Learning Optimal Card Ranking from Query Reformulation.

摘要: Mobile search has recently been shown to be the major contributor growing market. The key difference between mobile and desktop is that information presentation limited screen space of device. Thus, engines have adopted a new type result presentation, known as \textit{information cards}, in which each card presents summarized results from one domain/vertical, for given query, augment standard blue-links results. While it widely acknowledged cards are particularly suited user experience, also challenging optimize such sets. Typically, engagement metrics like query reformulation based on whole ranked list most traditional learning rank algorithms require per-item relevance labels. In this paper, we investigate possibility interpreting into effective labels query-card pairs. We inherit concept conventional learning-to-rank, propose pointwise, pairwise listwise interpretations reformulation. addition, learning-to-label strategy learns contribution card, with respect where contributions can used training ranking models. utilize state-of-the-art model demonstrate effectiveness proposed mechanisms large-scale data engine, showing models trained derived significantly outperform ones human judgment