Learning the Structure of Task-Driven Human-Human Dialogs