Query-expansion Approaches for Microblog Retrieval

The School of Information and Library Science at the University of North Carolina at Chapel Hill submitted three runs to the “Scenario B” task of the TREC 2015 Microblog Track. The task simulates a scenario where a user specifies a topic of interest in the form of a keyword query and the system produces daily updates with at most 100 tweets about the topic of interest. Systems were responsible for monitoring a stream of tweets and making daily predictions for a set of 250 interest profiles. Each interest profile was in the form of a short keyword query. Systems were asked to produce a ranking of at most 100 tweets per interest profile at the end of each day (shortly after midnight). The evaluation period extended a 10-day period from July 20 to July 29, 2015. All tweets published between 00:00:00 to 23:59:59 UTC were valid candidates for each day of the evaluation period. Our team submitted three runs for “Scenario B”. All runs were automatic runs and used the interest profile title field as the input query. We explored three di↵erent ways of enriching the query representation through query expansion. In two of our runs (UNCSILS WRM and UNCSILS TRM), we scored tweets proportional to the negative KL-divergence between a relevance model generated from an external collection and the tweet language model. These two runs mainly di↵er by the external collection used to generate a relevance model for interest profile query. In our UNCSILS WRM run, we used an external Wikipedia corpus, and in our UNCSILS TRM run, we used a corpus of tweets collected during a three-week period prior to the evaluation period. In our third run (UNCSILS HRM), we aimed to expand the query with related hashtags.