Knowledge representation: how AI stores facts, rules and relationships
Knowledge representation is the method by which AI systems store facts, rules, and relationships. It includes surrogates, ontological commitments, logic, knowledge graphs, and rule-based schemes. The chosen representation determines which inferences are possible and how efficiently they can be computed.
The five roles of a knowledge representation
A knowledge representation acts as a surrogate, a substitute for the thing it represents, so that an entity can reason about the world instead of acting in it. This is one of five roles identified by MIT CSAIL's research paper. The surrogate is imperfect, so incorrect inferences are inevitable when the representation does not fully capture the real world.
The same paper also describes a representation as a set of ontological commitments, an answer to the question of what terms to think about the world. It is also a fragmentary theory of intelligent reasoning that specifies which inferences are sanctioned and which are recommended.
Finally, a representation is a medium for efficient computation and a medium of human expression. These five roles place different demands on the properties a representation should have, and keeping them all in mind can clarify longstanding disputes in AI.
Logic and the representation of facts
Logic has been a central tool in AI since its early days. The Stanford Encyclopedia's logic in AI entry notes that logic provides techniques for analyzing the inferential properties of languages and can provide a specification for a programming language by mapping programs to the computations they should license.
In modern systems, declarative representations are treated as separate components from the reasoning machinery. The evolution of expert systems shows a shift from purely procedural rules to modular designs where background knowledge is stored separately and retrieved explicitly.
Nonmonotonic logic emerged from the need to formalize commonsense reasoning, where conclusions can be revised when new information arrives. This approach helps represent default assumptions without requiring an exhaustive list of exceptions.
Knowledge graphs and ontologies
A knowledge graph represents a network of real-world entities and the relationships between them. IBM's knowledge graph overview describes its three main components: nodes, edges, and labels, where a node can be an object, place, or person, an edge defines the relationship, and a label classifies the node appropriately.
Ontologies provide a formal representation of the entities in a graph and are often based on taxonomies. The Web Ontology Language (OWL) is a widely adopted ontology language supported by the World Wide Web Consortium, as IBM notes.
Protégé is a free, open-source OWL ontology editor used by organizations worldwide. Stanford University's Protégé page lists notable projects such as the World Health Organization's ICD-11, the National Cancer Institute Thesaurus, and many ontologies in the OBO Foundry. The tool supports the full ontology development lifecycle, from modeling to reasoning and querying.
Reasoning with knowledge and the frame problem
In a reasoning system, a knowledge base stores the structured facts, rules, and relationships, and an inference engine applies logic to reach a decision. IBM's AI reasoning page gives an example of an autonomous floor cleaner that uses a knowledge base about floor types and cleaning requirements to decide between vacuuming and mopping.
The frame problem, as described by the Stanford Encyclopedia's frame problem entry, is the challenge of representing the effects of actions in logic without having to list all the intuitively obvious non-effects. This requires a default assumption known as the commonsense law of inertia, which states that an action can be assumed not to change a property unless there is evidence to the contrary.
Reasoning strategies include deductive, abductive, inductive, commonsense, and probabilistic approaches. These methods are combined depending on the dataset and the target application.
| Role | Function |
|---|---|
| Surrogate | Substitute for the thing itself, enabling reasoning without acting |
| Ontological commitments | Answer to what terms to think about the world |
| Fragmentary theory of reasoning | Specifies sanctioned and recommended inferences |
| Medium for computation | Facilitates efficient computation and organizing information |
| Medium of human expression | Language for saying things about the world |
Frequently asked questions
What is knowledge representation?
Knowledge representation is a surrogate for the world, along with a set of ontological commitments and a theory of reasoning, as explained by MIT CSAIL.
What is an ontology?
An ontology is a formal representation of the entities in a domain, often based on taxonomies, and can be written in languages like OWL.
What is a knowledge graph?
A knowledge graph is a network of nodes and edges that stores entities and their relationships, often used to answer queries and support search.
What is the frame problem?
The frame problem is the challenge of representing the non-effects of actions in logic without listing them all explicitly.
How is Protégé used?
Protégé is a free, open-source OWL ontology editor that has been used for ICD-11 and other biomedical ontologies.