International Journal of Web Applications Rabeb Mbarek, Mohamed Tmar, Mohand Boughanem 8 1 2016 Pseudo-Relevance Feedback assumes that the top-ranked k documents of the initial retrieval are relevant, and then terms of these documents are used to re-weight the terms of the initial query (add new terms and/or change the weights of existing terms in the query). In this paper, we propose a new approach for query expansion for ad hoc search, by using an absorbing document which is the cross product of irrelevant documents. This document will be orthogonal to irrelevant ones. We show how this absorbing document can extract better expansion terms from the top-ranked k documents. The experiments show that our approach gives improvements for both collections, TREC-7 and TREC-8, over known models.