How does knowledge representation differ from a knowledge base?

The representation provides the language and interpretation; a knowledge base contains particular statements expressed through it.

Are symbolic representations the only kind used in AI?

No. AI also uses learned numerical representations. Their interpretation and reasoning properties differ from explicit symbolic structures.

Does a representation guarantee that its contents are true?

No. A well-defined model can still contain unsupported or incorrect assertions, so evidence and validation remain necessary.

AI Glossary / Graphs and symbolic AI / Intermediate

Knowledge representation

Definition

Knowledge representation is the way information about a domain is expressed so that a system can use it in reasoning and other intelligent tasks. It includes the chosen concepts, relationships, meanings and permitted inferences, rather than merely the physical format in which data is stored.

What it means

A representation stands in for selected aspects of the world. It might describe objects and properties, events and times, rules and exceptions, or uncertain relationships. Choosing it means deciding which distinctions the system can express directly and which it will leave out.

A graph alone is a data structure. Its nodes and edges become a knowledge representation when their meanings and interpretation are defined. Similarly, a list of numbers may represent coordinates, probabilities or identifiers depending on the surrounding specification.

Representations influence reasoning. A rule language makes conditional conclusions explicit; an ontology can support category relationships; a probabilistic model expresses uncertainty in a different way. No single representation automatically captures every useful aspect of a domain or makes every question easy to answer.

How it works

Start from the questions and distinctions the application needs. Define the objects, categories and relationships, together with the meanings of the symbols used to express them. Choose a formalism and specify how conclusions follow from its statements.

Populate the representation with supported content. Test whether ordinary situations, exceptions, missing information and change over time can be expressed without distorting their meaning. A representation that omits dates may be adequate for permanent categories but misleading for changing locations.

Evaluate both expressive ability and computational behavior. Richer languages can describe more, but reasoning and maintenance may become harder. Human readability also matters when specialists must inspect and correct the model. The design should support useful answers while keeping assumptions and limitations understandable.

Three practical examples

An equipment model distinguishes an individual device from the category to which it belongs.
A museum system represents an object's location together with the time for which that location is asserted.
A decision-support model distinguishes a definite rule from an uncertain association instead of expressing both as unconditional facts.

Think of it this way

Think of choosing a map's symbols and legend. A transit map emphasizes stations and connections, while a terrain map emphasizes elevation. Both represent the same area, but each supports different questions.

The analogy explains selective representation. A computational model additionally needs rules for interpreting its symbols and deriving answers; an attractive diagram by itself does not provide those rules.

Worked example

Why a changing location needs an explicit time

A fictional museum moves sculpture S from HallA to HallB. It records that S was in HallA on Day1 and in HallB on Day2. Both halls are within BuildingM.

A first representation stores LocatedIn(S, HallA) and LocatedIn(S, HallB), with no time. These statements cannot tell a query which location applied on which day. If the application interprets LocatedIn as one current location, the records can appear contradictory; if it interprets them as historical locations, it still cannot answer the date-specific question.

A revised representation uses At(S, HallA, Day1) and At(S, HallB, Day2). It separately records HallInBuilding(HallA, BuildingM) and HallInBuilding(HallB, BuildingM). A stated rule says that if an object is in a hall at time t, and that hall is in a building, then the object is in that building at t.

The system can derive At(S, BuildingM, Day1) and At(S, BuildingM, Day2). It can also retrieve the appropriate hall for each recorded day. It cannot conclude the hall on Day3 from these statements alone unless an additional persistence assumption or observation is supplied.

The symbol Day1 needs a defined interpretation, such as a designated observation time rather than an entire day of continuous occupancy. Likewise, the representation should not confuse the sculpture with a photograph of it. Syntax, meaning and inference all matter. This invented model is not a museum standard.

Illustrative case study

The fictional Willow Museum team first stores object locations as plain text labels. A question about a previous exhibition produces a mixture of current and historical answers, even though each source record was reasonable in its original context.

Willow introduces stable object identifiers and time-scoped location assertions. Curators review what each time label means, while developers implement the stated hall-to-building inference. The team tests a known historical date and a date with no supporting observation.

For the unresolved date, the interface reports the limitation instead of silently copying the last recorded hall. This fictional case improves answer meaning without inventing movement evidence.

Common misconceptions and mistakes

Knowledge representation is just choosing a file format. Meaning and inference are central, even when the storage format stays unchanged.
A graph automatically understands its relationships. Edge labels need defined semantics and appropriate processing.
More expressive is always better. Complexity can make reasoning, explanation or maintenance impractical.
An omitted statement is automatically false. That conclusion depends on the chosen semantics and completeness assumptions.

Business use

Good representations can make shared concepts clearer and support consistent answers across systems. They are useful when a project needs explicit relationships, traceable reasoning or integration of information from different sources.

Assess the representation against real questions rather than its visual appeal or terminology count. Include examples where information changes, sources disagree or an answer is unknown. Check whether subject experts can explain and correct the model. Preserving relevant distinctions can prevent ambiguous fields and maintenance problems.

Check your understanding

Why can the revised museum model answer where S was on Day2 but not automatically on Day3?

Answer: It contains a Day2 location assertion and rules that preserve that time scope. No Day3 observation or persistence rule has been supplied.

Related terms

Ontology
Semantic network
Knowledge base
Inference
Uncertainty
Representation learning
logical representation
rule
frame
temporal representation

Learn next

Represent a borrowed object, its owner and its temporary location. Write separate questions about ownership and location, then add a date change. Check that moving the object does not accidentally change ownership or erase the historical record.

Sources

Davis, Shrobe and Szolovits analyze representation as a surrogate, a commitment about the world and a medium for reasoning and expression: https://groups.csail.mit.edu/medg/ftp/psz/k-rep.html

W3C defines RDF's graph model, interpreted relationships and treatment of change: https://www.w3.org/TR/rdf11-concepts/

AIMA's first-order-logic material explains syntax, semantics and time-sensitive assertions: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter08.pdf

The sculpture model and Willow Museum case are original fictional illustrations.

Previous: Knowledge organization Next: Knowledge representation language