In this section, we describe how we formulate the WSD problem in game theoretic terms, extending our previous works [TRI 15 a, TRI 15b, TRI 17] with the use of transductive learning principles. In the next section, we describe how the interaction graph is constructed, in section 6.4.2 we describe how the strategy space of the players is initialized, then we introduce the payoff matrices of the games and the system dynamics.

Graph construction

The graph is constructed selecting from a text all the words that have an entry in a knowledge base such as WordNet [FEL 98], denoted by I = {1,...,N}, where N is the number of target words. From I, we constructed the NxN similarity matrix W where each element w_{tj} is the

similarity among words i and j. W can be exploited as a useful tool for graph-based algorithms, since it is treatable as a weighted adjacency matrix of a weighted graph.

A crucial factor for the graph construction is the choice of the similarity measure, sim(•,• R to weight the edges of the graph. For our

experiments, we used the Dice coefficient [DIC 45], since it has performed well on different datasets [TRI 15a, TRI 17]. This measure determines the strength of co-occurrence between two words and is computed as follows:

where c(i) is the total number of occurrences of word i in a large corpus

and c(i, j) is the co-occurrence of the words i and j in the same corpus.

This formulation is particularly useful to decrease the ranking of words that tend to co-occur frequently with many other words. For the experiments in this work, we used as corpus the British National Corpus [LEE 92]. The similarity graph W encodes the information of how two target words are similar, in a distributional semantics perspective [HAR 54].