Michael Mateas gave the keynote today at Intelligent Narrative Technologies 6. With reference (early on) to the Hero’s Journey, he presented a sort of “developer’s journey,” noting that indie developers (as seen at Indiecade) have been turning away from concern with structure and mechanics and toward narrative. He similarly encouraged those working in AI and narrative to turn from structuralist narratology and look at concrete traditions of narrative based in communities of practice.
I thought his repudiation of structuralist narratology was in some ways similar to someone in computer graphics objecting to the pixel or the polygon becuase pixels and polygons do not provide any guidance as to how to create beautiful images. He’s right, and if people are missing this, it’s worth pointing out. But the core problem is with one’s expectation for structural understanding. If people are interested (as Michael is) in modeling “internal processes and conflicts” mapping to “external conflicts” … what are these external conflicts made of? Aren’t they made of events? And wouldn’t it be great to have a solid model of events so more complex narrative phenomena can be built up out of them?
Michael gave us an array of options for work motivated by specific poetic ideas, and these (in keeping with his practice) were extremely grand plans, with a panpoly of dissertations needed to make real progress. His suggestions were that we build huge AI systems, implementing Beckett, Boal, detective novels, flash fiction, and the levels of intentionality in Virgina Woolf. I don’t object to attempting to deal with goals of this sort, but this tour of the interactive narrative candy store seemed to be proposing 23 different manned space missions.
My idea: Why not create & compose simple models of narrative aspects that are indeed culturally grounded and for which there is a poetics, but which seem lower-level (or easier to start on) than the ideas Michael has – aspects such as repetition (seen in Beckett, of course), ellipsis (and its relationship to suspense, as treated in Michael Young’s research), and the like?
These sorts of questions seem to have easier starting points and offer the potential to generalize to some extent across genres, while yielding specific insights as well.
A repetition or ellipsis system would be as culturally grounded as any one that Michael suggested, and it too could be “fully realized” and produce outputs.
My very simple system “Through the Park,” described in “Small-Scale Systems and Computational Creativity,” is an attempt to show how low the stairs are for those interested in investigating ellipsis. Although extremely simple, it is a real model of an aspect of narrative and has been useful to me in thinking about it.
Ian Horswill pointed out in the discussion after the talk that understanding repetition may not be easier than generating romance novels, flash fiction, or work like that of specific authors. This is true, but with repetition there are more concrete starting points. One can certainly start with a lightweight simulation of the world, and of narration, and then elaborate.
Richard Evans mentioned that the repetition in _The Odyssey_ and in Beckett’s plays doesn’t have much in common. I would say that this is true, but that they have _something_ in common – creating a coherent and distinctive texture of language. On the other hand, even within the _same work,_ or work by the same author, there will be different types of repetition that do different things. That’s what makes this aspect of narrative (or, really, textuality) a rich one. It seems to me that these different uses, within a single work or across different works, could be understood analytically and modeled computationally in useful ways. The focus on a single aspect, rather than a genre or community of practice, would make this more tractable and offer a foothold.
I should note that Michael and I agree on an important point: That work in INT and other AI work with narrative needs to be based on poetics, not just on narrative structures.
So, for instance, I think working with the poetics of repetition might be a good idea, or doing something of that sort. Michael is looking at the poetics of flash fiction and other coherent practices and genres. Although I prefer looking at smaller-scale narrative aspects, I do agree that attention to poetics is essential.
You seem to think that starting with lower-level narrative aspects will avoid the “moon shoot” flavor of the projects in my candy store. But procedurally modeling the aesthetics of ellipses, for example, could be just as deep and hard a problem as modeling the aesthetics of levels of intentionality in literary fiction. In Through The Park, you have carefully constructed phrases such that, when randomly conjoined, the reader is able to fill in emotionally powerfully implications in the ellipses. Much of the aesthetic value in the work is carried by the artistry and craftwork in these phrases, not in the process. You can imagine incrementally deepening this by having authors associate features with sentences that are used to determine which conjunctions, when separated by an ellipsis, produce an interesting aesthetic effect. Then you could create a taxonomy of different aesthetic effects achieved by ellipsis and associate these effects with the tags associated with conjoined phrases, using this to guide inclusion of phrases. Then you could explicitly model the semantics of phrases and use this to reason about the implications (including emotional responses) readers will draw in the gaps created by ellipses, and use this to ground a generative model. And pretty soon you’re off to the races with a moonshot sized project to procedurally model the poetics of ellipses.
What I’m concluding from your comments are:
1) I didn’t include any examples of the poetics of texture (ellipses, repetition), and I agree that textual texture contains many interesting phenomena worthy of attention.
2) The idea of starting with micro versions that minimize process and push much of the narrative phenomena of interest onto carefully authored content, and then incrementally adding more to the process model to increase generativity and opportunities for interaction is a great one. I think this could be applied to projects looking at Brechtian drama, or “ends-with-a-twist” flash fiction, or nested levels of intentionality in Woolf as well as to texture as you have done.
Michael, thanks for the reply. Having a blog conversation with you brings me back to the happy days of the 2000s.
I think “texture” or discourse-level variation is quite important. In Richard Evans’s talk today, he described a murder mystery and talked about how his AI influenced characters’ behavior in it. But an exceedingly important thing about a murder mystery, as far as I can tell, is that a character is murdered (in the underlying storyworld) but the act of this murder itself is not narrated. The narrator skips past this and only later do we figure out who committed the murder. This move (this ellipsis) is hard-coded in systems now, but I think it shows that the question of what to include and what to omit is a central one in narrative.
“Moon shots,” extremely ambitious projects, are not at all a bad idea. The issue I have is with how principled or structured a particular problem is and whether the focus is going to be productive. In developing a flash fiction author or a simulated Beckett, where do you start? In investigating repetition, you can begin with a poetics of repetition (e.g., Deleuze’s Difference and Repetition). Or you can begin with a concept such as suspense/surprise. There are concrete first steps. And, what does it mean to have a system that does 25% or 75% of flash fiction poetics or a similar amount of Beckett? Such a system might not do anything recognizable. On the other hand, a system that implements 25% or 75% of the poetics of repetition is indeed sensible to imagine, and I imagine it would do reasonable well at narrating even if it was not at the level of a master author.
I do recognize that I’m not the one with Ph.D. students who can focus on projects like these, so my suggestions, and my ability to work with students in these directions, will be limited by that. (I have too much else going, and too few resources in terms of research collaborators, to pursue ellipsis and extend the very slight initial project of “Through the Park.”) But I feel that these are reasonable directions to pursue, with a balance between certain aspects of generality across narratives and certain types of cultural and individual particularity.
A Michael Mateas related post not from 2007?!
What is this?!?!?!
Hey Nick – Thanks for your thoughtful comments.
The reason I brought up repetition in The Odyssey and in Endgame is because they are used so very differently in these two cases: in Endgame, repetition is used to underline the fact that the characters are trapped. In The Odyssey, repetition is used to provide moments of continuity and reassurance: no matter how bad things get, dawn is coming once again with her rosy red fingers.
I worry that any attempt to build a computer model of a literary technique (say, repetition), which does not focus on what the technique is for, is not going to capture the importance of that technique. My reason for bringing up The Odyssey and Endgame is that repetition is used in these works to serve such different narrative ends.
cheers, Richard
Richard, thanks for the reply and clarification.
On the one hand, I certainly think that serious creators of large-scale computational literary systems, and particularly those involving repetition, should study Homer and Beckett. It is a good idea to ground one’s systems in the literary practices of specific authors and those seen in specific works, because a particular technique, formally defined, does not have only one purpose or effect. For a Ph.D. project about literary repetition in English, it would be essential to study both Homer and Beckett.
On the other hand, I believe a bottom-up approach of simulating a technique, such as repetition, without deeply considering what it is for could still be worthwhile. An offhand project (for NaNoGenMo, for instance) that models repetition could lead to interesting discoveries, even if the creator doesn’t make the astute distinction that you did. It’s possible to build a system that repeats in various different ways and then see what the effect is by reading the output. And, one could build a repetition system first and later study Homer and Beckett.
I guess, then, that I generally agree — but I don’t think that one needs to consider what the technique is for initially and then begin building. I think one can begin by experimenting and find out later, through reading and experience, what the purpose of a technique is. And then wash, rinse … and repeat.