Semantic Web

Abstract

The Semantic Web is an extension of the World Wide Web in which data carries explicit, machine-readable meaning, so that programs can retrieve, combine, and reason over it rather than merely display pages. This article traces its descent from the semantic-network model of memory in cognitive psychology to its engineering realization as a stack of standards: the RDF triple, the ontology, and a query language that turn documents into a queryable knowledge graph. It explains how machines derive new facts by entailment, why shared ontologies let meaning travel between systems, and how linked data joins independent datasets into one global graph. It closes with the current fusion of knowledge graphs and large language models. Interactive demonstrations let the reader build RDF triples, watch entailment propagate through a class hierarchy, and run a pattern query over a small triple store.

Keywords: semantic web, knowledge representation, ontology, linked data, knowledge graph

The Semantic Web is a proposal to make the content of the Web understandable to machines, not only to people. On the ordinary Web a page is a document laid out for a human reader; a program can fetch it and display it but cannot, in general, tell what it means. The Semantic Web adds a second layer in which each fact is stated in a form a machine can interpret and combine — so that data from different sources can be linked, queried, and reasoned over as though it were one database (Berners-Lee, Hendler & Lassila, 2001). The intellectual core of that proposal is not new to computing: it is the idea, developed decades earlier in the psychology of memory, that meaning can be represented as a network of concepts joined by labelled relations.

What the Semantic Web Is

The Semantic Web is best understood by contrast with the Web it extends. The document Web is a space of pages connected by untyped links: a link records that one page points to another, but not why or how the two are related. A search engine can index the words on a page, yet the page's assertions — that a particular protein regulates a particular gene, that a city is the capital of a country — remain locked in prose that only a human can unpack. The Semantic Web replaces the untyped link with a typed one and the opaque page with an explicit statement, so that the relation itself is part of the data (Berners-Lee, Hendler & Lassila, 2001).

The unit of that explicit statement is the triple: a subject, a predicate, and an object, asserting that the subject stands in the named relation to the object. Paris — capitalOf — France is a triple; so is aspirin — treats — headache. Because the object of one triple can be the subject of another, triples chain together into a directed, labelled graph — a knowledge graph — in which nodes are things and edges are named relations. This is the same representational commitment that cognitive psychology made when it modelled human semantic memory as a network of concept nodes joined by labelled links, and the resemblance is not accidental: the Web's designers drew directly on that tradition (Berners-Lee, Hendler & Lassila, 2001).

From Semantic Networks to the Web

The graph idea entered psychology before it entered computing. The first computational model of human semantic memory represented word meanings as nodes in a network, connected by associative and categorical links, and simulated how a reader might retrieve and combine those meanings (Quillian, 1967). The model was made experimentally testable by storing category membership and properties on labelled links and predicting that verifying a sentence would take longer the more links a reader had to traverse — a prediction reaction-time data broadly confirmed (Collins & Quillian, 1969). When the strict hierarchy proved too rigid, it was replaced by spreading activation, in which activating one concept sends activation outward to its neighbours, decaying with distance (Collins & Loftus, 1975). The broader study of how such meaning is represented and retrieved is the subject of semantics and semantic memory.

Two properties of that psychological work carried directly into the engineering of the Semantic Web. First, meaning lives in the relations, not only the nodes: a concept is defined by its position in the web of links, so the links must be labelled and typed. Second, a network supports inference — from stored links a system can derive facts never explicitly stored, exactly as inheritance let canary acquire can fly from bird. The Semantic Web takes both commitments and asks what standards are needed to make them hold not inside one program's memory but across the whole Web, where the nodes are published by strangers who never coordinated. A parallel lexical resource, WordNet, organized tens of thousands of English words into a network of synonym sets linked by semantic relations, and became one of the first large machine-readable semantic networks in practical use (Miller, 1995).

The Layered Architecture

The Semantic Web is engineered as a stack of layers, each adding expressive power to the one below. At the base, every thing is named by a URI, a globally unique identifier, so that two datasets referring to the same entity can be recognized as doing so. On top of naming sits the Resource Description Framework (RDF), which fixes the triple as the universal data model: all Semantic Web data, however complex, decomposes into subject–predicate–object statements. The demonstration below lets the reader assemble triples and watch them accumulate into a graph.

Figure 1

The layered architecture of the Semantic Web

Above RDF sit the schema and ontology layers, which say what the vocabulary means, and above those a query layer and a logic layer that let programs retrieve and reason. RDF Schema introduces classes and the subClassOf and type relations; the Web Ontology Language (OWL) adds the richer constructs — property restrictions, disjointness, cardinality — whose formal grounding in description logic gives a reasoner well-defined entailments to compute (Horrocks, 2008). The query language SPARQL then treats the global graph as a database, matching patterns of triples against it. The layering is deliberate: each level is useful on its own, and a publisher can climb only as high as an application needs.

Ontologies and Shared Meaning

A triple is only as meaningful as the vocabulary its predicate is drawn from. If one dataset writes treats and another writes isIndicatedFor, a machine cannot combine them unless something declares the two relations equivalent. An ontology is that something: a formal, explicit specification of the concepts in a domain and the relations among them, shared so that independent systems commit to the same meanings (Gruber, 1993). The definition was later sharpened to emphasize that an ontology is a specification of a shared conceptualization — formal enough for a machine, explicit about its assumptions, and agreed among the parties who use it (Studer, Benjamins & Fensel, 1998).

The idea has a longer pedigree than computing. What counts as an ontology, and how a specification of concepts relates to the philosophical study of what exists, was worked out in careful detail before it became a Web standard (Guarino, Oberle & Staab, 2009). For the Semantic Web the payoff is reasoning: because an ontology states the rules that govern its classes and properties, a reasoner can derive facts that no one asserted directly. If the ontology says every canary is a songbird and every songbird a bird, then asserting that Tweety is a canary entails that Tweety is a bird, a songbird, and whatever else those classes inherit. The demonstration below lets the reader assert a single fact and watch the entailed facts propagate up a class hierarchy.

Linked Data and Knowledge Graphs

An ontology lets meaning travel within a domain; linked data lets it travel across the whole Web. Linked data is a set of practices for publishing RDF so that datasets interconnect: name things with dereferenceable URIs, serve useful RDF when a URI is looked up, and include links to other datasets' URIs so that a machine following them discovers more data (Bizer, Heath & Berners-Lee, 2009). Followed at scale, these practices weave independently published datasets — encyclopaedic, governmental, biomedical — into a single global graph in which a query can begin in one source and traverse into another. The demonstration below runs a pattern query over a small triple store, the operation SPARQL performs across that global graph.

The contemporary form of this vision is the knowledge graph: a large, general graph of entities and their relations, built by integrating many sources and used to answer queries and support applications (Hogan et al., 2021). Knowledge graphs are the direct descendants of the Semantic Web's data model, and the discipline has extended them with representation-learning methods that embed entities and relations as vectors, so that missing links can be predicted statistically rather than only derived logically (Ji et al., 2022). The graph structure the psychology of memory first proposed has, in this lineage, become industrial infrastructure.

Worked Example

To see how entailment turns a few asserted facts into many, consider a small ontology with a chain of four classes, each a subclass of the next: Canary is a subclass of Songbird, Songbird of Bird, and Bird of Animal. That is 3 asserted subClassOf triples. Because subClassOf is transitive, a reasoner computes its transitive closure: with four classes in a chain there are C(4, 2) = 6 ordered subclass pairs in total, so the closure adds 6 − 3 = 3 new subclass facts — Canary ⊑ Bird, Canary ⊑ Animal, and Songbird ⊑ Animal.

Now assert one instance fact: Tweety is of type Canary — 1 asserted type triple. The entailment rule for type over subClassOf says that an instance of a class is an instance of all its superclasses, so the reasoner derives that Tweety is also of type Songbird, Bird, and Animal — 3 new type facts. In total the ontology was given 3 + 1 = 4 asserted triples and the reasoner entailed a further 3 + 3 = 6, so the closed graph contains 10 triples. The ratio of derived to asserted knowledge grows sharply as hierarchies deepen and branch, which is precisely why explicit ontologies repay their cost — the same multiplication of stored into inferred facts that inheritance gave the hierarchical network model of memory. The InferenceDemo above performs exactly this propagation.

Discussion

Two decades on, the Semantic Web's record is mixed, and its own community has said so plainly. The full vision — a Web of formally described data over which agents reason autonomously — was only partly realized; the heavyweight ontology and logic layers proved harder to deploy at Web scale than the lightweight data-linking practices below them (Hitzler, 2021). What did succeed was linked data and, above it, the knowledge graph, which large organizations now run as core infrastructure. The retrospective judgment of the field's founders was already cautious at the halfway mark: the technology worked, but adoption depended on incentives to publish structured data that were slow to arrive (Shadbolt, Hall & Berners-Lee, 2006).

For cognitive psychology the interesting point is what the engineering effort revealed about the original idea. A semantic network is trivial to draw and deceptively hard to make interoperable: the moment two networks must be merged, every implicit assumption about what a link means has to be made explicit, which is the entire burden the ontology layer carries. The Semantic Web is, in effect, a large-scale test of whether the network theory of meaning can be made precise enough to share — and its partial success suggests that meaning-as-relations scales, while meaning-as-formal-logic scales only where the domain rewards the effort. Table 1 compares the layers of the architecture and what each contributes.

Table 1. The layers of the Semantic Web architecture and their analogues in the psychology of semantic memory. The mapping is close at the lower layers, where both are graphs of labelled relations, and looser at the ontology layer, where the Web demands a formal precision human memory does not.

Current Directions

The most active front is the meeting of knowledge graphs with large language models, two approaches to machine knowledge with complementary weaknesses. A language model holds vast, fluent, but unsourced and sometimes fabricated knowledge; a knowledge graph holds precise, verifiable, but incomplete and costly-to-build knowledge. Current research maps how each can repair the other — grounding a model's answers in a graph's verified facts, and using a model to help populate and extend a graph — a programme laid out in a recent roadmap for unifying the two (Pan et al., 2024). The knowledge graph, in this pairing, is where the Semantic Web's insistence on explicit, checkable meaning re-enters an era otherwise dominated by statistical models.

A second direction is representation learning over graphs themselves. Embedding entities and relations into continuous vector spaces lets systems predict plausible missing links and reconcile duplicate entities across datasets — statistical inference layered on top of the logical inference the ontology provides (Ji et al., 2022). The convergence is striking to anyone who knows the psychology: the field is rediscovering, at Web scale, that a graph of labelled relations and a high-dimensional vector space are two views of the same knowledge, the very reconciliation now under way between network and distributional models of human meaning.

Common Misconceptions

The Semantic Web is a different network from the Web.

It is a layer on the same Web, using the same URIs and HTTP. It adds machine-readable statements about resources; it does not replace the document Web (Berners-Lee, Hendler & Lassila, 2001).

An ontology is just a controlled vocabulary or a taxonomy.

A taxonomy only classifies; an ontology also states formal relations and constraints that license inference, which is what lets a reasoner derive new facts (Gruber, 1993; Horrocks, 2008).

The Semantic Web failed and is obsolete.

Its heavyweight logic layers saw limited uptake, but its data model lives on: linked data and knowledge graphs are the Semantic Web in production at enormous scale (Hitzler, 2021; Hogan et al., 2021).

Knowledge graphs are unrelated to the psychology of memory.

The triple graph is the same representation as the semantic network first proposed for human memory; the engineering inherited the psychology's core idea that meaning is carried by labelled relations (Quillian, 1967; Collins & Loftus, 1975).

Glossary

Class.

A named category in an ontology whose members share defining properties; classes form the subclass hierarchies over which a reasoner inherits facts.

Description logic.

A family of formal logics with decidable reasoning that provides the mathematical foundation of the Web Ontology Language.

Entailment.

A fact that follows necessarily from asserted facts under an ontology's rules, and that a reasoner can therefore derive without its being stated.

Instance.

An individual that is a member of a class; asserting an instance's type lets a reasoner infer its membership in every superclass of that class.

Knowledge graph.

A large graph of real-world entities and their relations, built by integrating many sources and queried as a unified store of facts.

Linked data.

A set of practices for publishing RDF with dereferenceable URIs and cross-dataset links, weaving independent datasets into one global graph.

Ontology.

A formal, explicit specification of a shared conceptualization: the classes, properties, and constraints of a domain, agreed so that independent systems interoperate.

RDF.

The Resource Description Framework, the standard data model of the Semantic Web, in which every statement is a subject-predicate-object triple.

Reasoner.

A program that computes the entailments of a set of RDF statements under a given ontology, making implicit facts explicit.

Semantic network.

A representation of knowledge as concept nodes joined by labelled relations; proposed for human memory and inherited by the Semantic Web's graph model.

SPARQL.

The standard query language of the Semantic Web, which retrieves data by matching graph patterns of triples against an RDF store.

Triple.

The atomic statement of RDF, asserting that a subject stands in a named predicate relation to an object; a single labelled edge in the graph.

URI.

A Uniform Resource Identifier, the globally unique name that lets independent datasets refer unambiguously to the same thing.

Web Ontology Language (OWL).

The standard ontology language of the Semantic Web, grounded in description logic, adding classes, property restrictions, and constraints beyond RDF Schema.

Key Researchers

Tim Berners-Lee

(b. 1955). Inventor of the World Wide Web and director of the World Wide Web Consortium; professor at MIT and the University of Oxford; the originator and chief advocate of the Semantic Web vision. ORCID - Wikipedia - W3C Page

Christian Bizer

Professor of information systems at the University of Mannheim; co-author of the linked-data principles and a founder of the DBpedia project that extracted a large knowledge graph from Wikipedia. Google Scholar - Faculty Page

Thomas R. Gruber

(b. 1959). Computer scientist and designer; author of the definition of an ontology that computer science adopted, and later a co-creator of the Siri intelligent assistant. Wikipedia

James A. Hendler

(b. 1957). Director of the Institute for Data Exploration and Applications at Rensselaer Polytechnic Institute; co-author of the founding Semantic Web paper and a leader in agent-based and web science. Wikipedia - Google Scholar - Faculty Page

Ian Horrocks

(b. 1958). Professor of computer science at the University of Oxford; a principal designer of the Web Ontology Language and of the description-logic reasoners that compute its entailments. Wikipedia - Google Scholar - Faculty Page

George A. Miller

(1920-2012). Founding cognitive psychologist; creator of WordNet at Princeton, whose network of synonym sets became one of the earliest large machine-readable semantic networks. Wikipedia

Nigel Shadbolt

(b. 1956). Principal of Jesus College and professor of computer science at the University of Oxford; chairman of the Open Data Institute and a leading figure in the linked-open-data movement. Wikipedia - Faculty Page

Frequently Asked Questions

What is the Semantic Web?

The Semantic Web is an extension of the World Wide Web that gives data explicit, machine-readable meaning, so programs can retrieve, combine, and reason over facts rather than only display documents. It represents knowledge as a graph of subject-predicate-object triples over which machines can compute (Berners-Lee, Hendler & Lassila, 2001).

How is the Semantic Web related to cognitive psychology?

Its core representation, a graph of concept nodes joined by labelled relations, is the semantic-network model that cognitive psychology first proposed for human memory, in which meaning is carried by the links between concepts and new facts are derived by inheritance (Quillian, 1967; Collins & Loftus, 1975).

What is an RDF triple?

An RDF triple is the atomic statement of the Semantic Web: a subject, a predicate, and an object, asserting that the subject stands in the named relation to the object. Because triples chain together, they form a directed, labelled graph: a knowledge graph (Berners-Lee, Hendler & Lassila, 2001).

What is an ontology in computer science?

An ontology is a formal, explicit specification of a shared conceptualization: the classes, properties, and constraints of a domain, agreed so that independent systems interpret data the same way and a reasoner can derive new facts (Gruber, 1993; Studer, Benjamins & Fensel, 1998).

What is linked data?

Linked data is a set of practices for publishing RDF so that datasets interconnect: name things with dereferenceable URIs, return useful RDF when they are looked up, and link to other datasets' URIs, weaving independent sources into one global graph (Bizer, Heath & Berners-Lee, 2009).

How does a machine reason on the Semantic Web?

A reasoner applies the formal rules an ontology states, such as class inheritance and property transitivity, to compute entailments: facts that follow necessarily from the asserted triples but were never stated, exactly as inheritance let a canary acquire the properties of a bird (Horrocks, 2008).

Did the Semantic Web succeed?

Partly. Its heavyweight ontology and logic layers saw limited uptake, but its lightweight data model succeeded enormously: linked data and knowledge graphs are the Semantic Web running in production at large organizations today (Hitzler, 2021; Hogan et al., 2021).

How do knowledge graphs relate to large language models?

They are complementary: language models hold fluent but unsourced knowledge, knowledge graphs hold precise but incomplete knowledge, and current research grounds model outputs in graph facts while using models to help build graphs (Pan et al., 2024).

References