Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Ontology Synthesis

What ontologies are

Across these materials, an ontology is best understood as an explicit, shared model of a conceptual domain: it names the kinds of things that matter, the relations among them, and axioms whose formally defined meaning lets software do more than string matching. notes/scigraph.md notes/owl-2-web-ontology-language.md

The ecosystem separates several layers that are often casually conflated. notes/rdf-resource-description-framework-1-1.md notes/owl-2-web-ontology-language.md RDF is the labeled directed multigraph foundation—subject–predicate–object statements identified mainly by IRIs—while Turtle, JSON-LD, RDF/XML, N-Triples, and the other concrete syntaxes are serializations of that abstract graph, not different conceptual models. notes/rdf-resource-description-framework-1-1.md notes/turtle.md notes/json-ld.md RDFS adds a deliberately small modeling vocabulary for classes, instances, subclass and subproperty hierarchies, global domains and ranges, labels, and comments. notes/rdf-schema-rdfs-1-1.md OWL 2 adds a more expressive ontology language for classes, properties, individuals, data values, restrictions, keys, property chains, cardinalities, and other axioms, with Direct Semantics for OWL 2 DL and RDF-Based Semantics for arbitrary OWL-as-RDF graphs. notes/owl-2-web-ontology-language.md Description logics supply the underlying mental model: concepts denote sets, roles denote ordered pairs, a TBox states terminological structure, an ABox states facts about individuals, and reasoners answer questions about subsumption, membership, relations, and consistency. notes/description-logics-dls.md

A taxonomy is therefore one useful shape of ontology, not the whole idea. notes/skos-simple-knowledge-organization-system.md SKOS intentionally represents thesauri, classifications, and subject-heading systems as URI-identified concepts with multilingual labels, broader/narrower/associative links, collections, and cross-scheme mappings, while explicitly distinguishing those concepts from OWL classes and formal facts about how the domain itself is arranged. notes/skos-simple-knowledge-organization-system.md Likewise, a knowledge graph can contain ontology-governed data, but “graph-shaped” alone does not imply formal ontology semantics: SciGraph, for example, loads an OWL-centered model into a useful but explicitly lossy Neo4j representation that cannot round-trip the ontology and is not itself an authoring or reasoning system. notes/scigraph.md

The recurring architecture is layered reuse. notes/bfo-basic-formal-ontology.md notes/common-core-ontologies-cco.md An upper ontology supplies domain-neutral categories—BFO deliberately excludes specialized scientific terms—while a mid-level layer such as CCO adds reusable notions such as agents, events, measurements, facilities, and information entities, and separately maintained domain extensions add the local subject matter. notes/bfo-basic-formal-ontology.md notes/common-core-ontologies-cco.md Other foundations make different commitments: DOLCE’s official release is a first-order-logic axiomatization, gist aims for a small enterprise-oriented core with extensive high-level disjointness, and gUFO is a lightweight OWL 2 DL implementation of a selected UFO subset. notes/dolce-descriptive-ontology-for-linguistic-and-cognitive-engineering.md notes/gist.md notes/gufo.md This makes ontology engineering less like inventing a universal dictionary and more like choosing explicit commitments at the right level, reusing stable identifiers and patterns, and keeping domain scope visible. notes/common-core-ontologies-cco.md notes/ontologydesignpatterns-org.md

Ontologies also sit beside, rather than replace, validation and query languages. notes/shacl-shapes-constraint-language.md notes/sparql-1-1.md SHACL and ShEx make expected RDF node shapes executable and report whether data conforms, whereas OWL reasoners derive consequences under formal semantics; SPARQL matches graph patterns and returns bindings, constructed graphs, existence answers, or service-defined descriptions. notes/shacl-shapes-constraint-language.md notes/shex-shape-expressions.md notes/sparql-1-1.md In practice, a healthy ontology project therefore has several artifacts and checks: identifiers and definitions, axioms, mappings, competency questions or queries, documentation, provenance, release metadata, and governance. notes/common-core-ontologies-cco.md notes/widoco.md notes/obo-foundry.md

The landscape

Languages and standards

Authoring, collaboration, and publication

  • Protégé is the general-purpose editor presented here: Desktop supports local OWL 2 modeling, reasoning, and querying, while WebProtégé provides collaborative browser editing and a Java plug-in ecosystem surrounds the desktop application. notes/protege.md
  • Different interfaces serve different modelers: Eddy uses the Graphol visual language with design-time validation and OWL 2 export, Fluent Editor uses controlled natural language, OWLGrEd emphasizes whole-ontology graphical editing, and VocBench supports collaborative multilingual OWL, SKOS/SKOS-XL, OntoLex-lemon, and generic RDF assets. notes/eddy.md notes/fluent-editor.md notes/owlgred.md notes/vocbench.md
  • Text-first engineering is viable: ELOT makes one Org document the formal source, documentation, hierarchy, query notebook, and diagram workspace, while Tawny-OWL constructs OWL programmatically in a Clojure DSL and yamlpyowl offers a YAML representation backed by Owlready2. notes/elot.md notes/tawny-owl.md notes/yamlpyowl.md
  • Automation closes the lifecycle: ROBOT covers conversion, extraction, merging, querying, reasoning, repair, reporting, templating, profile validation, and verification; OnToology reacts to GitHub changes by proposing generated documentation, diagrams, evaluation, and JSON-LD context; WIDOCO produces standards-aware documentation, provenance, visualizations, evaluation, and version changelogs. notes/robot.md notes/ontoology.md notes/widoco.md
  • Libraries embed ontology work in applications: OWL API and Apache Jena provide Java integration surfaces, Owlready2 exposes OWL entities and HermiT reasoning as Python objects, FunOWL mirrors OWL functional syntax in Python, and Horned OWL and Owlish provide Rust data structures and processing. notes/owl-api.md notes/apache-jena-ontology-api.md notes/owlready2.md notes/funowl.md notes/horned-owl.md notes/owlish.md

Reasoning, alignment, and access

  • Reasoners are chosen by semantic profile and workload, not by a single “best” ranking: ELK and CEL target tractable EL-family classification, Openllet provides broad OWL 2 DL services, HermiT implements OWL 2 Direct Semantics with hypertableau reasoning, and OWL-RL materializes RDFS/OWL 2 RL consequences by forward chaining in RDFLib. notes/elk.md notes/cel.md notes/openllet.md notes/hermit.md notes/owl-rl.md
  • Alignment is its own engineering problem: AgreementMakerLight emphasizes efficient automated matching, LogMap combines large-scale matching with reasoning and inconsistency repair, OntoAligner offers a modular classical/retrieval/LLM/RAG/embedding/ensemble pipeline, and the Alignment API supplies a shareable RDF correspondence format. notes/agreementmakerlight-aml.md notes/logmap.md notes/ontoaligner.md notes/alignment-api.md
  • Repositories and catalogs are entry points, not substitutes for evaluation: BioPortal supports biomedical browsing, annotation, recommendation, mappings, and APIs; Linked Open Vocabularies catalogs vocabularies and terms; BARTOC spans terminologies and registries; OBO Foundry couples a biological ontology catalog with community principles. DBpedia Archivo adds automated Web discovery, persistent snapshots, and minimum-viability signals for retrievability, licensing and interoperability, and consistency; its stars are not a quality ranking. notes/bioportal.md notes/linked-open-vocabularies.md notes/bartoc-basel-register-of-thesauri-ontologies-classifications.md notes/obo-foundry.md notes/dbpedia-archivo.md
  • Smaller access tools solve bounded jobs: OAK spans adapters, mappings, graph operations, validation, changes, and similarity; obonet reads OBO into a NetworkX multigraph; Tarql maps CSV rows to RDF through SPARQL 1.1 syntax; and SciGraph exposes OWL-centered data as Neo4j, vocabulary, annotation, and REST services. notes/oak.md notes/obonet.md notes/tarql.md notes/scigraph.md

Public knowledge bases and linked-data resources

  • General-purpose resources expose different kinds of knowledge: Wikidata is a community-editable structured-data hub with a SPARQL service; ConceptNet is an open multilingual semantic network with assertion-level source metadata and a JSON-LD API; and Cyc is presented as a commercial logic-and-common-sense reasoning platform whose fetched page supplies product positioning rather than technical ontology or evaluation detail. notes/wikidata.md notes/conceptnet.md notes/cyc.md
  • Integrative and historical resources require version context: KBpedia maps seven public sources into a shared ontology and graph; the fetched YAGO page describes a Wikipedia-, WordNet-, and GeoNames-derived knowledge base but identifies itself as an unmaintained old project page; Freebase is a frozen 2007–2015 dump with Wikidata mappings and no live API; and the archived NELL publication offers RDF releases with several provenance representations. notes/kbpedia.md notes/yago.md notes/freebase.md notes/nell-ontology-and-knowledge-base.md
  • Ecosystem-scale resources show connectivity and provenance, with limits: the Linked Open Data Cloud maps qualifying linked datasets and publishes diagram data; BTC 2019 preserves source-document context in more than two billion crawled quads but is heavily dominated by Wikidata; and the fetched Mosaic Knowledge Graphs source is only an archive wrapper, so it supports no capability claims about the underlying application. notes/the-linked-open-data-cloud.md notes/billion-triple-challenge-btc-2019-dataset.md notes/mosaic-knowledge-graphs.md

Alternative upper and mid-level foundations

  • SUMO combines an upper ontology with extensive formal and lexical infrastructure: it is written in SUO-KIF, mapped to all of WordNet, extended by domain ontologies, and distributed through tooling and translations for theorem proving, Neo4j, and OWL. notes/sumo-suggested-upper-merged-ontology.md
  • GFO and YAMATO emphasize different foundational theories and maturity profiles: GFO covers objects, processes, time, space, roles, functions, facts, and situations and offers a stable OWL core, while its documentation records incomplete tutorial and axiomatization material and possible inconsistency among parallel extensions; YAMATO highlights theories of quality and quantity, representations, and objects, processes, and events, with downloadable Hozo and OWL files and reported domain uses. notes/gfo-general-formal-ontology.md notes/yamato-yet-another-more-advanced-top-level-ontology.md
  • MILO is a bridge rather than another top level: it connects SUMO’s abstract categories to more specific domain vocabulary through KIF declarations, relation constraints, functions, and logical and temporal rules. notes/milo-mid-level-ontology.md

Users and use-cases

The use-cases are unusually broad: GO supplies evidence-backed cross-species gene-function annotations, Uberon integrates animal anatomy, and FoodOn standardizes foods and production processes in the OBO ecosystem. notes/gene-ontology-go.md notes/uberon.md notes/foodon.md FIBO supports shared financial meaning and cross-system federation, CIDOC CRM mediates heterogeneous cultural-heritage records, UCO standardizes cyber-investigation and security concepts, OEO supports energy-system analysis terminology, and Digital Buildings maps physical spaces, equipment, telemetry, and relationships for portable building analytics. notes/fibo.md notes/cidoc-crm-conceptual-reference-model.md notes/unified-cyber-ontology-uco.md notes/open-energy-ontology-oeo.md notes/digital-buildings-ontology-dbo.md Cross-domain vocabularies solve narrower reusable problems: PROV-O exchanges provenance, OWL-Time models instants, intervals, duration, and temporal position, Dublin Core supplies broad metadata properties, Schema.org publishes a large Web-oriented type hierarchy, and GoodRelations describes products and services for machine consumption. notes/prov-o.md notes/owl-time-time-ontology.md notes/dcmi-metadata-terms.md notes/schema-org-schemas.md notes/goodrelations.md

Ontologies + LLMs

The notes contain a small number of direct, documented LLM integrations and a much larger body of non-LLM graph, reasoning, and ML work; the distinction matters because an ontology or knowledge graph used for search, embeddings, or symbolic inference is not automatically an LLM application. notes/ontoemma.md notes/digital-buildings-ontology-dbo.md notes/materials-data-science-and-informatics-awesome-fair-data.md

Learning and construction

OntoLearner provides the clearest end-to-end ontology-learning account in the notes. notes/ontolearner.md It frames learning as term typing, taxonomy discovery, and non-taxonomic relation extraction, with a separate Text2Onto path for extracting terms and types from raw text. notes/ontolearner.md Its LLM-only learners rely on model-internal knowledge, retriever-only learners index training examples as embeddings, and RAG learners retrieve similar ontology examples as domain-specific few-shot context before generation; earlier task outputs may feed later tasks, but tasks can also run in parallel. notes/ontolearner.md Text2Onto additionally uses a Hugging Face generation backend to create synthetic documents, enriching prompts with term-typing, taxonomy, and non-taxonomic graph context so generated passages remain closer to the source ontology. notes/ontolearner.md The documented guidance is conditional rather than triumphant: pure LLMs are recommended mainly for general or well-known domains, RAG for specialized domains, and LLM/symbolic or multi-model ensembles for higher-stakes reliability. notes/ontolearner.md Clear prompts and structured output reduce but do not eliminate hallucination and inconsistency, so labels and relationships still need validation against representative held-out data, multiple metrics, domain criteria, and classical baselines. notes/ontolearner.md Large retrieval contexts can cause memory problems, and nonstandard ontologies can require custom extraction and ontology-specific blank-node filtering. For Text2Onto specifically, the documentation says it works best with instruction-tuned models and ontology context that fits the model window; it presents stricter structured-output validation, repair retries for missing labels, richer graph-context retrieval, and candidate reranking as possible improvements rather than framework-wide requirements. notes/ontolearner.md

A curated Semantic Web list names OntoGPT as a tool for populating semantic schemas from unstructured text with LLMs, but that note records only the directory entry and does not establish a detailed extraction method or evaluation result. notes/semantalytics-awesome-semantic-web.md The same list names SPARQLLM as allowing SPARQL queries to call search engines, LLMs, or vector databases during query execution; again, this is a linked resource description rather than a verified capability study in the harvested note. notes/semantalytics-awesome-semantic-web.md

Alignment and validation

OntoAligner implements direct LLM matching by turning candidate concept pairs—optionally with parent or child context—into yes/no prompts and mapping generated output back to match labels. notes/ontoaligner.md Its documented LLMs4OM RAG flow retrieves target candidates for each source concept and asks an LLM to verify each pair, with concept-only, parent-context, child-context, few-shot, and in-context-vector variants; ensembles can combine these signals with lexical, structural, retrieval, and knowledge-graph-embedding matchers. notes/ontoaligner.md Direct all-pairs LLM matching is quadratic and documented as appropriate only for small ontologies of roughly 200 concepts or fewer, while RAG narrows candidates and logit scoring reduces GPU cost; parser assumptions, context-encoder mismatches, and dataset-specific thresholds remain operational risks. notes/ontoaligner.md

LogMap’s README states that matching can use LLM validation and links LogMapLLM work describing LLMs as alignment oracles, while LogMap itself contributes scalable matching, reasoning, and inconsistency repair. notes/logmap.md That is an explicit project claim but not, in the harvested note, a detailed account of prompts, metrics, or validation guarantees, so it should not be stretched beyond “LLM-assisted mapping validation is supported and associated work is linked.” notes/logmap.md By contrast, OntoEMMA’s learned matcher is logistic regression or an AllenNLP neural network, with no LLM or RAG integration stated; calling every learned alignment pipeline “LLM-based” would erase an important category boundary. notes/ontoemma.md

Agentic authoring, linting, reasoning, querying, and editing

ELOT documents the most complete agentic ontology-engineering loop in the collection. notes/elot.md Its optional Emacs elot-gptel integration lets an LLM inspect resources and conventions, search labels, lint, issue SPARQL queries, invoke ROBOT-backed consistency, unsatisfiability, and explanation checks, mint identifiers under project policy, and edit or restructure the literate ontology source. notes/elot.md The mutation surface is deliberately guarded: writes are project-scoped, disabled by default, confirmation-gated, automatically linted and parsed, and rolled back on failed revalidation. notes/elot.md Those automatic mutation checks establish syntax and OWL parsability, not semantic consistency, so a separate consistency check is still required; moreover, a newly minted CURIE cannot be used as an axiom subject in the same batch. notes/elot.md The feature is also editor-specific in the documented matrix: Emacs is the reference implementation, while VS Code marks AI-assisted authoring, SPARQL, and diagrams as planned. notes/elot.md

Natural language to SPARQL and adjacent retrieval patterns

The Python ecosystem list explicitly describes LangChain’s GraphSparqlQAChain as generating SPARQL SELECT and UPDATE from natural language, executing against files, endpoints, or triple stores, and returning natural-language responses. notes/pysemtec-semantic-python-overview.md Because this evidence comes from a curated overview whose authors warn that it may be incomplete and biased, it establishes that the integration is listed and characterized that way, not its accuracy, safety, or production readiness. notes/pysemtec-semantic-python-overview.md A separate directory lists LoRiS for generated natural-language representations of SPARQL queries over Wikidata and DBpedia, but this is query verbalization in a resource list, not evidence of a general conversational ontology agent. notes/semantalytics-awesome-semantic-web.md

Two further pages merely point outward: OAK’s documentation index contains a guide titled “How to use Large Language Models (LLMs) with OAK,” and LinkML’s index contains “How to Generate AI prompts,” but neither harvested page states the guide’s method or capability. notes/oak.md notes/linkml.md A knowledge-graph directory similarly lists a book about putting graphs from structured and unstructured sources into LLM and RAG applications and says Termboard can use common chatbots, but the note is a catalog description rather than primary implementation evidence. notes/totogo-awesome-knowledge-graph.md

Finally, several tempting examples are explicitly KG-only or non-LLM in these notes: CCO recommends reusable graph patterns for mapping and querying, Digital Buildings uses its inheritance graph for navigation and search expansion, FAIR resources include graphs and conventional ML extraction, and OntoEMMA uses classical/neural matching without an LLM. notes/common-core-ontologies-cco.md notes/digital-buildings-ontology-dbo.md notes/materials-data-science-and-informatics-awesome-fair-data.md notes/ontoemma.md

Pitfalls

  • Open-world semantics defeats database-style intuition: absence of a fact does not imply its negation, different names need not identify different things, and SKOS deliberately defines few integrity conditions. notes/description-logics-dls.md notes/skos-simple-knowledge-organization-system.md
  • More expressivity is not free: richer description-logic operators raise inference complexity, OWL profiles restrict the language for particular implementation benefits, and tableau reasoning can become intractable through large disjunctions, cardinalities, existential expansion, or many different-individual assertions. notes/description-logics-dls.md notes/owl-2-web-ontology-language.md notes/openllet.md
  • Syntactic success is not semantic correctness: ELOT’s automatic LLM-edit validation checks linting and parsing but requires a separate consistency run, while OWL 2 warns that rule-based RL reasoning can be sound yet incomplete outside its structural and query conditions. notes/elot.md notes/owl-2-web-ontology-language.md
  • Representation conversions can be partial or lossy: SciGraph’s OWL-to-Neo4j mapping cannot round-trip and omits some SKOS arrangements, while obonet’s ontology edge direction conflicts with everyday NetworkX ancestor/descendant intuition and collapsing to a simple graph discards relationship distinctions. notes/scigraph.md notes/obonet.md
  • Identifiers and names require policy, not improvisation: RDF alone does not make producers agree on identifier semantics, while ELOT documents XML-name-invalid numeric counters, lossy slug schemes, and the principle that labels—not CURIEs—should carry human-readable meaning. notes/rdf-resource-description-framework-1-1.md notes/elot.md
  • Scope creep harms reuse: BFO deliberately excludes specialized scientific content, CCO tells adopters to publish domain extensions rather than expand the common core indefinitely, and Digital Buildings advises reusing types before proposing validated extensions. notes/bfo-basic-formal-ontology.md notes/common-core-ontologies-cco.md notes/digital-buildings-ontology-dbo.md
  • Version, profile, and tool compatibility are part of the model’s operational life: CCO warns that different OWL API and Protégé versions can produce spurious diffs, HermiT documents specific OWL API and plug-in compatibility boundaries, and gist had a major release that broke compatibility and required migrations. notes/common-core-ontologies-cco.md notes/hermit.md notes/gist.md
  • Domain transfer and generic parsing assumptions can fail: Digital Buildings marks some guidance as Google-campus-specific. Taken together, the OntoAligner and OntoLearner documentation shows that generic parsing or extraction depends on common RDF/OWL constructs and requires customization when an ontology represents labels, hierarchies, or other semantics differently. notes/digital-buildings-ontology-dbo.md notes/ontoaligner.md notes/ontolearner.md
  • Catalog inclusion is not endorsement or current maintenance: OBO Foundry’s homepage does not prove every listed ontology satisfies every principle, OntologyDesignPatterns.org says catalogue presence is not certification, and curated directories explicitly warn of abandoned, stalled, incomplete, or biased listings. notes/obo-foundry.md notes/ontologydesignpatterns-org.md notes/semantalytics-awesome-semantic-web.md notes/pysemtec-semantic-python-overview.md
  • LLM assistance adds scaling and trust problems rather than removing ontology engineering: Taken together, OntoAligner and OntoLearner document quadratic matching, dataset-specific thresholds, hallucination, context, validation, and baseline concerns. A separate authoring example, ELOT, keeps mutation confirmation, rollback, and separate semantic checks. notes/ontoaligner.md notes/ontolearner.md notes/elot.md

Where the idea came from

The word is older than the technology by about four hundred years, and the two senses have never fully separated. Aristotle named the subject in Book IV of the Metaphysics as “first philosophy,” the study of being qua being. The Latin ontologia first appears in Jacob Lorhard’s Ogdoas Scholastica in 1606, with priority contested, and was put into circulation by Christian Wolff in 1730, who defined it as the science of being in general as being, and who split general metaphysics from special metaphysics in a way that resembles the later split between upper and domain ontologies. notes/ontology-word-history.md notes/bfo-basic-formal-ontology.md

Quine made the idea tractable for engineers by moving the question. Instead of asking what exists, “On What There Is” asks what a theory must quantify over for its statements to be true. Ontology becomes a property of a language rather than a claim about reality, which is the form the engineering discipline inherited. notes/quine-ontological-commitment.md

Gruber’s 1993 definition — an explicit specification of a conceptualization — created the computer-science term, and Borst added the requirement that the conceptualization be shared. Guarino, Oberle and Staab later reconstructed conceptualization formally because the informal wording had produced years of argument. notes/gruber-ontology-definition.md Cyc, the era’s largest knowledge-engineering programme, is the field’s reference experiment on the cost side: the largest sustained attempt to answer how much of the world must be written down before a knowledge base pays off, at a person-century of effort by 1995. notes/cyc-lenat-1995.md

The 2001 Scientific American article moved the idea to web scale and set the design brief that RDF and OWL answer, with SPARQL arriving later to query the result. notes/semantic-web-2001-vision.md notes/rdf-resource-description-framework-1-1.md notes/owl-2-web-ontology-language.md notes/sparql-1-1.md The open-web version did not arrive; the machinery survived in governed domains and was rebranded around knowledge graphs after 2012. notes/semantic-web-retrospective.md notes/knowledge-graph-turn.md The retrospective’s own summary is that the survivors are “governed, funded, bounded domains where somebody has an incentive to curate.” notes/semantic-web-retrospective.md The corpus’s own success stories fit that pattern across gene function, anatomy, food, cultural heritage, finance, cyber investigation, energy, and buildings. notes/gene-ontology-go.md notes/uberon.md notes/foodon.md notes/cidoc-crm-conceptual-reference-model.md notes/fibo.md notes/unified-cyber-ontology-uco.md notes/open-energy-ontology-oeo.md notes/digital-buildings-ontology-dbo.md

Criticism worth answering

Three critiques have survived and each attacks a different layer. Doctorow’s “Metacrap” attacks the incentive layer: voluntary metadata fails because people lie, are lazy, are careless, and misjudge themselves, and because schemas and metrics are never neutral. His constructive residue is the distinction between observed signal and declared signal: he endorses implicit metadata, produced as a by-product of behaviour, citing Google’s use of link structure. notes/doctorow-metacrap.md

Shirky attacks the applicability layer. His target is narrower than usually reported: designing categories in advance to cover cases that have not happened yet, applied to unbounded corpora with uncoordinated cataloguers and untrained users. His library examples show that a classification optimises for its physical substrate rather than for concepts, and his Yahoo example shows designers reimposing that constraint after it disappeared. notes/shirky-ontology-is-overrated.md

Bowker and Star attack the consequences layer, and on the evidence of this corpus it is the critique the technical sources engage with least. Categories act on people. Their concept of torque names what happens when a person and a classification disagree and the classification wins. notes/bowker-star-sorting-things-out.md That argument becomes concrete and legal once an ontology is installed inside a state institution. notes/palantir-ontology-critique.md

These do not overturn the pitfalls already listed in this document. They add a category of pitfall the standards do not cover, because open-world semantics, profile complexity, and lossy conversion are all failures of the model against the world, while torque is a failure of the world against the model.

Operational ontologies: the other tradition

Palantir’s Ontology uses the word for something else, and it belongs to a different lineage than the W3C stack. Palantir’s own architecture documentation says it models the interconnected decisions of an enterprise rather than its data, and states outright that it is not a semantic layer. It pairs semantic elements, meaning object types, properties, and link types, with kinetic elements, meaning action types, functions, and interfaces, enforces security across data, logic, and action at once, and commits user edits back as data. notes/palantir-ontology.md

The contrast with OWL is instructive in both directions. OWL has model-theoretic semantics, an open-world assumption, reasoners that derive consequences, and public shared vocabularies, none of which the Palantir Ontology has. The Palantir Ontology has actions and permissions as first-class primitives, which OWL cannot express at all. The comparison is this corpus’s, not Palantir’s: its documentation never mentions OWL, the W3C, reasoners, or world assumptions. notes/palantir-ontology.md notes/description-logics-dls.md notes/owl-2-web-ontology-language.md A team choosing between them is choosing between deriving what follows and governing what may be done.

Ontologies after large language models

The corpus documents tooling that uses models to learn, align, and author ontologies, with consistent warnings about hallucination, quadratic matching cost, dataset-specific thresholds, and the gap between syntactic validity and semantic correctness. notes/ontolearner.md notes/ontoaligner.md notes/elot.md The broader argument about replacement is separate and worth keeping separate: whether a model can replace an ontology, build one, and use one are three different questions, and this corpus answers them differently. notes/llm-ontology-debate.md

Learning path

  1. Begin with the data model, not an editor: read the RDF note until the abstract graph is clearly separate from RDF/XML, then read Turtle and JSON-LD as contrasting concrete serializations. notes/rdf-resource-description-framework-1-1.md notes/turtle.md notes/json-ld.md
  2. Learn the lightest semantic layer: use RDFS to understand classes, instances, subclass hierarchies, labels, and the inference meaning of global domains and ranges. notes/rdf-schema-rdfs-1-1.md
  3. Learn when a vocabulary is not a formal domain theory: study SKOS labels, schemes, collections, and mappings, especially its explicit distinction between thesaurus concepts and OWL classes. notes/skos-simple-knowledge-organization-system.md
  4. Add formal ontology semantics: read the description-logics note for TBox/ABox, open-world semantics, and the expressivity/complexity trade-off, then use the OWL 2 overview to map those ideas onto language features, profiles, syntaxes, and semantics. notes/description-logics-dls.md notes/owl-2-web-ontology-language.md
  5. Author one small ontology in Protégé: create classes, properties, individuals, labels, and a few restrictions, then classify it and try Manchester class expressions in the DL Query tab. notes/protege.md notes/dl-query-class-expression.md
  6. Query data directly: work through SPARQL’s triple patterns, optional and alternative patterns, named graphs, property paths, SELECT, CONSTRUCT, ASK, ordering, aggregation, and limits. notes/sparql-1-1.md
  7. Separate inference from validation: compare an EL reasoner such as ELK with an OWL 2 DL reasoner such as Openllet, then encode data expectations in SHACL or ShEx and inspect machine-readable validation results. notes/elk.md notes/openllet.md notes/shacl-shapes-constraint-language.md notes/shex-shape-expressions.md
  8. Study layered reuse and governance: compare BFO, CCO, and a domain ontology such as GO, then read OBO Foundry’s principles and OntologyDesignPatterns.org’s pattern-oriented workflow before designing a larger model. notes/bfo-basic-formal-ontology.md notes/common-core-ontologies-cco.md notes/gene-ontology-go.md notes/obo-foundry.md notes/ontologydesignpatterns-org.md
  9. Automate repeatable engineering: use ROBOT for transformations and checks, WIDOCO for release documentation and provenance, and OnToology to understand repository-triggered generation and review. notes/robot.md notes/widoco.md notes/ontoology.md
  10. Embed ontologies in software: choose Owlready2 for a Python object model with reasoning, Jena or OWL API for Java, and OAK for a higher-level multi-adapter utility surface; inspect obonet only when the bounded OBO-to-NetworkX job fits. notes/owlready2.md notes/apache-jena-ontology-api.md notes/owl-api.md notes/oak.md notes/obonet.md
  11. Learn alignment after single-ontology work is comfortable: start with the matching problem and evaluation community, then compare AgreementMakerLight, LogMap’s logic-aware repair, and OntoAligner’s modular classical-to-LLM spectrum. notes/ontology-matching.md notes/ontology-alignment-evaluation-initiative-oaei.md notes/agreementmakerlight-aml.md notes/logmap.md notes/ontoaligner.md
  12. Approach LLMs last, with baselines and guards: reproduce OntoLearner’s LLM-only, retrieval, and RAG distinctions, compare OntoAligner’s retrieval-narrowed alignment, and study ELOT’s confirmation, rollback, lint, parse, and separate semantic-check boundaries before allowing model-written ontology changes. notes/ontolearner.md notes/ontoaligner.md notes/elot.md