Interactive Storytelling Tools for Writers

January 26th, 2022

Recently, whilst reading a long story about the huge investments companies are making in new technology, I wondered what it would be like if they realized the huge economic potential of interactive storytelling and decided to throw some loose change in that direction. What would I do, I mused, if SuperHyperMegaCorporation offered me a billion bucks to build a group for making interactive storytelling software?

Well, first, I’d have to refuse most of that money, because growing too fast only generates chaos. I saw that at Atari. Moreover, the ideas behind interactive storytelling are so unconventional that I’d need to educate a core team in the fundamentals, and people don’t learn on a rush schedule. Here are some thoughts on how I would proceed; I’m hoping that, when some bright person at SuperHyperMegaCorporation realizes that interactive storytelling is the Next Big Thing after games, they’ll steal these ideas.

Two primary strategies
The two primary strategies that I recognize for interactive storytelling are radically different. I call them the “AI strategy” and the “analytical strategy”. 

The AI Strategy
We select a genre of entertainment fiction (e.g., bodice ripper novels) and scan a few tens of thousands of such novels into our database.

Specimen of genre “bodice ripper"

Now comes the hard part: defining a set of analytical concepts that can express the ideas in these novels. Somehow we must reduce the components of these stories into computable terms. A good example of the overall concept can be found in the work of Vladimir Propp: Morphology of the Folktale. It’s a good start, but a close examination shows that his system is too small to truly handle even the simple folktales he addresses. 

For the specific case of bodice rippers, it’s easy to identify components such as “He kissed her” and “She blushed”. But these are only the most blatant of components. Most components are far more subtle, presenting the many subtleties of romantic drama. What components define the uncertainties that characters have about the intentions of each other? What components define the elaborate process of subtle enticement and tentative rejection? This will be an immensely difficult problem, and will be different for each broad genre.

Once we have assembled our set of dramatic components, we scan the entire database of Gothic romances, translating each one into a sequence of components. Then we assemble a table of probabilities of particular sequences of these components. For example, we may find that dramatic component E27 is followed by component R416 in 91% of its occurrances, but by component Z9 in 4% of the occurrences and component ß62 in 5% of the occurrences. (We could extend this table to sequences of more than two components for greater dramatic resolution.)

This table of probabilities need only be inverted to generate reactions to the player’s choices. The engine presents an initial component from its table, and offers the player a choice between all the components that follow that component. When the player makes a choice, the engine then considers each of the components that follow the player’s chosen component, selecting one (most likely the most probably succeeding component, but some randomness is permissible here.)

A second difficulty with the AI strategy is the translation of the dramatic components into language. This could be accomplished with Tinkertoy Text, but it would have be much richer, more fully developed Tinkertoy Text to be effective. We could throw a dozen writers at the problem.

Here’s a paragraph I wrote some years ago about various attempts to abstract dramatic components of storytelling:

There have been some such efforts. Vladimir Propp’s Morphology of the Folktale comes closest to our needs, but ultimately doesn’t work. Georges Polti’s The Thirty Six Dramatic Situations is too high-level to be of use. Stith Thompon’s The Folktale provides a great deal of useful information, but it is not organized enough to be of direct use. Barre Tolken’s The Dynamics of Folklore provides some interesting high-level analysis of folktales, but again, the information is not categorized well enough to be of use. At the opposite extreme is the Aarne-Thompson-Uther System  of motifs of folktales. Its flaw is that it catalogs Objects, not Processes. (Interestingly, Vladimir Propp made exactly this criticism nearly a century ago.) Of greater utility is An Atlas of Interpersonal Situations, but I warn you that the model this book presents is difficult to apply without thorough study.

Reducing text to a database of dramatic components is the killer problem in the application of AI to interactive storytelling. If I had tons of money to spend, I’d put together a group to try to figure out how to do this, but I would not expect it to produce anything useful for quite some time. It might be possible to identify some key components using a multi-dimensional analysis of similarity matrices — but multiple problems arise. First, what is the level of granularity of the elements to be compared? Sentences could well be too short to compare. Even paragraphs could lead to problems. Here’s an example of excerpts from two such novels. Are they the appropriate size for comparison? How closely do they match? Is the resolution of any similarity measurement adequate to determine the utility of fitting one into the same spot that the other fills? 

My purpose in this example is to demonstrate the enormous complexity of the problems we face here.

The Analytical Strategy
I have for 25 years pursued the analytical strategy in which an artist builds a storyworld using algorithms that specify the characters’ reactions to dramatic situations. In this, I failed; the task proved to be too big and too complicated for one person to carry out. Nevertheless, I consider this the best strategy for pursuing interactive storytelling. I developed a great deal of technology for this strategy, and I think I came close to success, but it was just too big a task for one person. A large group could probably pull it off. You can consult any of the voluminous literature on this website about the Storytron technology, or search the web for “storytron”.

Are the atoms of expression phrases or sentences?

Five primary technologies
As I have been telling people for decades, there are five primary technologies that interactive storytelling requires. 

#1: Faces
I’ve been telling people about the importance of emotionally expressive faces for many years. My own work is primitive; I’m not an artist. Nevertheless, there are many important points about the use of faces that are not addressed by any of the existing technologies. Here are some of the essays here on Erasmatazz discussing faces:

Erasmatron Faces
Facial Display Technology
Siboot Faces
Graphic Style for Siboot Faces
Two crummy essays on faces: Programming Moods, and More on Faces
Face Expressions
Alterations to the Face Editor
Pupils
New Face Design
Moods and Facial Expressions

#2: Personality Models
Most people underestimate the difficulty of creating good personality models. They just grab the most common personality model developed by psychologists, the OCEAN model. This is stupid. The OCEAN model describes real people, people like this:

Stories don’t have real people; stories have characters, who do things that real people would NEVER do. Seriously, how many real people do you know who would take the red pill?

Hence, we need a personality model tuned to the decisions that characters in a storyworld make. I have such a personality model, but I have never been satisfied with it.

It would be a fairly straightforward process to carry out a multi-dimensional analysis of similarity matrices of characters in fiction. This could be used to identify personality models of any desired size. I actually started such a project some years back, but my collaborators did not sustain their efforts and the project collapsed. 

#3: Language
This is the killer problem in interactive storytelling. There is no way to tell decent stories without language. Indeed, I believe that storytelling and the development of language are closely related. But we do not have software capable of comprehending language at the level necessary for dramatic purposes. Sure, this works:

“Siri, where’s the nearest Mexican restaurant?” 

But what about this:

Fred: “I can’t have sex with my wife because she’s got some sort of yeast infection.”
Bill: “Gee, she was fine last night while you were out bowling. Har!”

Think through the algorithms required for a computer to grasp the dramatic significance of that exchange.

My guess is that we’ll have to settle for complete sentences as the atoms of interaction in our storyworlds. We can make this less repetitive with Tinkertoy Text.

#4: Narrative Engine
This is not as intellectually challenging as the language problem, but it’s still a bear. My preferred approach (the analytical approach) requires the creative storytellers to express dramatic concepts in algorithmic form. Much of the Storytron technology was devoted to making this task manageable. 

#5: Development Environment
This is the software for generating the specifics of a storyworld. For Storytron, that development environment was called SWAT, and it was quite an impressive piece of software. I’m sure that it could be mined for useful ideas.

Staffing
Assembling the team required to build interactive storytelling will be a long process. To put it baldly, I do not believe that there are any people whom I could hire and have them get to work immediately; every new staffer would have to undergo a lengthy educational process to overcome the many misconceptions that inhibit progress in the field. I would want to hire at most a dozen people and put them through Interactive Storytelling Boot Camp, disabusing them of their misconceptions and getting them pointed in the right direction. I’d probably want to start with rather young people, who would have less to unlearn. Certainly I want to avoid anybody with lots of experience in game design.