ART - Dialogue Models and Dialogue Systems

2005-2006 Spring Semester - Half Module - 10 Credits

[ Summary | Objectives | Content and Recommended Readings | Slides | Assessment ]

Summary

Aims

• to provide a general overview of current research problems in the field of dialogue modeling and dialogue systems.
• to provide students with a basic understanding of dialogue modeling and dialogue systems. 
• to study the computational techniques that are employed in evaluating and automatic training of dialogue systems,
• to help students develop a critical appraisal of current research in this area.

Prerequisites and Corequisites

Objectives

• have a basic understanding of dialogue modeling and dialogue systems
• have a basic understanding of issues with evaluating dialogue systems
• understand the techniques employed in automatic training of dialogue systems and adapting them to individual users,
• be able to critically discuss recent research papers in the field.

Content

• Challenges in this research field,
• Overview of the remaining lectures.

• Fundamental theories of dialogue and dialogue interaction,
• Techniques/Approaches to building dialogue systems,

Readings

• J. Glass, J. Polifroni, S. Seneff and V. Zue, "Data Collection and Performance Evaluation of Spoken Dialogue Systems: The MIT Experience," Proc. ICSLP, Beijing, China October 2000.
• den Os, E.A., and Bloothooft, G. (1998), Evaluatin various dialogue systems with a single questionnaire: analysis of the ELSNET Olympics, Proc. LREC Granada, Spain, 1998.
• Polifroni, J., and G. Chung, "Promoting Portability in Dialogue Management"
• Polifroni, J., G. Chung, and S. Seneff, "Towards the Automatic Generation of Mixed-Initiative Dialogue Systems from Web Content"
• Glass, J., J. Polifroni, S. Seneff, and V. Zue, "Data Collection and Evaluation in Spoken Dialogue Systems: The MIT Experience"
• Bos, J. et al. "DIPPER: Description and Formalisation of an Information-State Update Dialogue System Architecture"
• VoiceXML Tutorials/VoiceXML Examples/VoiceXML Reference

• Representing Dialogue as an MDP or POMDP
• Dialogue System Strategies
• Examples of the use of RL for Dialogue manager optimization
• Current research challenges.

Readings

• Walker, M., "An Application of Reinforcement Learning to Dialogue Strategy Selection in a Spoken Dialogue System for Email"
• Singh, S., et al., "Optimizing Dialogue Management with Reinforcement Learning: Experiments with the NJFun System"
• Levin, E., et al., "Using Markov Decision Process for Learning Dialogue Strategies"

• An examination of recent research on user modeling in dialogue systems 
• The use of multi-attribute decision theory for user models
• Techniques for evaluating the effectiveness of user models
• Current research challenges

Readings

• Whittaker,S., Walker, M., Maloor, P., "Should I Tell All? An Experiment on Collaboration in Spoken Dialogue"
• Walker, M., et al., Generation and Evaluation of User Tailored Responses in MultiModal Dialogue

• An examination of recent research on stochastic generation techniques for dialogue systems 
• The use of boosting for sentence planning in dialogue systems
• The use of dialogue corpora for automatic training of generation modules
• Current research challenges. 

Readings

• Walker, M., Rambow, O., and Rogati, M., "Training a Sentence Planner for Spoken Dialogue Using Boosting"
• Walker, M., Rambow, O., and Rogati, M., "SPoT: A Trainable Sentence Planner"

Slides

• Week 1: Overview of Spoken Dialogue Systems
• Week 2: An Introduction to VoiceXML
• Weeks 3 and 4: Architectures for Spoken Dialogue Systems/Dialogue Management
• Week 5: Evaluation of Spoken Dialogue Systems
• Week 6: Reinforcement Learning in Dialogue Systems
• Week 7: User Modelling
• Weeks 11 and 12: Natural Language Generation (and sample paper outline)
• Week 13: Text-to-speech Synthesis

Lecturers

Resource Requirements

• None

Assessment

• 10%: Annotated bibliography prepared for paper (see below). Students are asked to find at least 10 sources (papers, books) that they think will be useful for their essay. They are asked to annotate each bibliography entry with some indication of why they thing the source will be useful. Due date: April 24th.
• 10%: Written questions in class.
• 80%: Essay, on one of the following four topics, each of which has been covered in class (4000 words in length, 20h work). Please use ACL style in your final essay. ACL style files, for LaTeX or Word, can be downloaded here. Due date: May 30th.

• Essay topics:
  
  
  • Dialogue System Design:
    Compare/contrast Galaxy Communicator with one or more other methods of dialogue management, e.g., the information-state approach as exemplified by Dipper or TRINDIKIT. Specific issues for comparison are control of the dialogue, the passing of information among modules, and ease of portability to other domains. How are each of these issues in dialogue management addressed by each system? How would you decide which of these approaches to use if you were going to build a dialogue system?
    
  • Reinforcement Learning:
    Summarize research using reinforcement learning in dialogue management. Compare and contrast what can be learned with reinforcement learning experiments using a simulated user and what can be learned using experiments with human users. Discuss the usefulness of what is being learned with each approach.
  • User Modeling:
    Summarize several uses of user modelling for spoken dialogue systems. Since user models represent several different aspects of the user, such as the expertise, knowledge, and preferences, discuss what system behavior is supported by different types of models. How can the system developer tell that the user models are improving the system's behaviour?
  • Spoken Language Generation:
    Discuss several different algorithms for generating spoken language that can be used in dialogue systems. Some of the algorithms currently in use are template-based (e.g., the Genesis system from MIT), and hybrid linguistic/statistical methods (e.g., Halogen, SPoT, SpaRKy). How would you decide which of these approaches to use if you were going to build a dialogue system?