Incorporating Relevance and Pseudo-relevance Feedback in the Markov Random Field Model

We present a new document retrieval approach combining relevance feedback, pseudo-relevance feedback, and Markov random field modeling of term interaction. Overall effectiveness of our combined model and the relative contribution from each component is evaluated on the GOV2 webpage collection. Given 0-5 feedback documents, we find each component contributes unique value to the overall ensemble, achieving significant improvement individually and in combination. Comparative evaluation in the 2008 TREC Relevance Feedback track further shows our complete system typically performs as well or better than peer systems.