Using “ Model ” Pseudo-Documents to Improve Searching-as-Learning and Search over Sessions

Traditional information retrieval systems are evaluated with the assumption that each query is independent. However, during their interactions with search engines, many users find themselves reformulating their queries in order to satisfy their information need. In this paper, we treat “Searching as Learning” as similar to Search over Sessions problem and build a system to improve search over sessions’ utility by first creating, for each session, pseudo-documents that we hypothesize to be model documents that relevant documents should look like. In a second step, we promote documents that bear more similarity to our model documents. We experiment with a few simple heuristics for obtaining a model document by concatenating different documents that were listed as relevant in previous interactions in a session. We show that using this simple re-ranking method can provide up to 42.59% increase over the nDCG@10 baseline for the 2013 TREC Session track data and up to 26% increase over the α-nDCG@10 baseline for the TREC 2011 Session track data.