Roger Schank is a big influence on AI. In comparison to many other voices, his writing is along the variety of traditional AI and cognitive science, but a close reading of his work suggests that, while his influence on AI concepts has been profound, his approach is more cautious than that of Newell and Simon, advocating computation as a means for testing a theory of cognition, rather than asserting that minds are computational.

This particular book is generally about cognitive science, but more specifically about the understanding of stories. When one hears the sentence “I was hungry so I went to the restaurant,” it is easy to figure out the meaning, given what we know about hunger and restaurants and whatnot. Computers lack this background knowledge, so Schank works to articulate how that knowledge might be captured and represented computationally. There are two ways of looking at this knowledge, the first is the perspective of scripts, and then there is the perspective of plans. Schank moves from scripts to plans, finding the latter more robust and powerful, but I think that scripts are much more promising, especially in context of social interaction and a situation-centric view of cognition.

Introduction

This book was written after Newell and Simon, and after Weizenbaum, but before further development in cognitive science. Schank reveals a belief in the overlap of problems between computers and humans, which is the conceptual apparatus. However, Schank does not necessarily assert that computational and cognitive concepts are the same (as does Newell). I think that the overlap is smaller than Schank suggests, but his caution is encouraging given the period of circumstances.

The foundation of this work is the Conceptual Dependence Theory, which is composed of five main rules, the first two of which are important and I have copied here: (p. 11)

1. For any two sentences that are identical in meaning, regardless of language, there should be only one representation.
2. Any information in a sentence that is implicit must be made explicit in the representation of meaning of that sentence.

These give a strongly linguistic foundation to language, and suggest that meaning and cognition should come through the understanding of concepts and sentences. I find this to be a dubious supposition. The assertions also suggest that the meaning of a sentence can be non-problematically separated from its form, which is also contestable. This suggests that it is even possible to express statements analytically and completely, without implicit information, which is also problematic. Schank organizes conceptual dependency as deriving from several primitive acts. These are:

1. ATRANS: Transfer of possession, ownership, or control
2. PTRANS: Change of the physical location of an object
3. PROPEL: Application of a physical force to an object
4. MOVE: Move of oneself or a part of oneself
5. GRASP: To grasp an object
6. INGEST: To “take in” or consume something into one’s body
7. EXPEL: To expel something from one’s body
8. MTRANS: Transfer of some mental information
9. MBUILD: Construction of new information
10. SPEAK: To produce sounds and say something
11. ATTEND: To focus one’s senses on something

Note that the phrasing of these asserts a kind of literalism and preoccupation with physical situations. On one hand, that is good, because it suggests some sort of anchoring in embodiment, but it also pre-loads the conceptual system with many anchored terms revolving around production, physical movement, and ownership. This does not, for example, suggest of ways to express emotions, build relationships, or change one’s mood or disposition. These could all be expressed within the system, but secondarily, whereas physical movement and ownership are built in at the first level.

The theories of cognition in AI must be specified fully. Schank gives an example of someone asking how to get to Coney Island, and being told to take the ‘N’ train to the last stop. These instructions are described as inadequate, at least for a computer, because all kinds of background knowledge are required to make sense of how to use the subway in the first place. This is a good example of the difference between situational and top-down models. Schank is proposing a top-down structure, where a general plan: go to Coney Island, is made up of smaller and smaller parts: take the ‘N’ train to the last station (which consists of going to the metro station, getting a fare card, going through the gate, getting on the train, etc etc etc). I agree that an AI simulation of cognition must know how to deal with this low level knowledge, but in a situational system, this knowledge really should be secondary. In context, the directions are certainly sufficient, and at the right level of abstraction, the finest granulation of instructions are not necessary.

Causal Chains

This chapter discusses ways of interpreting sentences with a distressing degree of literalism. The logic is used with a kind of causal chaining. Interpretation is described as a filling in of the blanks in a causal chain. For example, Schank gives the sentence “John cried because Mary said she loved Bill.” This sentence, with its face value taken at the most literal level, is absurd, John cried because Mary’s speaking. However, this is not the meaning of the sentence, at all. Schank argues that the reader constructs a causal chain behind the contents of the sentence, that Mary speaking to John transfers factual knowledge of Mary loving Bill to John, and this is what made John cry. The degree of chaining in this is ridiculous, as much as in the supermarket example given by Cohen, Morgan, and Pollack. I think that understanding of these sentences has much more to do with common usage an practices of use, or even in a sense of internalized “grand narratives” than causal chains. Schank gives a calculus of causation which is built from actions, states, reasons, and enabling.

Scripts

The chapter on scripts is rather hilarious from my perspective. To me, scripts are the most productive thing to be gained from an analysis of Schank’s book, however, in context, they are used only as a stepping stone to the discussion of plans in the next chapter.

Scripts are very useful structures to analyze. Scripts structure information that is relevant in the context of a particular situation, and organize new inputs and events in context. This is consistent with my understanding of models, and also with Goffman’s sense of framing. The discussion Schank gives is still preoccupied with story comprehension, especially as relates to included versus excluded information. For example, if someone is comprehending a set of sentences and knows what script is being followed, it will be easier for the reader to identify and contextualize the meaning of each sentence.

Schank explains scripts as being composed of props, roles, states, entry conditions, and resulting conditions. The goal of developing this formulation is the SAM program (Script Applier Mechanism), which understands (presumably) scripted stories, and is able to answer questions about them. The script described is the “restaurant script”, where a customer can go to a restaurant, order something, eat it, leave a tip, pay, and leave. Given gaps, it is still possible to piece together what might be happening in a story that abides by this script. Narratively, these are still very uninteresting, but I think they have the potential to be more meaningful. For instance, scripts could be annotated with other layers of meaning that give some sort of narrative value to how a character might act within the script.

Scripts are fitted with metadata, specifically that which describes how, based on events, the ssytem will recognize what scripts to use. Script headers describe the preconditions, instrumental relations, locales, and so on, used by scripts. This helps a script analysis program understand what script might be in operation at a given moment. This is not complete, but gives some background to the situation and ambiguity problems that come up with a situational model of interaction.

Schank moves to examine how to handle statements that are not immediately relevant to the script. Notably, he examines breaches and distractions. Distractions are not especially relevant to me, but breaches are extremely useful. Social situations are full of scripts and breaches in those scripts. From the perspective of simulation, when breaches occur, agents scramble to recontextualize and reground themselves in some sure footing of knowing how to interact. Scripts designate social rules, procedures, and conventions. Usually what is interesting narratively are the breaches. Frequently breaches allow scripts to interact simulaneously and play off each other. Schank gives a listing ways to handle unexpected inputs within scripts: (p. 53)

1. Does it specify or imply the absence of an enablement for an impending script action? (Obstacle)
2. Does it specify or imply that a completed action was done in an unusual manner, or to an object other than the one(s) instantiated in the script? (Error)
3. Does it specify an action which can be understood as a corrective resolution of an interference? (Prescription) This question would be activated when an obstacle is inferred from or described directly in the text.
4. Does it specify or imply the repetition of a previous action? (Loop) This is activated when an error is inferred from or described directly in the text.
5. Does it specify or imply emotional expression by the actor, likely to have been caused by an interference? (Reaction)
6. Does it specify or imply that the actor will have a new goal that has nothing to do with the original script? (Distraction)
7. Does it specify or imply the motivated abandonment of the script by the main actor? (Abandonment)

Note that emotional responses are “unexpected inputs.”

Schank poses scripts as a powerful component to cognition and to understanding the world. People adapt and transform scripts, but this does not mean that they fall under the general category of knowledge transfer (something we know to be flawed, eg Lave). This is precisely because scripts are known and learned from experience, and by being experienced. This dimension is not discussed (Schank may not even agree with it), but I believe this is a potent observation.

Plans

Turning to planning, Schank makes a claim here that I totally disagree with: that scripts come from plans. As described, plans are means of satisfying goals. In execution, plans make use of several low level behaviors. For example, the plan “USE(x) = D-KNOW(LOC(X)) + D-PROX(X) + D-CONT(X) + I-PREP(X) + DO”, where each of the D- expressions are subgoals that can be satisfied by other actions. The focus of discussion is still story comprehension, so the object is to understand the plans of story characters.

Goals

Schank introduces several types of goals:

1. S: Satisfaction: satisfying a basic need
2. E: Enjoyment: doing something for the sake of pleasure
3. A: Achievement: attaining some desirable outcome
4. P: Preservation: maintaining some desireable state
5. C: Crisis: responding to a sudden pressing emergency
6. I: Instrumental: a goal that realizes the precondition of another goal
7. D: Delta: effects a state change in the world

This general system of goals has been influential and used by others, notably Ortony, Clore, and Collins. This is still bound in understanding stories, but is reasonable as a general scheme of understanding motivation. The goal system alone is consistent with the idea of having conflicting goals.

Themes

Schank introduces the idea of themes, which make sense for story understanding, but are totally neglected within conventional AI. This asserts that goals and plans work in context of some broader theme, which guides the goals that occur and the plans to achieve them. In terms of stories, they are what a story might be about, for instance, success, interpersonal relationships, and so on. This serves as a filter or model to focus on select elements of a story world.