Dialogue Management Based on Information States

The main advantages of FST dialogue modelling are also its main disadvantages. Such systems must limit dialogue flexibility and user-initiative, in order to ensure the dialogue adheres to the script. This becomes a problem for handling dialogues for more complex tasks, where the course of the dialogue is harder to predict from the task structure alone. For an efficient dialogue, the system needs to be flexible and allow the user to exert initiatiave.

The information-state update paradigm of dialogue modelling has emerged as a general framework for modelling flexible dialogue. The information state contains the information that a dialogue participant has at a given point during the dialogue (i.e., IS is a structured context representation), and every utterance in the dialogue leads to one or more information state updates. In other words, the effects of communicative acts (dialogue moves) are described in terms of information state update rules. The ISU-based approach allows dialogue modelling for various degrees of task complexity, and allows the dialogue system designers to decide about the level of dialogue flexibility to be allowed: the differences are in the structure and contents of the information state, in the decision processes that use the IS as input, and in the update rules that manipulate it. Examples of ISU-based systems:

• GoDiS [Traum and Larsson2003], [Larsson and Traum2000], [Larsson2002a], [Traum
  2#2
  1999], [Traum
  2#2
  2000]: IS and dialogue modelling inspired by Ginzburg's Dialogue Gameboard Theory [Ginzburg1996]; utterance contents represented as slot+value pairs corresponding closely to database fields; dialogue management based on a stack of questions (or: issues) under discussion; like template-filling, with possibility to switch between tasks/templates (by question/issue-(re)raising and corresponding (re)loading of plans for different tasks) [Cooper and Larsson2003]; also various feedback/verification techniques used in GoDIS (no verification, ``implicit verification'', explicit verification) [Larsson2003]

• EDIS [Traum and Larsson2003], [Larsson and Traum2000], [Matheson
  2#2
  2000], [Traum
  2#2
  1999], [Traum
  2#2
  2000]: IS and dialogue modelling inspired by Poesio and Traum's theory of discourse/dialogue [Poesio and Traum1998], [Poesio and Traum1997]; follow-up and extension to multimodal dialogue [Traum and Rickel2002] utterance content modeled as a partial DRS; dialogue management based on obligations [Traum and Allen1994]; the crucial parts of the IS are dialogue history, intentions, obligations and social commitments [Matheson
  2#2
  2000]

• MIDAS [Traum
  2#2
  1999], [Traum
  2#2
  2000]: IS is essentially a DRS, containing additional speech act information; basically template-filling approach to dialogue management;

Various version of ISU based systems have been implemented using the TrindiKIT toolkit [Consortium2001]; see also the TRINDI project and the SIRIDUS project. Recently, CLT has developed a commercial ISU-based development toolkit, which is now being used in two research projects at CoLi (cf. DIALOG and TALK).

Reading: [Traum and Larsson2003] (and the older very similar paper [Larsson and Traum2000])