Upper Level Ontologies: The Contested Languages of Artificial Intelligence

Data exist in formats that are often incompatible and formalized only locally. Data-labeling standards

are made using general terms, are based on natural language, or are adopted using formalized but

limited classification systems. Such a lack of quality vocabularies for accessing and reasoning with

heterogeneous data in uniform ways makes it much harder to achieve semantic interoperability of data

across systems. Developed by researchers over the past several decades, one solution has been to

provide logical (computable) definitions using controlled vocabularies of preferred labels for describing

data combined with tags—a practice known as ontology-making.

A Google Ngram Viewer search for the word “ontology” shows use of the term increased steadily in the

1950s and skyrocketed after the 1980s. “Ontology” was used intermittently in early conversations about

artificial intelligence, information theory, and computer science but it was in the early 90s with the

publication of a series of papers by Thomas Gruber that ontology spread as a popular term for achieving

semantic interoperability among heterogeneous data. In his entry for “Ontology” in the Encyclopedia of

Database Systems, Gruber goes further, writing that “in practice, the languages of ontologies are closer

in expressive power to first-order logic than languages used to model databases” (Gruber would go on

to invent Siri at Google). A philosophical concern for centuries, ontology is now a practical concern for

many researchers who must grapple with data-driven labeling practices and technologies.

This paper provides a cultural history of a widely-used upper level ontology called the Basic Formal

Ontology (BFO). The research is based on long form ethnographic interview data collected from leading

ontologists at the National Center for Ontological Research (NCOR) as well as examples of the BFO’s

internal logics, including data visualizations using Protégé software, an analysis of the BFO’s XML syntax,

and NCOR member discussion data scraped from BFO public message boards using HTTrack. The BFO

proposes a new way to organize and communicate data between domains and is used by hundreds of

ontology-driven endeavors throughout the world. Yet, as a sociotechnical phenomenon, the BFO is

prone to several problems stemming from assumptions and biases in reasoning. For example, ontologybuilders

may disagree on shared terms or propose contradictory logics in the construction phase. Is

there evidence that ontologies typify logics or biases? What types of data do ontologies organize? How

are ontologies practically applied in social contexts? This paper will use the BFO as a case study to

explain the contested methods and theories in the history of ontology engineering.