Behaviors and Layers in Spoken Dialogue Systems

1
Behaviors and Layers in Spoken Dialogue Systems

• Pontus Wärnestål
• Linköping University
• Speech Technology, KTH
• Jan 21, 2005

2
Some Problems with Dialogue Systems
hard to design and build
low or no re-use
Practical Engineering Issues
complex architectures
symbol- grounding problem
Usability and Acceptance issues
Theoretical Issues
lack of harmony between theory and
architectures/models
lacking implementations of interaction-level
phenomena
long processing and delays
sense-think-act model fallacy
lack of robustness
3
Solution?

• There might be
• theoretical,
• practical, and
• usability-related
• findings triggered by a behavior-based approach
  to dialogue system construction

4
A Glance at AI/Robotics Research

• Reactivity
• Autonomous adaptation to dynamic (non-discrete)
  real-world problems
• Less top-down planning and representation
• More low-level behaviors emergence
• The behavior-based paradigm
• Resolved similar/analogous problems...but in
  limited scope

5
Behaviors and Layers in Dialogue Systems

• Towards more realistic dialogue systems?
• Natural interaction ? higher usability
• Behavior design ? less complex development
• Sounder cognitive theory, and taking
  psycholinguistic and communication theory into
  account
• So, what do we mean by layers and behaviors?

6
Layers in Software Engineering

• e.g. TCP/IP stack, operating systems
• Three main benefits (Garlan Shaw, 1994)
• design on different levels of abstraction
• enhancement/scalability/maintenance
• reusable layer interface standards

7
Layers as Behaviors

• The Subsumption architecture (Brooks, 1991)
• Low-level specification
• Emergent complex and adaptive behavior
• Mainly used for robots to avoid unknown obstacles
• Avoid obstacles ? dialogue
• a big leap
• what about top-down planning?

8
Subsumption Architecture
9
Reactive, Process Control and Content Layers

• Ymir (Thórisson, 1997)

10
Benefits

• Gestures and body language (including facial
  gestures) are integrated with the communication
  content, without artificial communication
  protocols
• Behaviors run concurrently at the expected times
  and without unnatural delays (reactive layer in
  the range of 150-500 ms)
• Miscommunication and speech overlaps are handled
  in a natural way, by using stops and restarts.

11
Interaction and Content Layers

• Separates interaction-level phenomena from
  content and context management (Lemon et al.,
  2003)
• Interaction-level phenomena
• realistic timing and turn-taking
• immediate grounding
• more realistic (continuous) feedback
• barge-in management
• NP selection (anaphora management).

12
Language Engineering Conclusions

• Reuse
• Scalability and maintenance
• Robustness
• Design methodology
• interaction-level phenomena seem easier to handle
  with low-level reactivity á la behavior-based
  design

13
Usability Conclusions

• Stress interaction and communication natural
  dialogue
• Increase dialogue system usability with better
  timing, continuous feedback, grounding,

14
Theoretical Conclusions

• Towards a theory of intelligence
• Incorporating Communication theory
• somewhat ignored in the classical approach(?)