Textbook of AI

Conceptual dependency theory (CD), introduced by Roger Schank at Stanford in 1972 and developed at Yale through the 1970s and 1980s, is a language-independent representation of meaning intended as a universal interlingua for natural-language understanding. Its central claim is that the meaning of any action sentence can be decomposed into a small inventory of primitive acts, each with a fixed argument structure (actor, object, instrument, source, destination, time, ...). The classic primitive set comprises eleven acts:

• PTRANS -- physical transfer of location ("John went to the store")
• ATRANS -- abstract transfer of possession or relationship ("John gave Mary a book")
• MTRANS -- mental transfer of information between minds or memory stores ("John told Mary", "John remembered")
• MBUILD -- mental construction of new information ("John decided", "John concluded")
• INGEST -- internalisation of an object ("John ate the apple", "John inhaled smoke")
• EXPEL -- the inverse of INGEST ("John exhaled")
• GRASP -- physical contact and grip ("John held the cup")
• MOVE -- a body-part movement ("John kicked")
• SPEAK -- production of sound ("John shouted")
• ATTEND -- direction of a sense organ ("John looked at the painting")
• PROPEL -- application of force to an object ("John pushed the door")

The motivation was that surface English (and natural languages generally) is heavily redundant. The sentences "John gave Mary a book", "Mary received a book from John", "John handed a book to Mary" and "Mary acquired a book from John" all express the same underlying ATRANS of book from John to Mary. By representing the meaning rather than the surface form, CD promised to support inferences that depend on meaning rather than syntactic form, to allow paraphrase and translation through an interlingual representation, and to expose causal structure: every primitive act has typical preconditions and consequences that drive forward inference.

CD was scaled up to discourse via two further constructs. Scripts (Schank and Abelson, 1977) are stereotyped event sequences that capture the structure of routine situations -- the canonical example being the restaurant script (entering, ordering, eating, paying, leaving). Plans capture goal-directed behaviour where no script applies. The Yale group built an extensive series of natural-language understanding programs on this foundation: MARGIE (paraphrase, 1973), SAM (Script Applier Mechanism, 1975), PAM (Plan Applier Mechanism, 1978), FRUMP (a fast newswire summariser), BORIS (story understanding, 1982) and IPP (Integrated Partial Parser).

Conceptual dependency did not scale to broad-coverage natural-language processing -- the primitive vocabulary was always too small for the open-ended diversity of human action concepts, and inference rules were brittle. By the late 1980s the CD programme was largely supplanted by statistical NLP and, later, by neural language models which represent meaning implicitly in continuous embedding spaces rather than in explicit primitive structures.

Even so, CD's methodological insights endure. The injunction to represent meaning, not surface form is the foundational commitment of all subsequent semantic parsing, abstract meaning representation (AMR), and frame semantics (Fillmore). Schank's later work on case-based reasoning descends directly from CD, and his cognitive architecture continues to influence research on commonsense reasoning in modern LLMs.