Query word deletion prediction

This work provides algorithms which perform substantially better than the baseline in predicting which word should be deleted from a reformulated query, for increasing query coverage in the context of web search on small high-quality collections.

Abstract

Web search query logs contain traces of users' search modifications. One strategy users employ is deleting terms, presumably to obtain greater coverage. It is useful to model and automate term deletion when arbitrary searches are conjunctively matched against a small hand constructed collection, such as a hand-built hierarchy, or collection of high-quality pages matched with key phrases. Queries with no matches can have words deleted till a match is obtained. We provide algorithms which perform substantially better than the baseline in predicting which word should be deleted from a reformulated query, for increasing query coverage in the context of web search on small high-quality collections.