“A term is known by the company it keeps”: On Selecting a Good Expansion Set in Pseudo-Relevance Feedback

Abstract

It is well known that pseudo-relevance feedback (PRF) improves the retrieval performance of Information Retrieval (IR) systems in general. However, a recent study by Cao et al [3] has shown that a non-negligible fraction of expansion terms used by PRF algorithms are harmful to the retrieval. In other words, a PRF algorithm would be better off if it were to use only a subset of the feedback terms. The challenge then is to find a good expansion set from the set of all candidate expansion terms. A natural approach to solve the problem is to make term independence assumption and use one or more term selection criteria or a statistical classifier to identify good expansion terms independent of each other. In this work, we challenge this approach and show empirically that a feedback term is neither good nor bad in itself in general; the behavior of a term depends very much on other expansion terms. Our finding implies that a good expansion set can not be found by making term independence assumption in general. As a principled solution to the problem, we propose spectral partitioning of expansion terms using a specific term-term interaction matrix. We demonstrate on several test collections that expansion terms can be partitioned into two sets and the best of the two sets gives substantial improvements in retrieval performance over model-based feedback.