Report on the TREC-4 Experiment: Combining Probabilistic and Vector-Space Schemes

This paper describes and evaluate a retrieval scheme combining the OKAPI probabilistic retrieval model with various vector-space schemes. In this study, each retrieval strategy represents both queries and documents using the same set of single terms ; however they weight them differently. To compine these search schemes, we do not apply a given combination operator on the retrieval status nor the rank of each retrieved record (e.g., sum, average, max., etc.). We think that each retrieval strategy may perform well for a set of queries and poorly for other requests. Thus, based on a given query's statistical characteristics, our search model first select the more appropriate retrieval scheme and then retrieves information based on the selected search mechanism. Since the selection procedure is done before any search operation, our approach has the advantage of limiting the search time to one retrieval algorithm instead of retrieving itmes using various retrieval schemes, and then combining the given results. In particular, this study adesses the following questions : can the statistical characteristics of a query be good predicators in an automatic selection procedure ; faced with the relativity high retrieval effectiveness achieved by the OKAPI model, can various vector-space schemes further improve the retrieval performance of the OKAPI approach, and can the learning results obtained with one tested collection be valid for anther corpus