Deterministic strategies for query (re)formulation in information retrieval

Abstract

Relevance feedback is a process whereby a user examines the documents selected by a retrieval system and provides feedback to the system as to their relevance. Such a feedback can then be used to formulate an optimal query with respect to the current information need of the user. This process of query (re)formulation can be based on probabilistic concepts, where Bayesian decision theory provides the framework for a decision rule, or ideas which instead employ deterministic strategies. The former class of techniques are limited by the fact that (strong) assumptions have to be made concerning the nature of the conditional orobabilitv density functions characterizing the data. In contrast, deterministic techniques, which do not require any explicit assumptions about the distribution of the various descriptor vaiues, can be adopted. Such methods would have the advantages of being "non-parametric" and robust (useful in a wide variety of contexts). A class of deterministic techniques has been advocated by Salton and the SMART project group at Cornell. However, that approach dws not obtain a new query that can be claimed to be optlmal in a certain sense. In this work, a deterministic method that obtains an optimal query according to a prescribed criterion is advanced. Furthermore, it will be demonstrated that such methods are applicable not only when there are two classes of relevance (relevant and non-relevant) but also when the feedback distinguishes documents according to several degrees of relevance.