Knowledge Graphs in the Landscape of Structured Knowledge: Origins, Definitions, and Theoretical Positioning

The Central Problem: Formalizing Heterogeneous World Knowledge

The modern discipline of knowledge representation confronts a question whose surface simplicity belies its depth: how can the heterogeneous, multi-relational, and continuously evolving body of human world knowledge be formalized in a way that simultaneously supports machine inference, human comprehension, and inter-system exchange? The answer that has consolidated across nearly two decades of practice is the *knowledge graph*—a representational artifact that has migrated from a niche research object of the Semantic Web community to a load-bearing component of industrial information systems powering search, recommendation, question answering, and increasingly, the structured memory of large language models. Yet the very success of the abstraction has tended to obscure its theoretical commitments. Many engineers treat a knowledge graph as little more than a graph-shaped database; many researchers treat it as a thin synonym for Description Logic. Neither framing is adequate. To deploy knowledge graphs responsibly, to extend their inference envelope, and to integrate them with the sub-symbolic architectures that now dominate artificial intelligence, one must first understand the precise object one is manipulating, the historical lineage that produced it, and the design space within which it competes with alternative representations.

This chapter serves as the conceptual scaffolding for the entire volume. It pursues four interlocking objectives. First, it reconstructs the intellectual lineage of the knowledge graph by tracing its descent from the semantic networks of early cognitive science, through the frame languages and Description Logics of the knowledge engineering era, into the Linked Open Data movement and the modern industrial-scale graphs of Google, Microsoft, and Wikimedia. Second, it supplies the formal apparatus that subsequent chapters will invoke without re-derivation: the labeled directed multigraph, the RDF/OWL stack, the dichotomy between open- and closed-world assumptions, and the entailment regimes that govern inference. Third, it situates the knowledge graph within a broader taxonomy of structured representations—comparing it against relational databases, property graphs, hypergraphs, and vector stores—so that practitioners acquire a defensible calculus for choosing among them. Fourth, it conducts an architectural autopsy of four canonical systems (DBpedia, Freebase, Wikidata, and ConceptNet) to demonstrate how theoretical commitments manifest in engineering decisions. The chapter closes by previewing the tripartite structure of the book and motivating the integration of symbolic and sub-symbolic paradigms that the later chapters will develop in depth.

Historical Evolution: From Semantic Networks to Industrial Knowledge Graphs

Cognitive and Linguistic Precursors

The conceptual ancestry of the knowledge graph is older than the term itself. Long before the advent of digital computation, the project of formalizing world knowledge had produced large-scale enumerative artifacts—Aristotle's categories, the Linnaean taxonomy of living things, Roget's thesaurus—that already exhibited the structural pattern of nodes connected by typed relations. The transition from these human-curated taxonomies to computational artifacts occurred in the 1960s, catalyzed by three converging research programs. Ross Quillian's semantic networks, introduced in his 1968 work on human semantic memory, proposed that conceptual knowledge could be represented as a graph whose nodes denote concepts and whose edges denote binary semantic relations such as *is-a*, *has-part*, or *leads-to*. The crucial innovation was the propagation of inheritance along the type hierarchy: properties asserted of a general concept were inherited by its specific instances, providing a primitive form of inference that foreshadowed later deductive mechanisms. Roger Schank's conceptual dependency theory, developed in the early 1970s, advanced a parallel agenda by decomposing natural language meaning into a small inventory of primitive semantic acts—*PTRANS* (physical transfer), *ATRANS* (abstract transfer), *MTRANS* (mental transfer)—whose composition was intended to capture the meaning of arbitrary sentences. Although Schank's specific formalism did not survive, his insistence on canonical semantic primitives left an enduring methodological imprint on knowledge engineering.

Frame Languages and the Knowledge Engineering Era

Marvin Minsky's 1974 paper "A Framework for Representing Knowledge" introduced the frame as a data structure for representing stereotyped situations. A frame is a record-like template whose slots describe the typical constituents of a concept and whose slot fillers may invoke other frames, enabling recursive composition. Frames differed from semantic networks less in their formal structure than in their emphasis on prototypical organization, default values, and procedural attachment. The frame paradigm dominated knowledge engineering throughout the late 1970s and 1980s, manifesting in production systems such as the KL-ONE family, Sowa's conceptual graphs, and the hybrid architectures deployed in expert systems ranging from MYCIN's medical diagnosis to XCON's computer configuration. KL-ONE, in particular, deserves attention as the precursor of modern Description Logic. Its concept language distinguished between a *terminological component* (TBox), which declared concept inclusions and role restrictions, and an *assertional component* (ABox), which asserted the membership of individuals in concepts. This separation, refined and formalized over the subsequent two decades, became the architectural backbone of the OWL standard.

The most ambitious exemplar of the knowledge engineering era was the Cyc project, initiated by Douglas Lenat in 1984. Cyc aimed to encode the implicit background knowledge—commonsense rules, default reasoning heuristics, taxonomic hierarchies—required for human-level inference. Over four decades, the project accumulated millions of hand-curated assertions expressed in the CycL language, a frame-based formalism with modal operators and microtheories for context-sensitive reasoning. Cyc's contribution was less a deployable system than a demonstration of the scale of the commonsense knowledge problem: even a sustained, well-funded effort spanning decades produced a coverage that, while impressive, remained brittle, expensive to maintain, and difficult to integrate with external resources. The lesson was not that large-scale knowledge engineering was futile but that the acquisition bottleneck had to be addressed through automation and open collaboration rather than solely through expert authoring.

The Semantic Web Vision and the RDF/OWL Stack

The next inflection point arose from the web rather than from artificial intelligence. Tim Berners-Lee's 1998 roadmap for the Semantic Web proposed that the existing document web be augmented with machine-processable metadata expressed in a standardized graph format, thereby enabling software agents to traverse, integrate, and reason over distributed knowledge. The first deliverable of this program was the Resource Description Framework (RDF), standardized by the World Wide Web Consortium in 1999 and revised in 2004. RDF's foundational commitment was minimalism: any assertion is reducible to a triple of the form ⟨subject, predicate, object⟩, where each component is identified by a Uniform Resource Identifier (URI) or, in restricted cases, a blank node or a typed literal. The triple can be read as an edge-labeled directed graph, and sets of triples constitute a directed multigraph in which multiple edges between the same pair of nodes are permitted, parallel edges carry distinct labels, and labels are themselves drawn from the URI space. This graph-theoretic reduction is theoretically elegant but practically consequential: it ensu