Language Model Document Priors based on Citation and Co-citation Analysis

Citation, an integral component of research papers, implies certain kind of relevance that is not well captured in current Information Retrieval (IR) researches. In this paper, we explore ingesting citation and co-citation analysis results into IR modeling process. We operationalize on going beyond the general uniform document prior assumption in language modeling framework through deriving document priors from papers citation counts, citation induced PageRank and co-citation clusters. We test multiple ways to estimate these priors and conduct extensive experiments on the iSearch test collection. Our results do not suggest significant improvements of using these priors over no prior baseline measured by mainstream retrieval effectiveness metrics. We analyze the possible reasons and suggest further directions in using bibliometric document priors to enhance IR.