Transition networks model dialog as a set of transitions (most simply achieved by speech acts of specified kinds) moving between discourse states. A straightforward transition network for paying a bill via a dialog is the following (taken from McTear's overview paper):
The drawbacks with this approach are now well-known and define the current main research directions in the field—a different transition network is necessary for every different application and there is no reuse of the description. It also does not describe what happens when misunderstandings occur so that the simple transition between states in the network is compromised. Flexibility has to be built in, and the main question is: how?
One move is to use the transition networks not for application-specific dialog paths but for abstract descriptions of the nature of the conversations themselves. This is done in the Conversational Roles model of Stein/Sitter which has transition networks for successfully asserting, questioning, clarifying, and so on; the original impetus for Stein and Sitter's model was the Action Model of Winograd and Flores (1986). An example of a Conversational Roles (COR) transition network is the following, the top-level schema for 'dialog'. This states that if there is to be a dialog between interlocutors A and B, they start in state 1 and have the specified possibiliites for navigating a path to any of the final states 7, 5, 6, 7' and 8 as shown around the outside of the diagram.
These transition networks are of necessity recursive, since, e.g., a misunderstanding can occur at any point, even when in the middle of dealing with a previous misunderstanding. This means that as well as basic transitions, which can be modelled or achieved by specific speech acts, there is the possibility of entering a further subnetwork for a similarly complex set of options: e.g., request (A, B) or evaluate (A,B). Almost all modern transition-based approaches are now recursive in this way and allow libraries of modular dialog behaviours to be built up. Approaches such as COR and more recent developments offer 'meta-knowledge' about how dialogs in general work rather than transition sequences for particular kinds of dialogs in particular kinds of situations, as the bill payling example above did.
