Crowdsourcing for Robustness in Web Search

Search systems are typically evaluated by averaging an effectiveness measure over a set of queries. However, this method does not capture the the robustness of the retrieval approach, as measured by its variability across queries. Robustness can be a critical retrieval property, especially in settings such as commercial search engines that must build user trust and maintain brand quality. This paper investigates two ways of integrating crowdsourcing into web search in order to increase robustness. First, we use crowd workers in query expansion; votes by crowd workers are used to determine candidate expansion terms that have broad coverage and high relatedness to query terms mitigating the risky nature of query expansion. Second, crowd workers are used to filter the top ranks of a ranked list in order to remove nonrelevant documents. We find that these methods increase robustness in search results. In addition, we discover that different evaluation measures lead to different optimal parameter settings when optimizing for robustness; precisionoriented metrics favor safer parameter settings while recalloriented metrics can handle riskier configurations that improve average performance.