Imaging and Information Retrieval: Variations on a Theme∗
Fábio Crestani, Fabrizio Sebastiani, C. J. van Rijsbergen · 2000
(Standard) Imaging is a method for the revision of probability functions originally proposed in the philosophy of language as a semantics for conditional logic. Recently, Standard Imaging and a variant of it, called General Imaging, have successfully been applied to the estimation of the probability of relevance in Information Retrieval (IR) by Crestani and van Rijsbergen. The experimental results they have obtained show that these methods perform better than a number of more established approaches, such as retrieval by joint or conditional probability. In this paper we report preliminary results on three main generalisations of these methods and their application to IR. These generalisations are orthogonal (and they may thus be freely combined), as they address three orthogonal issues in the probability kinematics of Imaging. The first generalisation, that we call Proportional Imaging, is a variation of General Imaging that is better suited to those cases in which similarity between “possible worlds” has a quantitative nature; this is indeed the case in the application of Imaging methods to IR, where keywords play the role of possible worlds. The idea that underlies Proportional Imaging is that the probability of an A-world w should be distributed to all A-worlds wi in a way that is proportional to the degree of similarity between w and the wi’s. ∗This work has been carried out in the context of the project FERMI 8134 “Formalisation and Experimentation in the Retrieval of Multimedia Information”, funded by the European Community under the ESPRIT Basic Research scheme.