Belle Tseng - Universal Web Search Relevance

With the fast penetration of the Web throughout the world, the number of search users has increased dramatically from many geographic locations. Search engines are now facing the problem of providing search results to many countries. Machine Learned Ranking (MLR) approach has shown successes in web search. With the increasing demand to develop effective ranking functions for many countries (domains), we face a big bottleneck of insufficient training data to build a learned ranker for each domain. In my talk, I will present two approaches to resolve this problem. The first is a tree-based adaptation that takes a ranking function from one domain and tunes it with a small amount of training data from the target domain. The second approach is a Dynamic Bayesian Network click model that combines small amounts of training data with click data to build an unbiased estimation of the search relevance. Finally, I will report our experiments in evaluating the two approaches on a large dataset from the Yahoo! Search query logs, and report our findings.