An Information Retrieval (IR) system ranks documents according to their predicted relevance to a formulated query. The prediction depends on the ranking algorithm adopted and on the assumptions about relevance underlying the algorithm. The main assumption is that there is one user, one information need for each query, one location where the user is, and no temporal dimension. But this assumption is unlikely: relevance is context-dependent. Exploiting the context in a way that does not require an high user effort may be effective in IR as suggested for example by Implicit Relevance Feedback techniques. The high number of factors to be considered by these techniques suggests the adoption of a theoretical framework which naturally incorporates multiple sources of evidence. Moreover, the information provided by the context might be a useful source of evidence in order to personalize the results returned to the user. Indeed, the information need arises and evolves in the present and past context of the user. Since the context changes in time, modeling the way in which the context evolves might contribute to achieve personalization. Starting from some recent reconsiderations of the geometry underlying IR and their contribution to modeling context, in this paper some issues which will be the starting point for my PhD research activity are discussed.
Author and article information
Emanuele Di Buccio
Department of Information Engineering
University of Padova
Via Gradenigo 6/B, 35131 Padova, ITALY