Relevance Ranking Using Kernels

This paper is concerned with relevance ranking in search, particularly that using term dependency information. It proposes a novel and unified approach to relevance ranking using the kernel technique in statistical learning. In the approach, the general ranking model is defined as a kernel function of query and document representations. A number of kernel functions are proposed as specific ranking models in the paper, including BM25 Kernel, LMIR Kernel, and KL Divergence Kernel. The approach has the following advantages. (1) The (general) model can effectively represent different types of term dependency information and thus can achieve high performance. (2) The model has strong connections with existing models such as BM25 and LMIR. (3) It has solid theoretical background. (4) The model can be efficiently computed. Experimental results on web search dataset and TREC datasets show that the proposed kernel approach outperforms MRF and other baseline methods for relevance ranking.