THE AGE OF INTELLIGENT MACHINES | All Work and No Play Makes HAL a Dull Program

September 24, 2001
Author:
Ray Kurzweil

Michael Lebowitz was a faculty member from 1980 to 1987 in the Department of Computer Science at Columbia University. His primary research interests lie in the areas of machine learning, natural-language processing, memory organization, and cognitive modeling. He also led a research group at Columbia that designed intelligent information systems that could read, remember, and learn from natural-language text. He has published many articles in a wide range of areas in artificial intelligence. He is currently a vice president of the Analytical Proprietary Trading unit at Morgan Stanley & Co. in New York City.

Introduction

The clear majority of applied work in artificial intelligence (AI) has involved practical problems in such areas as business, medicine, defense, and so forth. This is as it should be, but there is also room for the application of AI to the arts. AI has the potential to allow the creation of compelling new entertainment forms and to improve the creative process in existing art forms. In addition, such work can lead to important new insight into the creative process. In this article I shall discuss how AI can be applied to the area of story telling: both how it can be used in intelligent tools for writers and how it might lead to new forms of intelligent, interactive stories where a writer/creator can intimately involve a reader/user in the story.

At Columbia University my colleagues and I investigated the problem of developing systems that generate extended stories, those that continue over time. We selected the domain of “interpersonal melodrama,” one example of which is television soap opera. We felt that this is a good domain to look at both for scientific and practical reasons. Television melodramas are watched by large numbers of people over very long periods of time. From a scientific point of view, stories based on interpersonal relations have received much less attention than those that are more action oriented. Finally, stories of this sort combine complexity and creativity with a good deal of stereotypy. They help us get a handle on which parts of creativity are not too creative. Here is an outline of events in a typical melodrama (NBC’s “Days of Our Lives”).

Liz was married to Tony. Neither loved the other, and indeed, Liz was in love with Neil. However, unknown to either Tony or Neil, Stephano, Tony’s father, who wanted Liz to produce a grandson for him, threatened Liz that if she left Tony, he would kill Neil. Convinced that he was serious by a bomb that exploded near Neil, Liz told Neil that she did not love him, that she was still in love with Tony, and that he should forget about her. Neil was eventually convinced and married Marie. Later when Liz was finally free from Tony (because Stephano had died), Neil was not free to marry her, and their troubles went on.

As part of our research we developed a simple, prototype story-telling program, Universe. The program can create sets of characters appropriate for interpersonal melodrama and can generate simple plot outlines of about the complexity of this outline using the characters that it generates.1

Assisting Writers

In the immediate future it is unlikely that we will be able to create programs that produce melodrama with the richness of human writers. However, we can hope to provide tools to assist in the creation of such works. In particular, I feel that the creation of an automated assistant that can suggest possibilities to a writer is well within reach. We can create programs that help a writer develop meaningful characters and that suggest plot possibilities. I shall illustrate a simple version of one such program below.

In a sense, we can view a program of the sort I envision as a kind of intelligent story editor, the next step beyond text editing. Such a system would have to be embedded in a complete development system that would also keep track of the history of the story over time. The display below illustrates the kind of behavior I have in mind. It shows Universe running in an interactive mode in which instead of creating characters itself, it prompts a user for the relevant information. Items in italics were entered by the user.

Enter sex (m/f): m

Enter name: John Jones

Enter age group (ya/ma/old): ma

Cycling person John Jones

Enter year of next life event (cr = quit): 1970

Next event (m/d): m

Try to find spouse in universe (u/n): n

Enter name (list):

Cycling marriage between John Jones and Kathy Cole

Enter year of next marriage event (cr = quit):

Beginning finish up for John Jones

Pick traits (p/r): p

Possible traits to set: wealth, promiscuity, competence, niceness, self-confidence, guile, naivete, moodiness, physical attractiveness, intelligence

Enter trait and value (cr = quit): niceness 3

Enter trait and value (cr = quit): self-confidence 4

Enter trait and value (cr = quit): intelligence 3

Enter trait and value (cr = quit):

Optimal occupations are: sleazy-doctor

Enter occupation:

Selecting: sleazy-doctor

Trait (intelligence 3) already satisfied

For trait (self-confidence 4)

Choices (life-guard, swinger, big-eater): swinger

For trait (niceness 3)

Choices (taxidriver):

The AI aspects of the program arise in the representation of the information and where the program suggests possibilities for the character being created. These suggestions are intended to help make the character more believable and interesting by appropriately motivating the characteristics that the user wants. In the first section of the display, Universe prompts the user for a few vital statistics of the character being created: sex, name, age group (young adult, middle age, old). These are used to begin filling in a frame for the character John Jones.2 In all cases, if no answer is provided, the program will pick one.

After getting the basic facts, the program begins to “cycle” the character’s life, that is, to prompt the user for major life events, concentrating on marriages, divorces, the births of children, and deaths. Other parts of the program allow other sorts of history to be added. Note that by including past events and not just the current state of the world, Universe helps the user build a consistent and coherent set of characters. In the display the user specified that John Jones had gotten married (m) in 1970. The program knew that a spouse was necessary, and the user let the program pick her name (Kathy Cole).

The final section of the display is the most AI oriented. It involves filling in a personality profile for the character. Universe knows about a variety of different personality traits. Here the user suggests a set of values to describe John Jones: he is not very nice, is moderately self-confident, and moderately intelligent. Just specifying values, however, is not very satisfying, so the program assists the user in selecting various prototypical character types that will help explain the trait values. It allows the user to either accept a proffered prototype or provide one. It suggests that sleazy doctor would be a character type that meshes with these traits. It also suggests other prototypes that help explain the traits. Between the program and the user they come up with a character that is a swinging sleazy doctor and taxidriver. Needless to say, the quality of the system’s help depends on the quality of the prototype library. Such a description is the sort the writer would have to come up with anyway to make the character one a reader/user will believe.3

Tools like the one described here (with plot management included) can be of great use to writers of all sorts. They serve as partners by suggesting ideas but leave the real creativity to the writer. They will be particularly important in developing various kinds of interactive fiction but should also be of use in more traditional forms. One can easily imagine how an AI tool could greatly help the writers of a television serial that has been on the air for many years keep everything consistent and at the same time suggest new possibilities.

Personalized Stories

Beyond the creation of intelligent story-writing tools, the obvious next step in applying AI to story telling is the automatic generation of stories. With human writers available, it may not be clear why we would want to do this (beyond the goal of understanding how creative processes work). The reason is the same as for many other tasks: the application of AI techniques to story telling will lead to an art form that is personalized and interactive. Imagine a soap opera with characters that are exactly the kind you would find most interesting, not necessarily a set of characters that you create (after all, we are not all writers), but one molded to your own tastes. In addition, imagine that you can influence the unfolding of events-not necessarily directing the characters in detail, but influencing the story in more subtle ways. It is a compelling vision of a new fictional form. Stories in textual form should be exciting, and the addition of graphics and other computer-manipulated images should make it even more impressive.

Stories of this kind will evolve from two current lines of entertainment: interactive computer games and role-playing games. Several manufacturers are currently producing interactive computer games in which users work their way through a branching story, typically solving a series of puzzles along the way. Many such games are quite intriguing. However, all the various outcomes have to be handcrafted by a programmer and as a result are relatively limited in the world to be explored. In addition, programmers have tended to focus on action-oriented and mystery domains with little character development.

Such role-playing games as Dungeons and Dragons come closer to the model I have in mind. Here with a human game master a story can move through a very wide range of possibilities. Yet the game master is required, and these games too have tended to be quite action oriented (even though at least one was modeled after a television melodrama).

It is important to note that what we are after is more than simply games. We seek the depth and quality of fiction along with the interactive nature of games. Although the stories themselves would be generated by computer, the knowledge bases from which they worked-information about people, places, and plot devices-would presumably be built by human writers, although not by specifying every plot turn, as is done with interactive games.

To date there has not been a great deal of work on automatic story generation. Perhaps the best known work is that done by Jim Meehan in the mid 1970s when he developed Tale-Spin, a program that told children’s stories in the style of Aesop’s fables.4 Other workers have also been working on AI story telling.5 Our own work on Universe has taken us to the point where it can generate simple plot outlines by applying planning techniques to author goals.6

Below is a simple plot outline created by Universe. It uses characters created by the automatic version of the interactive program described in the previous section after being given as input an author goal, here to “churn” a relationship, that is, to keep two people in love from being happy together (a staple of melodrama). Since Teddy Bryks, the husband of one of the lovers, has an evil father, Universe selects a plot fragment that involves the father’s threatening the women not to betray her husband (which is loosely based at an abstract level on the outline above). The various elements of this plot are expanded using other plot fragments. By the time the first fragment ends, the spurned lover, Steven Kades, is involved with someone else, which creates further complications. Universe runs through a couple more of its library of ways to keep people apart.

The setting generated by the program as part of a set of characters is as follows: Nadine Burton is married to Teddy Bryks but is in love with Steven Kades. Gerald Bryks is Teddy’s father, and none too nice a guy. (Elaborative comments are enclosed in parentheses.)

Churn Nadine and Steven

Gerald Bryks threatens Nadine Burton Bryks: forget it (her affair with Steven). Nadine Burton Bryks tells Steven Kades that she doesn’t love him. Norma Ryan Bryks is worried about Steven Kades. Linda Einbinder Kimball seduces Steven Kades. Joe Kimball decides to expose Gerald Bryks. Joe Kimball tells the world about Gerald Bryks (and his evil doings). Nadine Burton Bryks and Teddy Bryks get divorced. Teddy Bryks tries to seduce Nadine Burton Bryks. Steven Kades gets frustrated (by this). Steven Kades tells Nadine Burton Bryks that he doesn’t love her. Nadine Burton Bryks seduces Steven Kades.

This story is certainly not intended to be great fiction. However, our program has shown us that even with a very limited library of plot fragments (about 60), through interactions of plots and characters, Universe can produce some rather clever plot outlines. A much larger library (perhaps by two orders of magnitude) should be practical and would be able to produce a large number of interesting stories, particularly if there are techniques to automatically create new fragments.7 No small part of the effect is based on the same trick that authors of standard fiction use: the reader’s imagination will enhance what is actually presented.

Our work on Universe has also given us some interesting insights into creativity, which I can only touch upon here. It appears that many parts of creativity involve primarily determining clever ways to apply previous situations (or abstracted versions of them) to new settings. In story telling, it is rarely necessary to create whole new ideas: one can take old ideas and combine them in new ways or apply new twists to them. This is certainly evident in television melodrama and is by no means bad: even plot turns that are familiar in the abstract can be quite interesting when applied to new situations. Although there are many other aspects to creativity, a crucial part is storing previous experiences in a way that they can be efficiently retrieved and applied in the future. This applies to both story telling and general day-to-day planning.8

We are just beginning the process of automatically generating stories. There are many problems involved in creating plot outlines, and beyond them there are issues in language generation, knowledge representation, knowledge-state assessment (who knows what when), memory organization and access, and user interaction, among many others that must be dealt with to achieve the sort of system envisioned at the beginning of this section.

It is no doubt appropriate for most AI work to address practical considerations. Indeed, most of our own research does just that. However, people should know that AI will enhance other aspects of their lives than just the workplace. AI can help creative people make better use of their talents and create interesting and entertaining new art forms. After all, if AI is going to help improve productivity, then it had better also help fill the leisure time thus created-and it can.

Notes

1. For further details, see M. Lebowitz, “Creating Characters in Story-Telling Universe,” Poetics 13 (1984): 171-194; and M. Lebowitz, “Story-Telling as Planning and Learning,” Poetics 14 (1985): 483-502.

2. For more on frames, see M. Minsky, “A Framework for Representing Knowledge,” in P. H. Winston, ed., The Psychology of Computer Vision (New York: McGraw-Hill, 1975).

3. For further details on Universe’s methods of character creation, see M. Lebowitz, “Creating Characters in a Story-Telling Universe,” Poetics 13 (1984):171-194.

4. See J. R. Meehan, “The Metanovel: Writing Stories by Computer, Yale University Department of Computer Science, technical report 74, 1976.

5. See N. Dehn, “Memory in Story Invention,” Proceedings of the Third Annual Conference of the Cognitive Science Society (Berkeley, Calif., 1981), pp. 213-215, M. Yazdani, “Generating Events in a Fictional World of Stories,” University of Exeter Computer Science Department, technical report R-113; and S. R. Turner and M. G. Dyer, “Thematic Knowledge, Episodic Memory, and Analogy in MINSTREL, A Story Invention System,” Proceedings of the Seventh Annual Conference of the Cognitive Science Society (Irvine, Calif., 1985), pp. 371-375.

6. See M. Lebowitz, “Story-Telling as Planning and Learning,” Poetics 14 (1985): 483-502.

7. See M. Lebowitz, “Story-Telling as Planning and Learning,” Poetics 14 (1985): 483-502.


Photo by Betsy Malcolm

Michael Lebowitz