Does pseudo-relevance feedback improve distributed information retrieval systems?

This paper presents a thorough analysis of the capabilities of the pseudo-relevance feedback (PRF) technique applied to distributed information retrieval (DIR). Previous studies have researched the application of PRF to improve the selection process of the best set of collections from a ranked list. This work emphasizes the effectiveness of PRF applied to the collection fusion problem. Usually, DIR systems apply PRF in the same way as traditional Information Retrieval systems. For each collection, local results are improved through PRF. A first question which arises is whether this local improvement is preserved in the final result. In addition, DIR systems merge the documents of rankings that are returned from a set of collections. Since a new global list of documents is available, we could use that list to apply PRF again, but on global level rather than on a local level. In order to apply global PRF, we have developed a merging approach called two-step RSV. Finally, we describe a number of experiments involving the two levels, local and global, of application of the PRF techniques.