Smoothing functions for automatic relevance feedback in information retrieval

Automatic relevance feedback, or query expansion, is a common technique in information retrieval systems; it uses the initial ordered list of documents at the system's output to improve its performance in future searches. The drawback of this improvement is an increase, sometimes excessive, of the computational burden. We propose a new form of relevance feedback based on the cluster hypothesis, which does not require cluster calculation. After explaining the general method, we describe a simplification applicable to vector based systems, with the advantages of query expansion methods but without having to carry out long calculations to establish the weights of the terms of the new query. The results of the first tests effected on the system, although non-conclusive, are highly promising. Some guidelines to improve the system are given.