PKUICST at TREC 2013 Microblog Track

This paper describes PKUICST’s entry into the TREC 2013 Microblog track. In this year of microblog track, we designed and conducted a series of experiments based on both our local crawled collection and the official track API. For runs with local crawled dataset, we exploited different retrieval models, such as TFIDF, Okapi BM25 and statistic language model and tuned optimal parameters for all these models with the dataset in TREC 2012. Furthermore, we attempted to combine these models to gain a better performance with the help of learning to rank framework. For runs with the official track API, we employed language model with two-stage pseudo feedback query expansion. In addition, a filtering component was adopted to refine the results retrieved by the expanded query. Experimental results demonstrate that our approach obtains convincing performances.