Proppian Projects
Working Notes, by Zach Tomaszewski
for ICS 699, Fall 2005, directed by Dr. Kim Binsted
Amazon lists 364 books citing Propp's Morphology of the Folktale. Most deal with studies of folklore, myth, film, and literary theory. Google Scholar also results in over 300 hits. Here are a few sample projects from the past few years.
Matea, Michael, and Phoebe Sengers. "Narrative Intelligence." 1998. http://liquidnarrative.csc.ncsu.edu:16080/classes/csc582/papers/mateasandsengers.pdf
Prada, Rui, Isabel Machado, & Ana Paiva. "TEATRIX: Virtual Environment for Story Creation." 2000. http://gaips.inesc.pt/gaips/shared/docs/Prada00Teatrix.pdf
Prada, Rui, Isabel Machado, & Ana Paiva. "Bringing Drama into a Virtual Stage." 2000.
http://gaips.inesc.pt/gaips/shared/docs/Machado00BringingDrama.pdf
- Teatrix: A learning environment to help students tell stories collaboratively, using autonomous agents following the dictates of their role and user-controlled agents. Based on a drama model, and an aim of merging acting, reading, and writing into one supported environment/process. Based on watching children acting in O Nosso Sonho school.
- Step One: Setup. Choose characters. 6 possible roles: hero, villain, helper, magician, beloved one, relatives. Setup scenes and items as well.
- Step Two: Story Creation. Each user sees world from placement of their characters. Those characters without a user follow rules of their role--for instance, villain's primary goal is to harm the hero. Users interact according to their role, w/ story emerging from interaction, often with a director in charge to produce coherent performance structure.
- Step Three: Story writing. Playback story, rewind, write about it or critique it.
- 2 parts of the system: World (modeling, ontology, animation, scene regions, change, etc) and Agents.
- Agents include:
- Mind -- behavior: knowledge of the world, goal, emotional state (preferences for actions), possible actions and probably consequences. When played by a character, mind is passive and character will not act, but will still respond and maintain goals and emotional state. Also filters sensory info based on importance/relevance.
- Body -- representation in the world. Physical characteristics, effects of emotional state.
- Effectors -- actions. 3 stages: meet preconditions and verify world state (character thinks its possible based on world view, but is it really?), execution, finalization (updating world).
- Sensors -- aquiring info about world state.
- Inventory -- item possession. One in hand, rest in pack.
Body and inventory optional.
- Director agent. God-like figure who determines when story is over; can be controlled by user. Sensors inall world locales, can insert new items and characters, and knows every action. Can send messages to all characters, and even control them.
- Server for initial creator, which also contains director, system clock, and resovles conflicts (two agents grab same object, etc). Client modules have complete info about the world. A local agent, and only clones of non-local agents (no minds, repeating actions sent by cloned source), based on coupling model of Java Match Maker library.
- System in use in schools, but has not yet been evaluated.
- Also mention of SAGA-Support and Guidance Architecture for guiding virtual world interaction and collaboration.
Callaway, Charles B. and James C. Lester. "Narrative Prose Generation." 2001?
http://research.csc.ncsu.edu/intellimedia/papers/npg-ijcai01.pdf
- Gaps exist between story generation (top down to individual actions, adding more detail) and nat lang processing (usually single sentences, only starting to do multi-paragraphs).
- Older story generation projects: TALE-SPIN (Meehan 1977), UNIVERSE (Lebowitz 1985). Recent story generation systems--MINSTREL (Turner 1994) and JOSEPH (Lang 1997) focus on character and plot. Missing decent linguistic structure.
- Most NLP focuses on explanation generation of scientific or instructional texts; not narrative generation.
- Developed AUTHOR, part of STORYBOOK--an end-to-end narrative prose generation system utilizing narrative planning, sentence planning, discourse history, lexical choice, revision, full-scale lexicon, and FUF/SURGE surface realizer to produce multi-page stories in the Little Red Riding Hood fairy tale domain. [Wow, tight domain!]
- Narrative domain includes certain features: dialog, personal pronouns and motivations, use of past, future, and present tenses, diversity of narrative writing styles, formatting (breaks, emphasis, etc). Each of these have sub-issues.
- Story planning mechanisms and story grammers; verses Proppian fabula (sum knowledge and facts about narrative world) and suzget (order of specifics about what an author presents in what order). AUTHOR uses latter.
- Some elements need to be left vague (too obvious; or as in mysteries)
- Details of paragraph and sentence structuring, avoiding repetitive sentences, etc.
- 74 fabula operators; 253 narrative stream primitives => inputs to the pipeline.
- Tested by turning off certain features of the pipeline (or substituting a finite state automaton) to generate 1 story each for 10 setups. Then had humans rate them in terms of style, grammer, flow, readability, logic, detail, believability, etc. Rated best with all, but lexical choice feature not that important (especially when subjects assumed the story was meant for children).
Theune, Mariet, et al. (Faas, Nijholt, Heylen) "The Virtual Storyteller: Story Creation by Intelligent Agents." 2003. http://wwwhome.cs.utwente.nl/~anijholt/artikelen/tidse2003.pdf
- Using agents at different levels: plot, narrative, presentation. Not purely scripted nor purely character-driven.
- Virtual Storyteller, from University of Twente in Netherlands. Attemping to create a VR storyteller--embodied, talking agent with gestures, etc.
- Autonomous characters, controlled -> plot, NLP -> narrative discourse, presented -> VR storyteller. And interactivity.
- 3 story levels (corresponding to fabula, story, text):
- Plot - events that occur in the fictional world; story content, casually related.
- Narrative - representation from a POV; reordering or omission of events
- Presentation - realization of story in a particular medium
These are usually done together, but could have one plot giving many narratives, and each narrative having many presentations.
- Plot requirements: consistent (events are naturally ordered, in accordance with world, and character's actions match their personality/motivation/previous actions); well-structured (beginning, introduction of problem; middle, climax; ending, success or failure). Necessary, but not sufficent--need conflict, suspense, interesting themes and characters, and good/exciting presentation.
- Existing approaches at plot generation
- Character-based generation. Bottom-up approach based on intelligent agents choosing actions based on internal state and perceptions of the virtual world. Basis for Talespin text story generation system. Used currently by Stern et al. (Catz and Dogz). Tends for consistent plots--matching characters' motivations. But often lack story structure, especially with primarily reactive characters. Better with strongly goal-driven characters, but not guaranteed--story cut short by death or easy success.
- Scripted plots. Characters have no autonomy, but driven by either authored or generated plot script. Generation usually relies on Propp. [See R. Lang "Declarative model for simple narratives" 1999, L Pemberton "Modular approach to story generation." 1989.] Often hollow characters, simply filling in functions in the story grammer. (Exception: Universe system where plot based on detailed character info. Lebowitz, "Planning Stories." 1987.) Limited interaction/replay value.
- Intermediate approaches. A script with room for improv and expansion.
- Script as part of characters' knowledge [Cavazza, Charles, Mead "Agents' interaction in virtual storytelling. Intelligent Virtual Agents 2001.] (prespecifying all script branches); Hayes-Roth, Van Gent, Huber, "Acting in character", 1997. (ensuring essential plot points)].
- Use of virtual director [Kelso, Weyhrauch, Bates. "Dramatic presence" 1993. (ensuring essential plot points); Young, "Notes on the use of plan structures in creation of interaction plot" 1999.].
- Just inserting small narrative fragments [Stern, "Virtual Babyz: Believable agents with narrative intelligence." 1999.].
- Leads to problems, namely character insconsistencies, etc [Mateas and Stern, "Toward integrating plot and characters for interactive drama" 2000.]
But forcing characters to adhere to or meet some existing plot. Instead, perhaps generate plot, guided by director, with characters. [Check out AAAI Fall Symposium]
- Storyteller architecture: characters, acting based on goals and world state. Check in with narrator, who has knowledge of story structure (and world state). Also a narrator and presenter. Built using Java Agent Development Environment (core agents) and Java Expert System Shell (for creating rule base, making rational agents) and Protege (free, for specifying ontology and knowledge-base, to design a story world). Presenter is an MSAgent coupled to the system.
- Director does not start with a plot, but generates it along the way. Controls the characters by Environmental (introduce new char/objects), Motivational (give a character a goal), Proscriptive (disallowing action) means. [Based on Blumberg and Gaylean "Multi-level control for animated autonomous agents..." Creating Personalities for Synthetic Actors 1997.]
- So director sets up environments and characters, then proscribes certain actions. (But doesn't prescribe, or force them.) So prevents an early killing, or creates a character/event to save the victime, etc. [Sounds like a lot of blocking to me.]
- Currently, characters quite simple. Also, director very loose with few restrictions. Have not yet implemented a Proppian story knowledge.
- Director could include a user-model to prefer one possible event over another. (N. Szilas. "A new approach to interactive drama", 2001.) Also experience with different character roles and plot structures, which architecture should allow.
- Narrator agent then takes story and turns the system events into natural language (ala Calloway and Lester's critiques). Currently only narrates as events happens; hopefully later will be able to narrate after story, at varying levels of detail, etc.
- Presenter then tells story. Hope to expand this later; also, explore virtual drama where characters have bodies and the narrator requires more theatre-oriented knowledge.
- Interactivity: 1) Let user choose characters and motivations. 2) Let the user take on some of the directing. 3) Let the user control a character ala RPG. In this case, how would the director maintain control?
- Related work that doesn't use pre-scripted plots: Szilas, where virtual narrator chooses the actions based on criteria, but actions come from story grammer nor from characters. Teatrix, though there the narrator aids children's creation process rather than creating the story itself. M. Lee (A Model of Story Generation, CS, UManchester Master's thesis, 1994.) text-based generation where characters plan within limits set by story grammer. No director, but does introduce new characters when plot gets stuck; top-down story creation from grammer rather than characters.
- Concl: semi-autonomous character-driven plots w/o inconsistencies due to pre-defined script. Yet director still provides for plot structuring.
- [Me: remember Meadows and narrator's POV.]
Braun, Norbert. "Automated Narration - The path to interactive storytelling."
Braun, Norbert. "Storytelling & Conversation to Improve the Fun Factor in Software Applications" Funology, From Usability to Enjoyment, Chapter 19. 2003.
- Aiming at top-down conversational user interfaces.
- Bates; Mateas & Stern are working bottom-up: agent-based narrative generation with believable agents. Agents don't know what's thrilling, nor how to incorporate user into their stories.
- Influening a story has broad meaning: generation of a story (Bringsjord & Ferrucci), guiding of what should happen (Bates, Strickland, Laurel, etc), to complete definition of all possiblities (ie, Joyce's Afternoon).
- By building top-down, we admit a story goal, and thus an end. Use Propp, select from possible stories, filling in scenes based on context constraints. Also possible to change underlying morphology.
- Laurel: Stories provide context.
- Story management vs. discourse management.
- Geist and IzA (kiosk) projects
- Non-Linear Storytelling API
- Coupled with conversational modelling
- Non-linear storytelling: telling a story with audience-impact on the storyline, but not the story goal (supposed end of story).
Grasbon, Dieter, and Norbert Braun. "A Morphological Approach to Interactive Storytelling."
Digital Storytelling Department, Computer Graphics Center, Darmstadt, Germany, 2001.
<http://netzspannung.org/version1/extensions/ cast01-proceedings/pdf/by_name/Grasbon.pdf
>