MATRIX at the TREC 2005 robust track

In the TREC 2005 robust retrieval track, we tested our adaptive retrieval model that automatically switches between the 2-Poisson model/adaptive vector space model and our initial predictive probabilistic context-based model depending on some query characteristics. Our 2-Poisson model uses the BM11 term weighting scheme with passage retrieval and pseudo-relevance feedback. The context-based model incorporates the term locations in a document for calculating the term weights. By doing this, different term weights are assigned to the same query term depending on its context and location in the document. We also use WordNet in the term selection process when doing pseudo-relevance feedback. The performance of our model is comparable to the median among all participants in the robust track on the whole query set including the title, descriptive and long queries.