Concepts
Lecture notes
Concepts preview:
1. Representation - frame, schema (pl. schemata), script, prototype,
scenario
Components - category, slot, filler, link (network)
Computation - matching, inheritance, spreading activation
Problem-solving - classification
2. Problem-solving (cont.) - language, taxonomy, explanation
Psychological plausibility - application, prototypes, stereotypes
Applications - WordNet, Cyc
Q: What is a concept/frame?
A1: A structure of propositions expected to be true of a given item
or situation (schema), e.g.,
Room
a kind of: place
subtypes: habitation, communal, utility
contents: ___ (?furniture)
access: ___ (?other-rooms)
instances: livingroom, gymnasium,
closet
A2: A generalized (idealized?) example of some item, situation or
event (prototype, paradigm), e.g.,
Foot
a kind of: body-part
a part of: leg
subparts: heel, ball, toes, toenails
properties: accepts socks and shoes
A3: A sequence of actions typically used to solve a problem or
reach a goal (script), e.g.,
Dine-out
a kind of: dining
related to: social-activity
location: restaurant
requirement: money
sequence:
get to restaurant
locate seat
order food
eat food
pay waiter
leave
subtypes: dine-fast-food
Each sequence entry is a sub-script
Q: What components do concepts have?
A1: defaults - slots and fillers
A2: links to other concepts - vertical (hierarchical, e.g.,
instances) and horizontal (related concepts)
Q: How are concepts associated?
A1: inheritance - rooms have locations because they are types
of places
A2: activation - dining out also activates concepts of social
activities
A3: exception - fast-food is different from other kinds of
dining out
A4: matching - a place with furniture in it is a room because
it matches the room concept
Q: What is the concept of a "ball?"
Ball
a kind of: physical object
a part of: sport-event
has shape: sphere
contains: gas
- baseballs and golf balls aren't gas-filled
- footballs aren't spherical
Possible solutions:
1. complicate the ball concept into a definition
2. take soccer balls and basketballs as typical and allow other
"balls" to overlap, i.e., to be a ball if they match at least one
of the last two requirements ("family resemblance" - Wittgenstein)
Q: When do concepts not apply properly?
A1: "Moses illusion" (Reder & Kusbit)
Q: How many animals of each kind did Moses take on the ark?
A: None, Noah was the ark-guy!
Moses and Noah are very similar.
A2: The picture problem (Smullyan, 4)
A man is looking at a portrait. When asked "Whose picture are
you looking at?" he replies:
Brothers and sisters have I none,
But this man's father is my father's son.
Most people surmise that the picture is of the man himself,
although it must be of his son. People instantiate "I" as
themselves in the viewing schema.
A3: The blow-up problem (Smullyan, 213)
The engineers' concept of a theory of rocketry includes a mathematical
model and a physically workable solution. Only the first component is
part of Von Neumann's concept - the physically workable solution is
weakly linked.
A4: The absent-minded professor (Smullyan, 227)
Taking off his tie invoked Hilbert's go-to-bed script.
Q: How are rules and concepts related?
A1: both admit defaults and exceptions
A2: concepts and rules may refer to each other
A3: rules connect items by IF-THEN relations - concepts are more general
Phil/Psych 256
Jan. 30, 1997
Q: How do scripts support inference?
A1: Scripts provide a basis for interpolating implicit information,
e.g.,
John went to a restaurant.
*He sat down.
*He looked at the menu.
He ordered chicken.
*He ate the chicken.
He left a large tip.
*He paid the check.
*He left the restaurant.
The *'d statements are easily inferred; people "recall" them even
when they're not given (Bower et al.)
Q: How do semantic networks support language understanding?
A1: horizontal links can explain priming effects, e.g., "sunset"
primes "sunrise," but not "bus" (Collins & Loftus)
A2: vertical links can explain latency in sentence verification
(Quillian et al.), e.g.,
1. A robin is a bird [*fastest]
2. Acorns grow on trees [fast]
3. A dog is a mammal [*slow]
- The hierarchy is zoological, but "A dog is an animal" is faster
than example 3.
- "A canary is a bird" is slower than 1.
(A robin is a more typical bird)
Q: What is an alternative to scientific taxonomy?
A1: folk taxonomy, e.g., Dyirbal (Lakoff)
bayi - men, most snakes, most fish, some birds,
most insects, the moon, some spears...
balan - women, some snakes, some fish, most
birds, dangerous things, the sun, fire...
balam - edible fruit, their trees, ferns, cigarettes,
wine...
bala - body parts, meat, most trees, language...
The breakdown reflects the "domain of experience" principle, e.g.,
bayi - human males
balan - human females
balam - edible plants
bala - everything else
Q: What is (proto)typicality?
A1: a "most representative" or average example of a concept set
- example-driven (Rosch)
Birds
robin
sparrow
bluejay
...
titmouse
emu
penguin
A2: a "most familiar" representative (Armstrong et al.)
("capitals" - Minsky)
Odd numbers
3
7
11
...
57
501
447
Q: How can new concepts be learned?
A1: averaging (birds) or definition (odd numbers)
A2: generalizing an old one - weakening/changing defaults, removing
slots, etc.
A3: specializing an old one - strengthening defaults, adding slots,
etc.
A4: "crossover" - interbreeding two old concepts, e.g., pulsar from
star and earthquake (Ferret - Maulden)
A5: education
Q: How can concepts explain things?
A1: in biology, a trait can be an adaptation -
adaptation
species has set of variable traits
species experiences selection pressures
pressures favor members of the species with
particular trait
members of species with trait tend to reproduce
better
most descendants of species will have trait
A2: a trait can also be a preadaptation, a vestige, or a result of
random drift
Q: What is a stereotype?
A1: a (negative) filter for information (social psych), e.g.,
"darkies," the UN conspiracy
A2: "opposed" locations in a conceptual network (Levi-Strauss), e.g.,
- giant: big, mean, pushy, obnoxious
(Titans - Xena: TWP)
- giant: gentle, childlike, clumsy
("Typhon" - Hercules: TLJ)
A3: note that stereotype has a more general meaning (Minsky)
Example: WordNet (lexicon)
Railcar (edited)
is a: wheeled vehicle
subtypes: baggage, cabin, freight, passenger
has part: suspension system for rails
member of: train
synonyms: auto, gondola, elevator, cable car
familiarity: very rare as a noun
definition: a wheeled vehicle adapted to railroad
example: "three cars had jumped the rails"
The first four lines capture general, conceptual information;
the last four capture specifically lexical information
Example: Cyc (?)
A1: (1984) Cyc was a frame-based system for capturing common
sense knowledge, using inheritance and matching as its main
inference mechanisms
Frame structure was ad hoc and became incoherent
A2: (1988) Cyc was re-done as a deduction system (several inference
engines) with "microtheories" or "contexts" - bundles of assertions
(in predicate logic) that share a common set of assumptions
Review of concepts:
1. Database - semantic network (horizontal and vertical links,
slots, fillers, defaults, exceptions)
2. Knowledgebase - inheritance, matching, spreading activation
3. Goals - classification, learning
4. Learning strategies - averaging, familiarity, specialization,
generalization, crossbreeding...
5. Works for learning, language, decisions, explanation
6. Accords with basic psychological data
Do look at the WordNet and Cyc Web pages!
Try out WordNet's WWW interface!
Next week:
- Analogies
- Gentner
Further materials
Return to Phil/Psych 256 home page