Recent Results on Fusion of Effective Retrieval Strategies in the Same Information Retrieval System1

Prior efforts have shown that data fusion techniques can be used to improve retrieval effectiveness under certain situations. Although the precise conditions necessary for fusion to improve retrieval have not been identified, it is widely believed that as long as component result sets used in fusion have higher relevant overlap than non-relevant overlap, improvements due to fusion can be observed. We show that this is not the case when systemic differences are held constant and different highly effective document retrieval strategies are fused within the same information retrieval system. Furthermore, our experiments have shown that the ratio of relevant to non-relevant overlap is a poor indicator of the likelihood of fusion’s effectiveness, and we propose an alternate hypothesis of what needs to happen in order for fusion to improve retrieval when standard voting/merging algorithms such as CombMNZ are employed.