This talk will share two early-stage projects leveraging high-dimensional data to understand human social interactions in naturalistic settings. In the first project, we investigate the development of peer interactions through longitudinal recordings of preschoolers' free play. Using vest-mounted video cameras, we capture rich, multimodal data streams including visual, audio, and spatial information. Our analysis employs advanced speech recognition and natural language processing techniques to extract and analyze sentence frames from this high-dimensional data. Initial findings reveal distinct patterns in peer conversations compared to adult-child interactions, highlighting unique opportunities for social learning. In the second project, we explore trajectories through topic space in everyday conversation. We are currently collecting annotations of topic transitions for the CANDOR corpus, an extensive multimodal dataset totaling over 850 hours of natural language conversations, to analyze the high-dimensional structure of open-ended conversation. Both projects employ recent machine learning techniques to uncover high-dimensional patterns in social dynamics, demonstrating the potential of large-scale, multimodal datasets to advance our understanding of social intelligence.
Copyright. All Rights Reserved.