Information Retrieval Approach based on Recursive Query Shifting

Abstract

In this paper, we introduce an information retrieval approach based on a Recursive Query Shifting. The idea is coming from the observation that precision value, in Precision/Recall curves, begins very high before downgrading and so no matter what the considered parameters: feature, matching measure, threshold, etc. In other words, the first results are commonly better than the other results coming later. Considering then the first returned result, recursively, as a query seems to contribute very well for improving the system accuracy. The idea is adopted firstly in the case of content based image retrieval and generalized for the case of documentary retrieval. The simple specificity is that the first returned image is considered as the new query, in the case of image retrieval, while the new query is extracted from the first returned document, in the case of text retrieval, for the reason that there is no documentary retrieval system with an entire document as a query. The proposed approach falls in the purview of mechanisms for pseudo relevance feedback with long- term learning helping to improve the interrogation protocol. It consists to reformulate the original submitted query recursively as an attempt for being closer to the user requirement as well as to the other collection elements. The approach is materialized into two algorithms namely QRM1 and QRM2. The experiments conducted on returned Google Scholar results as a text collection and COREL-1K images benchmark yield very promising results.