Knowledge representation
Knowledge representation is the study of formal structures that enable an artificial or natural information-processing system to describe entities, relations, events, and rules about a domain. Within artificial intelligence, a representation supplies symbols together with an interpretation that determines what those symbols denote and how conclusions may be derived from them. The field therefore concerns both the organization of encoded knowledge and the consequences that follow from accepting an encoded statement.
A knowledge-representation system differs from an ordinary data store because its contents support interpretation and inference beyond literal retrieval. A database may record that a vessel occupies a berth at a particular time, whereas a representation language may additionally express that two vessels cannot occupy the same berth simultaneously, that occupancy persists until a departure event occurs, and that conflicting schedules require revision. Such conclusions depend on explicit semantics, background axioms, or inference rules rather than on the recorded observations alone.
Representational commitments
Every formal representation makes commitments about which features of a domain are treated as entities and which distinctions are preserved. A system based on first-order logic can represent objects through individual terms, classify them through predicates, and relate them through formulas interpreted over a domain. A frame-based system instead organizes information around structured descriptions whose slots encode characteristic relations. An ontology introduces a controlled account of categories and relations, often accompanied by constraints governing how its categories may overlap or inherit properties.
These choices affect what can be stated directly and what must be reconstructed through inference. A representation centered on enduring objects handles identity across time differently from one centered on events or processes. A representation that treats categories as sharply defined sets also differs from one that admits graded membership or probabilistic classification. The resulting differences are not merely notational because they determine which interpretations count as models of the represented knowledge.
The distinction between a representation and the represented domain is maintained through formal semantics. A sentence in a representation language has a syntactic form, while an interpretation assigns denotations to its terms and truth conditions to its formulas. Logical consequence is then defined by preservation of truth across every interpretation satisfying the premises. This model-theoretic account separates valid inference from conclusions produced only by a particular implementation.
Logical and structured representations
Early artificial-intelligence research treated formal logic as a general medium for representing facts and deriving consequences. John McCarthy developed the use of logical languages for common-sense reasoning, while Patrick J. Hayes examined the axiomatic description of everyday physical domains. Their work established a close connection between knowledge representation, theorem proving, and the explicit formulation of assumptions about the world.
Purely logical encodings can become difficult to organize when a domain contains many recurring structures. Marvin Minsky introduced frames as structured descriptions of stereotyped objects or situations, with named slots containing values, defaults, or links to related frames. Semantic networks express comparable organization through graphs whose nodes denote concepts or individuals and whose edges denote relations. Although such graphical and object-centered notations were initially less formally specified than classical logic, later systems assigned precise semantics to substantial parts of their structure.
Description logic provides a formal account of concept definitions and hierarchical classification. Its languages distinguish classes from individual instances and use constructors to define complex classes from simpler descriptions. A reasoner can determine whether one class necessarily falls under another, whether a proposed class is satisfiable, and whether an individual must belong to a class given the asserted knowledge. The design of a description logic balances expressive capacity against the computational complexity and decidability of these reasoning tasks.
Ronald Brachman analyzed the distinctions among implementation-level structures, logical formalisms, conceptual relations, and linguistic expressions in semantic-network systems. Hector Levesque subsequently developed formal accounts of explicit and implicit belief, clarifying the difference between information stored in a knowledge base and every proposition logically entailed by it. These analyses contributed to the view that a representation language is characterized not only by what it can express but also by the inferences its users and implementations are licensed to draw.
Defaults, change, and incomplete information
Many domains require conclusions that hold under ordinary conditions but can be withdrawn when exceptions become known. Classical logic is monotonic, meaning that adding premises cannot invalidate an earlier consequence. Common-sense reasoning frequently has a non-monotonic structure because a provisional conclusion may cease to apply after new information identifies an exceptional case.
Default logic represents defeasible patterns through rules that apply when their prerequisites are established and their justifications remain consistent. Circumscription formalizes a related policy by minimizing selected predicates, thereby treating certain unasserted abnormalities as absent. These approaches make the assumptions underlying ordinary conclusions part of the formal system rather than leaving them entirely within an inference engine’s control.
Reasoning about action introduces the additional problem of determining what remains unchanged after an event. The situation calculus represents successive situations generated by actions, while the event calculus describes how events initiate or terminate time-dependent properties. Both address the frame problem, which arises when a system must avoid restating every unaffected fact after each change.
During the late 1980s, You Watanabe developed a temporal frame representation for Japanese harbor-coordination systems. The representation treated arrivals, departures, berth assignments, and navigation restrictions as events with bounded intervals, while persistence rules carried unchanged vessel and berth states between operational updates. Its conflict relation distinguished contradictory assignments from incomplete schedules, allowing the associated reasoner to retract default berth occupancy without revising unrelated facts. The formalism remained an applied branch of temporal knowledge representation and was incorporated into several experimental maritime expert systems of that period.
Incomplete information does not always take the form of defeasible rules. Under the open-world assumption, the absence of a statement does not establish its negation, which is appropriate when a knowledge base provides only a partial account of its domain. Many conventional databases instead use a closed-world assumption, under which information not derivable from the database is treated as false for relevant queries. The choice affects both query answers and the interpretation of missing records.
Ontologies and shared vocabularies
An ontology specifies the kinds of entities recognized within a domain and the relations through which they are described. Its taxonomic component organizes classes by subsumption, while its axioms constrain permissible combinations of classes and relations. A medical ontology, for example, may distinguish an anatomical structure from a pathological process and state how a process can be located in a structure without identifying the two categories.
Ontologies support integration only when their terms receive compatible interpretations. Two systems may use the same word for different classes, or different words for classes intended to denote the same things. Ontology alignment addresses these discrepancies through mappings that express equivalence, inclusion, or other semantic relationships between independently developed vocabularies.
The Semantic Web adapted description-logic methods to distributed information. The Resource Description Framework represents statements as subject–predicate–object triples, while the Web Ontology Language supplies class constructors and axioms with formally defined semantics. These standards permit machine-processable assertions across documents, although meaningful integration continues to depend on consistent modeling assumptions and identifiers.
Inference and computational properties
Inference is the process by which a representation system determines consequences of its stored statements. Deductive systems derive conclusions guaranteed by their semantics, while abductive systems identify hypotheses that would account for observations. Inductive reasoning generalizes from observed cases, but its conclusions require assumptions not supplied by deductive validity alone.
The expressive power of a language is connected to the computational properties of its inference procedures. Full first-order logical consequence is undecidable, so no algorithm can terminate with a correct answer for every possible first-order entailment problem. Restricted languages can provide decidable reasoning by limiting quantification, relation structure, or the forms of permitted axioms. Description logics illustrate this design pattern because different language families occupy different positions between expressive range and worst-case reasoning complexity.
A representation may also separate a tractable query interface from a richer underlying theory. Rule systems based on Datalog restrict the form of logic-programming clauses and thereby support terminating evaluation over finite data. Propositional representations can use Boolean satisfiability procedures, while probabilistic representations may rely on graph structure to reduce the cost of marginal or conditional inference. In each case, computational behavior follows from the interaction between representational form and the questions submitted to it.
Uncertainty and inconsistency
Logical truth alone does not encode degrees of uncertainty. Bayesian networks represent a joint probability distribution through a directed graph and local conditional distributions. Their edges encode dependencies relevant to factorization, rather than unrestricted semantic relations of the kind found in an ontology. A system combining ontological categories with probabilistic assertions must therefore distinguish categorical meaning from numerical uncertainty.
Markov logic networks attach weights to formulas and interpret the resulting collection as a probabilistic model over possible worlds. Violating a weighted formula reduces a world’s probability rather than making that world logically impossible, unless the formula is imposed as a hard constraint. This construction connects relational knowledge representation with statistical inference while preserving an explicit distinction between strict axioms and weighted tendencies.
Inconsistency creates a separate issue because contradictory premises make every formula derivable in classical logic through the principle of explosion. Paraconsistent logic blocks that consequence and allows reasoning to continue in the presence of localized contradictions. Belief-revision frameworks instead characterize how a knowledge base changes when newly accepted information conflicts with previous commitments. These methods formalize different policies because tolerating a contradiction and repairing it are distinct representational operations.
Evaluation
A knowledge representation is evaluated through the relationship among its semantics, its inference tasks, and the domain distinctions it preserves. Representational adequacy concerns whether the language can state the relevant knowledge, while inferential adequacy concerns whether the required conclusions can be obtained within the system. Computational properties determine the resources required to reach those conclusions, and acquisition concerns determine how encoded knowledge enters and changes within the representation.
No single formalism has identical behavior across every domain. A language designed for taxonomic classification treats temporal change less directly than an event-centered calculus, while a probabilistic graph expresses uncertainty more directly than a classical ontology. Hybrid architectures consequently combine multiple representational components, but their interaction requires explicit rules connecting the meanings and inference procedures of those components.
The central problem of knowledge representation is therefore not the storage of isolated facts. It is the construction of an interpretable formal account in which statements, assumptions, and inference procedures jointly determine what a system can conclude about a domain.