Representing Conversations for Scalable Overhearing