Predicting the addition of new concepts in a topic hierarchy

Ontologies often change through time, a process largely done manually by human editors. We discuss the task of automatically predicting when structural changes will occur in a given ontology. We first analyze the frequency of different types of structural changes in a large real-world ontology and then focus on the problem of predicting one specific type of structural change, namely the addition of a new category as a subcategory of an existing category, from which some of the existing instances are transferred into the new category. We show how the prediction of this type of structural change can be seen as a machine learning problem; the main challenge is to define a useful set of features. Experimental evaluation on a subset of the Open Directory Project hierarchy is provided. 1 INTRODUCTION An ontology may need to change through time, because the domain modelled by the ontology changes, or because the needs of its users change. Thus it is natural to ask whether such changes in an ontology can be predicted automatically as an aid to the people maintaining the ontology. In this paper we will limit ourselves to a simple class of lightweight ontologies, namely topic hierarchies. In these ontologies, the concepts are really topical categories and the only relationship between them is the is-a (parent-child) relationship which connects the categories into a tree. In addition, we assume that the ontology contains instances and that these instances are actually textual documents. Each document belongs to one or possibly several categories. A well-known example of such an ontology from the real world is the topic hierarchy of the Open Directory …