Knowledge Representation
Knowledge representation is the branch of artificial intelligence concerned with encoding facts, concepts, and relationships about the world in a form a machine can store, query, and reason over to derive conclusions that were never explicitly written down. Its central premise is that intelligent behavior often depends on having an explicit, inspectable model of what is true and what follows from it, rather than only on pattern matching. In their 1993 paper "What is a knowledge representation?" Randall Davis, Howard Shrobe, and Peter Szolovits described it as playing several roles at once: a surrogate for the world, a set of ontological commitments, a fragmentary theory of intelligent reasoning, a medium for efficient computation, and a medium for human expression.
Semantic networks, introduced in M. Ross Quillian's 1968 work on modeling word meaning, represent knowledge as a graph whose nodes are concepts and whose labeled edges are relations such as "is-a" and "has-part." The appeal was inheritance: if a canary is a bird and a bird can fly, then the canary inherits the ability without it being stored again. William Woods's 1975 paper "What's in a link" exposed the weakness of the informal approach, because the same arrow was being used for subclass, instance membership, and property assertion with no precise meaning. That ambiguity pushed the field toward representations with a formal semantics, where one can say exactly what a statement means and which conclusions are legitimate.
Frames, proposed by Marvin Minsky in his 1974 memo "A Framework for Representing Knowledge," bundle what is known about a stereotyped situation or object into a structure of slots with default values. A frame for a restaurant visit would have slots for the menu, the waiter, and the bill, filled by defaults until observation overrides them, so expectations arise automatically when a situation is recognized. Frames gave object-oriented programming and later ontology tools much of their vocabulary of classes, slots, and inheritance. The cost of defaults is nonmonotonic reasoning: the conclusion that this bird flies must be withdrawn when told it is a penguin, which a plain deductive logic does not allow, and handling exceptions cleanly spawned an entire literature on default and nonmonotonic logics.
Description logics are the formalization that grew out of structured inheritance networks, most directly Ronald Brachman's KL-ONE system from the late 1970s and early 1980s, and they define concepts, roles, and individuals within decidable fragments of first-order logic. A knowledge base separates the terminology (a TBox of concept definitions such as a parent being a person with at least one child) from assertions about individuals (an ABox). Reasoners then perform services like subsumption, deciding whether one concept necessarily contains another, consistency checking, and classification of individuals. The 2004 W3C standard OWL and its 2009 successor OWL 2 are built on this family, with OWL 2 DL corresponding to the logic SROIQ. Hector Levesque and Ronald Brachman's 1987 analysis of the tradeoff between expressiveness and tractability explains why the logics are deliberately restricted: richer languages quickly make inference undecidable or intractable.
The enduring lesson of the field is that every representation is a bargain between what can be said and what can be computed, and that no single scheme is best for everything. Rules, logic, probabilistic models, and graphs each make some inferences cheap and others awkward. A common misconception is that a knowledge graph is simply a database with a fashionable name; the difference is that a database stores facts under a fixed schema and answers queries about them, while a logically grounded representation also lets a reasoner infer facts that were never inserted. The opposite mistake is to expect general commonsense from hand-built knowledge bases: Doug Lenat's Cyc project, begun in 1984, spent decades encoding assertions by hand and demonstrated how hard coverage and brittleness remain.
Today the discipline is revived by the weaknesses of neural networks, which are fluent but opaque and can assert falsehoods with confidence. Retrieval systems that ground a language model in a knowledge graph, neurosymbolic reasoning, and learned graph embeddings all attempt to combine the flexibility of learned representations with the checkability of explicit ones. Enterprise knowledge graphs at search engines and in biomedicine are the clearest production use, and the open question is how to keep symbolic structure maintained and trustworthy at scale when much of the world's knowledge sits in unstructured text.
Read knowledge representation as the choice of a language for saying what is true about the world: the language decides which conclusions can be drawn mechanically and how costly they are, which is why frames, semantic networks, and description logics are best understood as different points on one expressiveness-versus-tractability curve.