State transition methods have been recommended for modelling dialogues that concern well-structured tasks that can be mapped directly on to a dialogue structure. However, other significant factors to be considered include the structure of the information to be transacted and the need for verification of the user's input as determined by the system's level of recognition accuracy (esp. for systems with spoken input).
The main advantage of dialogue scripting with FST is the tight control over the dialogue progress and the determinism of the resulting system. This has advantages for dealing with the (spoken) input and output at any state. An example of an FST-based dialogue system developed is the elevator-control system at CoLi [Auer
2#2
2002]. For more examples, see [McTear1998] and [McTear2002], or the MiLCA Lecture ``Finite State Techniques for Dialogue Processing'' http://www.coli.uni-sb.de/~stwa/Milca/Sommerschule/Dialog/html/ (login ``guest'', no password).
Various toolkits and authoring environments have been produced that provide assistance with the design of FST-based dialogue models, e.g. the CSLU's RAD (Rapid Application Developer) toolkit or the recently developed CLT DiaManT toolkit.
One step beyond FST modelling is the template-filling approach. The task is conceived of as filling a set of slots in a template, where the order in which slots are filled is not fixed beforehand. This allows for some degree of user initiative, concerning the order and the amount of information (slot-fillers) they provide in a single turn.
[McTear1998] critically evaluates the applicability of state transition models for different types of spoken dialogue. In [McTear2002], this discussion is revised and more elaborate, and also extended to template-filling approaches.
Reading: [McTear1998] (extended version [McTear2002])