Tyler T. Procko

What is an upper ontology?

An upper ontology (also called a top-level or foundational ontology) is a high-level, domain-independent ontology that provides basic categories and relationships applicable across all domains. These ontologies serve as a foundation for more specific domain ontologies, ensuring interoperability and consistency across different knowledge domains.

What is a coterie?

A coterie is a small group of people with similar interests. Because of the confusing nature of the Semantic Web, I have found some related "resource dumps". These are given here.

Resource Collections

Companies and Consulting Firms

Academic and Research Groups

Why did the Semantic Web fail?

Tim Berners-Lee has a new project for personal data governance called the Solid Project. In the FAQs page of that project (archive here), the Solid team mentions that Linked Data and the Semantic Web 'never took off'. I have researched in this domain for some time now. Through my research, and discussions with other expert practitioners (e.g., Nicholas del Rio, Jarno van Driel), it is apparent that the reason for the distinct failure of the Semantic Web as an ideal is due, in part, to the fact that the adoption of the W3C's standards was not clear or straightforward for data providers on the Web - it simply was not easy to do. The working groups pushing for the Semantic Web were primarily composed of academic types, not business people or developers, and so the more complex constructs, e.g., OWL, were not ever realized to any great extent on the Web. Also, the advent of social media sites replaced the vision of the Web 3.0 as a Web of interlinked concepts with interactive websites. That being said, the use of Linked Data is common in search engines, e.g., with the Schema.org vocabulary, which is used along with other microdata formats for more accurate searching, recommendations, etc.

What is OWL good for?

OWL is a manifestation of Description Logics (DLs); there are various DL profiles for OWL. For most use cases, OWL is exceptional overkill. Inasmuch as OWL is a DL implementation, it abides by the open-world assumption (OWA), which is extremely difficult for most people to reconcile with anything else they do, e.g., OOP, which is closed-world by nature. For example, in Java, we may define a class Human but never give it an attribute for a brain. In closed-world logic, it is inferred that the class of Humans do not possess brains. In an OWL Human class, unless it is explicitly stated that Humans do NOT have brains with an OWL constraint, it cannot be inferred that they do not. In other words, OWL never makes inferences without explicit assertions. This makes understanding its inferences on larger ontologies difficult; even veteran OWL modelers struggle to explain inferences. OWL is a very heavy-handed, complicated modeling language; and, unlike RDF / RDFs, which are rather simple, easily extended and only offer limited reasoning (e.g., subsumption inferencing), OWL is very difficult to extend, because, with each new OWL constraint added, the complexity of the OWA makes erroneous inferences exceptionally hard to diagnose. In my experience, OWL is only used to its full extent in the biomedical domain. Most people who work with an "OWL ontology" do not even come close to fully utilizing the DL underlying OWL: for the most part, they implement a class/relationship hierarchy (this can be done with RDFS), add annotation properties and perhaps a few necessary conditions (subClassOf constraints), and then they never run a reasoner. This is not an OWL ontology. In any case, for most use cases in business, OWL is far too complicated. Furthermore, OWL cannot be used to validate instance data against the ontology, unless one writes very specific SPARQL queries to do so; and, even then, "bad" instance triples cannot be rejected without a specific codebase written for this purpose. A language like the Shapes Constraint Language (SHACL) is a much more appropriate means of ontological modeling for business use cases, because every major graph database platform supports SHACL validation and it is a closed-world language. SHACL can be used to define an ontology by shapes, that is then used automatically on data ingestion to validate new triples. Inveterate OWL modelers like the founder of TopQuadrant have abandoned OWL in favor of SHACL. So, I point everyone I can to SHACL as an alternative to OWL.

How do you reconcile AI/ML with Semantic Web constructs like ontologies?

This will remain as a rather informal response. In the historical sense, ontologies and ML went hand-in-hand: ontologies were fed into ML, and vice versa. For instance, IBM's Watson touted ontology-based ML. In any case, as it stands now, in 2023 and beyond, ontologies are nearly irrelevant in practical use and the research landscape. A Google search for 'Python ML' will return thousands of posts from the same week; a Google search for 'Python ontology' will return posts from over a decade ago. Ontologies are complicated artifacts espoused by very few, while ML consistently explodes in popularity and use. Not many individuals can put forth the effort to abandon the closed-world logic driving literally everything we do as humans to fully comprehend the open-world logic of description logics, which ontologies (at least, OWL ontologies) subscribe to. But, collecting large amounts of data, training a model, getting a resultant vector matrix and using it? That is more approachable. And the research landscape is hot, with new articles being published literally by the minute. There simply is not the support for ontology work. Steadfast ontologists like Barry Smith are probably injured (understandably) by the exploding popularity of undocumented, unexplainable ML. Veterans of the Semantic Web space like Jens Lehmann have, in a sense, tipped the proverbial hat to the rise of ML and general AI. Whatever the thinking, ontologies still have some purpose. For instance, the heart of an ontology, its taxonomy, is essential in everything: models need to be classified to be found and used, as does training data. But for ontologists, true ontology work seems to be a concern only for biomedical clients. It's a bad game to be in.