CREATING A STORY-TELLING UNIVERSE

Michael Lebowitz
Columbia University, Department of Computer Science
New York, NY 10027

Abstract
Extended story generation, such as the creation of soap
opera stories, is a difficult and interesting problem for
Artificial Intelligence. We present here the first phase of
the development of a program, UNIVERSE, to tell such
stories. In particular, we introduce a method for creating
universes of characters appropriate for extended story
generation.
This method concentrates on the need to
keep story-telling universes consistent and coherent. We
describe "the information that must be maintained for
characters and interpersonal relations, and the use of
stereotypical information about people to help motivate
trait values.

1 Introduction
One of the most interesting forms of story telling is
extended story generation, the continuing serial. This can
occur as short fiction, movies such :is Star Wars and The
Empire Strikes Back, novel series, role-playing games
and, in one of its most popular forms, the television soap
opera.
The writer or writers of such serials face the
prodigious task of creating and keeping track of literally
dozens of intricately interrelated characters and numerous
past and present plots. Extended stories of this type may
well be forerunners of the kinds of complex, inleractive
stories that will ultimately be created as computers are
used as more than simple "word processors in the creation
of fiction.
In this paper we will describe the first phase of
development or a computer program, UNIVERSE. The
goals of this program are twofold. Initially, we expect to
develop a program that can serve as a "writer's a i d " for
extended story production, intimately, we hope to have
UNIVERSE tell "soap opera" type stories. The aspect of
UNIVEKSK to be discussed here is the creation of storytelling; universes (STUs) comprised of characters, their
histories, family relations, and interpersonal relationships.

2 Motivation in Character Creation

Good extended story telling is constrained by the need to
maintain consistency and coherence.
A story is
consistent if properties and events of the story world
(including properties of characters) do not contain any
internal contradictions. Bitter enemies should not become
close friends without explanation.
Coherence involves the idea that events should be
logically derivable, at least in retrospect, from the
information available to the reader. While it is neither
necessary nor desirable to provide the reader with enough
information to actively predict every event that will occur
in a story the events should make sense when they do
occur.
Most notably, actions should be based on the
personalities and backgrounds of the characters involved.
UNIVERSE creates a universe of characters before
beginning to tell a storv in order to maintain consistency
and coherence. It would be possible (and is sometimes

necessary) to create characters only on demand in plot

generation and fill in details of these characters only when
needed.
However, in order to keep the qualifies of
characters consistent, and to be able to insert information
into the story that will lead to later coherence (i.e. "drop
hints"), it is useful to have a substantial set of characters
in place when story telling begins.
Coherence provides benefits beyond merely protecting the
reader from untoward surprises. In large part, coherence
provides the "flavor" that makes serials so popular. It is
difficult to identify with characters whose actions seem
inconsistent. Part of the fun of soap operas is speculating
about what, will happen next, a game that is fun only if
plot actions are coherent.
The major relevant AI programs are Mechan's T A L E SI'IN [Meehan 7b] and Dehn's A U T H O R [Dehn 81]. Both
of these programs focus on the ability to tell planful
narratives involving small numbers of characters, and
usually focusing on a single character. In contrast we are
not concerned with fully representing every last detail of
every character and action, but rather with having
enough information available to tell consistent, coherent,
and hopefully interesting stories over long periods of time.

3 Person Frames
To maintain consistency in stories, we must consider the
information to be created and maintained about people.
This collection of information will be referred to as a
person frame. Previous work in this area, particularly
[Carbonell 79; Schank and Lebowitz 79], was concerned
with complex, plan and goal-based representations that
allowed a understander to explain a person's actions. We
will use here a somewhat simpler representation that
provides enough information to generate believable plots
by concentrating on personality traits, interpersonal
relations and, to some extent, goals.
The way to maintain coherence in character description
can be seen by considering how wc might answer a
question such as, "Why is Aaron so cold-hearted?" T w o
possible answers are, 'He's a door-to-door salesman" or
He got that way after his wife left h i m " . These answers
are illustrative of two ways to make characters coherent,
stereotypes and past events.
Stereotypes are common descriptions associated with
people in various classes such as occupations, social
groups or personal backgrounds. A stereotype conveys a
large amount of "default" information, and tends to
provide enough coherence to satisfactorily motivate an
aspect of a character's make-up. Further, the fact that
the person is a doctor provides an acceptable motivation
for any of these characteristics.
Stereotypes therefore comprise an important part of
character descriptions. The description of a character in
UNIVERSE contains a list of one or more stereotypes
from which the character inherits properties (unless
overridden). The stereotypes used here serve much the
same role for person descriptions as generalization-based
Memory Organization Points [Schank 82; Lebowitz 83) do
for events.
Figure 1 shows a few of the over 50
stereotypes currently included in UNIVERSE.

64

M. Lebowitz

doctor
warden
klutz
socialite
movie-fiend
Figure 1:

professor
flake
swineer
hari-krishna
sports-fan

store-owner
nasty-person
family-man
bowler
junkie

Some stereotypes used in UNIVERSE

Past events are used primarily to provide variety and

flavor and explain aspects of a character that do not fit
stereotypes. By including historical events in an STU, it
is possible to nave interesting variations in the various
characters, along with adequate justification to assure
coherence. (To achieve the coherence, it is, of course,
necessary to refer to the various past events so that the
reader is aware of them.) In the ultimate version of
'UNIVERSE, as the program tells stories it will naturally
affect the history of the characters involved.
Figure 2 a few of the plots that can be run historically by
UNIVERSE during STU creation.
lawsuit
consultant
big-loss
famous-rider

divorce
abuse-charges
threatened
break

Figure 2:

illness
robbery
rackets
takeover

Sample historical plots

The final important class of information needed about the

characters in an STU involves not individual characters,
but interpersonal relationships between characters.
Crucial to the selection of consistent plots is the need to
monitor the way characters relate to each other. To
achieve consistency of this sort, we must maintain
information about relationships between, potentially,
every pair of characters in our universe.
Figure 3 summarizes the information stored in a person
frame. Traits, goals and the role of marriages will be
discussed further in later sections of this paper.
Name, Stereotypes,
T r a i t modifications,
Individual goals, Interpersonal relations,
Marriages, History (list of events)
Figure 3:

Person frame information

3.1 Stereotypical frames

UNIVERSE stereotypes provide two basic types of
information
about
the characters they describe
descriptors of physical and personality traits (which we
will concentrate on* here), and typical goals. Personality
traits are particularly useful in selecting characters for a
given plot (or plots Tor a character). In a blackmail plot,
the blackmailer should not be a nice person, and the
person blackmailed, or his family, should have something
worth extorting.
Goals tend to be important in a somewhat different
context.
UNIVERSE will make use of certain plot
schemata, known as meta-plots, that are abstract in terms
of the events they describe. "Revenge" would be an
example. To "execute" such plots, it is often necessary to
have information about various characters' goals
-- particularly long-term goals in order to understand
what is important to them".
The selection of traits to be used in stereotypes is

certainly not self-evident. We need a selection of traits

wide enough to indicate which characters should
participate in which plots and how thev should fare in
such plots. The traits chosen for UNIVERSE are shown in
Figure 4. No claim is made that this listing is exhaustive.
It allows for a wide range of behaviors to be explained
and can be easily extended.
sex t
moodiness
wealth
religion
social-back

age
promiscuity
competence
race
time-used

Figure 4:

phys-att
niceness
intelligence
nationality

Person stereotype traits

3.2 Interpersonal relationships

While it is important to have individual characters
adequately represented, it is also necessary to represent
interpersonal relationships (IPRs) between characters.
We have chosen to do this with a set of numerical scales
based on those from the psychological experiments of
[Wish et al. 76), extended for AI use in jSchank and
Abelson 77].
This work posited three scales
-- positive-negative, intimate-distant and dominantsubmissive to describe the way people relate to each
other.
To these we have added a fourth scale,
attractedness, appropriate for soap opera universes. More
complete sets of scales are being considered for later
versions of UNIVERSE.
Our scales can include
asymmetric perceptions of each participant in an IPR.
As with individual character information, values of scales
can come either from stereotypes or past events. Past
events seem to be more important in shaping
interpersonal dynamics. While there are a reasonable
quantity of stereotypical IPRs ~ among family members
and people in occupations, for example - events such as
shared crises, conflicts at work, marital conflicts, and the
like, tend to be more significant in explaining how one
person relates to another.

4 Building Up an STU
Having specified the kind of information we would like to
maintain about the characters in a story-telling universe,
we are still left with the problem of how to generate that
information. We could specify properties of characters
and interpersonal relations randomly or only when needed
for a plot (a necessary mechanism in any case). However
to see how we might add coherence to our set of
characters, consider the universe of characters in a typical
television soao opera, Days of Our Lives. Figure b lists
some of the 40 or so current active characters in this NBC
soap opera.
T o m and Alice Horton family patriarchs
Mickey and Maggie son and wife
Sarah their daughter (by artificial insemination)
Stephano the bad guy
Tony his son
Renee -- Tony's half-sister
Liz Tonv's estranged wife
Don Liz s ex-husband (sort of, she was actually
still married to Tony then)
Marlena another of Don's ex-wives
Roman Marlena's fiancee
Anna Roman's first wife, presumed dead
Figure 5:

A sample soap opera universe

M. Lebowitz 65

The first thing we see from Figure 5 is that the relations

among the characters are quite complex (an informal
count showed at least 20.> identifiable interpersonal
relationships), and that many of the characters are related
to each other by past and present marriages of one sort or
another. Past and present husband-wife and parent-child
relationships add coherence to the character set in most
soap operas.
We would like to have this same sort of coherent intricacy
among the characters created by UNIVERSE. The most
feasible way to achieve this is to, in effect, do a simple
simulation of the past lives of the characters in the STU,
creating spouses, children and other assorted characters as
necessary.
We view marriages as a motivation for
characters rather than as arbitrary relationships that
should be established after we have created a set of
characters.
The simulation used by UNIVERSE involves cyclical
character creation. A queue is maintained of characters
who have been created during the simulation, but have
not had the details of their lives filled in. Each character
is, in turn, removed from the queue, and UNIVERSE
steps through his or her life, creating spouses and
children. wnen the present is reached further details of
the character, such as occupation and other descriptive
stereotypes are filled in.
The selection of the specific events that occur can be done
either by a user (in "writer's aid" mode) or somewhat
arbitrarily. The latter is not as unreasonable as it may
sound, as' the idea is often to have many interconnections
among characters, regardless of what they might be.

5 Adding More Background

Although the cyclic character creation process described
in the previous section provides the framework for the
STUs created in UNIVERSE, it is not quite enough by
itself. There are three main problems with the universes
created: 1) Almost all the characters are related in one
way or another; 2) Almost all the IPRs in the universe
have to do with marriages and children. 3) There is no
history, other than marriages and offspring, to the
characters, which is detrimental to coherence.
Fortunately, we can solve all three of these problems with
a single mechanism, one related to the main thrust of the
UNIVERSE program. The answer is to simply add
history to the characters' backgrounds by "executing" in
the past simplified versions of the plots UNIVERSE will
create for the present.
It is crucial to note that the past plots run by UNIVERSE
need not be executed in the same detail as during actual
story telling. Instead, it is enough to merely determine
how each plot affects the characters involved and the
IPRs among them, and modify their histories
appropriately. UNIVERSE can add coherence to its
current stones by referring to events in the past. In
addition, the mechanism for selecting the plots to be run
need not be the same as for real story telling. Instead, we
can select a given plot in order to obtain a desired [PR,
without requiring the same level of motivation as is
needed in real story telling.

6 Conclusion
We have shown here a method for generating extended
story-telling
universes,
consisting
of
characters,
interpersonal relationships and historical events. The
method of creating characters focuses on maintaining
consistency and coherence. The methods described are
flexible enough that they can be extended to address
additional problems, such as specific author intent, by

adding constraints in the character generation process.

The work described here provides the framework for the
extension of UNIVERSE to actual story telling.

Acknowledgements
Special thanks go to Susan Rachel Burstein who helped
develop many of the ideas described here. Comments by
Kathy McKeown on an earlier draft of the paper also
were extremely helpful. A more complete description of
the work described here is available from the author.

References
[Carbonell 70] Carbonell, J. G. Jr. Subjective
understanding: Computer models of belief systems.
Technical Report 150, Yale University Department of
Computer Science, 1979.
[Dehn 81] Dehn, N. Memory in story invention.
Proceedings of the Third Annual Conference of the
Cognitive Science Society, Berkeley, California, 1981, pp.
213- 215.
[Lebowitz 83] Lebowitz, M. "Generalization from
natural language text." Cognitive Scierice 7, 1 (1983), 1
- 40.
[Meehan 76] Meehan, J. R. The metanovel: Writing
stories by computer. Technical Report 74, Yale
University Department of Computer Science, 1976.
[Schank 82] Schank, R. C. Dynamic Memory: A
Theory of Reminding and Learning in Computers and
People. Cambridge University Press, New York, 1982.
[Schank and Abelson 77] Schank, R. C. and Abelson,
R. P. Scripts, Plans, Coals and Understanding.
Lawrence Erlbaum Associates, Hillsdale, New Jersey,
1977.
[Schank and Lebowitz 70] Schank, R. C. and
Lebowitz, M. Does a hippie own a hairdrier? Technical
Report 144, Yale University Department of Computer
Science, 1979.
[Wish et al. 76] Wish, M., Deutsh, M. and Kaplan, S.
"Perceived dimensions of interpersonal relations."
Journal of Personality and Social Psychology 88, 4
(1976), 409- 420.