A Simple Kernel Co-Occurrence-Based Enhancement for Pseudo-Relevance Feedback

摘要

In this article a kernel co-occurrence-based framework was proposed in which term co-occurrence information is integrated into the classic Rocchio model and a relevance model (RM3) to achieve enhanced retrieval performance. When selecting and weighting the candidate terms from feedback documents, we used a linear combination of model components to balance the influence of classic models and models that capture term co-occurrence information to achieve better performance. Thus, two kernel co-occurrence-based methods, KRoc and KRM3, are proposed. In particular, to better utilize the term co-occurrence information, we incorporated this information into the whole term weight formula by simultaneously refining both the factor of the term discriminating power and the factor of the within-document term weight in feedback documents to achieve better performance. The experimental results show that the proposed KRoc and KRM3 methods are effective and outperform the corresponding strong baseline methods in terms of the MAP and P@10 results on most collections used for testing. Meanwhile, our proposed methods are at least comparable to the state-of-the-art TF-PRF, PRoc, IF&FB, and MRF models. Additionally, we carefully analyzed the influence of σ on our proposed KRoc and KRM3methods, and an empirical rule for setting this parameter to achieve good performance is suggested.