Analysis of combining multiple query representations with varying lengths in a single engine

Abstract

We examine the issues of combining multiple query representations in a single IR engine. Differing query representations are used to retrieve different documents. Thus, when combining their results, improvements are observed in effectiveness. We use multiple TREC query representations (title, description and narrative) as a basis for experimentation. We examine several combination approaches presented in the literature (vector addition, CombSUM and CombMNZ) and present a new combination approach using query vector length normalization. We examine two query representation combination approaches (title + description and title + narrative) for 150 queries from TREC 6, 7 and 8 topics. Our QLN (Query Length Normalization) technique outperformed vector addition and data fusion approaches by as much as 32% and was on average 24% better. Additionally, QLN always outperformed the single best query representation in terms of effectiveness.