AI glossary · Basic · Data & evaluation · Language models
Knowledge graph
A knowledge graph stores knowledge like a web, with 'entities' such as people, places, events and concepts as points and the relationships between them as lines. This lets a computer read and reason about 'what is connected to what, and how'.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Foundations & applications
- Fields
- Data & evaluation, Language models
- Added
- 2026-10-07
How it works
The basic unit of a knowledge graph is a statement made of three parts: 'subject–relation–object'. For example: 'Seoul – is the capital of – South Korea'. RDF, a data format defined by the web standards body W3C, works exactly this way, and these three-part units are called 'triples'. As triples accumulate, they form a large web of points (entities) and named arrows (relations). Adding a concept system (ontology), such as 'a city is a kind of place', makes it possible to logically derive facts that are not written down directly.
How it has evolved
The idea of representing knowledge as a graph has been around for a long time, and the term 'knowledge graph' is said to have first been used in 1972 by the Austrian linguist Edgar W. Schneider. Later, standards such as RDF and OWL emerged from the 'Semantic Web' movement to build a web machines can read, and in 2007 general-purpose knowledge stores such as DBpedia, which organizes Wikipedia content, appeared. The 2020 paper 'Knowledge Graphs' by Hogan et al. served as an introductory text, bringing together the whole field of knowledge graphs, from data models and query languages to methods for building, refining and checking quality. The 2023 paper 'Unifying Large Language Models and Knowledge Graphs: A Roadmap' by Pan et al. divided the ways of combining LLMs and knowledge graphs into three: using knowledge graphs to enhance LLMs, using LLMs to help build knowledge graphs, and having the two complement each other and reason together.
Why it matters
LLMs know a lot, but that knowledge is blurrily mixed inside the model, which makes them prone to hallucination, stating facts incorrectly. Because knowledge graphs record facts as explicit connections, it is easy to check and explain the basis for an answer. That is why they have long been used in search engines, voice assistants and recommendation systems, and today they underpin approaches such as GraphRAG that supply accurate facts to LLMs. Their use is also growing in scientific fields that deal with complex relationships, such as genome and protein research.
Things to know
Knowledge graphs take a lot of effort to build and keep up to date. Facts about the world keep changing, but a graph cannot fill in new knowledge it does not contain on its own. Also, because what counts as an entity and which relation names to use must be agreed on in advance, aligning notation and criteria when merging data from different sources is tricky. Recently, research on having LLMs help with this construction work has been active.
Example
When you search for a famous person or city in a search engine, a separate box often appears with organized information such as birthplace, related people and affiliations. This kind of feature is based on a knowledge graph that links entities and relationships. When you ask a voice assistant a question that requires several steps, such as 'What other films has this movie's director made?', it follows the connections 'movie → director → other works'. Companies also link internal documents, products and customer information into a knowledge graph so that an LLM chatbot can refer to it and give well-grounded answers.