May 26, 2004
Continuing the theme of AI systems that use language: here’s a new paper by Rob Zubek at Northwestern, who has been thinking hard about how to make robust, richly interactive conversational characters. His PhD research is focused on building an architecture for structuring conversations as vast collections of reactions to player input, arranged in hierarchies, that compete to understand and respond to the player. Multiple possible threads of conversation all are listening simultaneously to what the player says at any time, and they each update their local probabilities of where they believe they are in the conversation. Assuming enough content is authored, this allows the conversation to have a variety of believable responses at any time, at varying levels of coherence. Thus the system can fail gracefully and perhaps move the conversation forward when the system has trouble understanding the player, or doesn’t have a good response.
Within this architecture, Rob is building The Breakup Conversation, in which the player is given the goal to successfully dump their significant other, played by the system. Sample dialog of The Breakup Conversation is included in the paper.
An interesting feature of such an approach is that “where you are” in the overall conversation can’t be pinpointed to single place; instead the conversational state at any one moment is the collection of potential directions the conversation can move in next.
Breaking up may be hard to do, but at least it’s becoming computationally inexpensive.