February 18, 2008

EP 5.1: The “Metanovel”

by Noah Wardrip-Fruin · , 6:25 am

In the fall of 1974, James Meehan was a graduate student at Yale University. He had an idea in mind for his dissertation topic, but didn’t know how to pursue it. The topic had been suggested to him by Alan Perlis — one of the most famous figures in U.S. computer science, who had become chair of Yale’s department a few years before — on the first day they met. But Perlis didn’t know how to move forward with the idea, either. In the preface to his dissertation, Meehan describes the idea this way:

A metanovel is a computer program that tells stories that only a computer could tell, stories of such complexity of detail that only a computer could handle, stories with more flexibility — even reversibility — of events and characters than a human could manage. A metanovel time-sharing system tells a story to many people at once, no two of whom read the same thing, because they have each expressed different interests in the events and characters they want to hear about, and because they may each desire a different style of storytelling. And yet, among all these readers, there is but one story — the Metanovel itself — and each reader is only following those threads of the story that interest him. (1976, ii)

Meehan’s dissertation didn’t set out to create a metanovel, but rather to make progress toward the possibility. As he notes, “If I’ve been successful, we’re a little closer than we were.” In the decades since, his work has become a landmark — the first project cited in nearly every discussion of story generation.

Computer systems had certainly produced stories as output before. The most famous of these was the “automatic novel writer” produced by Sheldon Klein (1971) and his collaborators at the University of Wisconsin, Madison. But the field has generally followed Meehan’s critique of Klein’s system — it was made up of explicit chunks of action, with a path through these chunks selected randomly.1 So, for example, Klein’s system includes a “rule for people arriving at Georges [sic] living room” (27). Meehan’s system, on the other hand, contains general rules for character movement (and reasons for movement) that operate no matter which spaces are available — living rooms, bed rooms, caves, or meadows.

More broadly, this points to the two main types of systems for creating variable fictions. Klein’s system, like quest flags, like Choose Your Own Adventure books, is composed of explicit chunks of story content. If one of these is changed, or removed, not much else changes. At most, if some story chunks are only accessible by passing through others, when a chunk is removed other chunks may become impossible to reach. Meehan’s system, on the other hand, models story as a relatively fine-grained set of processes and data that are used to generate story events. In the case of Tale-Spin, this is accomplished by creating a simulated world, processes for behavior in and of the world, and characters and objects that populate the world. As a result, changing one aspect of the simulation (such as a rule for character movement or the existence of a particular object) can lead to wildly different Tale-Spin fictions. This kind of flexible model has become a defining characteristic for the pursuit of story generation.

Basis for the model

The breakthrough that allowed Meehan to see a path forward toward the metanovel was exposure to the “scruffy” artificial intelligence ideas of Roger Schank and Robert Abelson. Schank had just arrived at Yale in the fall of 1974, and Meehan enrolled in his Natural Language Processing Seminar — which Meehan described as “a weekly battle of ideas” featuring favorite “guest speaker/victims” such as Abelson (iv).

Many of Schank’s theories were developed in the field of computational linguistics, where he argued that previous work was fundamentally flawed in its approach. For example, in the area of computer translation from one language to another, projects based on syntactic parsing and dictionary-style substitution had largely failed. Schank, instead, proposed a language-independent representation of semantic meaning, which he called “conceptual dependency” expressions. As Schank describes them in Conceptual Information Processing (1975a), such semantic representations could serve as an interlingua, so that the translation of a statement from any one language to another could be accomplished by translating that statement to and from the semantic representation. Such a representation could also serve as an internal meaning representation — or data format — for AI projects, helping in processes such as paraphrasing and inference-making, as demonstrated in systems built by Schank and his students. Tale-Spin followed this lead, operating in terms of conceptual dependency (CD) expressions.

But understanding the meaning of sentences, rather than simply their structure, requires understanding their context. “Mary hit John” probably means different things if John is Mary’s sparring partner or, on the other hand, if John is sitting at Mary’s blackjack table. This fact motivated a series of projects, directed by Schank, that attempted to bring together theorizing about how humans understand stories with the building of computer systems for the same task. The systems were viewed as experiments that would help refine and validate the theories, and a primary goal was developing theories for understanding the causal relationships between elements. The first step was a theory of “scripts” — that human knowledge for certain routine situations (e.g., going to a restaurant, catching a bus) exists as stereotyped sequences of common events. Computer systems were built that succeeded in interpreting simple stories (e.g., of auto accidents) by determining relationships between the sentences of the stories and elements of an internal script.

But Tale-Spin largely avoids scripts. As Meehan puts it, scripts are “so developed that they’re uninteresting: not great story material” (213). Scripts only allowed for theorizing, and system building, about the least novel elements of human life. The next step was to build systems around Abelson and Schank’s developing versions of the central AI concepts of plans and goals. This was taking place just as Schank arrived at Yale, and to Meehan the results looked like much better story material.2

The Tale-Spin system produces stories by changing the status of Schank and Abelson’s theories about planning behavior. In story understanding systems, these theories were data. They were used by the system processes as patterns to compare against the behavior described in stories being interpreted. But in Tale-Spin they become processes, used to create a world in which characters behave according to the theories. As Meehan puts it:

Tale-Spin includes a simulator of the real world: Turn it on and watch all the people. The purpose of the simulator is to model rational behavior; the people are supposed to act like real people. (107)

However, most Tale-Spin stories don’t feature “people” that are human characters. Instead, the people are largely birds, bears, bees, foxes, and so on. This was Schank’s suggestion, inspired by, as he puts it, “the fact that I had little kids at the time so I was making up [stories] and trying to see how [I] did it” (2006). It also served to compensate, as Meehan puts it, for the fact that “we weren’t going to get very sophisticated or elaborate output from this program” (2006). In other words, the decision to tell animal stories is an example of the common AI approach of defining a microworld in which the program will operate. Just as some AI researchers demonstrated programs that could understand the physics of a toy “blocks world” (allowing them to ignore issues such as relative strength of materials) Tale-Spin operated in a simplified animal world resembling the settings for Aesop’s fables (shaving complexity both from the possibilities of the world and from the potential motivations of the characters).

Tale-Spin today

Today Tale-Spin is one of the most widely-discussed digital fictions ever produced. It is not only a touchstone for computer science accounts of story generation, but also broadly cited in writing about digital literature and the future of fiction. At the same time, Tale-Spin, itself, seems permanently lost.

Ongoing discussion of this lost software is made possible by two facts. First, many examples of fictions produced by Tale-Spin are in circulation. These form the basis of most humanistic discussions of Tale-Spin, which tend to be dismissive. Second, Meehan’s dissertation gives detailed accounts of the operations of Tale-Spin’s processes, along with significant information about its data. This serves as the basis for most computer science discussions of Tale-Spin, which tend to treat the system as worthy of serious engagement.

For many years I was only exposed to examples of Tale-Spin fictions, which I found so uninteresting that I never learned more about the system. But once I began to learn about Tale-Spin’s processes I became fascinated — both with the system itself and with the differing attitudes between those who understood the system and those who only saw its output.

After reading Meehan’s dissertation I began to look for even more detailed sources of information and, especially, for a copy of the software itself. I had previously seen Micro Tale-Spin (1981), a smaller, pedagogical version. But this was so simplified that it lacked most of what interested me about Tale-Spin. Through Walt Scacchi, a former student of Meehan’s from his time at UC Irvine, I got an email introduction. I had a pleasant correspondence with Meehan, now at Google, who looked through his garage on my behalf — seeking traces of Tale-Spin — but his search came up empty.

However, I learned from Meehan about another iteration of the system, more complex than the “micro” version, created for the 1987 “Smart Machines” exhibit at the Boston Computer Museum. Unfortunately, the Computer Museum had since shut down, but most of their archives went to the Computer History Museum in California. Unfortunately, as Al Kossow (software curator for the museum) discovered, the archives only contained a record of the exhibition, not the software itself. Kossow also contacted Oliver Strimpel, who was in charge of the original “Smart Machines” exhibit, and confirmed that this version of Tale-Spin had probably never been archived.

In the field of digital media, such losses are unfortunately common. Even for the most influential projects, we depend on the authors to archive them. Only once libraries and museums get serious about preservation of software systems — rather than just the paper that comes with them or documents them — will this situation change.3

As a result, today I believe we will never again learn from Tale-Spin through direct interaction, as we can with Eliza/Doctor. But I still believe we can learn important lessons from looking at Tale-Spin’s processes and output, as recorded in Meehan’s dissertation, and the responses of those who have previously written about them. Some of the most interesting lessons emerge from an inversion of the Eliza effect.

Rather than a surface illusion of process complexity and intelligence, Tale-Spin creates a surface illusion of process simplicity and arbitrary action. It is far from alone in this. As a result, I believe understanding the more general “Tale-Spin effect” can provide insights for both authors and interpreters of digital fictions, as well as those who seek to understand computational systems broadly.


1This is one aspect of Meehan’s critique, other aspects of which focus on the mismatch between the behavior of Klein’s system and theories of the behavior of human storytellers.

2Schank echoes Meehan’s evaluation of the relative interest value of stories about scripts versus those about plans in Inside Computer Understanding (1981):

Many stories (particularly the more interesting ones) are about entirely new situations for which we have no available script. When we hear such stories we rely on our knowledge of how to plan out an action that will lead to the attainment of a goal. (33)

Schank and Meehan are not alone in linking art to the unscripted, the unexpected, and even the disorienting. Among fiction writers interested in link-based hypertext its potential for such effects has been cited as one of its potential benefits, though in more technically-oriented circles disorientation was seen as one of hypertext’s large potential problems. George Landow, in Hypertext 3.0 (2005), explores this difference through the work of writers such as Morse Peckham (author of Man’s Rage for Chaos).

3In the meantime, Nick Montfort and I have written a pamphlet for the Electronic Literature Organization, titled Acid-Free Bits, outlining some things that authors can do to make their work more likely to survive (Montfort and Wardrip-Fruin, 2004).