Word Sense Disambiguation (WSD) is a partic­ular problem of computational linguistics which consists in determining the correct sense for a given ambiguous word. It is well-known that su­pervised algorithms have obtained the best re­sults in public evaluations, but their accuracy is close related with the amount of hand-tagged data available. The construction of that kind of training data is difficult for real applications. The unsupervised WSD overcomes this draw­back by using clustering algorithms which do not need training data in order to determine the possible sense for a given ambiguous word.

This paper describes a simple technique for unsupervised sense induction for ambiguous words. The approach is based on a self term ex­pansion technique which constructs a set of co-ocurrence terms and, thereafter, it uses this set to expand the target dataset. The implemented system was performed in the task "SemEval-2007 Task 2: Evaluating Word Sense Induc­tion and Discrimination Systems"(Agirre and A., 2007). The aim of the task was to per­mit a comparison across sense-induction and dis­crimination systems. Moreover, the comparison with other supervised and knowledge-based sys­tems may be also done, since the test corpus was borrowed from the well known "English lexical-sample" task in SemEval-2007, with the usual training + test split.

The self term expansion method consists in replacing terms of a document by a set of co-related terms. The goal is to improve natu­ral language processing tasks such as cluster­ing narrow-domain short texts. This process may be done by mean of different ways, of­ten just by using a knowledge database. In information retrieval, for instance, the expan­sion of query terms is a very investigated topic which has shown to improve results with respect to when query expansion is not employed (Qiu and Frei, 1993; Ruge, 1992; R.Baeza-Yates and Ribeiro-Neto, 1999; Grefenstette, 1994; Rijsbergen, 1979).

The availability of Machine Readable Re­sources (MRR) like "Dictionaries", "Thesauri" and "Lexicons" has allowed to apply term expansion to other fields of natural language pro­cessing like WSD. In (Banerjee and Pedersen, 2002) we may see the typical example of using a external knowledge database for determining the correct sense of a word given in some con­text. In this approach, every word close to the one we would like to determine its correct sense is expanded with its different senses by using the WordNet lexicon (Fellbaum, 1998). Then, an overlapping factor is calculated in order to determine the correct sense of the ambiguous word. Different other approaches have made use of a similar procedure. By using dictionaries, the proposals presented in (Lesk, 1986; Wilks et al., 1990; Nancy and Veronis, 1990) are the most sucessful in WSD. Yarowsky (Yarowsky, 1992) used instead thesauri for their experiments. Fi­nally, in (Sussna, 1993; Resnik, 1995; Baner-jee and Pedersen, 2002) the use of lexicons in WSD has been investigated. Although in some cases the knowledge resource seems not to be used strictly for term expansion, the aplication of co-occurrence terms is included in their algo­rithms. Like in information retrieval, the appli­cation of term expansion in WSD by using co-related terms has shown to improve the baseline results if we carefully select the external resource to use, with a priori knowledge of the domain and the broadness of the corpus (wide or nar­row domain). Evenmore, we have to be sure that the Lexical Data Base (LDB) has been suitable constructed. Due to the last facts, we consider that the use of a self automatically constructed LDB (using the same test corpora), may be of high benefit. This assumption is based on the intrinsic properties extracted from the corpus it­self. Our proposal is related somehow with the investigations presented in (Schütze, 1998) and (Purandare and Pedersen, 2004), where words are also expanded with co-ocurrence terms for word sense discrimination. The main difference consists in the use of the same corpora for con­structing the co-ocurrence list.

Following we describe the self term expan­sion method used and, thereafter, the results obtained in the task #2 of Semeval 2007 com­petition.

In literature, co-ocurrence terms is the most common technique used for automatic construc­tion of LDBs (Grefenstette, 1994; Frakes and Baeza-Yates, 1992). A simple approach may use n-grams, which allows to predict a word from previous words in a sample of text. The fre­quency of each n-gram is calculated and then filtered according to some threshold. The re­sulting n-grams constitutes a LDB which may be used as an "expansion dictionary" for each term.

On the other hand, an information theory-based co-ocurrence measure is discussed in (Manning and Schuütze, 2003). This measure is named pointwise Mutual Information (MI), and its applications for finding collocations are analysed by determining the co-ocurrence de­gree among two terms. This may be done by cal­culating the ratio between the number of times that both terms appear together (in the same context and not necessarily in the same order) and the product of the number of times that each term ocurrs alone. Given two terms X1and X2, the pointwise mutual information be­tween X1 and X2 can be calculated as follows:

The numerator could be modified in order to take into account only bigrams, as presented in (Pinto et al., 2006), where an improvement of clustering short texts in narrow domains has been obtained.

We have used the pointwise MI for obtaining a co-ocurrence list from the same target dataset. This list is then used to expand every term of the original data. Since the co-ocurrence formula captures relations between related terms, it is possible to see that the self term expansion mag­nifies less the noisy than the meaninful informa­tion. Therefore, the execution of the clustering algorithm in the expanded corpus should out­perform the one executed over the non-expanded data.

In order to fully appreciate the self term ex­pansion method, in Table 1 we show the coocurrence list for some words related with the verb "kill" of the test corpus. Since the MI is calculated after preprocessing the corpus, we present the stemmed version of the terms.

total average and two baselines proposed by the task organizers. The first baseline (Baseline1) assumes that each ambiguous word has only one sense, whereas the second baseline (Baseline2) is a random assignation of senses. We are ranked as third place and our results are better scored than the other teams except for the best team score. However, given the similar values with the "Baselinel", we may assume that that team presented one cluster per ambiguous word as its result as the Baselinel did; whereas we obtained 9.03 senses per ambiguous word in average.

In Table 3 we show our ranking and the super­vised recall obtained (UPV-SI). We again show the best and worst team recalls. The total av­erage and one baseline is also presented (the other baseline obtained the same Fscore). In this case, the baseline tags each test instance with the most frequent sense obtained in a train split. We are ranked again in third place and our score is slightly above the baseline.

The results show that the technique employed have learned, since our simple approach ob­tained a better performance than the baselines, especially the one that have chosen the most fre­quent sense as baseline.

For the task #2 of Semeval 2007, a set of 100 ambiguous words (35 nouns and 65 verbs) were provided. We preprocessed this original dataset by eliminating stopwords and then applying the Porter stemmer (Porter, 1980). Thereafter, when we used the pointwise MI, we determined that the single ocurrence of each term should be at least three (see (Manning and Schütze, 2003)), whereas the maximum separation among the two terms was five. Finally, we selected the unsupervised KStar clustering method (Shin and Han, 2003) for our experiments, defining the average of similarities among all the sentences for a given ambiguous word as the stop criterion for this clustering method. The input similarity matrix for the clustering method was calculated by using the Jaccard coefficient.

The task organizers decided to use two differ­ent measures for evaluating the runs submitted to the task. The first measure is called unsuper-vised one, and it is based on the Fscore measure. Whereas the second measure is called supervised recall. For further information on how these measures are calculated refer to (Agirre et al., 2006a; Agirre et al., 2006b). Since these mea­sures give conflicting information, two different evaluation results are reported in this paper.

In Table 2 we may see our ranking and the Fs-core measure obtained (UPV-SI). We also show the best and worst team Fscores; as well as the

We have reported the performance of a single approach based on self term expansion. The technique uses the pointwise mutual information for calculating a set of co-ocurrence terms which then are used to expand the original dataset. Once the expansion has been done, the unsu-pervised KStar clustering method was used to induce the sense for the different ocurrences of each ambiguous word. We obtained the third place in the two measures proposed in the task. We will further investigate whether an improve­ment may be obtained by applying term selec­tion methods to the expanded corpus.