Virtual conversational agents, by and large, are notoriously ineffective at human-like conversation (albeit fun for humans to interact with). At worst, these agents’ responses come across as gibberish, and at best, they feel unnatural to interact with. Previously, the “best practices” for the construction of these agents has not centered around human-like speech behaviors, but rather around the appearance of human-like behavioral capabilities, the byproduct of machine-learning methods which essentially “train” the agent to mimic patterns which emerge from corpora of conversations.
The work I did at the USC Institute for Creative Technologies (ICT) this past summer in programming a virtual conversational agent centered around human-like conversational behaviors is a jumping off point for creating a virtual agent which portrays human-like conversational abilities for casual conversation. At ICT, we used an alternate method of constructing these virtual agents: we programmed dialogue policies modeled after the conversational behaviors of humans. The virtual agent I would like to create will have: a memory of what it has said and of what the user has said within a conversation; the functionality to recognize what the user has said; a method of using what the user has said in an effective manner; human-like speech utterances and conversational behaviors; and a way to remember this information for use in future conversations. It is my goal this year to create either a method of authoring such an agent, or to construct such an agent myself.
