## Definition **Ontologies** and **semantic networks** are structured forms of [[Knowledge Representation]] that model concepts and relations as labelled graphs. Foundational for the Semantic Web, biomedical knowledge bases, and many enterprise data integration systems. ## Semantic Networks A semantic network is a graph: - **Nodes** represent concepts (objects, categories, properties). - **Edges** represent relations (is-a, part-of, has-property). The classic example: WordNet, a lexical database where words connect via synonymy, hyponymy, meronymy. Strengths: visual, intuitive. Limitations: semantics often informal, hard to reason rigorously. ## Ontologies An ontology is a **formal, explicit specification of a shared conceptualisation** (Gruber, 1993). Three properties separate it from a semantic network: 1. **Formal semantics.** Underlying logic (typically a description logic). 2. **Explicit definitions.** Concepts and relations are precisely specified. 3. **Shared.** Designed for use by multiple agents or systems. ## Components - **Classes (concepts):** `Person`, `Vehicle`, `Disease`. - **Individuals:** `Alice`, `Tesla-Model-3`. - **Object properties:** `worksFor`, `livesIn`. - **Data properties:** `hasAge`, `hasName`. - **Axioms:** logical statements constraining the model (e.g., disjointness, inverse, transitivity). ## OWL and the Semantic Web **OWL** (Web Ontology Language, W3C standard) is the dominant ontology language. Variants: - **OWL Lite** — simplest, decidable. - **OWL DL** — balances expressivity and decidability; based on the SHOIN(D) description logic. - **OWL Full** — most expressive, undecidable. Storage and query: **RDF triples** (subject-predicate-object) in **triplestores** (GraphDB, Stardog, Apache Jena). Query language: **SPARQL**. ## Notable Ontologies - **SNOMED CT** — clinical terminology, ~350k concepts. - **Gene Ontology** — molecular biology. - **DBpedia / Wikidata** — encyclopedic knowledge graphs from Wikipedia. - **Schema.org** — web markup vocabulary. - **FOAF** — social relations ("Friend of a Friend"). ## Reasoning DL reasoners (HermiT, Pellet, ELK) perform: - **Classification** — compute the subsumption hierarchy. - **Consistency checking** — detect logically conflicting axioms. - **Instance retrieval** — find all instances of a class. ## Relevance in the LLM Era Knowledge graphs and ontologies have not been displaced by LLMs — they complement them: - **Grounded retrieval.** [[Retrieval-Augmented Generation]] over a knowledge graph constrains the LLM to factual entities. - **Auditing.** Ontology axioms can verify LLM outputs. - **Integration.** Enterprise data integration still relies on explicit schemas. ## Related - [[Knowledge Representation]] - [[First-Order Logic]] - [[Retrieval-Augmented Generation]]