalizable context representation for the meaning of a target word. Obviously, this approach suffers from the knowledge acquisition bottleneck in that there will never be enough training data to ensure a scalable result of such algorithms. The unsuper-vised alternative to WSD tries to alleviate the burden of manually sense tagging the corpora, by employ­ing algorithms that use different knowledge sources to determine the correct meaning in context. In fact, the "knowledge source usage" is another way to dis­tinguish among the WSD methods. Such methods call upon further processing of the text to be dis-ambiguated such as parsing and/or use handcrafted, semantically rich sense inventories such as Word­Net (Fellbaum, 1998). WSD methods in this cate­gory range from the very simple ranking based on counting the number of words occurring in both the target word's context and its sense definitions in a reference dictionary (Lesk, 1986) to the more elabo­rated approaches using the semantic lexicon's tax­onomies, (shallow) parsing, collocation discovery etc. (Stevenson and Wilks, 2001).

One of the central issues of any WSD implemen­tation is given by the context representation. The standard principle that is applied when trying to dis-ambiguate the meaning of a word is that the same word in similar contexts should have the same mean­ing. By and large, the context of a target word is ma­terialized by a collection of features among which are: the collocates of the target word, the part-of-speech (POS) of the target word, ±k words sur­rounding the target word and/or their POSes and so on. More often than not, the contexts similarity is es­timated by the distance in the feature vector space. Lin (1997) defines the local context of a target word by the collection of syntactic dependencies in which the word takes part. According to this notion of context, Lin assumes that two different words are likely to have similar meanings if they occur in identical local contexts.

What we will attempt here is to combine the two views of context similarity/identity versus meaning similarity/identity by using a dependency-like repre­sentation of the context as a lexical attraction model. More specifically, we will not consider any feature of the context and will try to maximize a meaning at­traction function over all linked words of a sentence. In section 2 we will describe SynWSD, an unsuper-vised, knowledge-based WSD algorithm and in sec­tions 3 and 4 we will present the application of Syn­WSD to two of SemEval-2007 "all words" tasks: English Coarse-Grained and English Fine-Grained. Finally, with section 5 we will conclude the article.

The syntactic context representation is not new in the realm of WSD algorithms. For instance, Lin (1997) used the dependency relations of the target word to specify its context and Stetina (1998) ex­tracted head-modifier relations to obtain the context pairs for each word of interest from a constituents tree. The syntactic representation of the context of a target word has one main advantage over the collec­tion of features method: the target word is related only with the relevant word(s) in its window and not with all the words and thus, many noisy cooc­currences are eliminated. Mel'cuk (1988) further strengthens the intuition of a syntactic context rep­resentation with his Meaning Text Model in which there is a deterministic translation from the surface syntactic dependency realization of the sentence to its deep syntactic one and therefore to the semantic representation.

To use a syntactic analysis as a context representa­tion, one needs a parser which will supply the WSD algorithm with the required analysis. Because we have intended to develop a language independent WSD algorithm and because there is no available, reliable dependency parser for Romanian, we have backed off to a simpler, easier to obtain dependency­like representation of a sentence: a slightly modified version of the lexical attraction models of (Yuret, 1998).

3. Log-Likelihood, ll(mi , m2) which is com­puted as in (Moore, 2004).

After the Viterbi path (best path) has been calcu­lated, every state (meaning) from V[i] (0 < i < | V| ) along this path is added to a final D vector. When the PWN20 sense inventory is used, the reverse of the generalization procedure is applied to each meaning recorded in D, thus coming back to the meanings of the words of S. Please note that an entry in D may contain more than one meaning especially in the case of PWN20 meanings for which there was not enough training data.

Coarse-grained English All-Words

LexPar and SynWSD were trained on an 1 million words corpus comprising the George Orwell's 1984 novel and the SemCor corpus (Miller et al., 1993).

Both texts have been POS-tagged (with MULTEXT-

East compliant POS tags) and lemmatized and the result was carefully checked by human judges to en­sure a correct annotation.

SynWSD was run with all the meaning attraction functions (dice, mi and ll) for all the sense in­ventories (PWN20, SUMO categories and IRST do­mains) and a combined result was submitted to the task organizers. The combined result was prepared in the following way:

1. for each sense inventory and for each token identifier, get the union of the meanings for each run (dice, mi and ll);

2. for each token identifier with its three union sets of PWN20 meanings, SUMO categories and IRST domains:

(a) for each PWN20 meaning mi in the union, if there is a SUMO category that maps onto it, increase mi's weight by 1;

(b) for each PWN20 meaning mi in the union, if there is a IRST domain that maps onto it, increase mi's weight by 1;

(c) from the set of weighted PWN20 mean­ings, select the subset C that best over­laps with a cluster. That is, the intersec­tion between the subset and the cluster has a maximal number of meanings for which the sum of weights is also the greatest; (d) output the lowest numbered meaning in C.

With this combination, the official F-measure of SynWSD is 0.65712 which places it into the 11thposition out of 16 competing systems.

Another possible combination is that of the inter­section which is obtained with the exact same steps as above, replacing the union operation with the in­tersection. When the PWN20 meanings set is void, we can make use of the most frequent sense (MFS) backoff strategy thus selecting the MFS of the cur­rent test word from PWN20. Working with the of­ficial key file and scoring software, the intersection combination with MFS backoff gives an F-measure of 0.78713 corresponding to the 6th best result. The same combination method but without MFS backoff achieves a precision of 0.80559 but at the cost of a very low F-measure (0.41492).

All-Words

For this task, LexPar and SynWSD were further trained on a 12 million POS tagged and lemmatized balanced corpus. The run that was submitted was the intersection combination with the MFS backoff strategy which obtained an F-measure of 0.527. This score puts our algorithm on the 8th position out of 14 competing systems. For the union combinator (the MFS backoff strategy is not applicable), the F-measure decreases to 0.445 (10th place). Finally, if we train SynWSD only on corpora from task#7, the union combinator leads to an F-measure of 0. 344.

SynWSD is a knowledge-based, unsupervised WSD algorithm that uses a dependency-like analysis of a sentence as a uniform context representation. It is a language independent algorithm that doesn't require any feature selection.

Our system can be improved in several ways. First, one can modify the generalization procedure in the case of PWN20 meanings in the sense of se­lecting a fixed set of top level hypernyms. The size of this set will directly affect the quality of meaning co-occurrence frequencies. Second, one may study the effect of a proper dependency parsing on the re­sults of the disambiguation process including here making use of the syntactic relations names and ori­entation.

Even if SynWSD rankings are not the best avail­able, we believe that the unsupervised approach to the WSD problem combined with different knowl­edge sources represents the future of these systems even if, at least during the last semantic evalua­tion exercise SensEval-3, the supervised systems achieved top rankings.