Edinburgh, Scotland
August 6-9, 2002
by Lloyd Rutledge
Site of NILE 2002 — University of Edinburgh John McIntyre Centre
I was definitely a geek (perhaps the geek) in a non-geek workshop, but fit in and enjoyed myself nonetheless. While our idea of narrative is much like that of the narrator of a factual video documentary, the types of narrative discussed in this workshop were different. The obvious difference is that we were the only ones discussing the automatic generation of narrative. Many talks involved narrative the emerges from the user's interaction with a well-crafted world model. Others involved teaching how to make good narrative, typically in the form of authoring tools. Some work involves the users themselves as characters in the narrations put before them, which our system does not do at all. Getting even further away from our work, others discussed the important of good narrative in material presented to children, whether it be human or machine-made. This last category came from storytellers, of which Scotland, and Edinburgh in particular, have an active community supporting the region's ancient tradition of the art. However much our work may differ from much of the workshop, there was definitely a place for our presentation and our ideas in this workshop.
I have a bound copy of the papers printed with the presentations, if anyone is interested. These are supposed to appear eventually. Videos were also made of the presentations, and these may become available in one way or another.
This talk introduced the workshop buzzword hard fun, which is engagement of the sort from entertainment and games, but applied to education. For hard fun, the means of engagement should be more than "sugar coating" (a buzzword introduced later) to dazzle the young student into sticking with the program. The means of engagement should be part of what is learned, and how it is learned — that is, there should be thematic coherence between the two. Having a computer interface convey and use learning in conveying a lesson to a student is what Clark presents as a means of implementing hard fun. This addresses the problem that e-learning is typically quite boring.
A key theme of this talk was emergent narrative, which is narrative that is not explicitly written by a human author but comes about because the user interacts with a complex enough world model such as those found in simulators or games. The advantage is that this provides many possible narrative progressions with less narrative creation work for the author. Emergent narrative is also more natural feeling than automatically generated generative, partly because the user acts as an unknowing narrative author, and partly because the user imagines some of the narrative beyond what comes directly from the interface. The disadvantage is that emergent narrative is still somewhat limited and "script-like" compared to human-authored narrative. But with work and insight you can make a good enough model that "computes" but still feels right to the user.
The primary implementation Clark presented was one he developed as part of a social project in Australia. It is a real live simulator for teen orphans, who have lacked role models from whom to pick up general survival and life management skills. Privacy regulations rightly prohibit formal case studies on wards of the state, but informal feedback has indicated that the teenagers get quite engaged by the program and that it has helped them manage their adult independence better. Using a computer program was not only helpful from its engagingly thematically coherent hard fun, but also because state-raised orphaned teens are quite skeptical of adults providing advice on daily living. This was one of several talks that showed how computers provide a unique, and even social, interaction for the students that is different from, while complementing rather than conflicting or competing with, the human teacher.
A ratio Clark presented from this work and as hearsay from the SimCity creator is that creating the world model from which the narrative emerges is 10% of the work. The remaining 90% is fine-tuning the model so that the narrative that emerges flows well enough. I imagine this applies to our work as well. Setting the the semantic relations in the Cuypers data store is complex enough. But fine-tuning these relations and how they are process so that a natural-enough narrative emerges on a regular enough basis would be 10 times that work. Of course, this extra work is only necessary for commercial-grade interfaces, so much less extra work would be required for us in our demos and feasibility case studies. However, and insight would could provide for fine-tuning automatically generated factual narrative would be a valuable contribution.
The questions period raised a wide variety of interesting topics. Many revolved around how encasing a message in the narrative of a memorable story makes the message itself more memorable. One example was how companies are investing in company culture, which is the professional development of stories conveying messages about how the company operates and what its values are. Essentially, it is the creation of fables and legends about the company, except that they are developed professionally by hired contractors rather than by casual chatter during coffee breaks.
My question was about how good emergent narratives could be as compared to authored narratives. His response was, of course, that authored narratives will always have far more potential for richness and subtlety, but there is a lot that emergent narrative can provide from the devoted crafting of a good model. Now I wonder, can Topia within Cuypers have an emergent narrative as well as a generated one? Can us crafting a good semantic model enable better narrative to emerge?
This work provided an authoring tool for grade-school children to write about their visits to museums. Apparently I should have drinken more coffee during the break, because I remember little about it! The rest of the audience seemed keen enough, however. It was one of several papers with authoring environments for kids that help them make good narrative structure.
Craig once had a six-month visit as CWI-INS2, so it was nice to see his work. He began this talk with "How many people play computer games more than they watch TV?". A small handful of arms raised, including Craig's. Craig's purpose is to make the world a better place for computer games. He does this by heading the Zero Game Studio at Visby, Sweden, which has been going on for 2-1/2 years and has seven people.
He discussed the importance of good narrative in games, and how the gaming industry is now working hard to put good narrative in their games to make them more engaging and to give the user the sense that he or she is part of something that is happening. Of course, emergent narrative is a big part of this. The game takes place in a world, and if that world's computed model is rich and well-crafted enough, an engaging, believable narrative will emerge from it. Craig also discussed the "three acts" of Hollywood film structure, and how they are repeated in games. The relation between ludology (game theory) and narralogy was also covered.
My question to Craig was the typical "How does this apply?" I asked if this field of research could apply outside the game industry similar to how game theory has applied to a wide variety of fields such as economics and evolutionary biology. He gave the standard response of artists and abstract mathematicians: that application doesn't matter. Doing it for the game, for the art, was important enough in its own right. Applications would fall out naturally (but its counterproductive to let concern of potential application guide, or ruin, the research). However, I think some of his larger-scale structuring of game (and film) narrative applies to our documentary narrative as well. The brief introductions and conclusions are important to us as well. Engaging the user via a sense of conflict approaching resolution is also common in human-authored documentary. Can this be generated?
Having wasted much of my youth playing Dungeons and Dragons, I rather enjoyed this talk. Ian is a world-class, long-standing gamesmaster. That is, he orchestrates role-playing games in which up to six players control their own characters interacting with a fantasy world he created. This is all done with paper and character models. Although no computers are used, the art of making a role-playing world applies quite a bit to worlds made for computer games. These aspects in common include the role of non-player characters and the progressive developement of the player characters.
I asked about the issues behind making a world with a multi-linear narrative. A role-playing world needs to be complex for engaging emergent narrative, but more so than with typical computer games, there is an underlying story narrative made by the gamesmaster that the characters are expected to, more or less, progress along while the emergent narrative fills in details. How many different paths and different possibilities can a gamesmaster allow? What happens with the paths not taken? Are they wasted?
He said his worlds tend to be details and strongly-linear, and he often has to nudge the characters along so that they stay close enough to the intended track. Some gamesmasters, however, make more loosely structured worlds with more paths. The paths followed thus have far fewer prepared details, and so these gamesmasters have to do a lot of improvisation. This dialog confirmed the idea that human-authored narrative is highly static and that making multi-linear human-authored narrative is very much work. In the generated and (hopefully) emergent narrative from Topia/Cuypers, the structure would have to be quite simple.
Presented here was a social simulator training children to handle bullies. While adults have traditionally considered bullied an essential character-building component of growing up, the current generation of educators has finally started respecting children's complaints and treating bullying as a serious and detrimental problem. This system helps train children in the intricate interaction of disarming bullies at the social level. It presents scenarios of children being bullied and allows the user to choose different responses and see how the situation changes with those responses. The narrative that emerges is helpful, because a primary characteristic of narrative is that it represents a progressive social event in which perspectives and desires of characters, and the resulting actions, work of each other and progressively change as a result.
This system also helps because of the social environment it sets up for the user. The user is interacting with a computer rather than an adult or peer, and this frees the child up to respond without fear of judgement. An appropriate balance between empathy and distance is also set up so that the child user can understand the social phenomenon well enough without feeling threatened. For us, this work is not all that instructive because it is emotion-focussed and because we do not engage the user as a character in the narrative.
Senga is a retired school teacher and a member of the local storytelling community. She has used storytelling quite a bit in her educating of young children. One of her interests is in the use of storytelling in local language to give the children an immediate identification with the story and to help build their appreciation of their own culture. Coming from the region of Fife, she took the English filter off her Fife speech for a few snippets of stories. She also gave a riddle in Scots, which was only slightly comprehendible as English (and perhaps only slightly less so as Dutch).
This was a very cozy and "thematically coherent" social event. We had dinner in a tiny basement bookstore/cafe with a focus on storytelling as a theme. After dinner, we were joined by Stuart McHardy, a singer, storyteller, scholar and Edinburgh phenomenon. He specializes in folktales with the most important aspects of truth: that they give life and body to ideas important to the culture around the tellers and listeners, and thus help hold the culture together across many generations. Thus, he would end each folktale, however implausible, with " ... and that's the absolute truth!" To appreciate how this applies to Topia and Cuypers, have a few beers.
Justine presented several tools from the MIT Media Lab for teaching narrative skills to young children. Primary among these was Sam, the virtual storytelling peer, which provided an avatar for sharing storytelling with young children. An immersive environment with motion detection and a wall-sized screen helped children to treat Sam the avatar as a peer. With some typically simple AI trickery, Sam could detect narrative transitions of the children's story telling to interact in a way to encourage improvement. Sam also told stories too. Sam's storytelling provided an example of good narrative. Sam's stories also involved the stories the children made, or as much of them as could be detected.
I asked about Sam's "emergent AI" that came from the trickery involved. Empty pauses in the child's speech trigger prompts. Identity of dolls handled and their positions in the toy castle gave Sam enough clues the the child's story that Sam could play off them. Indeed, there is little understanding on the part of Sam in what the child is doing, even in the AI sense. I also asked about the new category of being that Sam provided children. The phenomenon of children easily making a separate category for computers beyond humans and completely asocial machines was a repeated theme in the conference. The usefulness of this new sort of social creature for children was also discussed in this and other presentations.
Now I wonder if there is a similar new sort of being adults would accept from computers that automatically generate multimedia presentations of a more narratively complex nature than they currently do, such as what we hope for Topia/Cuypers? We want to avoid the typical computer oddities and incoherences as much as possible, but maybe there are ways to make the inevitable generated oddities more tolerable.
We good good feedback from the audience on this presentation, and got some interesting questions. Many were "wishlist" questions, wondering why we hadn't included certain presentational features that were simply out-of-scope for our research. One person asked about detecting sequence for presentation, which I had good answers for from our year 2000 pre-Cuypers publications.
Many questions were user-oriented, which is understandable given this audience. Clark asked: "What if the user changes his mind a little bit into the presentation?" Simple: go back to the start and make the appropriate changes to the request. Probably not the desired answer, but smoothing the user's interface for dynamism seems to me out of our scope, but not so much so as other wishlist questions. What about setting up more complex user goals? I mentioned the SRM-IMMPS work, but stated that while this function has been modeled it has never been demonstrated to significant complexity. There were user-oriented adaptation questions, such as modeling the user based on stored models, past interaction and clicks in the current presentation. I told him we may do some work on that in the not-too-far future. My favorite user question was about the requirements on meta-data for enabling adaption to the user. The questioner and I discussed the cycle of making a system that automatically generates presentations from meta-data, and only after this fully realizing what the meta-data needs are. Then you have to repeat part of the rather lengthy meta-data collection process.
Clark started a discussion about choosing different curator types as styles for the presentation on your selected topic. We made comparisons between what curators do in setting up an exhibition and what Cuypers could do in generating a presentation. We all discussed different idealogical goals between different curators, which lead to a comparison with Terminal Time. Can we encode and process the semanticsinto the rhetoric and narrative necessary for idealogical spins on presentating Rijksmuseum data?
Old-timers from CWI-INS2 are familiar with the work of Martin's lab DFKI and the work if its head Elisabeth André. This presentation was more of the same: user-disarming avatars having automatically generated conversations with other avatars. This presentation's novelty is that the avatars appear on different screens, providing "crosstalk" with the same sort of cozy feel of TV news anchors chatting with their remote correspondents. Furthermore, this system has user feedback buttons: "applause" and "boo!". These affect the conversation and the avatars' moods. Narrative is generated implicitly through the progression of the conversation. The demo videos were in German, of which I understand nothing, so it was hard to judge the quality of the conversation and the narrative underneath it.
In any event, there's not much here for us, unless we want a friendly-looking and rather chatty animated avatar narrator for our generated Rijksmuseum presentations. Perhaps the facial animators at CWI can draw up a Henk van Os avatar to replace Lynda's and Joost's voice-overs? Or perhaps Henk crosstalking with a swarmy London punk weekender tourist museum visitor avatar?.
Narrative writing tools for children may be the most common subject in this conference. This paper was most representative of this topic. HARRY is ELIZA combined with MS Word grammar check and applied to child's narrative. Children type stories and HARRY uses Eliza-like simple string detecting to provide narrative tips in an MS Work grammar check-like dialog box.
This provides insight to our work by being glue-text detection and suggestion rather than generation. Systems like this could provide narrative devices with associated text scripts that could be incorporated into Topia/Cuypers.
The interesting component of this talk was the developing Graphic Story Writer (GSW) tool, which provides children a graphic interface for manipulating icons for characters and story object through locales to help develop narrative in stories. This concept of a "narrative construction set" is interesting for us because it provides a set of narrative constructs that we can base ours on. However, the focus of this talk was how the GSW is being changed to 3D VR instead of 2D graphics. I skeptically asked if this wasn't adding meaningless sugarcoating (though I didn't put it quite like that) at the expense of moving away from the more interesting and important narrative structure itself. His answer, of course, is that VR does provide useful immersion
This paper discussed the generation of good narrative in the interaction given by presentation avatars. He presented some semiotic structures that he has used or was influence by, the most interesting of which was the semiotic structure of Russian fairy tales. He also analyzed some related work. Much of this was too story-focused to related to what we do.
PUPPET is another environment for helping children write stories, though an interesting and visually appealing one. It features a few characters and a simple world with enough foundations for a substantial set of simple stories. More meaningful was the child-friendly interface, and the video of children having interactions with the tool that were clearly meaningful to them. The children run the characters through a simple VR world much like a computer game. However, they also record the occurrences as well as their voice-overs, as narrator and as each characters. The children can also re-do their voice-overs later at will. Not much for us, but interesting in its own right.
While we were a close runner-up, this presentation definitely wins Geek of the Conference (with no offense to Chrystopher, who is quite a nice guy). Chrystopher presented the abstract mathematician's view of narrative, with characters as automata responding to stimuli. He presented 19 very LaTeXy slides of tables and formula in 30 minutes using a heavily bordered early 90s dvi presenter. Not my first choice for the 5:30pm slot on the last full day of a conference, which I'll leave as my excuse for not having more to say about it. Except for this: since we don't model characters, this talk doesn't apply much to our work.
Here are two more storytellers: Beth, from the USA, and Judy, from Edinburgh. Judy is also a computer science researcher at the university. They presented yet another tool for teaching narrative to children, called StoryStation. Their focus was not on the tool itself, however, it was on the childrens' attitude toward the tool and their teacher. This is an important topic for e-learning because tools aren't replacing teachers, they are a new categories of tools with a new role that complements the teacher. Among the several interesting insights is a reaffirmation that children appreciate interacting with a semi-social entity that does not judge them they way human peers and human teachers do.
This paper discusses how autistic children do, and don't, understand narrative in stories. They understand sequences of events. However, they do not readily understand character motivations, perspectives of different characters and what the emotion impacts of story events are on the characters. One could make the same observation about automatic presentation generators like Cuypers. By being able to introduce narrative, even of a factual documentary-style, into a presentation, it becomes more engaging. These aspects that autistic children don't readily understand are also complex to representa and process with a computer.
This is yet another yet another environment for teaching narrative. This being the last talk, I'm unfairly unable to tell the difference anymore!
This workshop is groming some of its presentations for becoming articles for a special issues of an e-learning jounral IJCEELL. Our work is too young for this publication, but I got a review for our paper anyway. It found the paper original and useful, but young and not geared for the e-learning context. I was given a printout of the review if anyone would like to read it.
Paul, the workshop's organizer, raised the topic of what to do with this workshop in the future. People suggested it could happen every year instead of every two years, but no one volunteered to hold it next year, so it looks like the next one will be held by Paul in two years. We also discussed having the also have the workshop as a single day workshop with a larger conference. Several people made suggestions. I suggested Hypertext 03, which I think would be a good fit. It is also a timely fit, since it happens in one year, thus halfway to the next Edinburgh half-week NILE. Finally, Paul proposed that NILE start an European Union-funded Network of Excellence, which he has been following up on in the days after the conference. This is worth tracking. Would we like to be a member if it happens?
Several of the tools presented gave tips, typically to children, on how to put in good what we would call "glue text". They detected areas where it was needed, and prompted for its insertion. The equivalent functionality on Topia/Cuypers would be to detect where certain narrative devices are needed in the progression of the presentation, and then to insert automatically the appropriate glue text. Perhaps an analysis of these systems would provide us with the devices and text needed.
The narrative device of introduction is one we've discussed already in Topia. When you go from one topic to another, you should introduce that new topic as part of a transition between the two topics so that the user understands a transitions is occurring and is ready for the new topic. In some narrative authoring tools presented here, I've seen similar interactions that encourage children to properly introduce new characters and occurrences in their narratives, something which children are not good at doing. Other suggestions that these systems give may also be similarly appropriate for Topia/Cuypers.
One narrative structure the conference enhance my understanding of is conflict and resolution. When you introduced conflict in the narrative and approach its resolution, this engages the user more. This is also done in the type of factual documentaries we wish to emulate. In museum presentations, you don't just present paintings of different genres, you discuss the forces at work pushing artists to develop from one genre to the next. This is often put in the form of a conflict that is resolved by developing the technique characteristic of a given genre. The conflict itself is demonstrating how it was lacking in the previous genre. Is there a way to automatically inference appropriate conflicts and resolutions from our semantic network?