February 28, 2006
Minstrel, Universe, and the Author
We’ve had quite a bit of writing about literary work using digital computation since the 1990s (e.g., Landow’s Hypertext; Ryan’s Possible Worlds, Artificial Intelligence, and Narrative Theory; Murray’s Hamlet on the Holodeck; Aarseth’s Cybertext). But there’s been surprisingly little written about two story generation systems that got their start in the mid-1980s: Minstrel and Universe. They’ve caught my attention recently, so I’m planning to write two or three posts (after this introductory post) about them. I’ll be contrasting them with the best-known story generator, Tale-Spin — which, if memory serves, is written about in all but one of the books mentioned above.
To us — 30 years after Tale-Spin‘s completion, and outside that context of AI research — Tale-Spin seems strange in its blindness to all but planning. But for understandable reasons the artificial intelligence community did not respond this way to Tale-Spin‘s limitations. After all (as outlined by people like Lucy Suchman and Phil Agre) planning was essentially the only account of intelligent action at that point in AI’s history. Rather, the common diagnosis was that Tale-Spin didn’t make enough plans. Specifically, Tale-Spin was critiqued for only accounting for the plans of characters, and not the plans of authors. The Minstrel and Universe systems sought to address this perceived lack in different ways.
Early publications about the Minstrel and Universe systems appeared in the mid-1980s, and bore a number of similarities. Not only did both emphasize the importance of authorial plans and goals (and cite Natalie Dehn’s 1981 publications about the importance of such goals), they also focused on the importance of then-recent models of how people understand stories. In particular, they focused on Roger Schank’s “Memory Organization Packages” (also called “Memory Organization Points”) and Michael Dyer’s “Thematic Abstraction Units” (also called “Thematic Affect Units”). No doubt this was in part due to the histories of the two systems’ architects. The primary designer of Minstrel, Scott Turner, was at that time a student of Dyer’s at UCLA, and Dyer had recently completed a dissertation at Yale influenced by Schank and Abelson’s ideas. (Dyer’s dissertation became the book In-Depth Understanding: A Computer Model of Integrated Processing for Narrative Comprehension, published by MIT Press in 1983.) The primary designer of Universe, Michael Lebowitz, had also recently written his dissertation at Yale under Schank’s supervision — contributing to Schank’s model of dynamic memory, especially in relation to story understanding. (Lebowitz’s dissertation was “Generalization and Memory in an Integrated Understanding System,” filed December 1980.)
Working with this shared background, it is not surprising that, on one level, the Minstrel and Universe models of authorial activity are quite similar. Both see the system as seeking to provide for the audience those elements identified as necessary for story understanding. Of especial importance, for both Minstrel and Universe, is the idea of the “point” or “theme” that the system is attempting to put forth — something notably absent in most of Tale-Spin‘s stories about characters trying to satisfy simple needs.
At the same time, the Minstrel and Universe systems are underpinned by rather different beliefs about how their processes relate to the behaviors of human authors. An examination of these differing beliefs is helpful in explaining the differing results achieved by the two efforts. And toward that is where my forthcoming posts will turn.
February 28th, 2006 at 6:00 pm
Very cool, Noah. I read Turner’s book on MINSTREL way back when, but not since working on interactive drama in earnest. I definitely need to re-read it — I’m looking forward to getting back into it with your posts here.
February 28th, 2006 at 7:18 pm
I read Turner’s book too, and totally failed to understand it. I’d welcome a clear explanation of how his system actually worked.
I’ve read a most of the existing of research theses on story generation now and I find that they all have very similar failings:
i) While they often show impressive output from their systems, they rarely show the the necessary INPUT to produce it.
ii) They demonstrate the system’s ability to produce ONE story, but if the system is worth its salt it should be able to produce MANY stories.
My fear is this: the systems are producing the one story that the programmer has cleverly crafted into the input. The story-telling is not being done by the program at all, but by the programmer who carefully engineers the input representation so that the system will “discover” the hidden story. Like a parent hiding eggs for an easter egg hunt, so the child can later find them. Don’t hide them too well, or else the game is no fun.
I am waiting to see a system with a world model that is rich enough for many stories to be told, including stories that the author did not expect. There are several things preventing this, I think:
1) We move on to quickly from one research system to the next. We only generate the necessary example to ensure publication, and then no-one ever runs it again.
2) Before we even begin to consider the “narrative modelling” problem, we are faced with major difficulties in common sense reasoning. It’s difficult to build models with all the complexity necessary to tell an ordinary story. Especially if we want to tell many stories, as we cannot focus our attention on implementing very specific bits.
March 1st, 2006 at 9:28 am
Glad you two are interested! I’m finishing up the first post right now, and getting the sense that it’ll probably take at least four posts to get through everything I want to say…
March 1st, 2006 at 9:38 am
Turner’s Minstrel, part 1
by noah @ 9:33 am
As mentioned in my previous post, the first publication about Minstrel appeared in t […]
March 4th, 2006 at 9:19 am
[…] erse, part 1
by noah @ 9:11 am
As mentioned in the first post of this series (1 2 3), the primary designer of Universe is Michael Lebowitz (also, […]
May 16th, 2006 at 11:04 pm
Noah, I’m not sure whether you know or remember but when I followed Jay Bolter to the Yale AI lab on sabbatical in 1985, I worked with Nathalie Dehn in Roger Schank’s group. It was then that we developed Stopryspace.
May 17th, 2006 at 8:21 am
Michael, great to hear from you!
I remember the version of the story from Of Two Minds, but it’s been a while since I read it, and I think that’s all I’ve heard. If you’d be up for expanding your thoughts here or in e-mail, that would be much appreciated — I’m going to be thinking and writing quite a bit about these topics over the next six months or so.
July 10th, 2006 at 6:44 am
I too have studies all the literature in Story Generation regard. I want to say that the best system at present: StoryBook by Callaway, does even produce only one crafted story, Little Red Riding Hood. It is unable to write stories other than this domain.
I tried myself also to develop such an algorithm through which this task of story generation could be done automatically for any sort of story and not for specific one.
I would like to listen views from you audience if you come up with any good suggesstions.
For that i think that system will require a bulk of input and will have internal structures to get its knowledge base updated with the changing world. But this task is hactic and no algo is on surface up till now.
A group at Hamburg university, Narratology Research Group, is still working to develop such an algorithm.
The irony of fate with me is that, This is my final year project and i havnt found any good approach to carry on so far. Looking for help out of google.
July 10th, 2006 at 9:02 am
Sana, you might want to check out Michael’s Terminal Time project, which takes a different (and, for me as an audience member, more successful) approach to the issues.
October 30th, 2007 at 8:41 am
[…] r I posted a series of thoughts about two story generation systems: Minstrel and Universe (1 2 3 4 5). I had some critical things to say about the Minstrel system, bu […]