Improving Search Engine Performance Based On Intention Comprehending Of Tail Query

Search engines are important tools for people to get information. The frequencies of queries follow the power-law distribution. We call the queries at the tail ”tail queries”,which also known as low frequency queries. From analysis on sampled real search engine data, we find that tail queries make up 70% of all unique queries and almost all users have issued tail queries. While tail queries have little user feedback and existing methods can not be used for tail queries. So tail queries are often difficult in search engines.By analyzing real search engine data, we find that a part of tail queries are not expressed correctly so that some relevant documents can not be returned by search engines.To solve this problem, we attempt to understand users’ information need by analyzing query refinement patterns. Further, we generate some reformulation candidates, and merge their results lists with the original results list to get a better one. The contributions of our paper are as follows:1. Analysis and prediction on query refinement patterns. Based on previous work,we classify query refinement patterns into 4 classes: New topic, Generalization,Specification and Parallel. We have proposed both prediction and classification methods and the precision is 79.29%.2. Evaluation of tail query performance. We have analyzed the relationship of the document relevance and its content presentation, extracted click features, match features and rank features, and trained an ensemble-learning based classifier.3. A fusion model to improve tail query performance. We try to find some well expressed query reformulations with similar search intent and incorporate their results lists to get a new one. Our method has introduced new documents but not re-ranked the original results. We have considered whether a tail query can be improved.What’s more, this method is also effective for non-tail queries.4. A system to improve tail query performance based on user intent understanding. We combine the prediction of query refinement patterns and fusion model. The effect gets more improvement.