LATE 2 EXAM PREPARATION SS2011

Note: These are *not* exam questions. These are just questions which
should help you revise for the exam. Please review *all* the slides
(*all* the topics). Do not assume that the exam will consist of only a
subset of the questions below.

Natural Language Interaction block

* What is a conversation turn? Explain the turn taking rules.

* Explain (and illustrate) what collaboration in dialogue means. What
  are collaborative responses?

* List and explain the Gricean maxims of cooperation.

* What are the characteristics of adjacency pairs in dialogue?

* What is the general global structure of a conversation? Give some
  typical examples of the respective parts.

* Explain the notion of a speech act.

* Explain the notion of common ground. When is a piece of information
  considered common ground between dialogue participants?

* Summarize the Clark's (1996) model of interpretation levels in
  conversation.

* What may be the reasons for grounding problems (why can an agent
  have difficulties grounding an utterance)? Explain and illustrate.

* Summarize the three verification strategies a dialogue system might
  implement.

* Compare the finite state machine-based, frame-based and
  informationstate update models of dialogue.

* Explain the concepts behind information state update based dialogue
  modelling. What kinds of rules have to be defined in such a model?

* What is the difference between fixed and mixed initiative? What
  advantages and disadvantaged has fixed system initiative in
  comparison to mixed initiative model?

* Explain the concepts of technical evaluation, usability evaluation,
  and customer evaluation. Give some (at least 3) questions that users
  may be asked to assess the usability of a system.

* What are the challenges in conducting a WoZ experiment? Give
  examples for methods that help to meet the basic requirements on a WoZ
  system.

* What are the basic kinds of information that go into the computation
  of the expected cumulative reward in reinforcement learning
  (according to Bellman's equation)?  You need not reproduce the
  equation literally, but can give an informal description.

* Given an ISU-based dialogue manager, like in the MP3-Player
  example of the lecture slides, but for train information, and a
  dialogue state where the system has understood that the user wants to
  go to Hamburg. If th next system ation/utterance is:

  S: Which day do you want to go to Hamburg?

  What could happen next, and how would this be modeled in the MDP
  framework? Give an informal description in terms of state change and
  reward.

* In Q-Learning, what is the motivation for the step-size parameter, and
  what is the standard way to handle it?