Relevance model revisited: With multiple document representations

Abstract

In this work, we extended Lavrenko's relevance model [6] and adapted it to the cases where an additional layer of document representation is appropriate. With this change, we are able to aggregate heterogeneous data sources and operate the model in different granularity levels. We demonstrated this idea with two applications. In the first task, we showed the feasibility of using a carefully-selected vocabulary as the query expansion source in a language model to enhance retrieval effectiveness. The proposed query refinement model outperformed the relevance model counterpart in terms of MAP by 17.6% under rigid relevance judgment. In the second task, we established a ranking scheme in a faceted search session to sort the facets based on their corresponding relevance to the query. The result showed that our approach improved the baseline performance by roughly 100% in terms of MAP. © 2010 Springer-Verlag.