Foundations of Knowledge Graphs

Foundations of Knowledge Graphs

The proliferation of large language models has reignited a fundamental debate in artificial intelligence: what role does structured knowledge play when neural networks can seemingly retrieve and synthesize information on demand? This chapter addresses that question directly by establishing what a knowledge graph actually is, how it differs from related concepts, and why its structured representation of reality remains indispensable even as transformer-based models achieve remarkable capabilities. Understanding these foundations is essential because knowledge graphs are not merely historical artifacts or database curiosities—they are active infrastructure powering search engines, recommendation systems, clinical decision support, and enterprise intelligence across industries. The central insight is that knowledge graphs provide explicit, interpretable, and machine-reasonable representations of domain knowledge that complement the implicit, statistical knowledge encoded in neural network weights.

Historical Evolution from Semantic Networks to Modern Knowledge Graphs

The concept of representing knowledge as interconnected graphs predates modern computing by decades, emerging from cognitive science and linguistics before finding computational expression in the mid-twentieth century.

The Cognitive Origins (1940s–1960s)

The intellectual foundation for knowledge graphs traces to psychologist Edward Tolman's work on cognitive maps in the 1930s, but the formalization began with Ross Quillian's semantic networks in 1966. Quillian's goal was to model human semantic memory computationally—he proposed that meaning emerges from relationships between concepts rather than from isolated definitions. In his doctoral work at Carnegie Mellon University, Quillian developed the Teachable Language Comprehender (TLC), which represented words as nodes with typed relationships to other words. When you looked up "bird," you didn't get a definition; you traversed a network connecting birds to wings, to animals, to properties like "can fly" with associated probabilities and exceptions.

This network representation was revolutionary because it captured something that flat dictionaries could not: the web of associations that give words their meaning in context. A robin is a bird, birds typically fly, but penguins are birds that don't—semantic networks could represent these exceptions through relationship types and constraints in ways that pure hierarchical classification could not.

The linguistic connection is important to understand: semantic networks borrowed heavily from dependency grammar and case grammar theories, which emphasized that verbs (actions) are the core organizing principle of sentences, with nouns filling grammatical roles like agent, patient, instrument, and location. Semantic networks translated these linguistic insights into graph structures where nodes represented concepts and edges represented case relationships.

Expert Systems and the Knowledge Revolution (1970s–1980s)

The 1970s brought the expert systems era, which demonstrated that capturing human expertise in rule-based systems could solve problems in narrow domains. Edward Feigenbaum's DENDRAL system (1965) for chemical structure analysis and Bruce Buchanan's MYCIN (1976) for medical diagnosis showed that domain knowledge could be encoded as facts and rules, then applied through inference engines to solve problems that would otherwise require human specialists.

The knowledge representation community during this period developed increasingly sophisticated formalisms. Frames, introduced by Marvin Minsky in 1974, provided a structured way to represent stereotypical situations—concepts with slots for attributes, default values, and inheritance hierarchies. The KL-ONE system (1978) from the University of Maryland pioneered description logics, which provided formal semantics for concept hierarchies and role restrictions that remain foundational to modern ontology languages.

The critical insight from expert systems was that knowledge itself was valuable—perhaps more valuable than the inference mechanisms that processed it. This realization led to the "knowledge bottleneck" problem: acquiring, validating, and maintaining domain knowledge at scale proved far more difficult than building inference engines. The Cyc project, initiated by Douglas Lenat in 1984, represented the most ambitious attempt to codify human common sense knowledge, aiming to encode millions of common sense assertions. Cyc remains active today, though its scope reveals the enormity of the knowledge acquisition challenge.

The Semantic Web and Ontology Boom (1990s–2000s)

The 1990s brought a shift toward standardized knowledge representation through the World Wide Web Consortium's Semantic Web initiative. Tim Berners-Lee's vision extended the web's document model to include machine-readable semantics, enabling software agents to navigate and reason about web content semantically. The core standards—RDF (Resource Description Framework), RDFS (RDF Schema), and OWL (Web Ontology Language)—provided standardized languages for expressing facts and vocabularies.

RDF's data model was elegant in its simplicity: everything was a triple consisting of subject, predicate, and object, where subjects and objects are identified by Uniform Resource Identifiers (URIs) and predicates define the relationship type. This tripartite structure directly prefigured modern knowledge graph implementations. The semantic web standards succeeded in establishing a shared language for knowledge representation, even if the vision of autonomous semantic web agents never fully materialized.

The ontology boom of this period saw industries develop standardized vocabularies for their domains. In medicine, the Gene Ontology (2000) standardized terms for gene function across species. In finance, XBRL enabled standardized financial reporting. In e-commerce, product categorization standards emerged. These ontologies demonstrated that communities could agree on shared vocabularies and that such agreement enabled interoperability and data integration at scale.

The Modern Knowledge Graph Era (2010s–Present)

Google's introduction of the Knowledge Graph in 2012 marked a pivotal moment, demonstrating that knowledge graphs could be both commercially viable and operationally effective at internet scale. Google's Knowledge Graph was not an academic exercise or a semantic web implementation—it was a production system indexing hundreds of millions of entities and relationships, powering search result enhancements that showed entity summaries alongside traditional page results.

What distinguished Google's approach was the scale of knowledge acquisition: rather than relying on manual ontology engineering, Google developed automated and semi-automated methods for extracting facts from web text, Wikipedia infoboxes, and structured data sources. The Knowledge Graph transformed search from pure keyword matching to entity recognition and relationship understanding. When you search for "tallest building," Google knows you're asking about buildings, understands "tallest" as a superlative comparison, and can retrieve the answer directly.

Amazon followed with its product knowledge graph, which integrated product attributes, categories, customer reviews, and related products into a unified structure supporting both search and recommendation. LinkedIn built the Economic Graph mapping professional skills, job postings, companies, and professional relationships to support job matching and economic analysis. Microsoft developed its Academic Graph for research intelligence. These implementations shared common characteristics: scale, diversity of data sources, integration with consumer-facing products, and continuous updates as new information became available.

The convergence of modern knowledge graphs represents an evolutionary synthesis of pr