Training a hierarchical classifier using inter document relationships

Concept hierarchies, also called taxonomies or directories, are widely used on the World Wide Web to organize and present large collections of Web pages. They were originally developed to help users locate relevant information by browsing. More recently, conceptual search engines such as KeyConcept have been developed that retrieve documents based upon the concepts they discuss in addition to the keywords they contain. Both applications require that documents be classified into appropriate concepts in a conceptual hierarchy. Most classification approaches use flat classifiers that treat each concept as independent, even when the concept space is hierarchically structured. In contrast, hierarchical text classification exploits the structural relationships between the concepts. In this paper, we explore the effectiveness of hierarchical classification for a large concept hierarchy. Since the quality of the classification is dependent on the quality and quantity of the training data, we evaluate the use of documents selected from subconcepts to address the sparseness of training data for the top-level classifiers and the use of document relationships to identify the most representative training documents. By selecting training documents using structural and similarity relationships, we achieve a statistically significant improvement of 39.8% (from 54.5% to 76.2%) in the accuracy of our classifier over that of the flat classifier for a large, 3-level concept hierarchy.