An Improvisational Approach to Acquire Social Interactions

To build agents that can engage users in more open-ended social contexts, research has increasingly been focused on data-driven approaches to reduce the requirement of extensive, hand-authored behavioral content creation. However, one fundamental challenge of data-driven approaches is acquiring the interaction data with sufficient variety that reflects the characteristics of open-ended social interactions. Previous work attempts to acquire social interaction data either from face-to-face interactions or human-agent interactions using a simulated environment. In this work, Active Analysis (AA), a theater rehearsal technique, was applied to collect diverse social strategies and interactions. In particular, this work integrated AA into a web-based crowdsourcing task that requires two crowd workers to conduct a bilateral multi-level multi-issue negotiation. Findings from a between-subject experiment with 200 crowd workers recruited from Amazon Mechanical Turk demonstrated that AA could facilitate the creativity of crowd workers and thus lead to social interaction data with greater variety. In addition, AA provides a means to control the diversity so that the coverage of the collected data is consistent with the goals of the application. The results presented in the paper lay a good foundation for future work on data-driven approaches to build socially interactive agents.