Overview of the TREC 2014 Session Track

Abstract : The TREC Session track ran for the fourth time in 2014. The track has the primary goal of providing test collections and evaluation measures for studying information retrieval over user sessions rather than one-time queries. These test collections are meant to be portable, reusable, statistically powerful, and open to anyone that wishes to work on the problem of retrieval over sessions. The experimental design of the track was similar to that of the previous three years [5, 6, 1]: sessions were real user sessions with a search engine that include queries, retrieved results, clicks, and dwell times; retrieval tasks were designed to study the effect of using session data in retrieval for only the mth query in a session. For the 2014 track, sessions were obtained from workers on Amazon's Mechanical Turk. As a result, the 2014 data includes far more sessions than previous years1,257 unique sessions as compared to around 100 for each of the previous three years. Apart from that, there is little different from the 2013 track [1]. This overview is organized as follows: in Section 2 we describe the tasks participants were to perform. In Section 3 we describe the corpus, topics, and sessions that comprise the test collection. Section 4 gives some information about submitted runs. In Section 5 we describe relevance judging and evaluation measures, and Sections 6 present evaluation results and analysis.