Query Dependent Time-Sensitive Ranking Model for Microblog Search

Previous works show that one main difference between web search and microblog search is that most microblog queries are time-sensitive. Therefore, many existing works based on one straightforward temporal assumption have tried to incorporate the temporal factors into ranking model to improve the retrieval effectiveness. However, our study show that temporal role in ranking is complicated and hard to be summarized into one straightforward assumption. In addition, temporal influence is different among queries. To address these problems, we propose a query-dependent time-sensitive microblog ranking model, which use learning to rank to combine both temporal and entity evidences into the ranking process as the basic ranking model. In order to leverage the query difference, the k most similar training queries are used to train the ranking model. Experimental results on the public TrecMicroblog2011 data set show that comparing with the existing time-sensitive models, our models can significantly improve the performance of microblog search.