Multiple Ranker Method in Document Retrieval

Abstract

In this paper, we propose a multiple-ranker approach to make learning to rank methods more effective for document retrieval application. In traditional learning to rank methods, a ranker is learned from a set of queries together with their corresponding document rankings labeled by experts, and it is then used to predict the document rankings for new queries. But in practice, user queries vary in large diversity, which makes the single ranker learned from a close set of data not representative. The single ranker cannot be guaranteed with the best ranking result for every single query, and this becomes the bottleneck of traditional learning to rank approaches. To address this problem, we propose a multi-ranker approach. We train multiple diverse rankers which can cover diverse categories of queries, instead of an isolated one, and take an ensemble of these rankers for final prediction. We verify the proposed multipleranker approach over real-world datasets. The experimental results indicate that the proposed approach can outperform existing ‘learning to rank’ methods significantly.

This work is supported by National Science Foundation of China under the grant 60673009, Tianjin Science and Technology Research Foundation under the grant 05YFGZGX24000 and Microsoft Research Asia Foundation.