Learning Situation Models for Understanding Activity

Human social activities follow loosely defined scripts in which individuals assume roles. Social conventions establish stereotypical skeletons for such scripts, with details and variations determined by the circumstances and personalities of the individuals who play the roles. Encoding social scripts in a formal representation would seem to hold the promise of building systems that provide services based on observation and understanding of human activity. Unfortunately, the cost and complexity of handcrafting models for the variety of ways in which scripts can be enacted makes such an approach infeasible. Clearly learning and development based on individual behaviour are necessary. In this paper, we present a method for building systems that develop social scripts through incremental supervised learning. We begin with a conceptual framework in which scripts for human activity are described as scenarios composed of actors and objects within a network of situations. We then present methods to develop situation networks by incrementally building on an existing stereotypical script. We present supervised learning techniques for learning to discriminate new situations and for learning to recognize new roles. We illustrate our approach with results from experiments with situation and role learning in an office environment.