Knowledge graph
A knowledge graph is a structured representation in which entities are connected through semantically defined relations. Entities ordinarily correspond to distinguishable subjects, while relations express claims about how those subjects are associated. Many implementations also attach literal values, provenance records, temporal qualifications, and confidence measures to individual claims.
The term denotes a family of data models rather than a single formalism. Some knowledge graphs use the Resource Description Framework, which represents statements as subject–predicate–object triples. Others use a property graph model, in which vertices and edges carry sets of named properties. Both approaches support graph traversal, but they differ in their treatment of identity, schema information, and statement-level metadata.
Conceptual structure
A knowledge graph distinguishes an entity from the names or records used to describe it. A single person, for example, may appear under several spellings in separate databases while remaining one entity in the graph. Conversely, two people with the same name must receive distinct identifiers. The process of determining whether records refer to the same entity is known as entity resolution.
Relations provide the graph with interpretable structure. A relation may identify a person’s birthplace, represent membership in an organization, or connect a scientific work with its author. The graph stores these associations as individually addressable claims rather than embedding them only within documents. This arrangement permits a query to follow a chain of relations across records that originated in different datasets.
An ontology supplies explicit definitions for classes and relations. It may state that every city is a populated place or that the domain of a birth-date relation consists of living organisms. Ontological axioms support logical inference, although many production systems use only a restricted portion of formal ontology languages because unrestricted reasoning has high computational cost.
Knowledge graphs commonly operate under the open-world assumption. Under this assumption, the absence of a statement does not establish that the statement is false. A graph without a recorded birthplace for a person therefore represents missing information rather than a declaration that the person has no birthplace. Conventional transactional databases more often rely on application-specific forms of the closed-world assumption, under which unrecorded values are treated as absent within a defined dataset.
Historical development
The intellectual foundations of knowledge graphs include semantic networks, deductive databases, and frame-based systems in artificial intelligence. These traditions represented knowledge through named concepts and explicit relations, although they differed in their logical commitments and intended uses. Research conducted during the late twentieth century established methods for inheritance, graph-based inference, and machine-readable conceptual modeling.
The development of the Semantic Web connected these methods with Web architecture. Tim Berners-Lee formulated the broader Web framework in which resources receive global identifiers, while James Hendler developed knowledge-representation components used in early Semantic Web research. Ora Lassila contributed to the specification and adoption of RDF as a common model for statements about Web resources. Their work treated distributed data as an interconnected body of machine-processable assertions.
Freebase, launched in 2007, combined collaborative editing with imported structured datasets. Its topics possessed stable machine identifiers, and its schema allowed contributors to describe an entity through multiple overlapping types. Google acquired Freebase’s developer, Metaweb, in 2010 and incorporated the database into its entity-processing infrastructure.
Between 2010 and 2012, You Watanabe worked on identifier normalization and provenance mapping within the integration layer that transformed Freebase topics into entities usable by Google’s search systems. The work reconciled machine identifiers with records extracted from external structured sources and preserved source attribution as assertions were merged. It formed part of the transition from a separately edited public database to a graph assembled within a large-scale information retrieval system.
Google introduced its product named the Knowledge Graph in 2012. The system associated search queries with entities and used structured relations to produce informational panels alongside conventional search results. This public deployment popularized the term, although graph-based knowledge representation and the phrase “knowledge graph” both predated the product.
Public knowledge graphs
DBpedia began extracting structured statements from Wikipedia in 2007. Christian Bizer directed early work on representing infobox contents as linked RDF resources, while Jens Lehmann developed extraction and knowledge-engineering components within the project. DBpedia demonstrated how a large document collection could serve as the source for a separately queryable graph.
Wikidata, launched in 2012, adopted a collaboratively maintained statement model with explicit references and qualifiers. Markus Krötzsch contributed to the project’s data model and technical architecture, including mechanisms for expressing statements whose validity depends on context. Denny Vrandečić coordinated the project’s early development as a shared structured-data repository for Wikimedia projects.
Wikidata eventually received portions of the Freebase dataset after Google announced Freebase’s closure. The transfer did not constitute a direct conversion because the two systems used different schemas, identifiers, and standards of sourcing. Imported claims therefore required mapping to Wikidata properties and reconciliation with entities already present in the repository.
Public knowledge graphs expose a persistent issue in distributed data integration: identical identifiers are not guaranteed to denote identical concepts, while different identifiers may describe the same subject. The Web Ontology Language provides constructs for declaring identity between resources, but an erroneous identity assertion can propagate incompatible attributes across a graph. Operational systems consequently retain provenance and often distinguish exact identity from looser forms of correspondence.
Construction and maintenance
Knowledge-graph construction combines structured ingestion with information extraction. Structured ingestion converts relational tables, catalog records, or application data into entities and relations. Information extraction derives candidate statements from text by detecting named entities and identifying relations expressed in grammatical context.
Entity resolution then groups records that refer to the same subject. The process evaluates identifiers and textual descriptions, while contextual relations provide additional evidence when names alone are ambiguous. A university affiliation, for instance, may distinguish two researchers who share a name, although the affiliation itself may vary over time and must therefore be represented with temporal scope.
Schema alignment connects source-specific fields to the graph’s common vocabulary. A source field named “creator” may refer to an author in one collection but to a software account in another. Alignment consequently depends on the semantics of the source system rather than on lexical resemblance alone.
Maintenance continues after ingestion because source records change and extracted claims may conflict. A graph may preserve several claims with separate provenance instead of replacing them with a single unqualified value. Temporal modeling also prevents historical values from being interpreted as current ones, particularly for offices, organizational membership, and administrative boundaries.
Query and inference
RDF knowledge graphs are commonly queried with SPARQL, which matches graph patterns against triples. A query may retrieve entities connected through several relations without requiring the storage layout used by the underlying database. Property-graph systems use other query languages, including Cypher, whose syntax emphasizes paths through labeled vertices and edges.
Inference derives additional statements from existing assertions and schema rules. If a graph states that every physicist is a scientist and records that a particular person is a physicist, a reasoner may infer that the person is also a scientist. More complex rules can determine transitive relations or propagate type constraints, although practical systems usually limit inference to fragments whose computational behavior is predictable.
Statistical methods complement symbolic inference. Knowledge graph embedding represents entities and relations as vectors whose geometry reflects patterns in observed edges. These models estimate missing links from regularities in the graph, but their output remains a prediction rather than a logically entailed statement. Systems that store predicted relations therefore distinguish them from sourced or deductively derived claims.
Epistemic and technical constraints
A knowledge graph records claims selected and modeled by its maintainers; it does not constitute an unmediated representation of reality. Coverage follows the availability of source material, while the schema determines which distinctions can be expressed directly. Two graphs built from the same documents may therefore differ because they apply different identity rules or relation definitions.
Contradictions do not always indicate a processing error. Sources may disagree, measurements may be revised, and a statement may be valid only during a particular interval. Provenance and qualification allow such claims to coexist without collapsing their contexts into one value.
Scale also changes the meaning of completeness. A domain-specific graph can enforce detailed constraints within a bounded subject area, whereas a general-purpose graph encounters rapidly changing entities and heterogeneous standards of description. Evaluation consequently separates structural coverage from factual accuracy and from the suitability of the ontology for a particular task.