New Framework Detects Team Communication Phases in VR

This paper presents a novel computational framework designed to analyze team communication dynamics within collaborative virtual reality (VR) settings. Traditional methods often oversimplify communication analysis by summarizing entire interactions or using fixed time windows, which can obscure critical shifts in team behavior. The new framework addresses this by employing late chunking and penalized Gaussian-kernel change-point detection on timestamped dialogue to pinpoint semantic transitions.Once these communication phases are identified, the system uses techniques like TF-IDF and non-negative matrix factorization to provide structured evidence for interpreting each phase. A local large language model assists in generating initial interpretations, which are then human-reviewed. The framework also aligns these detected phases with independently recorded interaction logs to understand corresponding task-action patterns, demonstrating its ability to reveal coherent and interpretable team structures while maintaining traceability to the original transcript data.