Textbook of AI

A script, introduced by Roger Schank and Robert Abelson in their 1977 book Scripts, Plans, Goals and Understanding, is a knowledge structure for stereotyped sequences of events. The canonical example is the restaurant script: enter, sit, look at menu, order, food arrives, eat, ask for bill, pay, leave, with sub-scripts (or tracks) for variants such as fast-food, fancy-restaurant, takeaway and café. Each script specifies typical roles (customer, waiter, cook, cashier), props (menu, table, food, money) and entry/exit conditions.

Function in language understanding

Scripts supplied default expectations against which novel descriptions of events were interpreted. A two-sentence pair like "John went to the restaurant. The bill was twelve dollars." can be understood, including the implicit ordering, eating, waiter interaction and paying , only by mapping it onto the restaurant script and inheriting the unstated steps. Schank's research group at Yale built systems including SAM (Script Applier Mechanism), PAM (Plan Applier Mechanism), FRUMP (newspaper-story summariser) and CYRUS (autobiographical memory) on this foundation.

Scripts dovetailed with Schank's Conceptual Dependency representation, a language-independent semantic formalism with primitive acts (PTRANS for physical transfer, MTRANS for information transfer, INGEST, ATRANS, etc.) and case roles, intended as the universal target language into which all natural-language input would be parsed.

Plans, goals and themes

Scripts, Plans, Goals and Understanding did not stop at scripts. Plans were used when a stereotyped script did not apply: a goal-directed search over actions to achieve a state. Goals were classified into types (achievement, preservation, satisfaction). Themes captured longer-term motivations (life themes, role themes). Together these layers gave a hierarchical theory of agent comprehension.

Limits and decline

Scripts proved difficult to scale. The world contains an enormous number of stereotyped sequences, each with many variants and exceptions, and there was no good way to acquire scripts automatically from data. Hand-engineering scripts for every domain became a bottleneck. The "script-and-frame" tradition was eclipsed in the 1990s by statistical NLP and again in the 2010s by deep learning.

Modern relevance

Large language models implicitly encode much of the same procedural and commonsense knowledge, they readily extend "John went to the restaurant. The bill was twelve dollars." with appropriate continuations, reason about waiters and tipping, and generalise across thousands of unseen scenario variants. The lesson is that the structure Schank identified is real and important; what was wrong was the assumption that it had to be programmed in by hand. Modern knowledge-grounded dialogue, commonsense reasoning benchmarks (e.g. HellaSwag, PIQA, ROCStories) and schema induction research can be read as a return to script-like questions, now answered by scale rather than knowledge engineering.