Cluster It! Semiautomatic Splitting and Naming of Classification Concepts

In this paper, we present a semiautomatic approach to split overpopulated classification concepts (i.e. classes) into subconcepts and propose suitable names for the new concepts. Our approach consists of three steps: In a first step, meaningful term clusters are created and presented to the user for further curation and selection of possible new subconcepts. A graph representation and simple tf-idf weighting is used to create the cluster suggestions. The term clusters are used as seeds for the subsequent content-based clustering of the documents using k-Means. At last, the resulting clusters are evaluated based on their correlation with the preselected term clusters and proper terms for the naming of the clusters are proposed. We show that this approach efficiently supports the maintainer while avoiding the usual quality problems of fully automatic clustering approaches, especially with respect to the handling of outliers and determination of the number of target clusters. The documents of the parent concept are directly assigned to the new subconcepts favoring high precision.