UMass Robust 2005: Using Mixtures of Relevance Models for Query Expansion
Donald Metzler, Fernando Díaz, Trevor Strohman, W. Bruce Croft · 2005
This paper describes the UMass TREC 2005 Robust Track experiments. For the 2005 Robust Track, we explore whether or not term proximity information and advanced pseudo- relevance feedback methods can be used to achieve good effectiveness on a challenging query set. All experiments used the Indri search engine indexed the full AQUAINT collection of 1,033,461 documents, used a Porter Stemmer and a stopword list of 418 common terms. All runs are automatic. We use Metzler's dependence model formulation to exploit term proximity information, which been shown to significantly improve effectiveness over simple bag of words models. The Indri query language can be used to express dependence model queries. Results indicate that both term proximity and pseudo-relevance are highly effective.