Information State Approach to Dialogue Management

As suggested by the scholar in the article, at the beginning of the 21st Century when the research was conducted, it was yet to have a common ground for the concepts and system design of the dialogue system, or specifically speaking, the dialogue manager. Is the Information state approach proposed in this paper being accepted widely in the research field? Is there any evolution of the terms and concepts in the past 10 years? What’s the up-to-date knowledge accumulated and what’s the state-of-art area of technology the academic and industry are using?

Based on a general understanding of the GoDiS system implementation process, theanalysis and partitioning of dialogue are still, to a large extend, done by a rule-based logic specification. Is such technique still viable and sufficient for nowadays real-time spoken dialogue system? Is there any better (in terms of efficiency) approach to classify the sematic of the utterance?

- How many other systems use the information state approach successfully, not connected to this research?
- Aren't there often more cases where a more specific-to-the-project dialogue manager is more efficient?
- What are the most common mistakes made by an information-state dialogue manager?


-       1) This paper has a great structure, where the ideas flow from section to section; I specifically appreciated how the authors tell us what is going to be discussed in future sections at the beginning, so we can mentally organize what we are reading as we are reading it.

-       2) I appreciate that the authors mention implementation as well, as opposed to simply discussing theoretical models (Section 2.2) and provide actual examples of how things would work in a dialogue management system (Section 2.4 and 2.6).

-       3) Finally, the discussions on actual systems that have been implemented (GoDIS and EDIS) and their architecture was particularly useful for understanding how everything is organized. I generally find that SDS papers assume that the SDS is a functioning black box so we never actually learn how one is implemented.


It seems the paper addresses more of a general software engineering problem, and one of code organization, than anything intrinsically SDS related.
What other solutions to the problems stated in the paper have been developed?
How domain extensible can a framework like TrindiKit be?

A. I can understand why the dialogue management system for an all-purpose tool like Siri or Cortana would be more than is needed for most smaller dialogue systems. However, if the efficiency stayed the same and the companies were willing to make their dialogue management systems open source, would that be a useful contribution to the field, or would it still be so specific that it would not be of much use to dialogue systems with other functionality?

B. What are discourse representation structures and how low-level are they? The set of formal representations implementing informational components are nearly at the data-structure level (at least at the data structure paradigm level). Is there a generally recognized paradigm for DSRs that everyone uses, or is this a general class of structures implemented in various ways?

C. How does the system decide if an utterance answers a question it may have in its agenda but hasn't asked yet? Does it just compare the utterance to every question and see if there's any overlap in topic?

1. The authors mention that information state-based theory of dialogue includes informational components, such as common ground, beliefs, intentions, etc.  How are these concepts extracted, and then represented in this approach?

2. Once a state is popped, can it be returned to later in the dialogue?  (For example, you might want to do this for error recovery purposes.)

3. Do the dialogue moves and update rules have to be tailored for each specific domain?


a) GoDis : Handles task accommodation : where an errand is not specified, but needs to be inferred from the user’s answer. Update Queries are updated / Update rules are generated on the basis of the data entered into the PLAN field.
* How does the system ­ Identify/differentiate between a question and an answer ?
b) EDIS : Information state based approach based on record representation used for coding information states * (What does this mean ??)

Analysis of results ­ What is : average return per episode? / no. of turns ­ K­L Divergence ? **