Enhancing Subject Metadata with Automated Weighting in the Medical Domain: A Comparison of Different Measures

Abstract

Subject metadata has an important role in supporting subject access to information resources in digital libraries. Existing subject indexing systems generally produce binary outcomes (either assign or not assign a subject descriptor) that do not adequately reflect the extent to which a document is associated with the assigned subject descriptors. An automated weighting mechanism for subject descriptors is needed to strengthen the role of the subject metadata field. This study compares five measures for automated weighting subject descriptors in documents. The performance of the measures is evaluated according to their ability to discriminate major descriptors from non-major ones. Experiments on a medical collection with 348,566 articles suggest that all the measures are able to rank the major descriptors significantly higher than the non-major ones. Jaccard coefficient, Cosine similarity, and KL Divergence showed better performance than Mutual Information and Log Likelihood Ratio. The findings of this study contribute to the development of weighted subject indexing systems that have direct applications for aiding users’ information seeking and improving knowledge organization in digital libraries.