“An ontology is a formal, explicit specification of a shared conceptualization” [1], reads one of the more widely accepted definitions--but how, in reality, do you obtain one? This paper cleverly combines the strengths of two different approaches to come up with a procedure that yields a high-quality ontology in a comparably short time, as suggested by a thorough user study that actually employs the suggested procedure.

It is well known that manual ontology construction (that is, sorting through the text base, recognizing the main categories and concepts, and structuring them into a personal hierarchy) is highly reliable but rather tedious. Automated ontology construction through semantic distance learning (that is, judging the similarity of concepts based on certain pre-structured sets of words), on the other hand, is fast but unreliable and, in many cases, lacks the ability to structure word pairs correctly (for example, “polar bear”).

The suggested procedure, supervised hierarchical clustering, starts with automatic clustering of concepts at the lowest conceptual level. The resulting sets of related concepts are then--and this is new--presented to the user in a program, Ontology Construction Panel (Ontocop), that allows the user to interactively restructure the proposed clustering, according to his or her individual preferences. This new and now personalized set of clusters is then taken as input for the next round of automatic clustering on the next higher level that is again improved through an interactive user session, and so forth.

Yang and Callan test the usability of their procedure with 12 users who are given the task of constructing an ontology for certain public comment datasets (containing email comments sent to public agencies on certain topics). The results are utterly convincing: within approximately 60 minutes, the users are able to successfully construct an ontology that comprises 300 to 1,000 concepts, with the highest possible quality scoring of 1.0 (F3 measure--recall and precision).

The paper is well written, easy to follow, and contains, to the extent necessary, formulas and pseudocode for the proposed algorithm. Given the persuasive user test results, I highly recommend this paper to all researchers and practitioners who deal with the problem of how to construct ontologies.