COLLAGEN: When Agents Collaborate with People

COLLAGEN: When Agents
Collaborate with People
Charles Rich and Candace L. Sidner
Presented by Daniel Schulman
What is Collagen?
A toolkit for building applications that use
collaborative discourse.
Concepts and Background
SharedPlans (Grosz & Sidner 90)
Mutual beliefs about the collaboration.
Focus stack (of discourse segments)
Keeps track of current attention.
Discourse interpretation algorithm
What does action do to focus stack?
Discourse generation algorithm
Focus stack + SharedPlans -> possible actions
Task Modelling
Domain-specific knowledge about the
collaboration.
Artificial discourse language (Sidner 94):
“Propose” and “accept” beliefs.
Beliefs include SHOULD (goals) and RECIPE
Recipe: Can produce steps to achieve a
goal.
For flexibility, Collagen uses generators.
Sample Application – Travel Advisor
Critique - Collagen
A very useful, well-thought out toolkit:
Doesn’t require natural language processing –
so it doesn’t require strong AI.
Application-independent design makes it very
flexible – ex. Could use it for a telephone-based
system.
Ability to add in application-specific code is
powerful.
Critique – The Paper
Primary purpose of paper:
To explain the architecture of Collagen.
Not much justification/evaluation of architectural
choices.
Very confusing explanations:
Tough to understand how pieces fit together.
What parts are Collagen, what’s app-specific?
Sample app doesn’t help – from wrong POV.
Critique – Future Work
Future work section is good, but:
Collagen is meant to be a toolkit, not a single
application.
To really evaluate it, it should be used to build
several different applications.
What kinds of collaboration is it good for? What
domains?