Reduction of the dimension of a document space using the fuzzified output of a Kohonen network

The vectors used in IR, whether to represent the documents or the terms, are high dimensional, and their dimensions increase as one approaches real problems. The algorithms used to manipulate them, however consume enormously increasing amounts of computational capacity as the said dimension grows. We used the Kohonen algorithm and a fuzzification module to perform a fuzzy clustering of the terms. The degrees of membership obtained were used to represent the terms and, by extension, the documents, yielding a smaller number of components but still endowed with meaning. To test the results, we use a topological classification of sets of transformed and untransformed vectors to check that the same structure underlies both.