USC/ISI at TREC 2011: Microblog Track

This paper describes the search system we developed for the inaugural TREC 2011 Microblog Track. Our system makes use of best-practice ranking techniques, including term, phrase, and proximity-based text matching via the Markov random field model, pseudo-relevance feedback using Latent Concept Expansion, and a feature-based ranking model that uses a simple, but effective learningto-rank model. We adapted each of these approaches to the specifics of the microblog search task, giving rise to a highly effective end-to-end search system. The official results from the TREC evaluation suggest that pseudorelevance feedback and learning-to-rank yield significant improvements in precision at early rank under different evaluation scenarios.