Learning to Blend by Relevance

Emergence of various vertical search engines highlights the fact that a single ranking technology cannot deal with the complexity and scale of search problems. For example, technology behind video and image search is very different from general web search. Their ranking functions share few features. Question answering websites (e.g., Yahoo! Answer) can make use of text matching and click features developed for general web, but they have unique page structures and rich user feedback, e.g., thumbs up and thumbs down ratings in Yahoo! answer, which greatly benefit their own ranking. Even for those features shared by answer and general web, the correlation between features and relevance could be very different. Therefore, dedicated functions are needed in order to better rank documents within individual domains. These dedicated functions are defined on distinct feature spaces. However, having one search box for each domain, is neither efficient nor scalable. Rather than typing the same query two times into both Yahoo! Search and Yahoo! Answer and retrieving two ranking lists, we would prefer putting it only once but receiving a comprehensive list of documents from both domains on the subject. This situation calls for new technology that blends documents from different sources into a single ranking list. Despite the content richness of the blended list, it has to be sorted by relevance none the less. We call such technology blending, which is the main subject of this paper.