Foundations of Knowledge Graphs: Concepts, History, and Architecture

The Knowledge Question: Why Structure Matters When Models Can Read

A language model in 2026 can produce a fluent paragraph on almost any topic, yet if you ask it for "all Nobel laureates in physics who held a dual appointment in a computer science department," it will often hallucinate plausible names. Statistical systems compress the world into parameter space; knowledge graphs preserve it in declared relationships. The distinction is not academic — it is the difference between a system that *sounds* knowledgeable and one that can *audit* its own claims. This chapter opens our investigation by asking three deceptively simple questions: What is a knowledge graph, where did the idea come from, and what are the architectural layers every working system must coordinate. The answers, it turns out, force us into a richer debate about what it means to represent knowledge at all.

Defining the Object: What Exactly Is a Knowledge Graph

A knowledge graph is a directed, labeled multigraph in which nodes represent entities of interest, edges represent typed relationships between those entities, and both nodes and edges carry semantic identifiers drawn from a defined vocabulary. The graph is *directed* because "Alice advises Bob" is not the same statement as "Bob advises Alice." It is *labeled* because "advises," "cites," and "co-authored-with" carry different semantics that the system must preserve and reason over. It is a *multigraph* because two entities can share more than one type of relationship simultaneously — Alice may both *advise* and *co-author with* Bob — and collapsing these into a single edge destroys information.

The formal substrate is the triple, a statement of the form (subject, predicate, object). Triples are the atomic unit of every knowledge graph. Given the triples `(Alice, advises, Bob)`, `(Bob, affiliatedWith, Stanford)`, and `(Stanford, locatedIn, California)`, we can answer "Where does Alice's advisee work?" by composing two edges — an operation called *multi-hop reasoning*. That compositional power is precisely what flat relational tables sacrifice, and it is the reason graph-structured representations have returned to center stage after decades of being eclipsed by pure statistical methods.

Three properties distinguish a knowledge graph from a generic graph database:

1. Semantic typing. Every node and edge carries a type drawn from an ontology or schema. Without types, a graph is merely a topology; with types, it becomes a model of a domain. 2. Identity and disambiguation. Each entity has a stable, globally unique identifier — often an Internationalized Resource Identifier (IRI) — that survives name changes and disambiguates same-named entities. The node `http://example.org/person/AliceChen` is not just the string "Alice Chen." 3. Open-world assumption. Most KGs do not assume the absence of facts simply because they are not present. If we do not know whether Alice advises Carol, we answer "unknown," not "no." This is a sharp departure from the closed-world semantics of classical databases and is essential for systems that must grow over time without invalidating existing queries.

These three properties — typed structure, stable identity, and open-world semantics — are not optional. Strip any of them away and you do not have a knowledge graph in the sense this book uses the term; you have a more limited artifact such as a property graph, a folksonomy, or an undirected social network.

The Distinction Between Knowledge Graphs and Knowledge Bases

The literature uses "knowledge graph" and "knowledge base" somewhat interchangeably, but a useful distinction has emerged. A knowledge base is the broader category of structured knowledge stores; a knowledge graph is a knowledge base that explicitly uses graph-structured storage as its primary representation. WordNet is a lexical knowledge base but not a graph in the operational sense. Cyc is a knowledge base with a graph-like internal structure. The Google Knowledge Graph, Wikidata, and DBpedia are knowledge graphs in the strict sense. Throughout this book we will use "KG" to mean knowledge graph and treat "KB" as the more general category.

Historical Evolution: From Semantic Networks to Hybrid Intelligence

The intellectual history of knowledge graphs is a story of repeated reinvention, each wave solving the limitations of its predecessor while introducing new constraints. Understanding this lineage matters because today's design choices are unintelligible without the problems they were forged to solve.

1960s — Quillian and the Semantic Network Hypothesis

Allan Collins and M. Ross Quillian proposed semantic networks in 1968 as a cognitive model of human associative memory. Quillian's hypothesis was that concepts are stored as nodes and that meaning emerges from the pattern of their connections. His system could answer questions like "Is a canary a bird?" by traversing the network through inheritance links. The contribution was not technological but conceptual: knowledge could be *declaratively* represented as relationships rather than as procedural code. The limitation was equally conceptual — semantic networks of this era lacked formal semantics, so two researchers could not agree on what "is-a" meant in their respective graphs.

1970s — Frames, Scripts, and the Limits of Association

Marvin Minsky's frame formalism (1975) added structure to nodes by packaging them as collections of slots and fillers. A `Bird` frame had slots for `color`, `habitat`, and `can-fly`. Roger Schank's scripts encoded event sequences. These ideas powered the first wave of natural language understanding systems and, eventually, expert systems. The lesson from this era was that pure associative structure is insufficient; *schemas* — patterns of expected attributes and values — are necessary for any system to reason about incomplete information.

1980s — Description Logics and the Birth of Formal Semantics

The most consequential development of the 1980s was the emergence of description logics (DLs), a family of formal languages designed to give semantic networks a precisely defined meaning. KL-ONE, LOOM, and CLASSIC established the pattern that persists to this day: a TBox (terminology box) defines classes and properties, an ABox (assertional box) populates those definitions with specific instances, and a reasoner computes the logical consequences of the whole. The tradeoff was stark: formal semantics meant provably correct inference but at the cost of expressiveness — every addition to a DL had to be balanced against decidability.

1990s — The Cyc Project and Common-Sense Bottlenecks

Doug Lenat's Cyc project attempted to encode the implicit background knowledge that humans take for granted. After three decades of work, Cyc remains the largest manually constructed ontology in existence, yet it has never achieved the broad utility its creators hoped for. The lesson was sobering: hand-curation does not scale to the open world. The web, with its billions of pages, exposed the futility of trying to write down every fact by hand.

1998–2006 — The Semantic Web and Linked Data

Tim Berners-Lee's 1998 proposal for a Semantic Web envisioned a layer of machine-readable structure on top of the existing web. The technical scaffolding — RDF (Resource Description Framework), RDFS (RDF Schema), and OWL (Web Ontology Language) — was standardized between 1999 and 2004. The Linked Data principles articulated in 2006 by Berners-Lee crystallized the idea that publishing data as dereferenceable URIs with typed relationships would produce a global knowledge graph. DBpedia, launched in 2007 as a structured extraction of Wikipedia infoboxes, was the first large-scale demonstration that the vision was feasible.

The Semantic Web era left a mixed legacy. On one hand, it produced durable standards (RDF, SPARQL, OWL) and a vibrant research community. On the other hand, adoption outside a