Improving Information Retrieval Performance by Experience Reuse

The goal to build a knowledge base, making "permanent" the user evaluations experiences on search results, constitutes the main motivation of this paper. In information retrieval systems, an experience is a search instance, represented by a search context and a set of evaluated document results. In this paper, we propose a search context model, characterized by the user's profile and query. Similarity functions between users' profiles and search contexts are defined, on the basis of the proposed model. The reuse process consists in expanding the initial user query by the documents terms contained in the similar instances. In contrast with the Rocchio method, these documents are those validated beforehand by users, having "similar" profiles as the current user, and being in a "similar" search situation. Our proposition enables to reduce system interactions with the user during his/her search session. These features were carried out in the COSYDOR (Cooperative System for Document Retrieval) project, based on Inrermedia (Oracle 8i). Tests and evaluations are carried out using the test corpus of TREC (Text Retrieval Conference). The results show, for first search iterations, a significant improvement of performance compared to that of Internsedia. This work is carried out within the framework of a regional project, in which the sight deficient users constitute our application case. This research represents a considerable contribution for these users, considering their use difficulties of current information retrieval systems (too interactive systems, accessibility problems, etc)