Comparing alternative data-driven ontological vistas of natural history

Traditionally, ontologies are created manually, based on human experts' view of the concepts and relations of the domain at hand. We present ongoing work on two approaches to the automatic construction of ontologies from a flat database of records, and compare them to a manually constructed ontology. The latter CIDOC-CRM ontology focuses on the organisation of concepts and relations. In contrast, the first automatic method, based on machine learning, focuses on mutual predictiveness between concepts, while the second automatic method, created with the aid of Wikipedia, stresses meaningful relations between concepts. The three ontologies show little overlap; their differences illustrate that a different focus during ontology construction can lead to radically different ontologies. We discuss the implications of these differences, and argue that the two alternative ontologies may be useful in higher-level information systems such as search engines.