Improving Tail Query Performance by Fusion Model

Tail queries, which occur with low frequency, make up a large fraction of unique queries and often affect a user's experience during Web searching. Because of the data sparseness problem, information that can be leveraged for tail queries is not sufficient. Hence, it is important and difficult to improve the tail query performance. According to our observation, 26% of the tail queries are not essentially scarce: they are expressed in an unusual way, but the information requirements are not rare. In this study, we improve the tail query performance by fusing the results from original query and the query reformulation candidates. Other than results re-ranking, new results can be introduced by the fusion model. We emphasize that queries that can be improved are not only bad queries, and we propose to extract features that predict whether the performance can be improved. Then, we utilize a learning-to-rank method, which is trained to directly optimize a retrieval metric, to fuse the documents and obtain a final results list. We conducted experiments using data from two popular Chinese search engines. The results indicate that our fusion method significantly improves the performance of the tail queries and outperforms the state-of-the-art approaches on the same reformulations. Experiments show that our method is effective for the non-tail queries as well.