Webis at TREC 2013-Session and Web Track

In this paper we give a brief overview of the Webis group’s participation in the TREC 2013 Session and Web tracks. All our runs are on the full ClueWeb12 and use the online Indri retrieval system hosted at CMU. As for the session track, our runs implement three main ideas that were slightly improved compared to our participation in 2012: (1) distinguishing low risk sessions where we want to involve session knowledge in the form of a conservative query expansion strategy (only few expansion terms based on keywords from previous queries and seen/clicked documents/titles/snippets) from those where we don’t, (2) conservative query expansion based on similar sessions from other users, (3) result list postprocessing to boost clicked documents of other users in similar sessions. As these techniques leave a lot of queries unchanged when not enough session knowledge is available, we do not expect large gains over all the sessions. As for the Web track, our runs exploit different strategies of segmenting the queries (i.e., identifying and highlighting concepts within the query as phrases to be contained in the results). Additionally to algorithmic segmentations based on our WWW 2011 and CIKM 2012 ideas, we had one run where we chose the segmentation according to a majority vote amongst five humans. In a last run, the results are constructed so as to be disjunct from the track’s baseline’s and our other runs’ results. Instead, we populate the result list with documents that different segmentations of the query would return top-ranked or that are deeper in the ranking for segmentations already chosen in previous runs. The underlying idea was to obtain at least some judgments for the top documents that other segmentations would bring up in their rankings. As most of the queries are rather short, we expect only slight improvements or no effect at all from the different segmentation strategies that are tailored to longer and more verbose queries.