Decision Support for Indexing and Retrieval of Information in Hypertext Systems

Abstract

This study introduces and evaluates the performance of two statistical models intended to support the automatic creation of a subject-based index containing links to hypertext documents. The fuzzy Bayes model makes strong dependence assumptions, and only considers the strongest evidence presented by single words occurring in a document, whereas the classic Bayes model makes strong independence assumptions and attempts to aggregate all the evidence. The links or index terms suggested by both indexing models overlapped greatly with those assigned by a human indexer. However, the probabilities calculated using the classic Bayes model were unstable because of data sparseness and severe violations of the independence assumptions. Subsequent analysis therefore focused on the fuzzy Bayes model. The latter analysis revealed that human experts rated index terms suggested by the model significantly higher than randomly selected index terms. When the index terms assigned by the fuzzy Bayes model were implemented as ...