Learning Query and Document Relevance from arnWeb-scale Click Graph

摘要

Click-through logs over query-document pairs provide richrnand valuable information for multiple tasks in informationrnretrieval. This paper proposes a vector propagation algorithmrnon the click graph to learn vector representations forrnboth queries and documents in the same semantic space.rnThe proposed approach incorporates both click and contentrninformation, and the produced vector representationsrncan directly improve ranking performance for queries andrndocuments that have been observed in the click log. Forrnnew queries and documents that are not in the click log, wernpropose a two-step framework to generate the vector representation,rnwhich signifcantly improves the coverage of ourrnvectors while maintaining the high quality. Experiments onrnWeb-scale search logs from a major commercial search enginerndemonstrate the eu000bectiveness and scalability of the proposedrnmethod. Evaluation results show that NDCG scoresrnare signifcantly improved against multiple baselines by usingrnthe proposed method both as a ranking model and as arnfeature in a learning-to-rank framework.