Query Ambiguity Identification Based on User Behavior Information

Query ambiguity identification is of vital importance for Web search related studies such as personalized search or diversified ranking. Different from existing solutions which usually require a supervised topic classification process, we propose a query ambiguity identification framework which takes user behavior features collected from click-through logs into consideration. Especially, besides the features collected from query level, we focus on how to tell the differences between clear and ambiguous queries via features extracted from multi-query sessions. Inspired by recent progresses in word representation researches, we propose a query representation approach named “query2vec” which constructs representations from the distributions of queries in query log context. Experiment results based on large scale commercial search engine logs show effectiveness of the proposed framework as well as the corresponding representation approach.