Improving Rankings in Small-Scale Web Search Using Click Implied Descriptions

When a searcher submits a query Q and clicks on document R in the corresponding result set, we may plausibly interpret the click as a vote that Q is a description of R. We call the Q and R pairing a ‘click description’. Click descriptions thus derived from search engine logs can be accumulated into surrogate documents and used to boost retrieval effectiveness in a similar fashion to anchor text. We investigate the usefulness of click description surrogate documents in processing queries for an external web site search service for four organisations. Using the mean reciprocal rank of best answers as the measure of performance, we show that, for popular queries, click description surrogates significantly outperform both anchor text surrogates and the original proprietary rankings. The amount of click data needed to achieve a high level of retrieval performance is surprisingly small for popular queries. Thanks to terms shared between queries, click description surrogates can answer queries for which no specific click data is available. We show a 92% improvement due to this effect for a set of lengthy, less popular queries. We also discuss issues such as spam rejection, unpopular queries, and how to combine click description scores with other evidence. We argue the potential of click descriptions in non-web applications where link and anchor text evidence is unavailable.