It is non-trivial to formulate a query that can precisely describe the goal of an informational search task. Query reformulation based on the query clustering approach addresses this issue by expanding a new query with related existing queries that were generated by other users. However, the query clustering approach is unable to cluster queries that are intrinsically related but neither contain common terms nor return common clicked Web page URLs. More importantly, it does not address the issue of ranking retrieved results according to their relevance to the search goal. In this paper, we present new query reformulation approach based on a novel probabilistic topic model to discovering the latent semantic relationships between the queries and the URLs. It can not only discover related queries that cannot be clustered by existing query clustering approaches but also rank retrieved results according to the similarities of probability distributions over the latent topics among the queries and the URLs. The results of our experiments have shown that this approach can significantly improve the performance of an informational search task in terms of search accuracy and search efficiency.
展开▼
