System fusion for improving performance in information retrieval systems

Abstract

The fusion of various retrieval strategies has long been suggested as a means of improving retrieval effectiveness. To date, testing of fusion has been done by combining result sets from widely disparate approaches that include an uncontrolled mixture of retrieval strategies and utilities. To isolate the effect of fusion on individual retrieval models, we have implemented probabilistic, vector-space and weighted Boolean models and tested the effect of fusion on these strategies in a systematic fashion. We also tested the effect of fusion on various query representations and have shown up to a 12% improvement in the average precision.