On the Effectiveness of Bayesian Network-based Models for Document Ranking
Xing Tan, Fanghong Jian, Jimmy Xiangji Huang · 2017
Theoretical soundness and technical feasibility of treating the problem of document ranking in IR as an inference problem in Bayesian Networks, was studied recently. A pilot framework was also proposed there. In this paper, we provide two implementations of the framework: BNBM25, the one based on BM25, and BNMATF, which is based on MATF, a recently proposed innovative ranking function. We empirically verify the effectiveness of these two implementations on several standard test collections. Positive, significant results are obtained. Potentials of this BN-based framework in addition to its verified effectiveness are also discussed. As a result of the study, we believe that the technique is promising, worthy of further analysis and application.