barsalou - 1992 - frames, concepts, and conceptual fields

8/2/2019 Barsalou - 1992 - Frames, Concepts, And Conceptual Fields

http://slidepdf.com/reader/full/barsalou-1992-frames-concepts-and-conceptual-fields 1/28

FRAMES, FIELDS ,

AND

CONTRASTS

New Essays in

Semantic and Lexical Organization

i

Edited by

ADRIENNE LEHRER

University of ArizonaEVA FEDER KITTAY

SUNY Stony Brook

1992

LAWRENCE ERLBAUM ASSOCIATES, PUBLISHERS

Hillsdale , New Jersey Hove and London

Frames, Concepts, and

Conceptual Fields

Law rence W - Barsalou

University of Chicago

In this chapter I propose that frames provide the fundamental representation ofknowledge in human cognition. In the first section, I raise problems with the

feature list representations often found in theories of knowledge , and I sketch the

solutions that frames provide to them. In the second section, I examine the three

fundamental components of frames: attribute-value sets, structural inv ariants,

and constraints . Because frames also represent the attributes, values, struc turalinvariants, and constraints within a frame , the mechanism that constructs frames

builds them recursively . The frame theory I propose borrows heavily rom pre-

vious frame theories, although its collection of representational components is

somewhat unique . Furthermore , frame theorists generally assume that rames are

rigid configurations of independent attributes , whereas I propose that rames are

dynamic relational structures whose form is flexible and context dependent. In

the third section, I illustrate how frames support a wide variety of representa-tional tasks central to conceptual processing in natural and artific ial intelligence.

Frames can represent exemp lars and propositions , prototy pes and membership ,

subordinates and taxonomies . Frames can also represent conceptual combina-

tions, event sequences, rules , and plans. In the fourth section, I show how frames

define the extent of conceptual ields and how they provide a powerful produc-

tive mechanism for generating speciic concepts within a field.

FEATURE LIST REPRESENTATIONS OF CATEGORIES

Before proceeding to a detailed discuss ion of frames, I i rst discuss their mostobvious competitor-feature list representations. Later, we see that frames reme-

2 1



38 BARSALOU , 1. FRAMES, CONCEPTS, AND CONCEPTUAL FIELDS 39

llhI

i

i

7

15

s

pi

2 i5

i 0 3

0z

Q2

<6

CO

g

>

15

.

, 5-9I E1

,5 2 S

< ar il QJ Oi

a * E

S 2

ir

ECo

2 s

\ a- o

52?LL. CD.Oil

5

I«K 1

li

particular sets of values locally . Consider the enables relation between San Diego

and suring in Fig. 1.7 . This relation specifies that a particular value of the

location attribu te constrains a particu lar value of the activity attribute . Another

value constraint sim ilarly relates Rockies and snow skiing . The requires relation s

between snow sk i and moun tains and between suring and ocean beach illustrate

somewhat more complex value con straints that cross levels with in the frame

representation . Similar to attibute constraints, value constraints may often repre-

sent statistical patterns and personal preferences , rather than necessary truths .

Contex tual Constraints. A distinc tion between contextual constraints and

optimizations is orthogonal to the d istinction between attribute constrain ts andvalue constraints . Contextual constraints occur when one aspect of a situation

constrains ano ther, such as physica l constraints in nature. For example, speed of

transportation constrains its duration over a fixed distance . Similarly, the activityof suring requires an ocean beach. Contextual constraints also reflect cultural

conventions . For example, people'

s income and the taxes they pay may bear a

relationship to one another. Similarly , swimm ing as an activity may require a

sw imsuit as clothing. In general , the various aspects of a particular situation arenot independent of one another. Instead, physical and cultural mechanisms place

con straints on combinations of compatible attribute values . As the precedingexamp les illustrate , contextual constraints can either be attribute constraints orvalue constraints.

Optimizations. Whereas contextual constraints reflect physical and culturalmechanisms , optimizations are con straints that reflect an agent

'

s goals . Consider

how the agent's goals in Fig. 1.7 constrain the values of variou s attributes . The

agent's goal of good exercise constrains the value of exertion in the activity frame

to be high. Sim ilarly, the agent'

s goals of short travel and low cost constrain the

duration and cost vaiables in the transporta tion frame to be short and low,

respectively. Although all of the optimizations shown in Fig. 1.7 are valueconstraints, optim izations can also be attibute constraints. For example, the

value of an agent'

s desire to achieve a goal generally constrains the value of theagent

'

s effort in pursuing it .

Whereas contextual con straints typically require that values satisfy them,

optim izations typically require that one value excel beyond all others. For exam -

ple , just about any kind of swimsuit will satisfy the contextua l constraint tha tclothing be worn while sw imm ing. In contrast , the cheapest form of transporta-tion optimizes the goal of inexpensive travel . Whereas people generally selectvalues that satisfy contextual constraints, they generally seek values that excelwhen optimizing goals .

Peop le often attempt to optim ize multiple goals simultaneously for an at-tribute. For transportation , someone might optim ize cost, speed , and comfortsimultaneously-not just cost alone . As a result , the optimal value may not be



52 BARSALOU

typ«

SIZE

SIZE

laroeBEAK

fDrown

COLOR

SIZE

COLOR

WATER )FOWLy

smal BEAK

B IRD BEAK

OCOMOT ION i

|5OCOMOTION SIZE

COLOR

upact duck:EAKshort

COCOMOTXJN

long

CK

FIG . 1.11. Example of us ing frames to represent subordinates in a

partia l taxonomy for bird .

Representing Conceptual Combinations

Frames readily represent conceptual combinations, which include adjective-noun compounds (e.g., red bird), noun-noun compounds (e.g., bird house),and more complex noun phrases (e.g., birds from Canada that live in Florida forthe winter). Complex propositional representations for discourses, such as theone in Fig . 1.9 , exhibit the ultimate in concep tual combination (Barsalou , 1992).

In their selective modif ication model . Smith, Osherson, R ips , and Keane

(1988) represent conceptual combinations with frames. On their view, when anadjective modifies a noun , the adjective replaces the default value for the rele -vant attribute in the noun

'

s frame. In red bird, for example, red replaces color'

s

default of brown in the frame for bird . All other nonmodified attributes in the

1. FRAMES , CONCEPTS , AND CONCEPTUAL FIELDS 53

noun'

s frame retain their defaults (e.g., size retains its default of small). How-ever, selective modification fails to account for interactions that often occur

between modified and unmodified attributes (Medin & Shoben, 1988 ; Murphy ,

1988, 1990). In white bird, for example, size as well as color is likely to bemodified, because white birds are generally large , thereby overiding the defaultof small for size.

The examp les in Fig. 1.12 illustrate how frames epresent conceptual com -

binations and the interactions that often accompany them . As can be seen, one

meaning of bird house can be represented by integrating frames for bird and

house into the agent and location attribu tes in the frame for live (cf . Gleitman &

Gleitman, 1970). Constraints capture the interactive inferences that peop le are

Jr tocwo tT *- - . fy

(UVE)

typesizeX smal

I

agent

LIVE

LOCAT ION

<DO ( < (Sza

sma l

medum

arge

APARTMENT> ( afd)<-SE!- .none

-i smal

FIG. 1.12. Examp les of using frames to represent conceptual com-

binations for bird house and apartment dog .



58 BARSA LOU

/ JGlNEV asoeclBvan t

Rule 3 Riie 4ule 2ule 1CAUSEAUSEAUSEAUSE

outcomg? g fiondton Cw tcomeomej) (Sondno rO cputcome

f

(jtrokeT) C oke2 stroke 3 stroke 4

1 . FRAM ES, CONCEPTS, AND CONCEPTUAL FIELDS 59

FIG. 1.15. Using a frame to integwte the rules that produce the event

sequence for engine cycle in Fig . 1.13.

Representing Rules 1 through 4 in this manner differs in important ways fromproduction systems . Whereas these rules are integrated in a frame representation ,they would be independen t entities in a production system . As a consequence ,

the frame representation is more economical . Each stroke line is only representedonce in Fig. 1.15 but would have to represented twice in a production system :

once as a condition, and once as an outcome (similar to Rules 1 through 4

preceding). Moreover, the frame representation provides a more coherent andgloba l account of the engine cycle , because relations among the four rules areshown explicitly. However, an individua l rule can still fire separately of theothers. Rule 3 ru le could fire , if its condition occurs, even if Rules 1 and 2 have

not . Additionally , automatic firing of rules could be represen ted as the extent towhich the relevan t causa l relations becom e strengthened in memory (Shiffrin &

Schneider , 1977). or as the extent to which exemplars are integrated with ru les

(similar to Figs . 1.8 and 1.10; Logan, 1988). Nevetheless, the representation inFig . 1.15 suggests that rules embedded in frames should be harder to access thanisolated rules in production systems, if surrounding rules in an integrated frame

provide interference . Evidence from Carlson and Yaure'

s (1990) blocking condi-

tion dem onstrates such difficulty.

Figure 1.16 illustrates two further examples of frames representing ru les. A t

the top , an intuitive view of combustion is represented as a causal relationbetw een two sets of states defined over fuel, air , and heat source . At the bottom,

the epitome of a rule , modus ponens , is represented as a relation between twosets of states defined over X-» Y , X , and Y. As these examples illustrate, frames

can represent a wide variety of rules .

CAUSE

7

cona tion ) r outcome

aspactnorma l

fue)

aspec tcarbon

COI USTIONambient aspecttemperature

slaeHot- aspec tir

expanded aspect

Heat slals aspec tpresentource

Y

premsex conc lusion

aspec ttrue

Y

false

aapee ltrueMODUS

false

aspec t ,true

-""v urknown '3Pecl

Y

FIG . 1.16. Examples of using frames to represent intu it ive rules forcombus tion and modu s ponens .

Representing Plans

Frames are cen tra l to the initial planning of events (Barsalou, 1991). When

people plan events such as trips , pu rchases , social events , and repairs , they oftenbegin by partially activating a frame for the event being planned. In planning avacation

, people might activate the partial frame in Fig . 1.17 .

The primary activity during the initial planning of an event is to instantiateframe attibutes. As illustrated in Fig . 1.17 , a planner might select snow skiing as

barsalou - 1992 - frames, concepts, and conceptual fields

Documents