Knowledge representation is the part of artificial intelligence that studies how to encode information about the world so that a computer can store it, retrieve it and reason with it to solve problems. The field is often called knowledge representation and reasoning because a representation is only useful if a program can draw conclusions from it. A good representation captures what matters about a domain, such as medicine or law, in a precise form that supports questions, explanations and decisions.
History of knowledge representation
The roots of the field lie in logic and philosophy, which have long tried to formalize reasoning. In the 1950s and 1960s, early AI programs used logic and search. In the late 1960s, Ross Quillian proposed semantic networks, graphs of concepts linked by relations, and in 1974 Marvin Minsky proposed frames, structures that bundle typical properties of a situation. The 1970s and 1980s brought rule-based expert systems and large hand-built projects such as Cyc, begun by Douglas Lenat in 1984 to capture everyday common sense. From the 1990s, description logics and ontologies were developed, and the Semantic Web movement led the World Wide Web Consortium to standardize RDF and OWL. Since 2012, knowledge graphs have become widely used in search engines and business systems.
Main methods of knowledge representation
Several approaches are common, each suited to different needs.
- Logic: propositional and first-order logic state facts and rules precisely and allow rigorous deduction.
- Production rules: if-then statements, as used in expert systems, that fire when their conditions are met.
- Semantic networks: graphs in which nodes are concepts and edges are relations such as is-a or part-of.
- Frames: structured records with slots and default values for typical objects or situations.
- Ontologies: formal vocabularies that define the classes, properties and relations in a domain, often written in OWL.
- Knowledge graphs: large networks of entities and relationships, often stored as subject-predicate-object triples.
More recently, researchers also use learned representations, such as embeddings, which place concepts in a numerical space where similar meanings lie close together.
How knowledge representation works
Building a representation begins with choosing the concepts and relations that matter in a domain and deciding how to write them in a formal language. The knowledge is then stored in a knowledge base, and a reasoning component, described in the entry on reasoning engines, uses it to answer queries. If a knowledge base states that every mammal is warm-blooded and that a whale is a mammal, a reasoner can infer that a whale is warm-blooded without being told. Good representations also make it easy to add new facts, check consistency and explain how an answer was reached.
Expressiveness versus efficiency
A central insight of the field, developed by Hector Levesque and Ronald Brachman in the 1980s, is that there is a trade-off between how much a language can express and how easily a computer can reason with it. A very expressive language can describe subtle facts, such as statements about beliefs or time, but reasoning with it may be slow or even impossible to complete in general. A simpler language limits what can be said but allows fast, guaranteed answers. Description logics were designed to balance the two. Choosing a representation therefore means matching the language to the questions the system must answer.
Applications of knowledge representation
Knowledge representation underlies knowledge graphs used by search engines to show direct answers about people, places and organizations. It supports biomedical ontologies that standardize terms for diseases, genes and drugs, helping researchers combine data from many sources. It powers configuration tools, regulatory compliance checkers, question answering systems and robotic planners that must understand objects and actions. In the Semantic Web, data on different sites can be linked through shared vocabularies. It also supports retrieval-augmented generation systems that ground language models in structured facts.
Challenges and limitations
Making knowledge explicit is expensive. Human experts often cannot state what they know, and common sense is vast and hard to list. Hand-built knowledge bases can be incomplete, inconsistent or out of date, and merging knowledge from different sources runs into conflicts in meaning. Representations also struggle with uncertainty, exceptions and changing situations, which led to extensions such as probabilistic and non-monotonic logics. Machine learning offers a complementary route by learning patterns from data, and many current efforts combine the two so that learned systems can use explicit knowledge for accuracy and explanation.
Knowledge representation and large language models
Large language models store a great deal of knowledge implicitly, in the numerical weights learned from text, but that knowledge is hard to inspect, update or verify. Explicit representations offer the opposite strengths. Facts in a knowledge graph can be checked, corrected and traced to a source, and a change takes effect immediately. For this reason, many systems pair the two, using a knowledge graph or database to supply reliable facts to a language model, or using a model to extract facts from text to build a graph. Research on neuro-symbolic AI aims to combine the flexibility of learned models with the precision of structured knowledge.
Examples in everyday technology
Knowledge representation is often invisible but common. When a search engine shows a panel with a person’s birth date and notable works, it is reading from a knowledge graph. Online stores use product ontologies to organize categories and attributes so that filters work. Hospitals use standardized medical vocabularies so that records from different systems can be compared. Voice assistants use structured knowledge about places, contacts and devices to carry out requests. Even a simple spreadsheet that lists items and their properties is a small form of representation, though lacking the formal rules that allow automatic reasoning.
Key terms in knowledge representation
A few words appear throughout the field. A concept or class is a category of things, such as Person or Disease. An instance is a specific member of a class, such as a particular patient. A relation or property links things, as in treats or located-in. An axiom is a statement accepted as true in a knowledge base, and an inference is a new statement derived from existing ones. A triple is a single fact written as subject, predicate and object, the basic unit of RDF. Understanding these terms makes it easier to read ontology documentation and to design a knowledge base that other people and programs can use.
FAQs
What is the difference between data and knowledge representation? Data are raw facts, while a knowledge representation also encodes meaning and relationships in a form that supports reasoning.
What is an ontology? It is a formal description of the concepts in a domain and how they relate, used to share and reason over knowledge.
Is a knowledge graph the same as a database? A knowledge graph is a kind of database that stores entities and relationships explicitly and usually supports reasoning over them.
More in Artificial Intelligence
Key facts
- Knowledge representation and reasoning studies how to encode information so a computer can use it to solve complex tasks.1
- Common formalisms include logic, semantic networks, frames, production rules and ontologies.1
- The W3C standards RDF and OWL define widely used ways to represent knowledge on the web.2
Related concepts
Quick checkWhat are two trade-offs that make knowledge representation difficult?Show answer
Expressive power versus efficient reasoning, since richer languages describe more but make automatic inference slower or undecidable.
Sources
- Wikipedia, Knowledge representation and reasoning, https://en.wikipedia.org/wiki/Knowledge_representation_and_reasoning
- W3C, OWL 2 Web Ontology Language Overview, https://www.w3.org/TR/owl2-overview/
- Stanford Encyclopedia of Philosophy, Artificial Intelligence, https://plato.stanford.edu/entries/artificial-intelligence/
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
