Co-Learning Ranking for Query-Based Retrieval

摘要

In this paper, we propose a novel blending ranking model, named Co-Learning ranking, in which two ranked results produced by two basic rankers interact with each other adequately and are combined linearly with a pair of appropriate weights. Specifically, in the interaction process, a reinforcement strategy is proposed to boost the performance of each ranked results. In addition, an automatic combination method is designed to detect the better-performance ranked result and assign a higher weight to it automatically. The Co-Learning ranking model is applied to the document ranking problem in query-based retrieval, and evaluated on the TAC 2009 and TAC 2011 datasets. Experimental results show that our model has higher precision than basic ranked results and better stability than linear combination.