My TEAM has worked on advanced dialogs for FranceTelecom for around 10 years. Since the beginning of the development of the Artimis software
[IJCAI97], we tried to transfer the very advanced concepts from research to industry. Among these concepts, we had world modelling and rational behaviour
modelling and reasoning. Our rational behaviour theory is an agent cognitive model, using several advanced concepts such as intention, belief and action, and
expressed in a logical language. An inference engine running this theory implements the rational dialoguing agent.

Unfortunately, we failed because of the difficult appropriation by service developers of these advanced concepts. 

This led us to release a simpler design structure, a kind of agent behaviour finite state machine. This finite state machine defines the various successive states of
a dialoguing agent to dynamically build his dialog reaction. Here is an example of a sequence of state for a dialoguing agent: interpretation, contextual
interpretation and link with already detected user intentions, selection of one user intention, various reaction parts (feedback, help, correction, request
negotiation, answer including cooperative answer, etc. 

The structure of design is thus a finite state machine but compared to the classical finite state machine induced by the current standards like VXML, 1/ its level
of abstraction is higher and 2/ each state of the finite state machine does not correspond to one reaction of the dialoguing agent, but a calculation stage of this
reaction. 

Thus, it is in the way of using the concept of finite state machine that one can adjust the more or less advanced level of a dialogue. 

The software solution on which we work today, for service deployment like call routing, for the advanced service experimentations such as automatic SAV, for
researches like the optimization of dialogue strategy, implements this simple concept of finite state machine. Using traditional techniques like VXML, Java, and
JSP, this solution allows using the structure of finite state machine for simple design of dialogue as well as advanced dialogue. Very close to the proposals of the
meta-language group in the VXML Forum, this solution uses an XML description of a finite state machine which contains at the same time internal states of the
dialoguing agent and visible states at the dialogue level (those dialogue state generates dynamically VXML pages). 

 

[IJCAI 97] Sadek D., Bretier P., Panaget F., « Artimis: Natural dialogue meets rational agency », 15th International Joint Conference on Artificial Intelligence
IJCAI'97, <?xml:namespace prefix = st1 ns = "urn:schemas-microsoft-com:office:smarttags" />Nagoya, Japan, 1997, p. 1030-1035