16

Automated Story Generation

In her seminal book Hamlet on the Holodeck, Janet Murray expands on the concept of ancient stories (such as the Iliad and the Odyssey) working as highly formularized communication systems. As she recounts, other examples, such as Russian fairy and folk tales, bear solid evidence that a few dozen basic plot events can generate hundreds and even thousands of stories. The storyteller of old—armed with “meta-data” about his story functions or morphemes—could easily shuffle his deck of “story cards,” redress them as necessary for the current location or audience, and produce a performance-specific, site-specific story to tell. Evaluating the audience immersion and understanding of the story (are they laughing? are they crying? etc.), the storyteller could refine and alter the telling as necessary, on the fly.

In theory, robust software AI should be able to accomplish the same thing: building stories on the fly, based on pedagogical needs, user reaction, input, etc. Audio libraries of phrases and phonemes could be swapped in to assemble fresh and original dialogue and voiceover narration (a sort of “mad libs” approach); video snippets could be shuffled around to create content; a real-time 3D engine could load up a new game level and customize the level for story needs. (Game levels are a standard videogame convention dividing segments or movements of the experience: packaging up a terrain, sets, events, and nonplayer characters (NPCs) that a user must engage with in order to advance.)

However, this would require a much finer granularization of story content than has been discussed earlier in the book: we have to go well beyond the classic ideas of 3-act structures and inciting incidents, setups and payoffs, ticking clocks, and the like.

To aid this effort, characters would need to be separated from plot events. As Janet Murray points out, oral storytellers would do just this: a clown figure, or a damsel-in-distress, could be pulled into a story when necessary, and given its basic behaviors, the stock character would then find a way into the current story actions and movements. (The storyteller would place this character into the story at this moment because of a desired emotional or story-arc objective.)

These stock players would be independent agents, entering and leaving story events as required. Naturally, backstory histories would need continual updating, and progress toward both story events and learning objectives would need monitoring and evaluation.

If this doesn’t sound hard enough, then add the crucial element of user interactivity. The more freedom and response gradation allowed the player, the harder all this assemblage of story and learning content will become. What happens when users stray from defined story paths, or test the limits of the system, or just behave in “irrational” ways? Will the automated story-generation system be intelligent enough to work around (and even try to correct) these problems—or will it have to construct increasingly artificial roadblocks, eventually undermining user confidence in the integrity of the interactivity and the “realism” of the simulation? In the worst-case scenario, will the automatic story generator end up creating incoherent storylines and irrelevant learning content?

These are only some of the questions and challenges facing researchers attempting to create automatic story generators. The advantages to creating such a generator, particularly in the context of creating fresh story-driven content for simulations, should be obvious. Simulation systems could respond to new data, news events, studies, and learning points, and immediately generate new scenarios. Users having trouble with the pedagogy of a simulation could return to the environment again, and be confronted with new storylines, rather than rehashing the same old plot turns.

The system we’ve just described would seem to be a long way off. Although story AI systems, similar to what’s been described, have been tested in highly circumscribed, “miniaturized” story worlds, no automatic story generator has been able to author a truly usable, real-world simulation or videogame.

However, smaller steps toward this goal are being taken. One such approach is the Interactive Drama Architecture (IDA) being proposed by Brian Magerko, a researcher out of the University of Michigan (and a collaborator on the Leaders project). Magerko accepts as a given that fully automated storytelling isn’t yet executable. However, it may be possible to substitute an “omniscient story director agent” mechanism for the traditional “director”; and this “director-agent” can, along with the original human storyteller, collaborate in creating an interactive story on the fly. The trick is in giving a user maximum freedom within the environment, while still respecting the construct of the story and the essential plot points and outcomes designed by the storyteller.

According to Magerko, an author begins by creating a “story space” which would include the following:

Authors need to create narratives that are topological, rather than strictly linear. Represented visually, plot points become nodes in an event topology. The plot is no longer an action-by-action line, but a skeletal framework, with as few plot constraints as possible.

NPCs within this world can become semi-autonomous, providing that they have been given specific objectives to undertake within the topological narrative. This will make NPC behavior more believable and the environment more immersive: instead of being puppets (as NPCs so often are in videogames), they become unpredictable characters with real motivations confronting obstacles to their objectives.

The omniscient director can change the objectives of the NPCs, depending on story progress and user input. Providing the director understands the state of the story world at any moment, and has a good grounding in believable objectives and transformation arcs, a truly rich, interactive, dynamic story space can theoretically be achieved—without the human storyteller stepping in to tweak scenarios and restructure the narrative.

In a sense, the omniscient director in the IDA becomes an on-scene Dungeon Master or man in the loop, affecting the pacing, the story, and the emotional experience, based on user interaction and psychology.

All this presumes that an ontology of interactions has already been developed, with an encoded syntax for interactions between all game agents (be they NPCs, environments, event triggers, or users).

While not qualifying as pure automated story generation (and not pretending to do so), this approach offers a new paradigm for immersive storytelling that uses all the classical tools of Hollywood-style narrative and still stresses the primacy of story narrative in a simulation experience. Users, however, should experience dynamic, highly responsive story worlds with the feeling that they share fully in the story creation, rather than feeling narrative and plot events imposed on them, impossible to budge.

Magerko, as part of an interactive drama team at the University of Michigan, has created a story space called Haunt, built around the Unreal game engine (see Chapter Twenty-Four for discussion of real-time 3D game engines). As of this writing, Haunt has undergone two iterations, and successfully balances autonomous NPCs, maximum user “freeplay,” and dramatic developments and turning points, albeit in a very defined and specific environment. Magerko’s work can be explored further at http://www.magerko.org.

The University of Michigan is not the only school to explore the automating of story content. Research in the arena of automated story generation is now a hot topic at different university programs, given that we now seem so close to arriving at tools that can achieve this. However, progress toward this goal is likely to be incremental, and for now, offers more hope than immediate usability.

Some will argue that “machine-driven” story-intensive simulations will necessarily be soulless and mechanical, and unlikely to ever feel immersive and real in the way that a great movie or great videogame can. But this may be like arguing that the only way to build an aesthetically beautiful and satisfyingly functional automobile is for the designer to hand-build each unit.

Even the most automated story generators will continue to require the spark of human imagination and ingenuity. And if automated story generators are truly to work, they are likely to require that human authors dig down even deeper into the source of their creativity, in order to define a story space that AI routines can shape and manage. In addition, when brought into the interactive realm, where there will be one or more human users (and perhaps an instructor-in-the-loop), automated stories (at their best) should feel absolutely unique, authentic, and original. They’ll feel human, because everything about them is human.

SUMMARY

Janet Murray’s Hamlet on the Holodeck suggests the possibility of an AI “cyberbard” automating the process of story generation within an interactive simulation, thus creating the possibility of greater replayability, greater user customization and greater user immersion. To date, progress toward achieving this has been quite modest. Various approaches, including the Interactive Drama Architecture (IDA) proposed by Brian Magerko, begin to build bridges toward this illusive goal. IDA suggests the creation of an “omniscient director” who can operate as a kind of on-site Dungeon Master for a simulation experience. Research in this arena should continue to be monitored, and incremental progress toward automating story generation is something we should expect in the decades ahead. The daunting nature of this endeavor—aimed at the illusive core of creativity—is much more difficult than increasing CPU cycles or accelerating graphics processing. No matter how much progress is made, the “human storyteller” will stay central to the conception, creation, and composition of immersive story experiences.