|
Main
Description
Status
Research
Publications
Demo
People
Funding
Links
|
Research
Our work concentrates on several research themes:
- When users are not AI experts, what kinds of help do they need to
teach new knowledge to an intelligent system?
We have conducted a number of user studies
and some of the central problems we found are:
- How can we help users handle the formal
representations that our systems use?
Building on past research on natural language generation, we have developed an
English-based
editor that allows
users to select meaningful strings of a paraphrase and
search the knowledge
base for sensible alternative replacements.
- How do users know they are doing the right thing?
When a system understands how each individual piece of
knowledge relates to others
then it can figure out how new knowledge fits, and can
figure out how to interact
with the user when new knowledge is inconsistent or when
other necessary knowledge
is missing. Our approach is to build acquisition
interfaces that can
automatically derive Interdependency Models from a
knowledge base and use them to
guide the acquisition dialogue.
We used this approach in
EXPECT,
EMeD (EXPECT's
Method Editor),
KANAL, and
- How can we help users make complex modifications
to a knowledge
base that require many related small changes?
We have developed dialogue planning tools that represent
typical sequences of changes that users make and use them to
dynamically generate follow-up questions to users.
We used this approach in
KA Scripts,
PSMTool/Constable, and
our new work on
acquisition
dialog planning.
- How can we help users figure out where to start?
Because it is hard for users to start entering knowledge
from scratch, we assume that the system has some
general background knowledge that the user populates
with domain-specific information.
We have built several declarative domain theories for this
purpose, including
PLANET (a PLAn
semantic NET).
- How can a user turn informal and possibly disconnected information
into increasingly more formalized and operational knowledge?
A new research area is on tools to develop
Resilient Hyper Knowledge Bases (RHKBs),
which capture the rationale and development of knowledge
bases from the original sources of information used.
These tools will guide users step by step to
create increasingly more formal representations of knowledge.
We are using some natural language processing techniques
to help users enter knowledge from textual sources.
Our research is motivated by real applications, which often
revolve around planning and process models:
|
