Abstract by Nancy Fulda
EVE: Emotive Adversarial Ensembles for Mixed Initiative Dialog
Machine learning researchers often represent dialog as a strict mapping from input text to output text, but psychological and cognitive science literature shows that humans consistently model each other during communicative processes. Building upon a dialogic view of conversation and working in conjunction with Amazon's Alexa Prize support team, we present EVE, a conversational AI architecture that treats communication as a cooperative effort between unique individuals. Rather than attempting to imitate a human, EVE embraces its identity as a synthetic being, actively seeking information to expand its world-view while still accessing live news streams, fact repositories, and other online APIs in response to user queries. By modeling the expectations of its conversation parter, EVE is able to respond to user utterances in meaningful ways. The completed system will compete against other university teams for a one million dollar grand prize.