Comparison of different Collection Fusion Models in Distributed Information Retrieval

Distributed information retrieval comes into play when a user wants to get information from di erent sources in parallel. One of the challenges of this topic is the Collection Fusion problem: The distinct result lists of the underlying information retrieval systems (IR) have to be fused to give a global relevance-ranked result list according to the user's information need. In this paper several Collection Fusion models were scrutinized which obey certain restrictions. They should not use more parameters than are given by common collections of distributed digital libraries so that they could be deployed in a real-life distributed retrieval environment. The selected models are evaluated in a test system with an interface to the IR systems of Managing Gigabytes (MG) and Oracle. The quantitative analysis of the models was performed using test queries, test documents and relevance evaluations from TREC. Two test runs were executed, one with all four collections stored in MG, another with three collections in MG and one in Oracle. The results of these test runs are discussed with respect to the test setup.