There exists significant evidence that word sense disambiguation is important for a variety of nat­ural language processing tasks: machine transla­tion, information retrieval, grammatical analysis, speech and text processing (Veronis, 2004). How­ever, the "fixed-list" of senses paradigm, where the senses of a target word is a closed list of defini­tions coming from a standard dictionary (Agirre et al., 2006), was long ago abandoned. The rea­son is that sense lists, such as WordNet (Miller, 1995), miss many senses, especially domain-specific ones (Pantel and Lin, 2002). The miss­ing concepts are not recognised. Moreover, senses cannot be easily related to their use in context.

Word sense induction methods can be divided into vector-space models and graph based ones. In a vector-space model, each context of a target word is represented as a feature vector, e.g. fre­quency of cooccurring words (Katz and Gies-brecht, 2006). Context vectors are clustered and the resulting clusters represent the induced senses.

Recently, graph-based methods have been em­ployed for word sense induction (Agirre and Soroa, 2007). Typically, graph-based methods represent each context word of the target word as a vertex. Two vertices are connected via an edge if they cooccur in one or more instances. Once the cooccurrence graph has been constructed, dif­ferent graph clustering algorithms are applied to partition the graph. Each cluster (partition) con­sists of a set of words that are semantically related to the particular sense (Veronis, 2004). The poten­tial advantage of graph-based methods is that they can combine both local and global cooccurrence information (Agirre et al., 2006).

Klapaftis and Manandhar (2008) presented a graph-based approach that represents pairs of words as vertices instead of single words. They claimed that single words might appear with more than one senses of the target word, while they hy­pothesize that a pair of words is unambiguous. Hard-clustering the graph will potentially identify less conflating senses of the target word.

In this paper, we relax the above hypothesis be­cause in some cases a single word is unambiguous. We present a method that generates two-word ver­tices only when a single word vertex is unambigu­ous. If the word is judged as unambiguous, then it is represented as a single-word vertex. Otherwise, it is represented as a pair-of-words vertex.

The approach of Klapaftis and Manandhar (2008) achieved good results in both evaluation settings of the SemEval-2007 task. A test in­stance is disambiguated towards one of the in­duced senses if one or more pairs of words rep­resenting that sense cooccur in the test instance. This creates a sparsity problem, because a cooc­currence of two words is generally less likely than the occurrence of a single word. We expect our ap­proach to address the data sparsity problem with­out conflating the induced senses.

In this section we present our word sense in­duction and disambiguation algorithms. Figure 1 shows an example showing how the sense in­duction algorithm works: The left side of part I shows the context nouns of four snippets con­taining the target noun "chip". The most rele­vant of these nouns are represented as single word vertices (part II). Note that "customer" was not judged to be significantly relevant. In addition, the system introduced several vertices represent­ing pairs of nouns. For example, note the vertex "company.potato". The set of sentences contain­ing the context word "company" was judged as very different from the set of sentences contain­ing "company" and "potato". Thus, our system hypothesizes that probably "company" and "com­pany .potato" are relevant to different senses of "chip", and allows them to be clustered accord­ingly. Vertices whose content nouns or pairs of nouns cooccur in some snippet are connected with an edge (part III and right side of part I). Edge weights depend upon the conditional probabilities of the occurrence frequencies of the vertex con­tents in a large corpus, e.g. 102,6 in part III. Hard-clustering the graph produces the induced senses of "chip": (a) potato crisp, and (b) microchip.

In the following subsections, the system is de­scribed in detail. Figure 2 shows a block diagram overview of the sense induction system. It consists of three main components: (a) corpus preprocess­ing, (b) graph construction, and (c) clustering.

In a number of different stages, the system uses a reference corpus to count occurrences of word or word pairs. It is chosen to be large because fre­quencies of words in a large corpus are more sig­nificant statistically. Ideally we would use the web or another large repository, but for the purposes of the SemEval-2010 task we used the union of all snippets of all target words.

The induced senses are used to sense-tag each test instance of the target word (snippet). Given a snip­pet, each induced cluster is assigned a score equal to the number of its vertex contents (single or pairs of words) occurring in the snippet. The instance is assigned to the sense with the highest score or with equal weights to all highest scoring senses.

The algorithm depends upon 4 parameters: Pi thresholds frequencies and P3 collocation weights. P2 is the LL threshold and P4 the similarity thresh­old for discarding pair-of-nouns vertices.

We chose Pi G {5,10,15}, P2 G {2,3,4,5,10,15,25,35}, P3 G {0.2,0.3,0.4} andP4 G {0.2,0.4,0.6,0.8}. The parameter tun­ing was done using the trial data of the SemEval-2010 task and on the noun data of correspond­ing SemEval-2007 task. Parameters were tuned by choosing the maximum supervised recall. For both data sets, the chosen parameter values were Pi - 10, P3 ~ 0.4 and P4 ~ 0.8. Due to the size difference of the datasets, for the Semeval-2010 trial data P2 ~ 3, while for the SemEval-2007 noun data P2 ~ 10. The latter was adopted because the size of training data was announced to be large. We induced senses on the training data and then disambiguated the test data instances.

Three different measures, V-Measure, F-Score, and supervised recall on word sense disambigua­tion task, were used for evaluation. V-Measure and F-Score are unsupervised. Supervised recall was measured on two different data splits. Table 1 shows the performance of our system, UoY, for all measures and in comparison with the best, worst and average performing system and the random and most frequent sense (MFS) baselines. Results are shown for all words, and nouns and verbs only.

We presented a graph-based approach for word sense induction and disambiguation. Our ap­proach represents as a graph vertex an unambigu­ous unit: (a) a single word, if it is judged as unam­biguous, or (b) a pair of words, otherwise. Graph edges model the cooccurrences of the content of the vertices that they join. Hard-clustering the graph induces a set of senses. To disambiguate a test instance, we assign it to the induced sense whose vertices contents occur mostly in the in­stance. Results show that our system achieves very high recall and V-measure performance, higher than both baselines. It achieves low F-Scores due to the large number of induced senses.