Future of Games: Generating Natural Language For Game Dialogue

Marilyn Walker, Director of the University of California, Santa Cruz's Games and Playable Media program's natural language and dialogue systems division, presented a talk at the school's Future of Games symposium today that delved into how the group hopes to get closer to generative dialogue that can respond to player interaction. Bringing up a screenshot of Mass Effect 2 -- widely praised for being the most compelling example of interactive narrative in this console generation -- she said while games have evolved, "for dialogue it still looks pretty much like it did 20 years ago." While a game's story can be compelling, it may not truly reach players. "Are these dialogue choices meaningful? Is this the way we would want to play the story? Are we getting engaged and getting into the story? Is the solution to write more dialogue and having more dialogue trees?" she asked. She brought up a quote from BioWare's James Ohlen, in which he said that forthcoming MMO The Old Republic has "more content than every other BioWare game put together." That's quite a lot of writing, she noted. "What kinds of different games could we make if we could generate NPC dialogue?" With machine-generated dialogue, "we could personalize what the NPC says to this game interaction," Walker said. Through play, "we could learn what the player's personality was" and speak to it. "Human storytelling is so fundamental it's easy to draw out that human transportation," but "the games that we have now are not even near being able to mimic in any interesting way what really happens in human oral narrative." Procedural dialogue would also help designers "build more games faster with completely different tools." The goal is to "make games really have meaning, and speak to your heart. We're nowhere near games doing that because we're not drawing in what's personal for you when you're playing a game." We won't find a solution for this, she feels, "without throwing the full architecture of procedural language generation at this problem." 
Fortunately, work in the sciences has laid down a foundation that can be worked with, she said. "I've done a lot of work with the big five model of personality from social psychology,"which describes people in terms of five basic traits, Walker said. This affords a "scientific basis for character language behavior" with "50 years of studies" behind it, she said. "Even though the environment and situation has an influence... there are these very general patterns you see that seem to be dispositional and there is evidence that they have a biological basis." And when you look at film studies literature on character archetypes, Walker said, "you do not get anything like the models you would get from the psychology. You get nothing from which to build a conversational model, but if you look at what's actually authored" -- meaning the films themselves -- "you might be able to analyze and build classes of characters." The question, then, is "if by creating models of characters we can increase the creativity of authors of interactive fiction." The team is currently scraping dialogue from film scripts and developing potential character archetypes -- such as, for example, "Alfred Hitchcock leading ladies" -- which couldEve be applied in games. "We're looking and seeing if we can extract interesting parameter models," she said. Which can be used to create a caricature, "a snapshot or imprint of a character; compare against a norm, and the accentuate the things that are different."