Towards adaptive structured Dirichlet smoothing model for digital resource objects

Abstract

Digital resource objects (DRO) are among the most valuable resources that store the accumulated knowledge of humankind. Nowadays, many organisations aim to make these resources available to users. Basically, Dirichlet smoothing (DS) model is widely used to retrieve DRO documents. DS model uses a smoothing parameter μ which plays a strong role in finding the value of the unseen terms to avoid zero probability value. For documents of equal length, the value of μ is set as a constant value although its value depends on the length of a document. In DROs, almost all documents are of different length, and each metadata unit in a document also has a different length. Hence, it is not appropriate to predefine the μ parameter with a constant value and uses it for different search space. This leads to difficulty in accessing and retrieving the DRO documents. To solve fixed smoothing-parameter value problem in DRO’s retrieval, and make DROs more accessible, Adaptive Dirichlet Smoothing (ADS) and Adaptive Structured Dirichlet Smoothing (ASDS) models are proposed to improve the performance of the DRO’s retrieval by estimating the smoothing parameter automatically. The proposed ASDS model comprises the ADS model together with an existing DS model. Experimental results on CHiC2013 collections show that the proposed models have the ability to retrieve the most relevant results (documents or metadata units) related to a particular query and reduce the zero-probability values compared with state-of-the-art traditional methods particularly on DROs. Moreover, t-test result is used to prove that the performance of the proposed models is statistically significant.