WordNet [1] is an online lexical reference that uses psycholinguistically derived relationships to organize approximately 115,000 synonym sets, containing 150,000 English words. WordNet has been attractive to a number of artificial intelligence (AI) projects because it is large, free, and machine-readable. Problems arise, however, when attempting to use synonym sets and hyponym (is a) relationships between these words as a true ontology. To take the simplest example, WordNet does not distinguish individuals (for example, Red Cross,) from types (for example, organization) from roles (for example, causal agent), but logically these concepts and the relationships between them are quite different.

The authors have applied their OntoClean methodology, for developing coherent and consistent ontologies, to WordNet. The OntoClean methodology [2] defines the kinds of concepts and conceptual relationships an ontology should have, and a process for building an ontology. In this paper, the authors first describe their application of the methodology to identify problems with the critical top levels of WordNets noun hierarchy. They cleaned up these problems by moving items that do not belong at the top level, and by introducing missing concepts to produce a more conceptually coherent top-level hierarchy they call Dolce (a descriptive ontology for linguistic and cognitive engineering).

The paper is detailed and well written. Reading the principles of OntoClean is useful, but not essential. The examples are concrete and relatively intuitive. The paper would have been improved if it included a demonstration of actual benefits from using the improved WordNet hierarchy. Reference is made to a terminology integration project in the fishery domain, but no details are given.