Pseudo Topic Analysis for Boosting Pseudo Relevance Feedback

Abstract

Traditional Pseudo Relevance Feedback (PRF) approaches fail to mode real-world intricate user activities. They naively assume that the first-pass top-ranked search results, i.e. the pseudo relevant set, have potentially relevant aspects for the user query. It is make the major challenge in PRF lies in how to get the reliability relevant feedback contents for the user real information need. Actually, there are two problems should not be ignored: (1) the assumed relevant documents are intertwined with the relevant and the non-relevant content, which influence the reliability of the expansion resource and can not concentrate in the real relevant portion; (2) even if the assumed relevant documents are real relevant to the user query, but they are always semantic redundance with various forms because the peculiarity of natural language expression. Furthermore, it will aggravate the ‘query drift’ problem. To alleviate these problems, in this paper, we propose a novel PRF approach by diversifying feedback source, which main aim is to converge the relatively single semantic as well as diversity relevant information from the pseudo relevant set. The key idea behind our PRF approach is to construct an abstract pseudo content obtained from topical networks modeling over the set of top-ranked documents to represent the feedback documents, so as to cover as diverse aspects of the feedback set as possible in a small semantic granularity. Experimental results conducted in real datasets indicate that the proposed strategies show great promise for searching more reliable feedback source by helping to achieve query and search result diversity without giving up precision.