Understanding Understanding

1
Understanding Understanding

• Deadly pastimes suggest adventurousness.

2
Early AI 1950 - 60

• Game Playing / Problem Solving
• Chess
• Checkers
• Go
• General Problem Solver (GPS) Newell Simon
• Restricted domains Did not require general
  language competence.

3
AI and Language 1960s

• Joseph Weizenbaums ELIZA
• A digital Rogerian psychotherapist.
• PARRY (Kenneth Colby)
• Answers questions from a paranoid perspective.
• RACTER
• Generates stories.
• Other Programs (link)

4
LISP and AI

• AI programs use LISP (List Processor) to
  manipulate language.
• invented by John McCarthy
• A typical LISP expression
• (car (dog cat mouse))
• Evaluates to
• Dog
• But this really is syntax, not semantics

5
Characteristics of these programs

• Limited domain
• Syntax processing
• Breakdowns occur, even in the target domain.

6
Microworlds AI

• Early programs dont come close to having
  understanding.
• Need to add some knowledge
• Circumscribe the domain to limit amount of
  information needed in a database
• A microworld is a limited body of knowledge.

7
Example SHRDLU

• SHRDLU (Winograd, 1972)
• Manipulates blocks
• Contains separate but interacting programs to
  deal with sentence syntax and block information.
• Understands the block microworld.

8
Bar-Hillels Test Sentence

• The box is in the pen.
• What kind of box?
• What size box?
• What kind of pen?
• What size pen?
• How do you answer these questions?

9
Language understanding requires common sense.

• Everyday knowledge
• Understanding context
• Making the right assumptions
• Tolerating inconsistency
• Recognizing change.

10
HALs conversations

• Dave has drawn a picture of Dr. Hunter
• Can you hold it a bit closer?
• What are the ambiguities?
• How are they resolved?

11
HALs conversation II

• HAL Thats Dr. Hunter, isnt it?
• Dave Yes
• Why didnt Dave say No, thats a sketch of Dr.
  Hunter.

12
HALs Conversation III

• HAL The radio is dead.
• Question
• Did HAL think that the radio was a living thing?

13
Dealing with Ambiguity

• More Examples
• Resolving ambiguity requires common sense.

14
Dealing with inconsistency

• A contradiction is something of the form P and
  not P.
• Consider
• There are no vampires.
• Dracula is a vampire.
• A computer with common sense has to tolerate
  inconsistency.

15
Another example of inconsistency

• All unmarried adult males are bachelors.
• Father Gregory is an unmarried adult male.
• Therefore, Father Gregory is a bachelor.
• But hes a priest.

16
Intelligent action requires common sense.

• What would a car that drives for you have to
  know?
• The Yale Shooting Problem
• Assume that an individual is initially alive,
  then the gun is loaded, then he waits for a
  while, and then he is shot with the gun.
• Infer that the individual is not alive.
• What assumptions must you make?

17
Can we infuse a digital computer with common
sense?

• Lenat thinks it can be done.
• The three steps
• Prime the knowledge pump.
• This takes a while in humans and machines.
• Construct the ability to communicate in a natural
  language
• Give the computer the ability to explore and
  experiment on its own.
• The CYC Project