Iterative Estimation of Document Relevance Score for Pseudo-Relevance Feedback

Abstract

Pseudo-relevance feedback (PRF) is an effective technique for improving the retrieval performance through updating the query model using the top retrieved documents. Previous work shows that estimating the effectiveness of feedback documents can substantially affect the PRF performance. Following the recent studies on theoretical analysis of PRF models, in this paper, we introduce a new constraint which states that the documents containing more informative terms for PRF should have higher relevance scores. Furthermore, we provide a general iterative algorithm that can be applied to any PRF model to ensure the satisfaction of the proposed constraint. In this regard, the algorithm computes the feedback weight of terms and the relevance score of feedback documents, simultaneously. To study the effectiveness of the proposed algorithm, we modify the log-logistic feedback model, a state-of-the-art PRF model, as a case study. Our experiments on three TREC collections demonstrate that the modified log-logistic significantly outperforms competitive baselines, with up to \(12\%\) MAP improvement over the original log-logistic model.