I’ve used terms like ontology, taxonomy, metadata, and vocabulary in my data modeling training and consulting, but my knowledge lacked context. Context includes not just a broader perspective than data modeling, but knowledge of how metadata standards build on a controlled vocabulary, a taxonomy builds on metadata standards, a thesaurus builds on a taxonomy, an ontology builds on a thesaurus, and a knowledge graph builds on an ontology. I found this understanding of the dependencies among each part of Jessica Talisman’s Ontology Pipeline to be a key takeaway from her book.
The book describes how to build semantic knowledge systems that both people and machines can understand. The book addresses a problem that has become much more important with the growth of AI: organizations have enormous amounts of data, but often lack a consistent way to define what that data means and how different concepts relate to one another. This book shows how techniques librarians have used for decades can provide structure for a modern data architecture.
Chapter 1 begins with controlled vocabularies, the foundation of the Ontology Pipeline. Jessica explains that organizations already have vocabularies scattered throughout databases, documents, applications, search logs, and employee terminology. The first job is to find those terms, reconcile duplicates and synonyms, define them clearly, and establish preferred terminology. A controlled vocabulary creates agreement about language, which we need before moving on to more sophisticated semantic structures. The chapter also connects this work to AI, showing how consistent terminology improves search, classification, interoperability, and the ability of AI systems to interpret an organization’s language correctly. Before machines can reason about concepts, organizations first need to agree on their terminology and definitions.
Chapter 2 moves to metadata schemas and explains how metadata turns vocabulary into structured descriptions of information. Jessica draws heavily from library science and introduces metadata application profiles, or MAPs, as a way to connect controlled vocabularies, schemas, governance, and business systems. Rather than viewing metadata as labels attached to data, she treats it as semantic infrastructure. One of the chapter’s stronger arguments is that many problems companies try to solve with additional integration technology are really problems of meaning. Metadata application profiles provide a semantic bridge between systems by defining elements, relationships, allowable values, ownership, and governance.
Chapter 3 takes the controlled vocabulary and organizes it into a taxonomy. Here, the emphasis shifts from defining concepts to establishing hierarchical relationships among them. Jessica walks through domain research, coverage models, top-level categories, definitions, alternative labels, and the “is a” test for determining legitimate parent-child relationships. She then introduces SKOS as a way to make taxonomies portable and machine-readable. The chapter stresses that good encoding cannot compensate for poor thinking. A taxonomy must first accurately represent the domain. When done well, it becomes infrastructure for navigation, classification, search, RAG, and AI systems that need to move between broader and narrower concepts.
Chapter 4 extends the taxonomy into a thesaurus. A taxonomy primarily provides vertical relationships, while a thesaurus adds synonyms, alternative labels, equivalence relationships, and associations between concepts in different parts of the hierarchy. This gives systems a richer understanding of meaning. Jessica shows how SKOS can support these relationships and uses real examples such as EuroVoc, AGROVOC, and MeSH to demonstrate that thesauri are already important knowledge infrastructure. For AI, these relationships can improve query expansion, entity recognition, retrieval, classification, and RAG by helping systems understand that different words may refer to the same concept and that related concepts may be relevant even when the words themselves are different.
Chapter 5 is the largest chapter, focusing on ontologies. Jessica distinguishes an ontology from a taxonomy or database schema by emphasizing classes, properties, axioms, logical relationships, and machine reasoning. She introduces RDF, RDFS, OWL, ontology design principles, competency questions, governance, validation, and the use of standard and custom ontologies. The chapter also explains how ontologies can support structured retrieval, GraphRAG, AI system registries, and organizational memory.
Chapter 6 brings all of the previous stages together in the knowledge graph. Jessica argues that a knowledge graph should be viewed as an architecture, not a product. Its components can include vocabularies, ontologies, knowledge bases, metadata schemas, reasoners, NLP pipelines, vector retrieval, and graph analytics. The chapter examines enterprise, domain, and linked data knowledge graph patterns, along with implementation, governance, and semantic debt. The knowledge graph is never complete, though; it has to evolve with the domain it represents.
I love this excerpt from Chapter 5 that explains why it’s important to distinguish each level and how taxonomies and ontologies differ:
The Ontology Pipeline describes the systematic process of constructing knowledge infrastructure through defined stages: controlled vocabularies, metadata schemas, taxonomies, thesauri, ontologies (where we are now), and knowledge graphs. Each stage produces a distinct deliverable with its own structure, standards, and engineering value.
The Pipeline insists on calling each stage by its own name, because the current moment in industry discourse has a specific pathology. “Ontology” is being applied indiscriminately to flat keyword lists, product category trees, JSON config files, and any structured artifact dressed up for a pitch deck.
If a SKOS taxonomy is already an ontology (just a lightweight one), an organization that has built a taxonomy may believe it has “done” ontology work and can proceed to building AI systems that require formally modeled domain knowledge. It cannot. A taxonomy categorizes. An ontology—whether expressed in RDFS and/or OWL—models the semantic structure of a domain: what exists, how things relate, and what those relationships mean.
If the Pipeline stages are clearly named, the gap is visible and can be planned for. If everything is an ontology of some weight, the gap disappears in the language, even though it persists in the infrastructure—and that invisible gap is where projects fail. The spectrum tells you where you are. The Pipeline tells you how to build. We are interested in both building and semantic expressivity.
Data Modeling Deep Dive
Learn data modeling concepts, techniques, and real-world applications – November 9-11, 2026.
