Estimating Semantic Similarity between Expanded Query and Tweet Content for Microblog Retrieval

This paper reports the systems we submitted to the Microblog Track shared in TREC 2014 which focuses on ad hoc retrieval (i.e., retrieving top 1, 000 relevant tweet for every given topic). To address this task, we adopted a two-stage framework, i.e., firstly, we performed query expansion (i.e., expanding relevant inforamtion using pseudo-relevance feedback and Google search engine results) to retrieve more relevant tweets, then extracted several effective semantic features (e.g., Jansen-Shannon Distance, Overlap Similarity, Lucene Score, etc) from retrieved results and built ranking model using supervised machine learning algorithms with the aid of these features to perform re-ranking. Our systems ranked 3th out of 21 teams.