Effective pseudo-relevance for Microblog retrieval

Microblog services such as Twitter have become a part of daily life for many users, with thousands of documents published each second. Microblog documents are often too short, overwhelming in their use of informal language and hard to understand due to a lack of contextual clues. Retrieving relevant documents from microblogs is somewhat challenging because of its nature and the massive scale of the data. However, microblog retrieval models suffer from a vocabulary mismatch problem that leads to insufficient performance. In this paper, we address microblog retrieval limitations by proposing a pseudo-relevance feedback model. Our model considers discriminative expansion to meet user interests. Experimental results on TREC 2011 and 2012 microblog datasets show that our model demonstrates significant improvements over the baseline models.