Improving the Computation of Relevance Degrees in a Linguistic Information Retrieval System