Writing and Representation

Philip E. Agre
Department of Information Studies
University of California, Los Angeles
Los Angeles, California 90095-1520
USA

pagre@ucla.edu
http://polaris.gseis.ucla.edu/pagre/

This is a chapter in Michael Mateas and Phoebe Sengers, eds, Narrative Intelligence, Amsterdam: John Benjamins, 2003.

Please do not quote from this version, which probably differs in small ways from the version that appears in print.

8300 words.

[Author's note: This is a heavily abridged and revised version of a paper that I circulated informally in 1989. I have not tried to bring it up to date. Some parts of the argument are developed further in (Agre 1997). Another subsequent work that treats these issues is (Hutchins 1995).]

Introduction

The notion of representation obviously labors under a long philosophical history (Judovitz 1988, Rorty 1979, Silvers 1989), not to mention the history of art (Hagen 1986, Wallis 1984), literature (Auerbach 1953, Brodsky 1987, Krieger 1987), and historiography (Hartog 1988, White 1973). These days, though, it also labors under an appreciable technical history, handed down through a practice of building computer systems that construct, maintain, and manipulate "representations" (Brachman and Levesque 1985, Haugeland 1981). And the philosophical and computational issues interact. I often find that philosophy helps to interpret the difficulties that arise in my technical practice. And I want to believe that technical practice can help philosophy. In writing the stories that follow, I have explored some places where technical questions align with philosophical answers. I don't yet know how to convert these answers back into technical practice.

I disagree with two widespread ideas about representation, namely "semantics" and "world models". The main tradition of semantics holds -- presupposes -- that a representation has a "meaning" or a "content" independent of the identity, location, attitudes, or activities of any particular agent (but cf. Barwise and Perry 1983). This meaning or content is often understood as a systematic, objective relationship between the representation itself and states of affairs obtaining in the world. A world model is a component of a physically realized computational system, an object whose internal structure stands in a systematic, objective, analogical relationship to states of affairs presumed to obtain in the world. Some computational process maintains the model as the world changes; reasoning about the world involves inspecting and manipulating the model. I will argue that both semantics and world models overlook central features of representations and their use.

On the positive side, I offer two suggestions about how people use symbolic representations. The first is that people interpret symbolic representations in making sense of particular situations. Interpretation is a situated activity.

(Whatever the form of the representation, whether written or spoken or displayed on a video monitor, I shall speak of its configuration of symbols as a "text". The "representation" is the actual material object: the sheet of paper, the speech signal, or the video image.)

My second suggestion is that what a given text is talking about is a fresh problem in every next setting (Amerine and Bilmes 1988, Zimmerman 1970). The work of relating a text to a concrete setting -- looking around, poking into things, trying out alternative interpretations, watching someone else, getting help -- will generally be both "mental" and "physical", though it is best not to distinguish. Relating a text to a concrete setting takes work because the text might be relevant to the situation in a great variety of ways. The text has a great deal of "play", so that much of one's interpretive effort must wait until the time comes. This is the opposite of extracting a "meaning" from a text as soon as it arrives. The point is not that interpretation is wholly unconstrained by the text; rather, interpretation is constrained jointly by the text and by the circumstances in which it is interpreted. The only way to explain the point is through examples.

(1) I'm trying to find a friend's house in a heavily wooded mountainous rural area. I had received directions to the house by e-mail and printed them out so I could carry them in the car. I knew the main road, but I knew nothing about the residential streets leaving it. The last two paragraphs read:

About a mile up from the intersection, look on the left for Elk Tree Road -- it's a dirt road with a little bus stop at the end. Follow Elk Tree past the first left (Elk Tree WAY) to the mailboxes and take the middle of the three-way fork on the right, Upper Elk Tree ROAD.

My place is the first one on the left, #27. Park on the left shoulder near my red Honda, and come down the steps and up the stairs to my front deck. (If you went to the main door, you'd get my landlords, not me.)

To make a long story short, the middle branch of the three-way fork immediately branched into Lower Elk Tree Road steeply down to the left and Upper Elk Tree Road steeply up on the right. A large hand-painted sign for Upper Elk Tree Road clearly marks the split, but I was already looking for the first place on the left. Writing this now, I realize that it makes perfect sense for Lower Elk Tree Road to branch steeply down and Upper Elk Tree Road to branch steeply up, but at the time I didn't give the right branch any thought at all since I simply did not see the branch as a branch. If you're standing there looking at the branch as a branch then it's hard to imagine how it looked to me, but it never occurred to me that I had another choice of roads to make. The sign must have occupied a reasonable portion of my visual field, but I was already looking for a house, not a sign, and to the left, not the right. In short, I saw what I assumed the instructions were telling me to see. Told about a "three-way fork to the right" and unable immediately to see such a thing in that place, I worked, successfully, to see it. Told about "the first [place] on the left", I unquestioningly saw what the instructions implicitly told me was there, namely the road on which my friend's house was located.

(2) A friend recently taught me to fold origami paper cranes. In walking me through the various steps, she often had to explain by some combination of words, pointing, and demonstrations where and how to fold next. The intermediate forms don't look much like cranes, and the paper keeps taking on unexpected new identities as you fold it, even after you've gotten reasonably proficient. These forms can be hard to talk about because they're importantly asymmetrical in nonobvious ways. Even demonstrations are only of limited use if you can't see the asymmetries for yourself.

In the course of her explanations my friend said things like "put your finger in the pocket", "fold it back to make a boat", and "make the legs skinnier". Making each of these metaphors refer to parts and aspects of the folded paper always took considerable effort, even though it was always wholly evident in retrospect. Much of my friend's job was to get me to look at my partly-folded origami crane in the right way, so that certain parts and aspects would stand out as units for me. She was teaching me the skill of seeing my paper as having a pocket, a boat, or legs. Although I got better at this skill, it never stopped taking work. The work only became more routine.

(3) The final example comes from my experience teaching people to program computers. If you're comfortable in front of a computer terminal, it's easy to teach the wrong things. You've got all kinds of theory, but theory doesn't help someone who hasn't yet gotten the idea of being "in" the editor. So I sit the student at the keyboard and tell them very concretely what to look at and what to type. As they get comfortable, my instructions grow more abstract. For example, I might say "Type open paren then ...", only to see them type the letters "o p e n" and "p a r e n". In that case I have to point out where the parenthesis keys are. Later, though, I can say things like, "Let's define a function called ...". When they're learning to read code, I have to point out that there are conventions about indentation that result in common types of code having characteristic shapes. And I don't explain abstraction hierarchies until I explain that two hunks of code that look alike are often good candidates for a common abstraction.

As these examples illustrate, my prototype of representation is natural language, whether as spoken utterances, written texts, or mental thoughts. In each case, figuring out what in the situation the text was talking about took work: creative improvisation, reference to artifacts, and interactions with others. The work consisted in relating natural language to concrete situations: identifying the things the words were mentioning, seeing materials under metaphorical descriptions, and heuristically associating visual patterns with verbalized technical abstractions. And the work required to make sense of "take the middle of the three-way fork" or "put your finger in the pocket" or "define a function" might differ greatly in different settings or under different conditions.

I want to concentrate on representations written on paper. The image of standing in a kitchen or on a street with a written text in your hand is a good reminder that relating texts to circumstances requires work and that this work requires understanding in some measure what you're doing. You have to understand what you're doing since the text certainly doesn't (cf. Searle 1981). Obviously representations often influence your actions, but you don't understand what you're doing in virtue of owning them. This point is supposed to apply equally to all forms of representation, not just writing. The idea that understanding does not reside in representations is difficult and consequential. The paper's later stories will explore this idea in the context of "internal language".

Five fairly independent notes follow.

The homunculus and the orbiculus is an assault on the notion of a world model, placing it in a philosophical tradition of trying to explain the human ability to act competently in the world by pretending that the relationship between person and world is reproduced inside the person's head. This sort of explanation is seductive because it plays to the principal strength of current computational technology: building abstractions inside of computers that are almost entirely cut off from the outside world. But it doesn't work well in practice.

Writing as bad and good metaphor for representation contrasts two ways in which written texts might be regarded as prototypes of representation. The first, "bad" way focuses (however tacitly) on certain physical properties of written texts. This bad understanding of writing is central to the main tradition of computing. The second, "good", way concentrates on more abstract aspects of written texts: both the text and your surroundings are outside of you, and so it takes work to see the world as being what the text is talking about.

A story about photocopier supplies also concerns written instructions. A secretary is justifiably annoyed because somebody has put laser-printer dry toner in the photocopier. What happened and why? At issue is the selective use people must make of the representational materials that surround them. I defend the culprit, arguing that the phrase "dry toner" on a bottle is not "ambiguous" in a way that anybody could be expected to notice, even if evidence serving to identify and resolve the ambiguity is readily available. One would like photocopier users to be "careful", but it's hard to formulate the demand in an actionable way, given that such a problem could hide behind any of the vast number of unarticulated assumptions that form the background of any such activity.

A story about some instructions at a performance in an art gallery is another story about instructions, this time a single sentence that one person hollered to a group of others. These instructions did not function well because the recipients could not see, or even imagine, what in the setting the instructions could be talking about.

A story about my routines for reading the Sunday Globe is a story about an instruction I issued to myself in the course of reading the newspaper one Sunday morning. The pattern of activity in which the instruction participated had long since become routine. Nonetheless, the way in which the instruction went wrong reveals that it was something like an natural language imperative and not, for example, a computer program. In this case, a change in the environment led me to make a different sense of an ambiguous phrase than I used to.

The stories are not supposed to prove any general propositions. Instead, they invite you to be aware of similar phenomena in your own experience of everyday representation-use. Computational investigations and awareness of everyday life can influence one another. Parallel pursuit of these two kinds of inquiry will, I believe, lead to deeper understandings of why our life is the way it is and why machines can take certain forms and not others.

The homunculus and the orbiculus

In the old days, philosophers accused one another of believing in someone called a homunculus -- from Latin, roughly "little person". For example, one philosopher's account of perception might involve the mental construction of an entity that "resembled" the thing-perceived. Another philosopher would object that this entity did nothing to explain perception since it required a mental person, the homunculus, to look at it. Computational ideas appeal to these philosophers because they can imagine "discharging" the homunculus by, for example, decomposing it into a hierarchy of ever-dumber subsystems (Dennett 1978: 124).

But the argument about homunculi distracts from a deeper issue. If the homunculus repeats in miniature certain acts of its host, where does it conduct these acts? The little person lives in a little world -- the host's surroundings reconstructed in his or her head. This little world deserves a Latin word of its own. Let us call it the orbiculus. One way to say "world" is orbis terrarum, roughly "earthly sphere". But orbis, I am told, extends metaphorically in the same ways as "world" in English: one might speak of the world of a peasant or a movie director, meaning roughly their existential world, "the world they live in" (more literally, their sphere). So the orbiculus is your world copied into your head.

AI is full of orbiculi. A "world model" is precisely an orbiculus; it's a model of the world inside your head. Or consider the slogan of vision as "inverse optics": visual processing takes a retinal image and reconstructs the world that produced it (Hurlbert and Poggio 1988). You'll also find an orbiculus almost anywhere you see an AI person talk about "reasoning about X". This X might be solid objects, time-extended processes, problem-solving situations, communicative interactions, or any of a hundred other things. "Reasoning about" X suggests a purely internal cognitive process, as opposed to more active phrases like "using" or "participating in" X. AI research on "reasoning about X" requires representations of X. These representations need to encode all the salient details of X so that computational processes can efficiently recover and manipulate them. In practice, the algorithms performing these abstract manipulations tend to require a choice between restrictive assumptions and computational intractability (see Brachman and Levesque 1984, Hopcroft and Krafft 1987).

If you prefer the phrase "using X" over "reasoning about X", an AI person will ask you, "But we can reason about things that aren't right in front of us, can't we?" AI's version of mentalism offers seductive answers to many questions, and this is one of them. According to mentalism, reasoning about a derailleur proceeds in the same way regardless of whether the derailleur is in front of you or across town. Regardless of where the derailleur is located, you reason about it by building and consulting an orbicular derailleur-model. If having the derailleur present helps you, it is only by helping you build your model.

This is, of course, contrary to common experience. As we all know, the first several times you try to reason about a derailleur (1) it has to be sitting right in front of you and (2) you have to be able to look around it, poke at it, and take it apart (Chapman and Agre 1986). I've disassembled and reassembled several derailleurs. Yet without a derailleur in front of me, or at least a good diagram, I cannot explain how a derailleur changes gears, or even list the parts involved. Experts can, but not an amateur like me. Why aren't several disassemblies and reassemblies of derailleurs enough to build a mental model of them?

Maybe the computational complexity of reasoning with realistic world models is trying to tell us something. Maybe what you learn when you gain experience with a derailleur or a city or a recipe is more specific. Perhaps it is more biased to the specific things you've had to remember in the course of the activity. Perhaps it is more closely tied to your goals at particular moments of the activity. Perhaps it is more organized around the experience of the individual situations that arise in the course of the activity. These are difficult ideas. The question is complicated and messy and poorly worked out. But that's to be expected. Expecting it to be easy is a sign of addiction to the easy answers of the orbiculi.

Writing as bad and good metaphor for representation

Within the technologically informed human sciences, cognition is almost universally understood to involve the mental manipulation of assemblages of symbols called representations. These representations represent the individual's world -- they are the orbiculus. The vast majority of this research assumes symbolic representations to have certain properties. They are:

object-like (neither events nor processes),

passive (not possessing any sort of agency themselves),

static (not apt to undergo any reconfiguration, decay, or effacement, except through an outside process or a destructive act of some agent),

structured (composed of discrete, indivisible elements whose arrangement is significant in some fashion),

visible (can be inspected without thereby being modified), and

portable (capable of being transported to anyone or anything that might use them without thereby being altered or degraded).

This conception of representation-as-writing is topical for several reasons. The connectionist movement has lent urgency to the seeming conflict between symbolic manipulation and the relatively simple, uniform, statically and locally connected, highly parallel hardware of the human brain (Fodor and Pylyshyn 1988, Hutchins 1986, Rumelhart et al 1986). Anthropologists such as Goody (1986), Ong (1982), Harris (1980, 1987), and Latour (1986) have challenged views of cognition that make universal principles out of psychological and social phenomena found only in literate cultures.

This section has three purposes. First I argue that symbolic representation in artificial intelligence is, historically, modeled on written texts, as opposed to (say) photographs or spoken utterances. Then I describe how writing is a bad metaphor for symbolic representation. These arguments implicate prevalent technical methods. Finally I describe how writing is a good metaphor for symbolic representation. These arguments suggest new technical directions.

Representation as writing

Roy Harris (1987), among others, has argued that ideas about representation in philosophy and linguistics have been biased by writing. He observes that these fields have emphasized those aspects of human utterances that appear in a conventional written representation. One might read in a textbook a sentence such as,

Suppose that John says to Mary, "Please close the window".

• his tone of voice,
• his articulation of the various phonemes,
• the shape of his intonation,
• the timing of the various elements within the utterance,
• the timing of his utterance relative to other actions and events,
• whether he and Mary have a history of interactions over this window,
• his position relative to Mary and the window,
• his posture,
• his gestures,
• his facial expression,
• the direction of his gaze,
• whether and when he has caught Mary's gaze.

Given that these aspects of speech regularly affect the import of utterances, a written sentence must be considered a poor representation of a spoken utterance. But most philosophical and linguistic analyses have proceeded on the basis of this idealized representation, a tradition that AI has carried on. The point is not that these fields talk about writing; only that they concentrate on the aspects of representation that writing normally captures. As a result, theories will naturally tend to lean on distinctions that writing captures, and not on the many distinctions it doesn't.

Among the many routes by which the writing metaphor entered AI practice, one moment stands out: Newell and Simon's (1963, 1972) invention of symbolic programming. Most of Newell and Simon's domains, especially in their earlier work, have been domains like cryptarithmetic in which the "world" consists of a sheet of scratch paper. Newell's production system models do not contain separate mechanisms for the scratch paper and for the agent's "short-term memory". Newell and Simon invented symbolic programming in order to implement the sorts of structures and operations that their models specified. List structures, like scratch paper, and like the symbolic structures of all subsequent AI programming languages, have many properties of writing and few properties of speech. People invented writing because there's nothing in their heads that's anything like paper.

Writing in the head

AI research is often caught in a pattern whereby mechanisms that seem extremely "expressive", "powerful", and "general" refuse to scale up. Let's return to the properties that AI has ascribed to symbolic representations -- object-like, passive, static, structured, visible, and portable -- and consider how they lead to difficulties of scaling and implementation.

Symbolic programming languages endow their datastructures with all six properties of writing. They implement these properties using pointers that metaphorically make objects visible to processes. Pointers connect the components of structures. One transports a structure by "passing" a pointer to it. Structures only change when processes change them. Pointers do not obey any locality beyond that of their own connectivity. Thus they are eminently reconfigurable.

Pointers cause two sorts of difficulties. First, they require their implementation medium to be infinitely reconfigurable (Chapman 1991: 35-41). They fight both against the locality of physical space and against the inertia of physical machinery and its interconnections. On serial machines we observe this difficulty in the complexities of dynamic storage management. On parallel machines we observe it in the complexities of shared-memory management. Pointers also cause algorithmic complexity. Just as you can write symbols on paper in any order, a pointer can point at anything. As a result, algorithms for the manipulation of symbolic pointer-structures often suffer from the combinatorial arbitrariness of the objects they reason with.

If writing is a good metaphor for symbolic representation, then, it is not because we have things in our heads that are object-like, passive, static, structured, visible, and portable. These properties of writing don't help us to understand human use of symbolic representation in general because they are the properties of written texts that are specific to written texts. Far from picking out the essence of symbolic representation, they dwell on the physical activity of using a written text: inscribing, gathering, comparing, storing, and destroying. The invention of writing was important precisely because it permitted these useful forms of activity.

Writing as representation

How, then, can writing serve as a model of symbolic representation? Imagine that you're using a recipe -- that is, a recipe on paper -- to help you cook dinner. Or perhaps you're using some directions -- again, on paper -- to help you get to a party. The paper has a paradoxical position. Even though it's a physical object with a definite size, mass, and location, it plays its role -- at least qua representation -- entirely through your interpretation. And even though it seems to offer opinions about the particulars of the situation, it only does so because you figure out what in your surroundings it's talking about.

The paper in your hand is both part of the material situation and doubly removed from it. It underdetermines the sense you make of it because it is separate from you -- after all, someone else in the same situation would probably do something different. And it underdetermines what in the situation it picks out because you are separate from your surroundings -- after all, it would probably be useful in other situations as well. A similar argument then applies, not just to notes written on paper, but to all symbolic representations: their meaning must be completed in the act of use (Ingarden 1973). The world does not come innately parcelled out into the categories we find mentioned in written texts. Instead, people use representations to help them make sense of particular situations. What a given text is talking about is a fresh problem in every next setting.

(For those who care about such things, this is what Jacques Derrida means by the word "writing". For introductions to Derrida's philosophy see Culler (1982) and Norris (1982). I have also been influenced by Garfinkel's (1984 [1967]) ethnomethodological ideas about the indexicality of representation use. For an introduction see Heritage (1984).)

This idea has many consequences for computation. The relationship between internal processes and internal symbolic representations is qualitatively the same as the relationship between a person and a sheet of paper. The idea of cognition as a process operating on symbolic representations is therefore of less help than we had hoped. We can put symbolic representations inside our robots, but the hard problems remain.

What other forms might a theory of symbolic representation take? Vygotsky (1978 [1934]; for an introduction see Wertsch 1985) suggests that cognitive skills involving mental representation arise from the everyday use of physical representations, many of which are embedded in patterned social relationships. These skills are diverse because the everyday forms of representation use are diverse: speaking and hearing give rise to internal speech, making and looking at pictures give rise to visualization, and so forth. And the "internal" and "external" uses of representations tend to blur together in practice. But the internal processes differ from the external processes because the insides of our heads aren't like the outsides. Instead, they are shaped by the kind of machinery we have in our heads. For example, internalized speech is far more consequential than internalized writing because our brains are better suited for reproducing speech than writing. If this is true, then many properties of speech that philosophy and linguistics have marginalized, such as tone of voice and intonation (and their pragmatic import in particular contexts of use), will have to be readmitted to our theoretical center stage.

A story about photocopier supplies

Back when I worked at counter jobs, I discovered that people are oblivious to signs. You could put a big red sign with stars and arrows

THE MACHINE WILL BE BACK UP AT MIDNIGHT

Here is an example of this effect.

Date: Tue, 17 May 1988 17:05 EDT
From: J...
To: All-AI
Subject: Xerox copiers and Lazer Printers

Due to someone's IGNORANCE, CARELESSNESS, or LACK OF PATIENCE, SOMEONE PUT DRY IMAGER FOR THE LAZER MACHINE INTO THE XEROX MACHINE on the 8th floor. These supplies, although both dry imager, ARE NOT INTERCHANGEABLE!!! It says on the box it comes in (and on the bottle itself) which machine it is for.

We were warned by the Xerox people before that if this happened again, they may discontinue servicing our machines -- not to mention the cost of having it corrected (or maybe having to get a replacement).

If there is a problem with the xerox or lazer machine on the 7th floor and you do not know how to correct it or have a question, see D.... She is in charge of the overall care of those machines as I am the machines on the 8th floor (I can be located at ...).

Because of one person's lack of resourcefulness (he/she could of went down to the 7th floor or seen me or D), we are all suffering! If it is after hours and you are not sure what to do, it is better to do nothing than ruin a machine.

--J

By the late 1980s, however, many rooms with photocopiers in them also had laser printers in them. (Laser printers did not yet have disposable print cartridges with their own toner supplies.) And in the copier/printer room on the 8th floor of the MIT AI Lab, the two types of dry toner were incompatible. So what happened when the copier ran out of dry toner? Exactly the same thing as before. Except that the bottle marked "dry toner" one happens to come across first might or might not be the correct dry toner.

Does this happen because people are too lazy to check whether it's laser or copier toner? After all, as J points out, "It says on the box it comes in (and on the bottle itself) which machine it is for". Do the people just go irresponsibly ahead putting laser toner in the copier figuring it's probably OK? Do they proceed despite a conscious uncertainty?

Although we must appreciate J's situation, I think all these hypotheses are unnecessary. Put yourself in the place of someone to whom the photocopier is asking for dry toner, and suppose that you had not yet known that the copier and printer employed two incompatible types of dry toner that could be confused for one another. You made dozens of separate moves in answering the copier's call to be resupplied with toner, and any of those moves could be mistaken in dozens of different ways. The actual problem, namely that the bottle marked "dry toner" was not actually the correct substance, is pretty obscure, as if somebody parked a car nearly identical to yours a couple spaces down. In the case of the wrong car, some discrepancy would probably force itself upon your attention before you got too far. Yes, evidence of the mistake was readily available, but were you really supposed to list all the things you might be doing wrong and go looking for evidence to rule out each one? That would be impossible.

The problem, in short, is only obvious in retrospect. Having been warned that toner comes in two types, one should probably start to check. But nobody is born with that knowledge. Someone could become perfectly proficient at replacing toner in copiers and still run afoul of this difficulty, simply because types of toner had never become an issue for them. The arrival of the laser printer would invalidate one of the innumerable implicit background assumptions of their toner-changing routine, but it is hard to articulate the general policy that could have informed them of this. Read the fine print on every label every time? But the world is full of representations; how do you decide when to stop reading them all and start doing something? The phrase "dry toner", to us, having been informed of the problem, is ambiguous: it could mean "laser dry toner" or "copier dry toner". But that ambiguity is only consciously and morally an ambiguity for us, the well-informed. And for all we know, "dry toner" could also be ambiguous in an unlimited variety of other ways.

In the end, all we can do is stop moralizing and fix the problem. Make the toner bottles so different that it's physically impossible to install the wrong stuff. Train everyone. Or lock up the toner.

A story about some instructions at a performance in an art gallery

One day I went to the MIT Media Lab to see a performance. The performance took place in a windowless room that's about thirty feet square with a high ceiling. When we arrived the doors hadn't yet opened, so the audience milled about outside. Finally the time came and the ticket-seller wandered into the lobby and yelled something like, "OK the doors are open". Then as people were drifting toward the door she moved into the middle of the crowd and yelled a complicated set of instructions that went something like this:

You can sit on the chairs or you can go to the back wall and look through the windows or you can go up on the balcony, but don't lean against the side wall.

I found this amusing, and adopting a gently ironic imitation of the register and diction of the yeller I said something like, "you can stand between the monsters but don't sit on the toadstools". Not many people got the joke, I'm afraid, and especially not the yeller. I found this interesting in itself. The yeller was obviously familiar with the room, and she presumably had no problem visualizing what she was talking about. She evinced no awareness that others might be having a problem.

Once inside, there was an audible rush to attach the words to parts of the room, which was dark and full of peculiar wooden structures. The "side wall" is immediately there on your right, verifiable by the readily visible chairs along it, and the "balcony" could be found along the back wall with a little scanning. The "windows" weren't at all obvious; they were windows in the wall supporting the balcony, behind which one could stand. I suggested we go for the balcony; although no stairs were immediately visible it was obvious where they should be and others were already headed that way ahead of us.

In this story, the peculiar relationship between the instructions and the physical setting disrupted an aspect of language understanding that normally goes unremarked. The situation resembles the mnemonist's method of loci: if someone tells a story that happens in a familiar space -- or a space for which you have a cultural model, such as a story set in a generic Western kitchen -- the objects and actions in the story get "placed" in a way that can be either quasi-visual or kinaesthetic or both. If you can't put the elements of the story in their places then you won't be able to hang onto them, just as we had trouble hanging onto the balcony and windows in the ticket-seller's instructions.

This story connects to a larger theme about language. I want to believe that an utterance has no meaning outside the particular concrete setting where it's used. But then how can we can talk about things that are distant or hypothetical? Consider the examples that linguists ask you to evaluate out of context, like those sentences where you're supposed to indicate whether the pronoun can refer to (a) John or (b) Bill. Often I've found an interpretation not-OK until I work out a hypothetical context that makes it OK.

I suspect that such exercises get consistent results only because of cultural conventions about the default contexts. The willingness to evaluate decontextualized sentences at all is culturally specific. The cognitive anthropologist A. R. Luria (1976), for example, found his informants refusing to answer syllogistic reasoning tests until they knew the particulars of each sentence -- the equivalents of the John and Bill who remain so comfortable as ciphers in my own culture. These people haven't learned the language game of decontextualized grammar and syllogism quizzes, and don't care to.

Heath (1983) describes how this capacity for decontextualization arises. She found that middle-class parents use rituals like bedtime-story-reading to introduce children to decontextualized letters, words, and forms of speech. Children who don't get this training have a harder time relating to decontextualized school exercises. Decontextualization is a complex and culturally specific skill laid on top of the more natural ability to relate language to familiar contexts. This has led to two quite different phenomena -- one of them more fundamental and universal than the other -- being run together and confused, with the later, more articulated phenomenon (the ability to perform certain tricks with decontextualized representations) getting all the credit.

A story about my routines for reading the Sunday Globe

This is a story about the indeterminacy of plans. The plan was something I said to myself as part of a settled routine. And I have a theory about the role of plans in settled routines. A routine might start out being mediated by an imperative utterance, such as a command you subarticulate to yourself. As the routine settles down, I hypothesize, it will still exhibit all the underdetermination, ambiguity, and indexicality of the original utterance, long after you've lost all awareness of any English being involved. Just because the activity has gotten "compiled" -- as computer people would say -- doesn't mean that the connection between the plan and the concrete situation of using it becomes any less problematic. Why? Perhaps you never stop saying the plan-text to yourself; perhaps as you routinize your actions you routinize your interpretive process as well.

Here, then, is the story. The Boston Globe recently began an expanded arts section in their Sunday edition. Called "Arts Etc", it gathers all the Sunday movie, arts, book reviews, arts schedules and advertising, and high-brow cultural commentary. This section doesn't have clearly delineated departments except for the final few pages, which are marked off for book reviews. The book review department, in fact, is wholly unchanged from the pre-Arts-Etc Sunday Globe. The first of the book review pages has its own banner and distinctive format, and all of the longer book reviews begin on that page and continue inside, where there are also shorter reviews and lists of best sellers and so forth.

Now, when reading the newspaper I will often come across the continuation of an article that looks interesting even though I hadn't noticed it when I was reading the page on which it began. So I'll have to back up to the earlier page to read the beginning of the article.

Last Sunday, then, I was reading a book review in the Globe. In particular, I was in the interior of the book review section, having followed an article from the book review section's first page (which, let us say, was page C15), when I came upon a headline about an author I was interested in reading about. Focusing on this headline, I found that it was a continuation. Whereupon, oddly, I turned to page C1 -- i.e., the front page of the whole arts section -- and not to page C15 -- i.e., the first page of the book review section. I knew that all the book reviews began on C15, not C1, but I turned to C1 anyway. When I got C1 in front of me, it was not at all what I expected; momentarily confused, I figured out that I should turn to C15 instead.

Saying "C1" and "C15" is of course misleading. I don't think I knew that it was section "C" or page 15. My mistake, I think, turned on my never having reflected on the odd relationship between the two pages: the book review section was a clear "part" of the arts section, but the arts section didn't have any other clear "parts". Both page C1 and page C15 were the "front" of something -- the arts section and the book review section, respectively. I had long been familiar with the Sunday Globe book review section's format, and its design and layout did little to make it look like a part of the superordinate arts section.

Here's what I think happened. When I went to turn to the beginning of the review I wanted to read, I turned to "the front", perhaps even "the front of the section". I'm not sure what I mean by double-quoting those two English phrases, but I want to mean something fairly literal. That is, I think I made my mistake because I was saying a phrase to myself in English, the phrase was ambiguous, and I interpreted it wrongly.

I've been reading newspapers for at least fifteen years and Boston Sunday Globe book reviews for almost ten years. Stumbling upon the continuation of an article and wanting to find its beginning is a routine I've been through many hundreds of times. And at least a couple dozen of these episodes must have occurred during my reading of the Sunday Globe's book reviews.

Why did I know to turn to the "front" of the section? Not all articles in the rest of the paper start at fronts of sections. I think at some point I noticed, and articulated to myself, that the book reviews start at the front of the book review section. So this was not the first time I've said "go to the front of the section" to myself in my head in such a situation. This internal uttering-to-myself and the actions I typically take in consequence of it must certainly have worn deep grooves in my brain by now. You might think that it was so thoroughly "compiled" that it no longer resembled English. Yet still it was capable of this very language-like underspecification of the situation. I still had to figure out what the English phrase was referring to in this specific concrete situation, and even though this figuring occurred perfectly automatically, smoothly, and routinely, it was still problematic. I could still get it wrong.

So "the front of the section" was ambiguous in this situation. But all of this still doesn't explain why, on the particular moment in question, it led me to turn to C1 rather than C15. Before the Globe reformatted and publicized its "Arts Etc", I had never given any particular thought to the idea of the Sunday Globe having an "arts section". In fact I clearly recall the first Sunday of the new section: the front page of the paper -- i.e., A1 -- had an ad for it, and despite the silly name I decided to give it a fair try. In fact it contained a useful article about Soviet dissident films, a topic that interested me. The matter of "the Globe's new arts section" -- in exactly those words -- had thus been on my mind. I don't want to conclude that the arts-section interpretation was "stronger" than the book-review-section interpretation, but whatever is operating as we constantly use background information to "fill in the details" of utterances when determining their relevance to particular concrete situations was operating here as well.

What I'm trying to figure out

Implicit in these stories are some ideas about representation and action.

1) Writing model of representation. By "writing", I mean that the representation is something apart from you. It is a resource in situated action (Suchman 1987). You have to make sense of it in each next situation. There are no complete, systematic, guaranteed rules for this making-sense. And when you do manage to figure out what in some situation a representation is talking about, there is no way to finish listing what about the situation enabled you to do this.

2) Dependency model of routine evolution. All forms of activity are snapshots in the evolution of routines. The routines themselves are intertwined with the patterns of society and with the layout of particular places. The model says: when you think a new thought in some situation, you connect it back to its premises (Stallman and Sussman 1977). When you believe those same premises again in some future situation, you automatically call up the conclusion.

This sounds like a lot of work. If it were really a fresh challenge to make sense of "turn left" or "open the bottom drawer" on every moment, how would we ever do anything? But I don't think it's that bad. What you have in practice is a patchwork of routinized methods. In some past situation someone said "take the next left", and you took certain specific concrete actions: you looked around, you searched for particular shapes and colors, and maybe you walked around and looked some more. You had your own reasons for doing all these things, and all of your actions and their reasons got connected, so now they're ready and waiting to happen again in new situations. In subsequent situations some of the reasons might not have applied, so instead you took other actions, and these themselves led to new connections.

Perhaps after enough of this you develop a sufficient repertoire of routines to apply "turn left" to almost all of the left-turn situations you encounter in the average day. You've developed habits of interpretation. So whenever you tell yourself -- or someone else tells you -- "turn left", you'll be able to do it right away, "automatically", without any hesitation or difficult figuring-out.

These routines, like all routines, will evolve (Agre 1985a, 1985b). Very often the evolution of a routine involving a representation will permit you to undertake the activity without the representation being present in physical form. For example, using a recipe ten times might let you make the dish without the recipe. One precondition for this effect seems to be that you understand the reasons for the recipe's instructions, but this turns out to be a complicated idea. In any event, there's a sense in which the representation never goes away. Even when you're routinely deciding to turn left or add salt, you're still -- in some sense I wish I understood -- saying the utterance "turn left" or "salt to taste" to yourself, and you're still interpreting it just like any other natural-language utterance that you need to relate to a concrete situation.

Acknowledgements

Conversations with Marty Hiller and Jeff Shrager helped me think about what people do with instructions and directions. David Chapman and Beth Preston carefully read a draft of this paper. Michael Mateas helped with the final revisions.

Patricia Morison helped with Latin etymology.

The "Upper Elk Tree Road" directions in the introduction and the computer message in "A story about photocopier supplies" are reproduced with the permission of their authors. I have altered them slightly to suppress identities and remove some comments on other matters.

References

Agre, Philip E. (1985a). The structures of everyday life. Working Paper 267, MIT Artificial Intelligence Laboratory.

Agre, Philip E. (1985b). Routines. AI Memo 828, MIT Artificial Intelligence Laboratory.

Agre, Philip E. (1997). Computation and Human Experience. Cambridge: Cambridge University Press.

Amerine, R. & J. Bilmes (1988). Following instructions. Human Studies, 11, 327-339.

Atkinson, J. Maxwell & John Heritage (Eds.) (1984). Structures of Social Action: Studies in Conversation Analysis. Cambridge: Cambridge University Press.

Auerbach, Erich (1953). Mimesis: The Representation of Reality in Western Literature. Translated by W. R. Trask. Princeton: Princeton University Press.

Barwise, Jon & John Perry (1983). Situations and Attitudes. Cambridge: MIT Press.

Brachman, Ronald J. & Hector J. Levesque (1984). The tractability of subsumption in frame-based description languages. Proceedings of the National Conference on Artificial Intelligence, Austin, TX, pages 34-37.

Brachman, Ronald J. & Hector J. Levesque (Eds.) (1985). Readings in Knowledge Representation. Los Altos, CA: Morgan Kaufmann.

Brodsky, Claudia J. (1987). The Imposition of Form: Studies in Narrative Representation and Knowledge. Princeton: Princeton University Press.

Chapman, David (1991). Vision, Instruction, and Action. Cambridge: MIT Press.

Chapman, David & Philip E. Agre (1986). Abstract reasoning as emergent from concrete activity. In M. P. Georgeff & A. L. Lansky (Eds.), Reasoning about Actions and Plans. Los Altos, CA: Morgan Kaufmann.

Culler, Jonathan (1982). On Deconstruction: Theory and Criticism after Structuralism. Ithaca: Cornell University Press.

Dennett, Daniel (1978). Brainstorms: Philosophical Essays on Mind and Psychology. Montgomery, VT: Bradford.

Derrida, Jacques (1976). Of Grammatology. Translated by G. C. Spivak. Baltimore: Johns Hopkins University Press.

Fodor, Jerry & Zenon Pylyshyn (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28, 3-72.

Garfinkel, Harold (1984). Studies in Ethnomethodology. Oxford: Polity Press. Originally published in 1967.

Goody, Jack (1986). The Logic of Writing and the Organization of Society. Cambridge: Cambridge University Press.

Hagen, Margaret (1986). Varieties of Realism: Geometries of Representational Art. Cambridge: Cambridge University Press.

Harris, Roy (1980). The Language-Makers. Ithaca: Cornell University Press.

Harris, Roy (1987). The Language Machine. Ithaca: Cornell University Press.

Hartog, Francois (1988). The Mirror of Herodotus: The Representation of the Other in the Writing of History. Translated by J. Lloyd, Berkeley: University of California Press.

Haugeland, John (1981). Mind Design: Philosophy, Psychology, Artificial Intelligence. Cambridge: MIT Press, 1981.

Heath, Shirley Brice (1983). Ways with Words: Language, Life, and Work in Communities and Classrooms. Cambridge: Cambridge University Press.

Heritage, John (1984). Garfinkel and Ethnomethodology. Cambridge: Polity Press.

Hopcroft, John E. & Dean B. Krafft (1987). The challenge of robotics for computer science, in Jacob T. Schwartz & Chee-Keng Yap, eds, Algorithmic and Geometric Aspects of Robotics, Volume 1. Hillsdale, NJ: Erlbaum.

Hurlbert, Anya & Tomaso Poggio (1988). Making machines (and artificial intelligence) see. Daedalus, 117, 213-239.

Hutchins, Edwin (1986). Mediation and automatization. Quarterly Newsletter of the Laboratory of Comparative Human Cognition, 8, 47-58.

Hutchins, Edwin (1995). Cognition in the Wild. Cambridge: MIT Press.

Ingarden, Roman (1973). The Cognition of the Literary Work of Art. Translated by R. A. Crowley & K. R. Olson. Evanston: Northwestern University Press.

Judovitz, Dalia (1988). Subjectivity and Representation in Descartes: The Origins of Modernity. Cambridge University Press.

Krieger, Murray (Ed.) (1987). The Aims of Representation: Subject, Text, History. New York: Columbia University Press.

Latour, Bruno (1986). Visualisation and cognition: Thinking with eyes and hands. Knowledge and Society, 6, 1-40.

Levinson, Stephen C. (1983). Pragmatics. Cambridge: Cambridge University Press.

Luria, A. R. (1976). Cognitive Development: Its Cultural and Social Foundations. Translated by M. Lopez-Morillas & L. Solotaroff, edited by M. Cole, Cambridge: Harvard University Press.

Newell, Allen & Herbert A. Simon (1963). GPS, a program that simulates human thought. In E. A. Feigenbaum & J. Feldman (Eds.), Computers and Thought. New York: McGraw-Hill.

Newell, Allen & Herbert Simon (1972). Human Problem Solving. Englewood Cliffs, NJ: Prentice-Hall.

Norris, Christopher (1982). Deconstruction: Theory and Practice. London: Methuen.

Ong, Walter J. (1982). Orality and Literacy: The Technologizing of the Word. London: Methuen.

Rorty, Richard (1979). Philosophy and the Mirror of Nature. Princeton: Princeton University Press.

Rumelhart, David E., Paul Smolensky, James L. McLelland, & Geoffrey E. Hinton (1986). Schemata and sequential thought processes in PDP models. In J. L. McLelland & D. E. Rumelhart, eds, Parallel Sequential Processing: Exploration in the Microstructure of Cognition, Volume 2: Psychological and Biological Models. Cambridge: MIT Press.

Searle, John R. (1981). Minds, brains, and programs. In J. Haugeland (Ed.), Mind Design: Philosophy, Psychology, Artificial Intelligence. Cambridge: MIT Press.

Silvers, Stuart (1989). Rerepresentation: Readings in the Philosophy of Mental Representation. Dordrecht: Kluwer.

Stallman, Richard M. & Gerald Jay Sussman (1977). Forward reasoning and dependency-directed backtracking in a system for computer-aided circuit analysis. Artificial Intelligence, 9, 135-196.

Suchman, Lucy A. (1987). Plans and Situated Actions: The Problem of Human-Machine Communication. Cambridge: Cambridge University Press.

Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Edited by M. Cole, V. John-Steiner, S. Scribner & E. Souberman. Cambridge: Harvard University Press. Originally published in Russian in 1934.

Wallis, Brian (Ed.) (1984). Art after Modernism: Rethinking Representation. New York: New Museum of Contemporary Art.

Wertsch, James W. (1985). Vygotsky and the Social Formation of Mind. Cambridge: Harvard University Press.

White, Hayden (1973). Metahistory: The Historical Imagination in Nineteenth-Century Europe. Baltimore: Johns Hopkins University Press.

Zimmerman, Don H. (1973). The practicalities of rule use. In J. D. Douglas (Ed.), Understanding Everyday Life: Toward the Reconstruction of Sociological Knowledge. London: Routledge and Kegan Paul.