Development of Speech Dialogue Systems for Social AI in Cooperative Game Environments

Abstract

There is increasing interest in developing more human-like artificial intelligence (AI) capable of natural social interaction. Previous research has suggested ideas about what it means to be "life-like" AI, and some studies have attempted to test these hypotheses using game environments. In this paper, we introduce the development of the Speech Dialogue System for a "Social AI", which communicates and interacts autonomously with a human player in cooperative game environments (in this case a social survival game called "Don't Starve Together"). Based on our hypothesis that the AI should contain specific components to be perceived as more human-like, we conducted a series of pilot tests to develop the Social AI using a data-driven approach. After finishing the pilot tests, we identified six components to add or revise, based on participant interactions and feedback. These components mainly include features of the Speech Dialogue System that pertain to the interplay of AI behavior with contextual factors of the social environment ("the game state"). In future work, we intend to improve the Social AI based on these findings. The research here highlights the use of cooperative game environments for data-driven development of speech dialogue systems for artificial agents.