Schank's Conceptual Dependency Theory and Action Primitives

Roger Schank’s Conceptual Dependency (CD) theory revolutionized early artificial intelligence by proposing a method to represent the underlying meaning of natural language completely independent of its grammatical syntax. Developed in the late 1960s and 1970s, the theory posits that any two sentences with the same meaning, regardless of the language or words used, should have the exact same internal representation. To achieve this, Schank introduced a small set of "action primitives"—fundamental building blocks of action—that stripped away the surface structure of sentences to capture pure semantic intent, allowing computers to infer unwritten truths and understand context.

Prior to Schank's work, natural language processing relied heavily on syntactic parsing, which focused on grammar rules, parts of speech, and sentence structure. This approach often failed because human language is inherently ambiguous, and different grammatical structures can convey the exact same meaning. For example, "John gave Mary a book," "Mary received a book from John," and "John handed a book over to Mary" all look different syntactically. Schank realized that for an AI to truly comprehend these sentences, it needed to bypass the specific verbs used ("gave," "received," "handed") and represent the single conceptual event occurring beneath the words.

To eliminate this dependency on syntax, CD theory introduced the concept of a "conceptualization," which centers around a core set of eleven action primitives. Schank argued that thousands of verbs in human language could be broken down into combinations of these eleven basic actions, categorized into physical, mental, and instrumental acts.

The physical primitives handle actions in the tangible world:

  • ATRANS: The transfer of an abstract relationship, such as possession, ownership, or control (e.g., give, buy, steal).
  • PTRANS: The transfer of the physical location of an object (e.g., go, walk, drive, put).
  • PROPEL: The application of a physical force to an object (e.g., push, pull, throw, kick).
  • MOVE: The movement of a body part by its owner (e.g., raise a hand, scratch, nod).
  • GRASP: The physical grasping of an object by an actor (e.g., hold, grab, clutch).
  • INGEST: The taking in of an object by an animal to the inside of that animal (e.g., eat, drink, smoke, breathe).
  • EXPEL: The expulsion of an object from the body of an animal into the physical world (e.g., sweat, spit, cry).

The mental and instrumental primitives account for internal processes and communication:

  • MTRANS: The transfer of mental information between animals or within an animal's own memory (e.g., tell, read, remember, see).
  • MBUILD: The construction of new information from old information by an actor (e.g., decide, conclude, imagine, consider).
  • SPEAK: The actions of producing sounds (e.g., say, scream, purr).
  • ATTEND: The action of directing a sense organ toward a stimulus (e.g., listen, look, smell).

By using these primitives, CD theory represents actions through conceptual roles rather than grammatical ones. Instead of identifying a subject, verb, and direct object, a CD representation identifies an Actor, an Action (one of the primitives), an Object, and a Direction (Recipient or Location).

Returning to the earlier example, both "John gave a book to Mary" and "Mary took a book from John" are reduced to the exact same primitive structure:

  • Action: ATRANS (transfer of possession)
  • Actor: John (or Mary, depending on who initiated the action)
  • Object: Book
  • Direction (To): Mary
  • Direction (From): John

Because the action is mapped to ATRANS, the AI system does not care about the syntax of the original English sentence. Furthermore, each action primitive comes with a built-in set of inference rules. When an AI encounters an ATRANS event, it automatically infers that the "To" entity now possesses the object, and the "From" entity no longer possesses it. If a sentence states, "John ate a frog," the verb "ate" triggers the INGEST primitive, automatically allowing the AI to infer that the frog is now inside John and is no longer in its previous location.

Ultimately, Roger Schank's Conceptual Dependency theory detached meaning from syntax by treating words merely as pointers to deeper conceptual structures. By decomposing complex verbs into eleven universal action primitives, CD theory provided artificial intelligence with a standardized, language-independent framework capable of logical inference, memory retrieval, and genuine semantic comprehension.