A task-oriented dialog system is modeled as an active dialog partner participating in a mixed-initiative dialog. If a dialog system is designed to provide the user with the capability which enables the user to communicate without much effort, then the resultant dialog is said to have conversational efficiency. However, previous researches on dialog systems didn't have much consideration on the efficiency of conversation, which is one of the important factors that greatly influence human dialogs.
In this thesis, we present strategies to make a dialog system achieve conversational efficiency by allowing a dialog to have coherence and informativity. First, coherence provides the capability of changing the system's execution flow according to the current dialog flow. We achieved this goal through finding out the focus of a dialog and maintaining it. Second, the informativity of a dialog system means the capability of informing the user not too much information while the given information should easily be linked to the user's existing knowledge. We achieved informativity by considering the user's level of expertise in the domain, i.e., adapting a user model to system utterances.
The coherence and informativity of a conversation can increase the efficiency of a dialog as follows. First, since the user is less likely to start up a clarification subdialog, the total number of interactions can be reduced. Second, linguistic devices such as ellpsis and pro-forms can be used since the contents of consecutive utterances are interrelated.
In this thesis, we also proposed the evaluation measure to evaluate the efficiency of a dialog, and evaluated our model by this measure in the hotel reservation domain. Resultant dialogs showed that the number of interactions can be reduced, and it was possible to apply ellipsis and pro-forms to them.