Transcript Title - MIT Computer Science and Artificial Intelligence Laboratory

Collecting Grist for the
Analogical Mill
Patrick H. Winston
Ford Professor of Artificial Intelligence and
Computer Science
17 Dec 2007
AFOSR MURI Meeting
1
Cultural Prediction & Blunder Stopping
•
Stories capture cultural beliefs & values
•
•
Folktales, Myths, Morality Tales, Religious Texts, Urban
Legends
Analogical Framework
•
•
Predict role-models; individual & group behavior
across cultures from stories
Story Database
•
Automatic NLP
A bottleneck
Textual
Materials
Structured
Representations
Experimenter
17 Dec 2007
AFOSR MURI Meeting
2
The Story Workbench
• Modular
Workflow
• Easily Extendible
1. User enters
Natural Language
• Cross-Platform (99% pure Java)
• Open Source
• Large, dedicated User-base
2. Workbench
makes its
best guess
• Commercial-Quality
4. User enters
corrections or
adds detail
3. Problems are
highlighted
17 Dec 2007
AFOSR MURI Meeting
3
A Rich Suite of Representations
Listening on the Radio to:
Where we are going
• Context
(last game of the World Series)
• Location
(pitcher is on the mound, baseman and crowd are in
front of the pitcher)
• Events (and their order)
(The look, the windup, then the pitch)
• Causality
(pitcher’s pitch causes the ball to fly forward)
• Discourse
Classical
pitch
(pitcher is rumored to have taken steroids, he’s quoted
as denying it)
• Emotions
(to)
(The pitcher is calm, the crowd is going wild)
pitcher
batter
ball
17 Dec 2007
• Motivations
(The pitcher wants to win the game, wants to strike out
this batter)
AFOSR MURI Meeting
4
Accomplishments to Date
• Implemented Story Workbench foundation
•
•
Syntax-level NLP representations & parsers
Facility for user correction
• Interfaced successfully with Forbus Group
•
Ability to assign CYC propositions to text (semantic interpretation)
• Started beta testing
•
•
17 Dec 2007
Word Sense Disambiguation as a benchmark semantic annotation task
Annotation of Discourse structure by Gibson & Kraemer at MIT BCS
AFOSR MURI Meeting
5
Activities in the Near-term
• Implement next layers of representation and parsers
•
•
•
•
•
Entities
Events
Discourse Relations
Mental States and Motivations
Moral of the Story
• Collaborate with Medin at Northwestern to infer beliefs
from stories and predict Menominee and Itza behavior
17 Dec 2007
AFOSR MURI Meeting
6