[Crowd Leader: Ville Miettinen] Can the Crowd Help Make Virtual Personalities Cool Again?
Once upon a time, interacting with artificial intelligence was pure science fiction: something only handsome crime fighters and fictional spacemen got to do.
Today, we interact with virtual actors so often that we barely notice them. But while these virtual personas are seemingly everywhere, they are not very lifelike and hardly ever as cool as their sci-fi ancestors (for the record I preferred HAL to KITT). There is always a point when you realize your new paperclip friend isn’t really listening to you at all, and the relationship goes downhill. Luckily for those of us who spend too much time in front of a computer, crowdsourcing may soon be able to make these virtual actors much more human.
Over at the University of Florida, Brent Rossin and Benjamin Lok have developed the Virtual People Factory. Before you get your hopes up, it’s nowhere near as sick or menacing as it sounds. The virtual people this factory produces are not psychopathic androids hoping to replace you and your friends. They are (hopefully) friendly virtual patients designed to help train students in the medical, psychological and military professions.
While virtual training aids like these are not exactly a new idea, the way the Virtual People Factory creates them is. Using a mixture of crowdsourcing and play-testing, the virtual people are created in a process its inventors succinctly call Human Centered Distributed Conversational Modeling. Open for academics to create their own custom virtual person, the factory works by first having an expert establish a basic set of questions that the virtual person might be asked along with the answers they should give. This set of questions and appropriate answers is called the “corpus”.
The virtual person is then distributed to a crowd of students who ask it questions as if it were a patient. When the provisional corpus fails to give an appropriate answer, or the student asks it a question that it doesn’t have a reply for, the fault is logged. The questions the students asked along with the faults are then sent back to the expert who accounts for the missing information. It is sent back out to the crowd of students for more testing. This way the corpus grows with each cycle and becomes ever more refined. After sufficient cycles the Virtual Person should be able to answer almost any (relevant) question a student might ask no matter how it’s phrased.
Crowds with benefits The benefits of this model are substantial. Compared to a system that uses just one or two experts to create all of the various questions and responses, the crowdsourced virtual people only require 15 hours of expert time. Non-crowdsourced versions can take 200 hours and more. Also, the corpus generated for the crowdsourced versions are much larger.
The applications could go far beyond training for medical and military students. Job seekers could practice interviews with crowdsourced interviewers (or sensitive HR staff could practice firing crowdsourced employees). Awkward teenagers could practice asking out virtual girlfriends and then practice breaking up with them when things got too serious (although I think Sierra might have covered this in the Leisure Suit Larry series).
While it’s clearly a system with huge scope for application, it seems to me they’re missing out on some of crowdsourcing’s potential. Referring all of the work back to one or two experts is an obvious limitation of the model. Perhaps weighting student inputs according to their experience could defer some of this work. Using machine learning to reference the student responses against a database of relevant knowledge could also take some of the workload off the experts.
Maybe they could even team up with a game studio to produce characters that can respond to any kind of question. The crowd could be used to generate all of the questions and answers with a minimum of oversight. With a broader project like that insights might be gained on how to improve the Virtual People Factory’s core product. And if not, it might at least give gamers a few more hours of gaming before the virtual actors stop listening.