If you've read a choose-your-own-adventure book or played a Twine game, then you've read interactive fiction. It's like any other story, but you get to choose how it unfolds by making different choices at certain plot points. In the experiments conducted by Georgia Institute of Technology, Scheherazade-IF read a couple hundred stories written by humans about either going to the movies or robbing a bank, and used them to create interactive stories of its own.
Interestingly, even when asked to write simple stories, the AI came up with unexpected results.
"When we first tried the date at the movie theatre, we expected to see things about buying concessions, finding seats, etc," said Mark Riedl, who co-authored a paper on Scheherazade-IF's AI (he also worked on that AI that generates Super Mario levels). "The system learned that people should hold hands and kiss. This was great, because it is the sort of thing that we geeky AI researchers might have overlooked if we had to program the 'script' for going to a movie theater."
"Scheherazade-IF treats a crowd of people as a massively distributed knowledge base from which to digest new information"
Scheherazade-IF doesn't understand what the stories it's reading are about but it recognizes patterns. When it sees a bunch of sentences from different people that seem to refer to the same event—buying popcorn‚ for example—it recognizes that there must be something important going on. It can also learn that some events tend to occur before other events, resulting in the kind of story map you can see on page three of the research paper, "Crowdsourcing Open Interactive Narrative."
To test their AI's writing skills, researchers compared Scheherazade-IF to a version of the system that produced random stories, and one that had a "perfect" script programmed by a single human storyteller. They then asked people to play the different versions and report how often the story didn't make sense. In some cases, Scheherazade-IF performed just as well as the human-programmed version, and in other cases it did about 83 percent as well. Some scenarios were easier to learn than others but it always did much better than the randomly created stories.
For the movie date scenario, for example, players on average found three errors in the human-created interactive story, five in Scheherazade-IF's interactive story, and 15 in the randomly created interactive story.
For now, the researches are paying people to write the stories Scheherazade-IF is learning from. The stories can be written in natural language, but with some limitations. Each event must be one sentence, for example, and compound sentences—two independent clauses joined by, say, an "and" or "but"—aren't allowed.
"Humans are pretty good storytellers and possess a lot of real-world knowledge," Riedl said. "Scheherazade-IF treats a crowd of people as a massively distributed knowledge base from which to digest new information."
At the moment, paying people to write stories in a way that Scheherazade-IF can understand is cheap and effective, but in the future the AI could theoretically develop to the point where it can read a bunch of Star Wars fan fiction to generate a Star Wars interactive story.
Frontpage Image: Rob Nguyen/Flickr