As empirically demonstrated by the last SensEval and SemEval exercises, assigning the appropriate meaning to words in context has resisted all at­tempts to be successfully addressed. In fact, super­vised word-based WSD systems are very depen­dent of the corpora used for training and testing the system (Escudero et al., 2000). One possible reason could be the use of inappropriate level of abstraction.

Most supervised systems simply model each polysemous word as a classification problem where each class corresponds to a particular synset of the word. But, WordNet (WN) has been widely criticized for being a sense repository that often provides too fine-grained sense distinctions for higher level applications like Machine Translation or Question & Answering. In fact, WSD at this level of granularity has resisted all attempts of in­ferring robust broad-coverage models. It seems that many word-sense distinctions are too subtle to be captured by automatic systems with the cur­rent small volumes of word-sense annotated ex­amples.

Thus, some research has been focused on deriv­ing different word-sense groupings to overcome the fine-grained distinctions of WN (Hearst and Schütze, 1993), (Peters et al., 1998), (Mihalcea and Moldovan, 2001), (Agirre and LopezDeLa-Calle, 2003), (Navigli, 2006) and (Snow et al., 2007). That is, they provide methods for grouping senses of the same word, thus producing coarser word sense groupings for better disambiguation.

In contrast, some research have been focused on using predefined sets of sense-groupings for learn­ing class-based classifiers for WSD (Segond et al., 1997), (Ciaramita and Johnson, 2003), (Villarejo et al., 2005), (Curran, 2005), (Kohomban and Lee, 2005) and (Ciaramita and Altun, 2006). That is, grouping senses of different words into the same explicit and comprehensive semantic class. Most of the later approaches used the original Lexico­graphical Files of WN (more recently called Su-perSenses) as very coarse-grained sense distinc­tions.

We suspect that selecting the appropriate level of abstraction could be on between both levels. Thus, we use the semantic classes modeled by the Basic Level Concepts (BLC) (Izquierdo et al., 2007). Our previous research using BLC empiri­cally demonstrated that this automatically derived set of meanings groups senses into an adequate level of abstraction in order to perform class-based Word Sense Disambiguation (WSD) (Izquierdo et al., 2009). Now, we also show that class-based WSD allows to successfully incorporate monose­mous examples from the domain text. In fact, the robustness of our class-based WSD approach is shown by our system that just uses the Sem­Cor examples (SC). It performs without any kind of domain adaptation as the Most Frequent Sense (MFS) baseline.

This paper describes our participation in SemEval-2010 Task 17 (Agirre et al., 2010). In section 2 semantic classes used and selection al­gorithm used to obtain them automatically from WordNet are described. In section 3 the technique employed to extract monosemous examples from background data is described. Section 4 explains the general approach of our system, and the ex­periments designed, and finally, in section 5, the results and some analysis are shown.

The set of semantic classes used in this work are the Basic Level Concepts (BLC) (Izquierdo et al., 2007). These concepts are small sets of mean­ings representing the whole nominal and verbal part of WN. BLC can be obtained by a very simple method that uses basic structural WordNet proper­ties. In fact, the algorithm only considers the rel­ative number of relations of each synset along the hypernymy chain. The process follows a bottom-up approach using the chain of hypernymy rela­tions. For each synset in WN, the process selects as its BLC the first local maximum according to the relative number of relations. The local maxi­mum is the synset in the hypernymy chain having more relations than its immediate hyponym and immediate hypernym. For synsets having multi­ple hypernyms, the path having the local maxi­mum with higher number of relations is selected. Usually, this process finishes having a number of preliminary BLC. Figure 1 shows an example of selection of a BLC. The figure represents the hy­pernymy hierarchy of WordNet, with circles rep­resenting synsets, and links between them repre­senting hypernym relations. The algorithm selects the D synset as BLC for J, due to D is the first maximum in the hypernymy chain, according to the number of relations (F has 2 hyponyms, D has 3, and A has 2, so D is the first maximum).

Figure 1 : Example of BLC selection

Obviously, while ascending through this chain, more synsets are subsumed by each concept. The process finishes checking if the number of con­cepts subsumed by the preliminary list of BLC is higher than a certain threshold. For those BLC not representing enough concepts according to the threshold, the process selects the next local max­imum following the hypernymy hierarchy. Thus, depending on the type of relations considered to be counted and the threshold established, different sets of BLC can be easily obtained for each WN version.

We have selected the set which considers WN version 3.0, the total number of relations per synset, and a minimum threshold of 20 concepts to filter out not representative BLC (BLC-20). This set has shown to reach good performance on previ­ous SensEval and SemEval exercices (Izquierdo et al., 2009). There are 649 different BLC for nouns on WordNet 3.0, and 616 for verbs. Table 2 shows the three most frequent BLC per POS, with the number of synsets subsumed by each concept, and its WordNet gloss.

We did not applied any kind of specific domain adaptation technique to our class-based supervised system. In order to adapt our supervised system to the environmental domain we only increased the training data with new examples of the domain. To acquire these examples, we used the environmen­tal domain background documents provided by the organizers. Specifically, we used the 122 background documents. TreeTagger has been used to preprocess the documents, performing PoS tag­ging and lemmatization. Since the background documents are not semantically annotated, and our supervised system needs labeled data, we have se­lected only the monosemous words occurring in the documents. In this way, we have obtained au­tomatically a large set of examples annotated with BLC. Table 3 presents the total number of training examples extracted from SemCor (SC) and from the background documents (BG). As expected, by this method a large number of monosemous ex­amples can be obtained for nouns and verbs. Also as expected, verbs are much less productive than nouns. However, all these background examples correspond to a reduced set of 7,646 monosemous words.

Table 3 lists the ten most frequent monosemous nouns and verbs occurring in the background doc­uments. Note that all these examples are monose­mous according to BLC-20 semantic classes.

Our system applies a supervised machine learn­ing approach. We apply a feature extractor to represent the training examples of the examples acquired from SemCor and the background doc­uments. Then, a machine learning engine uses the annotated examples to train a set of classi­fiers. Support Vector Machines (SVM) have been proven to be robust and very competitive in many NLP tasks, and in WSD in particular (Marquez et al., 2006). We used the SVM-Light implementa­tion (Joachims, 1998).

We create a classifier for each semantic class. This approach has several advantages compared to word-based approach. The training data per clas­sifier is increased (we can use examples of dif­ferent target words for a single classifier, when­ever all examples belong to the same semantic class), the polysemy is reduced (some different word senses can be collapsed into the same se­mantic class), and, finally, semantic classes pro­vide higher levels of abstraction.

For each polysemous word occurring in the test corpus, we obtain its potential BLC-20 classes. Then, we only apply the classifiers corresponding to the BLC-20 classes of the polysemous word. Fi­nally, our system simply selects the BLC-20 class with the greater prediction.

In order to obtain the correct WordNet 3.0 synset required by the task, we apply a simple heuristic that has shown to be robust and accurate (Kohomban and Lee, 2005). Our classifiers ob­tain first the semantic class, and then, the synset of the first WordNet sense that fits with the semantic class is assigned to the word.

We selected a simple feature set widely used in many WSD systems. In particular, we use a win­dow of five tokens around the target word to ex­tract word forms, lemmas; bigrams and trigrams of word forms and lemmas; trigrams of PoS tags, and also the most frequent BLC-20 semantic class of the target word in the training corpus.

Our system is fully described in (Izquierdo et al., 2009). The novelty introduced here is the use of semantic classes to obtain monosemous exam­ples from the domain corpus.

Following the same framework (BLC-20 se­mantic architecture and basic set of features) we designed three runs, each one using a different training corpus.

• SC: only training examples extracted from SemCor

• BG: only monosemous examples extracted from the background data

• SCBG: training examples extracted from SemCor and monosemous background data

The first run shows the behavior of a supervised system trained on a general corpus, and tested in a specific domain. The second one analyzes the con­tribution of the monosemous examples extracted from the background data. Finally, the third run studies the robustness of the approach when com­bining the training examples from SemCor and from the background.

A total of 29 runs has been submitted for the En­glish All-words WSD on a Specific Domain. Ta­ble 5 shows the ranking results of our three runs with respect to the other participants. The figures for the first sense (1 sense) and random sense (Ran­dom) baselines are included.

In general, the results obtained are not very high. The best system only achieves a precision of 0.570, and the first sense baseline reaches a preci­sion of 0.505. This shows that the task is hard to solve, and the domain adaptation of WSD systems is not an easy task.

Interestingly, our worst result is obtained by the system using only the monosemous background examples (BG). This system ranks 23th with a Pre­cision and Recall of 0.380 (0.385 for nouns and 0.366 for verbs). The system using only SemCor (SC) ranks 6th with Precision and Recall of 0.505 (0.527 for nouns and 0.443 for verbs). This is also the performance of the first sense baseline. As ex­pected, the best result of our three runs is obtained when combining the examples from SemCor and the background (SCBG). This supervised system obtains the 5th position with a Precision and Re­call of 0.513 (0.534 for nouns, 0.454 for verbs) which is slightly above the baseline.

Possibly, the reason of low performance of the BG system is the high correlation between the fea­tures of the target word and its semantic class. In this case, these features correspond to the monose­mous word while when testing corresponds to the target word. However, it also seems that class-based systems are robust enough to incorporate large sets of monosemous examples from the do­main text. In fact, to our knowledge, this is the first time that a supervised WSD algorithm have been successfully adapted to an specific domain. Fur­thermore, our system trained only on SemCor also achieves a good performance, reaching the first sense baseline, showing that class-based WSD ap­proaches seem to be robust to domain variations.

This paper has been supported by the Euro­pean Union under the project KYOTO (FP7 ICT-211423), the Valencian Region Government un­der PROMETEO project for excellence groups and the Spanish Government under the projects

KN0W2 (TIN2009-14715-C04-04) and TEXT-MESS-2 (TIN2009-13391-C04-04).