Concept
Knowledge Graph
A structured representation of real-world entities and their relationships as a directed graph - enabling machines to reason over factual knowledge, answer questions, and make inferences by traversing a web of interconnected facts.
Added May 18, 2026
Knowledge graphs represent facts as triples: (subject, predicate, object). "(Eiffel Tower, locatedIn, Paris)", "(Albert Einstein, bornIn, Ulm)", "(Python, isA, ProgrammingLanguage)". The entities are nodes in the graph; the predicates (relationships) are directed, typed edges. This triple structure is simple but powerful: it allows complex knowledge to be represented as a navigable network of atomic facts that machines can query, reason over, and complete.
Large-scale knowledge graphs have become foundational infrastructure for AI applications. Google's Knowledge Graph (underpins search answer boxes), Freebase, Wikidata, DBpedia, YAGO, and domain-specific graphs (UniProt for proteins, DrugBank for drug-gene interactions, the Unified Medical Language System) each contain billions of triples encoding structured world knowledge.
Knowledge graph completion is one of the central tasks: predicting missing links in an incomplete graph. Given a knowledge graph where many true facts are unrecorded ("Eiffel Tower (?, ?) ?"), learn to score potential triples by their likelihood of being true. This is the setting for knowledge graph embedding methods: TransE, RotatE, ComplEx, DistMult, and others that learn vector representations for entities and relations such that true triples score higher than false ones under a geometric scoring function.
GNNs are increasingly used for knowledge graph tasks. Relational GNNs like R-GCN (Relational GCN) aggregate messages separately for each relation type, building node representations that encode the node's role in each type of relationship. More expressive models like CompGCN combine entity and relation embeddings during aggregation. These GNN-based methods tend to outperform pure embedding methods on transductive tasks and have better generalisation properties.
Knowledge-augmented language models integrate knowledge graphs with LLMs: retrieval from knowledge graphs provides structured factual grounding that reduces hallucination. Systems like KGQA (knowledge graph question answering) parse natural language questions into structured SPARQL queries that execute against the knowledge graph, returning verified facts rather than LLM-generated text. This hybrid approach leverages LLMs' language understanding and knowledge graphs' factual reliability.
Ontologies define the schema of a knowledge graph: the types of entities, the types of relationships, and the logical constraints on how they can combine. OWL (Web Ontology Language) and RDF (Resource Description Framework) are the W3C standards for representing knowledge graph schemas. SPARQL is the query language for querying RDF knowledge graphs, enabling complex pattern matching and inference.
Analogy
An encyclopaedia, reimagined as a network rather than a book. A traditional encyclopaedia entry on Paris describes the city in prose. A knowledge graph stores the same information as structured facts: Paris-isCapitalOf-France, Paris-population-2.1M, Paris-contains-EiffelTower. The network structure allows machines to answer questions by traversing links: "What country contains the Eiffel Tower?" - follow the Eiffel Tower node to Paris via contains, then follow Paris to France via isCapitalOf. No language understanding required - just graph traversal.
Real-world example
Google's Knowledge Graph contains billions of facts about entities (people, places, organisations, concepts). When you search for "Albert Einstein", the Knowledge Panel on the right side of Google search results is populated by querying the Knowledge Graph: birth date, birthplace, fields, notable works, spouse, and a structured fact summary. This information is more reliable than extracting it from web text because it is curated, structured, and verified - structured knowledge rather than potentially incorrect language model recall.
Why it matters
Knowledge graphs are the structured, verifiable alternative to language model parametric memory. Where LLMs encode facts implicitly in weights and can hallucinate, knowledge graphs store facts explicitly as queryable triples with deterministic, verifiable retrieval. Understanding knowledge graphs is essential for building AI systems that need reliable factual accuracy, for graph-based question answering, and for any domain (medicine, law, science) where knowledge provenance and consistency matter more than linguistic fluency.
In the news
No recent coverage - check back later.
Related concepts