QU at TREC-2013: Expansion Experiments for Microblog Ad hoc Search

Abstract

In the first appearance of Qatar University (QU) at Text REtrieval Conference (TREC), our submitted microblog runs explored different ways of expanding the context of both queries and tweets to overcome the sparsity and lack of context problems. Since the task is real-time, we have also considered the temporal aspect, once combined with tweet expansion technique, and another separately as a scoring factor. Our explored ideas were all unsupervised and only used internal resources (i.e., the provided API service with only access to the tweets). For query expansion, we have used pseudo relevance feedback to include terms from the top-ranked retrieved tweets. Based on experiments on previous TREC collections, an aggressive expansion with 30 terms or more provided the best improvement. For tweet expansion, a 2-step relevance modeling approach was leveraged to temporally and lexically expand a tweet. To further explore the effect of considering the time dimension in scoring tweets, we also developed a temporal re-scoring function used to favor tweets that are closer in time to the query over tweets that might be more lexically relevant but are posted further apart in time from the query. We also conducted post-TREC experiments in which we worked on enhancing the query expansion and temporal re-scoring approaches. Resuls released by TREC have shown that the temporal re-scoring run was the most effective run among all of our submitted ones. As for the post-TREC experiments, our results have shown that the enhanced query expansion and temporal re-scoring approaches had notable improvements on retrieval effectiveness.