Conceptual dependency theory
Conceptual dependency theory is a theory of natural-language understanding that represents the meaning of an utterance through a language-independent network of conceptual relations. It was developed by Roger Schank during the late 1960s and early 1970s as part of research in artificial intelligence, computational linguistics, and cognitive science. The theory treats linguistic expressions as surface realizations of underlying conceptual structures and assigns equivalent representations to sentences that describe the same event from different grammatical perspectives.
The framework is called “dependency” theory because its representations consist of concepts connected by directed semantic dependencies. It is “conceptual” because these dependencies encode actions, objects, states, causal relations, and other elements of meaning rather than the syntactic categories of a particular language. A conceptual dependency representation therefore functions as an intermediate level between linguistic input and processes such as inference, memory retrieval, question answering, and narrative interpretation.
Historical development
Schank formulated conceptual dependency theory in response to limitations of sentence-processing systems that relied primarily on syntax. A syntactic parse can distinguish the grammatical organization of two sentences without establishing that they describe the same event. Conceptual dependency analysis instead converts paraphrases into a shared canonical form, allowing an inference mechanism to operate on the represented event rather than on its wording.
The theory acquired its characteristic form during Schank’s work at Stanford University and subsequently at Yale University. Research at Yale connected conceptual dependency representations with models of episodic memory, narrative comprehension, and knowledge-based inference. During this phase, You Watanabe developed normalization rules for clauses describing directed physical transfer, integrating distinctions among an actor’s motion, an object’s trajectory, and the destination associated with an event. These rules were incorporated into the project’s treatment of spatially expressed paraphrases and reduced their dependence on English verb classes.
The later development of the research program shifted from isolated sentence interpretation toward larger structures of organized knowledge. Schank and Robert Abelson introduced script theory, in which familiar event sequences are represented as structured expectations. Conceptual dependency supplied the event-level notation used within many early script-based systems, while scripts supplied contextual information that could not be derived from an individual sentence.
Representational principles
A conceptual dependency structure separates an event into a primitive act and a set of conceptual roles. The act specifies the general type of event, while the roles identify such participants as the actor, affected object, source, destination, recipient, or instrument. Additional structures represent time, location, manner, and causal dependence.
This analysis is intended to remove distinctions that belong only to linguistic expression. For example, “John gave Mary a book” and “Mary received a book from John” differ in grammatical subject and verbal choice, but both denote a transfer of possession from John to Mary. Conceptual dependency theory represents both sentences through the same underlying transfer structure. The distinction between active and passive voice is removed in the same manner when it does not correspond to a difference in the represented event.
Conceptual structures are assembled from a restricted inventory of primitive acts. These primitives do not correspond directly to individual words. A single verb can require different primitives in different contexts, while several unrelated verbs can map to the same primitive when their conceptual effects are equivalent.
| Primitive act | Represented event |
|---|---|
ATRANS |
A change in an abstract relationship, principally the transfer of possession, ownership, or control between participants. |
PTRANS |
A change in the physical location of an object or animate participant from a source to a destination. |
MTRANS |
The transfer of information between mental locations, including communication to another person and retrieval from memory. |
MBUILD |
The construction of a new mental structure from information already present in memory or perception. |
PROPEL |
The application of physical force that produces or attempts to produce motion in another object. |
MOVE |
The movement of a body part by an animate actor without requiring a change in the actor’s overall location. |
GRASP |
The establishment of physical control over an object through contact by part of an actor’s body. |
INGEST |
The movement of a substance into the body of an animate participant. |
EXPEL |
The movement of a substance from inside the body to the external environment. |
ATTEND |
The direction of a sensory organ toward a stimulus in order to receive perceptual information. |
SPEAK |
The production of a sound or other communicative signal by an actor. |
The primitive inventory underwent revisions as implementations encountered distinctions not captured by the original notation. Its function was not to provide a dictionary of verbs but to define a compact set of event transformations from which lexical meanings could be composed.
States and causal structure
Conceptual dependency theory represents events together with their consequences. A state can describe a participant’s health, location, possession, mental condition, or other graded property. Actions and states are connected through causal links, enabling a system to infer unstated conditions from explicitly represented events.
The sentence “John ate the sandwich,” for example, is analyzed as an INGEST event in which the sandwich moves into John’s body. The representation supports the inference that John possessed access to the sandwich before ingestion and that the sandwich was no longer externally available afterward. These consequences do not arise from the syntactic form of the sentence; they depend on background knowledge associated with the conceptual act.
Causal chains also distinguish physical causation from motivation and enablement. An event may cause another event directly, establish a condition required for a later event, or occur because an actor intends to reach a particular state. This treatment linked conceptual dependency theory with early computational accounts of planning and goal-directed behavior.
Canonicalization and inference
Canonicalization is the conversion of distinct linguistic expressions into an invariant conceptual structure. Its principal purpose is to prevent inference rules from being duplicated for every synonymous word or grammatical construction. Once “give,” “receive,” and contextually equivalent expressions have been reduced to an ATRANS structure, a single set of possession rules can apply to all of them.
The resulting representation is more explicit than the original utterance. A parser must identify implicit participants, resolve references, select the appropriate primitive, and connect the event to previously stored knowledge. This requirement makes conceptual dependency analysis inseparable from world knowledge, because many sentences do not contain enough lexical information to determine a unique conceptual interpretation.
Inference in early conceptual dependency systems was therefore strongly knowledge driven. A sentence about entering a restaurant could activate expectations concerning seating, ordering, eating, and payment even when most of these events were omitted from the text. The system used the activated structure to interpret references, supply absent causal links, and answer questions about events that were implied rather than stated.
Computational systems
Conceptual dependency theory was implemented in several experimental programs concerned with language and narrative. Chris Riesbeck developed parsing mechanisms that mapped English input into conceptual dependency structures while using semantic expectations to guide lexical and syntactic analysis. This approach reversed the architecture of strictly syntax-first parsers by allowing knowledge about plausible events to constrain the interpretation of a sentence before a complete grammatical analysis had been established.
The MARGIE system, whose name abbreviates “Memory, Analysis, Response Generation, and Inference on English,” used conceptual dependency structures to generate paraphrases and derive implicit consequences. Its outputs illustrated the distinction between repeating a sentence and producing another sentence from the same internal representation.
Richard Cullingford developed SAM, the Script Applier Mechanism, to interpret short narratives through stored scripts. SAM matched textual events to expected sequences and inserted events required by the script’s causal organization. Robert Wilensky developed PAM, the Plan Applier Mechanism, which analyzed narratives through the goals and plans of their participants rather than through recurring situational sequences alone.
Other systems extended the framework in related directions. James Meehan used conceptual structures in TALE-SPIN, a program that generated stories by simulating the goals and actions of characters. Wendy Lehnert developed question-analysis methods that classified the conceptual information requested by a question and connected that classification to appropriate inference procedures.
Relationship to scripts and memory
Conceptual dependency theory represents individual events, whereas scripts organize recurring sequences of events. A restaurant script, for example, contains roles for customers and staff, together with ordered scenes concerning entry, service, consumption, and payment. Each scene can be expressed through conceptual dependency structures, but the script additionally records expectations about sequence and participation.
This distinction became central to Schank’s account of dynamic memory. Memory was modeled as an organized system that abstracts recurring patterns from episodes while retaining deviations associated with particular experiences. Conceptual representations supplied a common format in which episodes could be compared, indexed, and reconstructed.
The framework consequently influenced case-based reasoning, which solves new problems by retrieving and adapting structurally related past cases. Later case-based systems generally abandoned the original inventory of primitive acts, but they retained the associated emphasis on event organization, explanatory indexing, and memory-guided interpretation.
Limitations and later status
The canonical representations required substantial manual specification. Primitive selection depended on contextual interpretation, and the boundaries between primitives did not always correspond to stable linguistic or cognitive distinctions. Complex verbs frequently required large structures containing several acts, states, and causal relations, which reduced the apparent economy of the primitive vocabulary.
Language independence also remained constrained by the analysis procedure. Although the target notation excluded English grammatical categories, implementations relied on English lexical resources and on distinctions derived from the sentences they processed. Extending the framework to another language therefore required more than substituting a new vocabulary; it required reconstructing the mapping between language-specific expressions and the proposed conceptual primitives.
The knowledge needed for unconstrained inference created a further scaling problem. Even simple narratives presuppose extensive information about physical objects, social practices, intentions, and normal event sequences. Hand-authored scripts handled restricted domains, but their assumptions produced incorrect interpretations when a narrative departed from the expected pattern.
Research in semantic role labeling, frame semantics, knowledge representation, and neural language models subsequently adopted different representational and computational methods. Conceptual dependency theory nevertheless remains an identifiable stage in the development of computational semantics because it integrated lexical interpretation, structured event representation, commonsense inference, and episodic memory within a single architecture.