Peters

When to send system-mediated interruptions within collaborative multi-human-machine environments has been widely debated in the development of interruption management systems. Unfortunately, these studies do not address when to send interruptions in multi-user, multitasking scenarios or predictors of interruptibility within communication tasks. This paper addresses the issue of predicting interruptibility within these interactions with special attention to which users are engaged in which tasks or task engagement and where users are within a current task or task structure as predictors of interruptibility. Using natural human speech from these interactions, we attempt to model task engagement and task structure to predict candidate points of interruptions. The motivation for these models and their performance in a multi-user, multitasking environment are discussed as proposals in developing communication interruption management systems. To model task structure, a task breakpoint model is proposed which performs with a 90% accuracy within a multi-user, multitasking dataset. Integrating this task breakpoint model into a real-time interaction results in an average accuracy of 93% using the proposed task breakpoint model and a rule-based model. To determine the current task in which users are engaged or task engagement, a proposed task topic model performs with an accuracy between 76-88% depending on the topic within the dataset.