Applying the Query Change Retrieval Model on Session Search-Georgetown at TREC 2013 Session Track

In TREC 2013 Session Track, we experiment the Query Change Model (QCM) for session search on the new document collection ClueWeb12 CatB. We use structured query formulation and pseudo-relevance feedback for RL1. The QCM approach is used in RL2 and it studies query change between adjacent queries and models session search as a Markov Decision Process. We further add information from other sessions by a majority vote of cross-session clicked data to the model in RL3. Comparing the retrieval accuracy in RL1 with that in RL2 and RL3, we obtain 46.1% and 46.6% improvements, respectively. We present an analysis and discussion on the dataset difference between ClueWeb12 Cat A and Cat B, the difference between TREC2012 and TREC2013 session, and their impact on the retrieval accuracy.