Improving query suggestion through noise filtering and query length prediction

Liang Wu, Bin Cao, Yuanchun Zhou, Jianhui Li

Abstract: Clustering-based methods are commonly used in Web search engines for query suggestion. Clustering is useful in reducing the sparseness of data. However, it also introduces noises and ignores the sequential information of query refinements in search sessions. In this paper, we propose to improve cluster based query suggestion from two perspectives: filtering out unrelated query candidates and predicting the refinement direction. We observe two major refinements behaviors. One is to simplify the original query and the other is to specify it. Both could be modeled by predicting the length (number of terms) of queries when candidates are being ranked. Two experimental results on the real query logs of a commercial search engine demonstrate the effectiveness of the proposed approaches.