Natural Semantics in Artificial Intelligence

kmerican

Journal

of Computational Linguir

tics

M i c r o f i c h e 3

Jaime

R. Carbonell andLAl'lan M.

Collins

Bolt Besanek and Newman I n c .

Cambridge, Massachusetts

ABSTRACT

This papar d i s c u ~ s e e human semantic knowledge and proceesing i n terms of the SCHOLAR system. In one major s e c t i o n we d i s c u s s

t h e imprecision, the incompleteness, the open-endedness, and the u n c e r t a i n t y of psopl,eqs knowledge. In t h e o t h e r major s e c t i o n we d i e c u s s strategies people use t o make d i f f e r e n t types of deductive, n e g a t i v e , and functional i n f e r e n c e s , and the way u n c e r t a i n t i e s

combine i n these i n f e r e n c e s .

Irapreeision can occur either i n memory or i n c a m u n i c a t i o n .

SCHOLAR can have precf ae values or fuzzy v a l u e s stored, and its procedures can, t o eome extent, deal w i t h fuzzy questions when

precise valtues

are

stored, and w i t h precise queltions, when fuzzy

values are stored. Embedding allows i n f o t i o n to be specified i n the data base t o any level of d e t a i l

or

preci s f o n . But SCHOLAR

o n l y camunicatea the nwst important i n f o t i o n on any topic (as

rneasured by importmce t a g s ) , unless more i n f o t i o n i s requested.

1.t should a l s o be p o s s i b l e by using importance tags to a d j u s t what

+jon SCHOLAR communicates, i n accord with the sophistication and, i n t e r e s t s of the l i s t e n e r .

Inference atrategiee t h a t are appropriate when the complete set of object attributes, or values, i s known (i .em, in a closed

world) do not apply when knowledge is incomplete e , i n an open w r l d )

.

There are a variety of u n c e r t ~ i n i n f e r c n c e s t h a t

people use to circumvent the hales i n their knowledge, which are being progr e d i n SCHOLAR.

There is a _set o f transitive relations

--

superordinate,

superpart, s l a r i t y , p r a i m i t y

,

subordinate, and subpart relations

Currently SCHO o n l y handles superordinate i n f e r e n c e s ( e m g

.

,

the Llanos has a rainy season because it is a savanna) and super- part inferences (e .g,

,

the language in Rio is Portuguese because

R i o is part of Brazil). ~ e d u c t i v e inferences can be more or l e a s

c e r t a i n (similarity inferences are l i k e s u p t o r d i n a t e inferences. but l e a s c e r t a i n ) and can have restrictions on t h e i r use ( o n l y certain attributes transfer an s u p e r p a r t ) .

~ b knowledge n is incomplete, it is not safe t o assume that something is n o t true just because i t i s not s t o r e d . Thus an in- ference is necessary to decide when to say ' N o w and when to say

'I d o n B t know,@ There is a eamp1Pcated s e t sf strategies in

SCHOLAR to find vatious kinds of contradictions that people case to say *NO,

"

If a c o n t r a d i c t i o n c a n n o t be f o u n d , a n o t h e r nega8ive inference, c a l l e d the Wlack-aF-knawledgea inference, f s tried*

When enough is known a b u t an a b j e c h it is possible to c o n c l u d e that someaing is not true ut mat object ow t h e ground8 that if it were true, it m u l d be stared,

another class of w c e r t a i n i n f e r e n c e s depends an i l l - d e f ined

knovledge of functional d a t e n a i n a n t s

,

e .g.

,

that climate depends

O R l a t i t u d e and a l t i t u d e . D % f ferent ways t h a t p o p l e use f m c t i o n a l knowledge irnm1ve f m c t i a n a l ealeulatisms ( a . g , , if a place h a s a

p a r t i c u l a r latitude, i t probably has a particular elihate), func-

tional analogies ( e , g , , if a place is like another place in latitude and altitude, it probably has the s m e c h i m t e )

,

and ts answer Why q u e s t i o n s (egg,, a place h g a particular e l f m t e because of its

latitude and altitude),

D i f f e r e n t inf erenees can esIwBine in d i f f e r e n t ways, Somtims one strategy may call another strategy to h i n d an anewer, men

d i f f e r e n t inferences i n d e p e n d e n t l y reach the s m e or d i f f e r m t con-- elusions, a e y coItnbiwe to increase or d s r e a s s certainty, T h e pm-

TABLE OF CONTENTS

2 . The S c h o l a r System a s a n E n v i r o n m e n t t o Study Natural

~ ~ m a ~ ~ i ~ ~ . ~ ~ . ~ . ~ 6 . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . ~ ~ ~ ~ ~ ~ ~ ~ ~ . ~ . . ~ . ~ ~ . ~ . ~ ~ ~

3 . 1 Imprecision or F u z z i n e s s . . o a . . . o o O O o a I ) D o O o ~ O o O o O O O o O ~ ~

3.2 Incompleteness, Embedding, and R~levancy... 12

3 - 3 The R e f e r e n c e Problem and Context

...

13

3,4 Closed versus Open Worlds o O o ~ o ~ o ~ o ~ ~ O ~ D o D e O o 15 o ~ d ~ O O D o

3.5 The True-False ~ichotorny and ~ u a n t i f i c a t i o n

..,...

17

4 , N a t u r a l I n f e u e n @ e s . , . . o . . . , P 0 D o o m o e O O o 0 O O ~ B O o O O O o o . ~ . o . o . 0 1 8

4 . 1 Deductive ~ ~ f ~ ~ e ~ ~ e ~ . . o O Q . o O . O ~ O o O O ~ ~ O O ~ o O ~ O O O O O O . O o ~ ~ 4 . 2 Negative I n f e r ~ ~ ~ e s . . o . . o . o o O o O o D o o O O o O o O o o O o o a o O ~ o . ~ ~

4 - 4 Inductive I n f e r e n c e s o 0 . 0 . 0 0 D ~ 0 0 0 D o D . 0 . . o ~ a ~ O O O o O O D ~ 0 2 8

% , I n t r o d u c t i o n

In t h i s paper we will discuss how to eepresent and process

i n f o m a t i o i l in a computer in ways t h a t are natural to p e o p l e .

This does n o t mean doing away completely w i t h representations and procedures which computers have traditionally used, but adding

new representations and procedures which they have not used.

People o f t e n store and communicate imprecise, incomplete, and unquantified info t i o n ; they a f t e n assert truth o r falsity

in relative terms; and they seldom seem to use r i g o r o u s l o g i c in

t h e i r i n f e r e n t i a l processes. Because of these c o n d i t i o n s , people s , e e m to have an almost i n f i n i t e i n f o tion processing c a p a c i t y ,

w i t h i n f e r e n c e making and problem s o l v i n g abilities more r e f i n e d

and far more f l e x i b l e than any e x i s t i n g computer program.

Now man we study a e s e h m a n eapabiLities in order to make

our machines show similar perfbnnance? A combination of

approaches is perhaps best. O b s e r v a t i o n of people

'

s behavior, introspection, some experimentation, protocol _analysf s, and

synthesis of computer programs can a l l be vhluable t e c h n i l ~ u e s . A 6

recent paper (CalZins, mrnock and P a s s a f i m e ) d i s c u s s e s a tech- nique for combining protocol analysis with program synthesis as

a p p l i e d to t u t o r i a l d i a l o g u e s . The synthesis directs what to

analyze, and the strategies observed in the analysis are evaluated

the SCHOLAR

system

in this

way as a

vehicle

f o r

experimentation

w i t h

natural

semnties.

Before

we

d i s c u s s

some

of

t h e

major

problems

in natural

semantics, we

w i l l

b r i e f l y describe

the

SCHOLAR system,

since

it

i e

t h e

e n v i r o m e n t

f o r

o u r

research.

A word

of

c a u t i o n

though:

we

are

o n l y

t r y i n g

to

develop

some

i n s i g h t s ,

w i t h o u t

attempting

to

be

exhaustive.

More

q u e s t i o n s

w i l . 1

be raised

than

qnewers

provided.

There

are

many

observable

t h i n g s

people do

t h a t

we

do

n o t

h o w

how to

simulate,

Semantics

.-

1x1

this

s e c t i o n

we

w i l l

d i s c u s s ,

very

b r i e f l y ,

same

p e r t i n e n t

aspects

o f SCBO a

mixed-initiative

i n s t r u c t i o n a l

system.

More

d e b i l e d

dilseussions

a r e

pmvided

i n

CarboneIlP 3 p and Wamock

14

and C o l l i n s ,

.

Several

data

bases

c u r r e n t l y

e x i s t :

m e

is

&out

t h e geography

of

South

America,

anoeher

a b o u t t h e ARPA n e t w o r k , and a t h i r d

about

a t e x t - e d i t i n g

system called

NLS.

SCHOLAR s knowledge about any subject matter is

in

the form

of

a

s t a t i c

semantic

network

of

f a c t s , concepts,

and

procedures.

T h i s

12

i a

a

modified and

extended

nemork

a

la

Q u i l l i n

and

has a r i c h

~ i a l o g u e

w i t h

SCHOLAR take8

place

$XI a aeubaet

of ~ n g l i s h

t h a t

is limited

mainly

by

SCHOLAR%

currently

primitive

syntactic

capabi

l i t i e s .

In tutorial

fashion,

t h e

system

uses

its

semantic

n e t w o r k

to

generate

t h e

material

it presents,

t h e questions

it

asks,

and

the

corrections

it

makes.

At

any

time

SCHOLAR

accepts

questions and

r

ests

f r o m

the

s t u d e n t ,

and

generates

responses

based

on its

semantic

network, making

c a l c u l a t i o n s

and

f nferences

of

d i f f e r e n t

types t h a t i t h a s

been

progaamed

to

handle.

The

dialogue

is

unanticipated,

and i s dependent

on

t h e

student's

rsspanses,

questions,

and

requests.

Figure 1

presents

a

sample

protocol

of

an

i n t e r a c t i o n

w i t h SCHOLAR:

the

person

types

a f t e r

t h e

a s t e r i s k and

returns c o n t r o l

to the system

w i t h a

second

a s t e r i s k ,

F i g u r e 2

shows

some excerpts

of SCHOLAR'S

semantic

network.

Properties,

none of which

are

obligatory, can

have

as

values

( s i n g l e

words

_{u s u a l l y}

_English

_words

d e f i n e d

elsewhere

in

the

n e t w o r k ) ,

n

_ers,

_{different types}

_of

_{l i r t s ,}

_and

_{o t h e r p r o p e r t i e s .}

Attributes

are usually English words,

but

there

is

a s e t

of

special

a t t r i b u t e s

f o r

i

r t a n t r e l a t i o n s ,

like

SUPERC

( f o r

superconcept

or

_{superardinate,} e.g.

,

Lima

is

a

city and

a capf

t a l )

,

SUPERP

(for

superpart,

e.g.,

Lima

i a

a

p a r t

of

P e r u

and

South

~ m e r i c a ) , SUPERA

( f o r superattribute,

e , g , , f e r t i l e

refers

to

s o i l

*PROBABLY SPANISH*

"TELL ME SOFIETHING ABOUT PERU* PERU IS A COUNTRY,

IT IS IN WESTERN SOUTH RSCCA w

IT IS ON THE COAST OF THE P A C I F I C ,

THE ITRL IS L a

THE MOUNTAINS ARE W OUS

.

THE REGIONS ARE THE ZON mNGLE, THE ATaC DESERT, AND

THE A N D E S ,

NOW YOU CAN EITHER A N S m R MY QUESTION OR ASK A NEW ONE, *THE A OF ARGENTINA IS PIP(90,000 SQUARE MILES*.

APPROX TELY TRUE, BUT A M O W EXACT VALUE I$ ~ , 2 0 0 , 0 0 0 ,

THE EASTERN BORDERING COUNTRY OF PEhU IS

----

NO.

THE SOUTHERN BORDEMNG COUNTRY OF PERU SfS C H I L E , THE EASTERN BORDERING COUNTRY OF PERU IS BRAZIL,

6 T 1s THE MOST mPOR'PCWT D I F F E CE BE EM PERU AND P

THE P R f Y D I F F E m N C E 4115':

PERU IS IN WESTERN SOUTH RICA.

BUT PAmGWAY IS IN CENTRAL SOUTH RICA,

F i g u r e 11, A Sample Dialogue mtwean S & - I O U R and a S t u d e n t .

PTAL

SUPERC

( r

0 ) CITY

PLACE ( X 01

OP (I: 0 ) GOVElRNmNT Z E D / m

(T

4 ) COUNTRY STATE

PLES (I 2 ) ($EOR BUEN MONTEVIDEO

l3RAS I L I A GEORGECS(FOW BW(P"rA Q U I T O SAPaTfAGO ASUNCTON U / P A Z WASHINGTON)

FERTILE

CONTRA (I 0) BA SUPERA (I 0) SOfE

P E W

SUPERC (I 0) CO

S U P E X (I 1 B ) LOCATION (I 0)

IN (I 0 )

~ O U T N / M E R I C A (I. O ) STE EN

ON (I 0)

COAST (I 0)

OF (I 0) P A C I F I C

IATITWDE (I 4 )

CE (I 0) -18 0 LONG1 E (1 5)

GE (1 0) - 8 2 -68 ~ R D E R I N G / C O ~ T R I E S (I: 1 )

N O R T ~ ~ (I 1) (SL COLOMBIA ECU EASTERN (I 11 IBWiZfL

SOVTHEASTEHaM (1 1) B O L I V I A

SOUTHERN (1 2 ) CHILE e-nTaE (I 1) E X ~

C I T I E S (1 2 )

PR1PaCI:PAE (I 0) (,$L LIMA ClhEEWO AREQUIPA TRUJILLO C H I C I A Y O CUZCO)

LIMA

SUPEW ( I 0 ) C f W CAPITAL SUPERC (I 1 B) PERU SOUTH/ LOCATION (I 0 )

IN (I 0 ) PERU

and capit.1 to countrian and s t a t e s )

,

CONTRA (for contradihflon, e .go

barren contradicts fertile and democracy c o n t r a d i c t s d i c t a t o r s h i p )

,

8

case-structure a t t r i b u t e s l i k e a g e n t and instrument (see Fillmbre )

,

and various o a e r t s ,

The e n t r y f o r l o c a t i o n under P e r u in Figure. 2 i l l u s t r a t e s an

important aspect of SCWOIARBs semantic n e t m 9 k called Under t h e attribute l o c a t i o n there is t h e value S o u t h

plus several s & a t t r i b u t e s ng which is h r d e e f n g countries.

But mder bordering c o u n t r i e s there are s - a t t r i b u t e s like S o r t h e r n

and e a s t e r n , some af w h i c h have several v a l u e s , E&eddfng

describes t h e a b i l i t y $a go dom as deep as n s e e s s a q to describe

a property i n more or less d e t a i l .

In the data base there are a l e 0 tags, s u c h as the ( X 0) a f t e r Pocation and the (1 1) a f t e r b r d e r i n g csmtries, m e a e tags are

e a l l e d or t a g s ( 1 - t a g s ) , and t h e y vary

f r o m 0 to 6 , The lower t h e tag, the mre b p r t a t the piece sf in fa t i o n is, The tags add up a s you ga dam through Pmer e&edded Bevels. One of t h e ways SCHO uses 1 - a g s is to decide w h a t +s r e l e v a n t to say at any g i v e n ti-,

In the rest of this paper, we will d i ~ c u s s how we are u s i n g

SCHOLkR to cope:

wi*

gome of the problems in n a t u r a l s e m a n t i c s . However, Lhero are still many nathral-aewantics problems we have

3 ,

In this section we discuss some a s p e c t s of natural semantic

i n f o m a t i o n and its r e l a t i o n to a r t f F i c i a l intelligence.

3 , f

Imprecise language is an essential characteristic of human PO

comnrun&cation. As Lyons says, *Far from b e i n g a defect a s

some philosophers have suggested, referential 'impreciseness'

...

makes language a more efficient means of communication.

"

T a l k i n g

ut a tall person or a blue-green object does not require

precise s p e c i f i c a t i o n of h e i g h t or spectral characteristics. The imprecision may occur either in communication or storage. If we

say that a colleague receives a large salary, we

may

or may n o t h o w the f i g u r e .

SCHOLAR c u r r e n t l y stores areas and populations in n

form, but it can respond to the fuzzy question "Is Montevideo

large?" w i t h a p e r t i n e n t answer l i k e : 'It is n o t one of the largest cities in South America, but it is the largest c i t y in Uruguay,

"

Here SCWOLaR has found tm s u p e q a r t s , South meriea

and Uruguay* and then campred M n t e v i d w to o t h e r c i t i e s in each

w i t h qespect to population.

However, it is mrce co n f o r people to @tore v a l u e s

t h a t are irarprecise or f l f u z z y ' , what 2adeb19 calls 'linguistic'

' r i c h g , etc. It

seems

to us t h a t

one

must be a b l e to store

e i t h e r precise values

or

f u z z y

values

interchangeably. (In f a c t , SCHO has fuzzy

values

as

well

as

preoiae values

a t o t e d , e . g . ,

t h a t t h e Brazilian Highlands has a large population.) F u r t h e r -

more,

the

procedures

t h a t a c t

upon therre values

must be

f l e x i b l e

enough to d e a l

with

e i t h e r .

3 2

Imprecise statements $re

of

t e n

motivated by incomplete

s p e c i f i c a t i o n .

S i n c e a l l

specifications

c a n be

refined,

they

are

essentially

incomplete. We

store

what

is

necessary, and

if

we

store more,

we o n l y comunieate what is pertinent. SCHOLAR

does this through its I - t a g s . If it

is

asked ' T e l l

me

about Perm," it only gims a f e w

s a l i e n t

f a c t s .

F u r t h e r s p e c i f i c a t i o n

can

be

added by

_{r e f i n i n g e x i s t i n g}

values. l o r a x m p l e , i n s t e a d

of 'blue

we can have 'Navy

bluev,

OB.

'quite

dark

MaPry b l u e " ,tee.

Furbher

specification

can

a l s o

be

added by giving plow properties w i t h

a t t r i b u t e s

somenhat

ortkogonal to previous ones. An example of this

is

'tall man'

veraua

'tall,

h e a v m n

wearing glasses0.

P r o p e r t i e s c a n be

Somawhat related to incompleteness and relevancy is the eference

problem

(see

Olsonl')

Referring to a

colleague,

we

nay 'define8 him a s t h e father

of

Jack and Jill,

or

t h e

author

of

t h a t paper

on

self-referential o t a t m e n t s , or tihe t a l l t h i n f e l l - ufth glasses. We decide on

some

specification _de.pending on the

:ontext, i n c l u d i n g our assumptions about t h e person

we

are talking to. People usually s p e c i f y o n l y to the degree that is needed.

In this sense, every partial specification is a ' d e f i n i t i o n ' .

The problem of context pemades matualb senam8;%es.

D e f i n i t i o n s and s p e c i f i c a t i o n s , anaphoric references, what and

h o w to answer, a l l depend

on

context. Furthe re, there u s u a l l y

e o - e x i s t a range of contexts from overall c o n t e x t to s h o r t - t e r n r u n n i n g c o n t e x t s . For example. at a given time, SCHOLAR may

have the contexts South America, Argentina and Buenos A i r e s r each

w i t h some dwamically a d j u s t a b l e

If

fe. _What is releva*. at any

given t h e depends

on

this contextual hierarchy.

A s t a r t toward making references specific to t h e listener

is possible in a SCHOLAR-type system by u s i n g I-tags (see Collins,

Warnock, and ~ a s s a f

iumeb)

.

_The likelihood t h a t another

pereon

will k n o w about any concept i e raugkily pmwrtional to t h e

importance of

the

c o n c e p t , a s measured by the 1 - t a g s , with

sstfmate .the s o p h i s t i c a t i o n of a permn based

on

the level of tags

of the concepts he m e n t i o n s in h i s conversation. This estimate

then can i n f l u e n c e t h e d e s c r i p t i o n one uses in referring to Some concept. For example, to an unsophisticated listener one might

refer to the " c a p i t a l of Argentina" rather t h a n 'Buenos Aires, a

because t h e I - t a g s f o r the concepts *capitalw and "Argentinaw are lamr than *ose _{f o r} 'Buenos A i r e s F w a s masured from a c o n t e x t such a8 geographyc

In t h e f u g u r e we want to have adjustable c o n t e x t s in S C R O L M ,

eo

mat

it c a n t a l k about ~e ARPA n e t w o r k , say, "from a communi-

c a t i o n s p o i n t of viewm to one p e r s o n and *from a progr f n g p o i n t a%

viewm

to another person. What this e n t a i l s

is

a temporary

a l t e r a t i o n of t h e r e l a t i v e v a l u e s of I-tags t h r o u g h o u t t h e

s a a n t i c n e b o r k , T h w e concepts t h a t are referred to u n d e r $Pae

concept w c o m m i c a t i o n U (such as message c a p a c i t y b i t - r a t e , e e

.

) should be temporarily increased in importance wherever they occur

d a b base, f o r the person i n t e r e s t e d in communication. A

c o r r e s p o n d i n g chmge must be made f o r the p e r s o n i n t e r e s t e d in ing OP any a a e r concept

or

s e t of concepts. This k i n d of sensitivity to tihe interests and background of the person, and

the kind of s e n s i t i v i t y (described above) to the s a p h i s t i c a t i o n

3.4

In

some realm o f discourse such as an a i r l i n e resrsmatians

17 1 5

.ystam

(Hoods )

,

a blocks world (Winograd

F,

or a

lunar

rocks

catalogue (Woods, Kaplan, and ~ a s h - ~ e ~ e r l ~ )

,

there is a closed a e t of objects, a t t r i b u t e s , and rmluel~ to deal w i t h . However,

.ip most real world domains such as those faced by SIR ( ~ a ~ h a e l " ) ,

2

TLC

(quillian'2)

at SCHOLAR (Carbonell )

,

there are open s e t 8 of objects, attributes, and values. It turns o u t t h a t the procedures

and

even the rulae of inferdnce that can be applied are different

i n

closed

and open worlds,

The d i s t i n c t i o n between closed and open assts is one of exhaustiveness and not one of s i z e , For ex le,, the set of

s t a t e s (e.g., Iowa)

.

which is a cYbeed s e t for most people, is

probably larger than the set of cattle breeds ( e a g e r Aolatein), nfrich is an o s e t . However, & p a s e t s t m d to be l k r g e r i n %

general than closed s e t s .

The d i s t i n c t i o n

is

important in a variety of ways. _For example, if there are no basaleic r o c k s stored 191 a _ehasad a t a

base, then it makes sense to aay *Nom to the quelrtian %ere any basaltic rocks brought back?* Bag if no oolcanoea are

stared

for the U. S,, it does not f o l l m that the aur

to the question "Are there any mlcandee

in

the U. S . 7 . A more

appropriate answer i s *I dont.t know.' P u t t h e re, it makes

w i a

l e a s t a l m i n m c ~ n c m t R a t f s n ,

But

it makes no sense to ask

what is t h e m a l l e a t c i t y in B r a z i l or the l e a a t f

in t h e U. S, It uould be an appmpriate strategy for d e c i d i n g

how many flights f r o m Boston to Chitago are nonstop, to consider each flight and count how m n y make O stops. But it would n o t

be an appropriate strategy to consider each person smred in a

1 h P t e d h t a base (such as h s b v e ) , in orden to answer the

question *How many people

in

the U. S. are over 30 years old?"

Within open worlds there are c b s e d s e t s , so t h a t a question l i k e 'How many states are on the Pacific?' makes sense whereae 'How

many cities

are

on

the

P a c i f i c ? " does not. SCAOLAR dais w i t h thie by distinguishing exhaustive s e t s from n~n-exhaustive sets.

We

_{w i l l}

d i s c u s s in S e c t i o n 4 how SdlHOmR Begins to deal with open mrld semantics. We essmtial p o i n t here is that the well- defined pmcadums t h a t are appropriate for a closed world s h p l y

do _{n o t carry}

over

to

an

open world. U n f o r t u n a t e l y , m a t of h

knowledge is open-ended, and

w

people have complex s t r a t e g i e s f o r d e a l i n g with uncertainty and f a c i n g problears such as h o w to apply

new

a t t r i b u t e s

ar

values Lo objects where they haven't applied

in

3 3

me two-valued logic that undatliee t h e propasitional calculus and related approachee to inference c a ~ o t encompass natural ssltwa cs. The tfouble arises because

truth

varies in

degree,

in

t h e , in range,

in

c e r t a i n t y , and in p o i n t o f

view

of

the _observer, when it

is

applied to real-world objects. We will briefly examine some

of

the implications of the multivalued n a t u r e

or

t g u % P a f o r n a t u r a l smaatica,

olVc logic uses quantif i c e t i o n to d i s t i n g u i s h between

t h e universal and the particular, e.g., between "All men are

m r t d W

and

amme

men b v e m r t s , "" But *ere is no allowance

mde for the degrees of t r u t h as between say "Some nren have wartsR and .Some men have ears,* even though o n l y a f r a c t i o n have warts

and a &t a l l haw ears, P w p l e w i l l i n f e r t h a t Mewbn had _ears (given no

info

t i o n to t h e contrary as w i t h Van Gogh)

,

b u t will

not i n f e r *at Nets&an had warts, The i n f e r e n c e

in

*e

fomer

case t r e a t s t h e particular l i k e the universal, because almost all men have ears. The m o r e generally t r w a statement

is,

the

more c e r t a i n t y people a s s i g n to such an inference. There j u s t are n o t many universal t r u t h s to be found o u t i n the cold, c r u e l world,

and so p a ~ p l e make the Best of it,

Degree of truth mries not o n l y w i t h respect ta fuzzy

respects. The aky i s blue, but not all the t h e . The y e l l o w of a l e m n Pa less variable

an

the yellow sf corn, which sometimes b r d e r s O M w h i t e , IBsaton is c o l d in the w i n t e r , but it i s not m%a

cold from the p a i n t of view of an Eskimo. Nixan t o l d us t h a t he didn't know a b o u t the cover-up of Watergate, b u t o n e is only

=re

or lees c a t a i n t h a t he didn't know. What these examples are

designed to show

is

e a t people are u n c e r t a i n about the t r u t h of

any p r o p a i t i o n for a v a r i e t y of reasons. Sometimes people seem to merge a l l the many sources of u n c e r t a i n t y together, but

somethes they can distinguish d i f f e r e n t aspects sf a e i r u n c e r t a i n t y w i t h respect to a s i n g l e proposition.

SCMOLM does not

wow

have m y means f a r representing

u n c e r t a i n t y , b u t t h e n a t u r a l w a y to add such info t i a n is i n

tags s t o r e d along w i t h t h e I-tags.

J u s t

ae w i t h I-tagsp U-tags

can

apply at a l l edded levels of a e data b a s e . B e a u s e we

have ~ t a r t e d on p r s g r ng uncertain in'ferences (discussed below),

it has be~orrme desfraB%e ka represent the u n d e r l y i n g uncertainty

in

cIatxa base as well., %n order ta e v a l u t e how c e r t a i n a n y

inference m y be,

4 , Natural Pfiferenees

W e c l a s s i f y h u m senoantic i n f e r e n c e s i n t o Pour major t y p e s :

5

and Quillian7 and C o l l i n s , Carbonell, and Warnock We do

not

argue

mat

these describe all the inferential

strategies

that people use, but o n l y some of the major v a r i e t i e s . The d i f f e r e n t strategies described

are being

i m p l m e n t e d as subroutines in

SCWO

m i l e

we

think that people have a large eet of such strategies, the n er is probably less than one hundred.

Therefore, d e s p i t e the inelegance of such an approach, we do not

regard it as an e n d l e s s t a s k to encompass the bag of i n f e r e n t i a l t r i c k s a person uses,

In Figure 3 we have included excerpts from tape-recorded dialogues between h n tutares and s t u d m t s to i l l u s t r a t e some

of the more complicated strategies p e o p l e use, and t h e ways they

ca&ine togethere We will discuss examples i n d i v i d u a l l y

be low,

4.1 Deductive Inferences

There are several

transitive

relations

that

people use

frequently to i n f e r that a property of one t h i n g may be a property of the

other.

These include superordinate, superpart, s i m i l a r i t y , proximity, s rdinate, and eubpart r e l a t i o n s .

O f the above types SCHO

now

h a n a e s o n l y s u p r a r d i n a t e

and superpart infersnces, mich are the mast co n. For example,

There

is

some

j u n g l e

in

here

( p o i n t s

to Venezuela) but

this breaks

i n t o

a

savanna

around the O r i n o c o ,

Oh r i g h t , t h a t

is

where they grow t h e coffee up there? I don't t h i n k t h a t the

savanna

is

used f o r growing

coffee. The trouble is the

savanna

h a s a rainy season

and *u can't count on r a i n in general. But I don't know. T h i s area around Sao Paula

is

caf fee regionr and it is sort of g e t t i n g

i n t o

the savanna region there.

A r e there any other areas where o i l is found o t h e r than

Venezuela?

N o t p a r t i c u l a r l y . There

is

same

a i l

o f f s h o r e there b u t in g e n e r a l o i l comes from Venezuela. Venezuela i e t h e

o n l y one t h a t b making any money in o i l .

Is t h e C h a c ~ t h e cattle c a u n t q ?

I

know the c a t t l e c o u n t r y is down there,

I

t h i n k

it's

mre

sheep c o u n t r y . I t @ , e l i k e w e s t e r n Texas

80 i n same sense 1 guess

i t u s

c a t t l e country.

&nd t h e

n o ~ t h e r n

p a r t

oT

Argentina has

a

l a r g e

sort

of semi-arid p l a i n t h a t

e x t e n d s

i n t o

Paraguay. And that's a p l a i n s area h a t

is r e l a t i v e l y

unpopulated.

Because it" p r e t t y drya

f i r s t l o o k under Llanos and failing to find the

info

there, will look under Llanosv S U P E X (for superordinate), which

is savanna, and its SUPEW (far superpart), which is Venezuela

and Colombia. A rainy season i s a property of savannas and so the superordinate i n f e r e n c e

p r o v i e s

the answer. The superpart inference is less general because it i s restricted to certain a t t l t l b u t e s such as climate, language, and topography. One would

not want to conclude that the capital of Massachusetts is Washington, D. C., j u s t because Massachusetts is part of the United Shates, Because moat properties of a sdperordinate or

superpart are o n l y generally true, and hot u n i v e r s a l l y true,

e x c e p t i o n s must be stored to preclude an i n c o r r e c t inference (~a~heel'~).

similarity and proximity inferences parallel the eupemrdihate

and superpart inferences, but they carry less c e r t a i n t y . An ex- ample of a person u s i n g a pmximity i n f e r e n c e i s sham in t h e

latter part of the tutor's response

in

E

The tutor f i r s t said t h a t a savanna c o u l d n o t be used f o r growing caf f e e , b u t then he backed o f f this conelusion because of the

proximity of the large Brazilian savanna to the coffee-growing region there. To illustrate a similarity inference: if one

knows a wallaby is like a kangaroo, Only a m a l l e r , then one will

infer that a wallaby probably has a pouch. W e plan to add

will a l s o b e u s e f u l

in

making functioaal analogies

which

are discussed below. The tecently added map f a c i l i t y (Harnock and

~ollins'') which ties together visual and semantic representations,

makes proximity inferences possible, but they are still a way off.

Subordinate and subpart i n f erencee follow a somewhat d i f f e r e n t pattern f r o m the athers discussed. If asked whether South rfca

produces any o i l , a person

will

answer "YesM because Venezuela,

which

is

part of South America, produces oil. But one does n o t

want

to conclude t h a t South merica is h o t because the _mazon

jungle is. We haven't warked out the d e t a i l s of the r e s t r i c t i o n s

on these i n f e r e n c e s a s y e t .

There are o t h e r t r a n s i t i v e r e l a t i o n s that are used to mke deductive i n f e r e n c e s but they are not as p r e v a l e n t a s t h e ones o u t l i n e d here,

Negative i n f o r m a t i o n , such as t h e fact that men do not have

wheels, is n o t u s u a l l y stared b u t rather i n f e r r e d . In a closed

world t h i a p r e s e n t s no prablemt it is r e a e o n d l e to assume t h a t

if

something it? n o t stored, then it is n o t true. In f a c t r e a r l y versions of SCHOLAR say * N o w if asked "fs

o i l

a product of B r a z i l ? "

t

it

is

n o t

true.

People seem to have complex strategies f o r

d e c i d i n g

when

to say @ N o w and

when

to say

@I

don't know. ~4 We

have recently

been

implementing t h e s e

in

SCHOLAR.

One k i n d

of

negative

inference

now in

SCHOLAR

is

a

simple

contradiction procedure. It r e l i e s

on

contradictory values

stored w i t h various concepts: for example, barren c o n t r a d i c t s

fertile, and democracy contradicts dictatorship. Suppose

SCHOLAR i a a s k e d 1 s the Pampas barren?* It would f i n d the s o i l

of the Pampas 1s f e r t i l e , anti since _{f e r t i l e} c o n t r a d i c t s barren, it w u l d say *No. The s o i l of t h e Pampas

is

f e r t i l e . "

There

is

an h p o N a n t class of c o n t r a d i c t i o n s t h a t are n o t

subsumed under t h e procedure

ve. For

example, conrtider the

q u e s t i o ~ *Is Buenos A i r e s a c i t y in Brazil?. The fact t h a t

Buenos Aires

is

n o t among t h e c i t i e s of B r a z i l

is

no reason to

say "Norw because there are eitiea

in

B r a z i l , such as Cor

which are not stored. But there are three f a c t s that ether

make a

contradiction

possible$

(1) Buenos A i r e s

is

located

in

Argentina, (2) c i t i e s only have one l o c a t i o n , and 3 ) Argentina and Brazil are mutually exclusive. We can i l l u s t r a t e tne

necessity

for

conditions ( 2 ) and ( 3 ) : (2) even though Portuguese

is the language of Portugal, it is a l s d the language of B r a z i l

e , language can have mbre than one location]; ( 3 ) even though

America and Brazil

are

not

mutually

exclusive). Making an

incorrect negative

inference about

cities

w i t h

more

than

one

location (e.g. Kansas C i t y ) or different c i t i e s w i t h the same

name

(Rome,

New

York,

and

Rome,

Italy)

is

precluded by

storing

bath locations specifically, just

as

with deductive inferences

.

The strategy we

have

worked o u t and implemented to find different

contradictions of t h i s kind

is

f a i r l y complex.

Failure to find a c o n t r a d i c t i o n leads

to anather

k i n d of negative

i n f e r e n c e

people -

use

which

we c a l l

the lack-of -knowledge

5

i n f e r e n c e

( C o l l i n s , Carbonell and Warnock

.

Ex

le 2

of

F i g u r e 3 shows the tutor using this strategy. The baeiB of t h e

t u b r w s i n f e r e n c e is this: since he knows as muck a b u t o t h e r

Sou*

merican

countries

as he knows about Venezuela, it

is

a

p l a u s i b l e b u t u n c e r t a i n i n f e r e n c e that. if other countries produced a i l ,

he

would h o w about L t . ( h i s conclusion was at l e a s t

somavhat wrong, because there are

in

fact several other aountries in South A m e ~ i c a that produce o i l . though far those countries oil is not nearly so i m p o r t a n t a% it is f o r Venezuela.)

Sach a s t r a t e g y

is

c u r r e n t l y b e i n g implemented

in

SCHO

in

the following way: If asked a q u e s t i o n l i k e *Is o i l a product of

Uruguay?" where no a i l

is

stored, SCHOLAR can look f o r oil under

a i d l a r object8 k e g . , Venezuela

or

B r a z i l ) or objects w i t h the

Venezuela (say w i t h an 1 -tag of 3 ) and i f it has enough

t i o n stored a b u t Uruguay (up to

an

I-mg of 8, s a y )

to know about oil

if

it were a t a l l important, then it can i n f e r t h a t Uruguay probably has no

I

.

The degree of c e r t a i n t y

expressed

in

the answer ~ h o u l d depend

on

the difference in I - t a g s

en

the depth of what it k n m s ut Uruguay and the l e v e l at

which o i l

is

stored w i t h s M l a r ob j e c t a . If SCHOLAR can f i n d no

similar objects

t h a t

have property

in

question, as with "Is sand a product of Uruguay?" the appropriate answer is something

like *I donot know _whether sand

is

a product of any country in

South America." The 'heaakk-of-knowledge inference is based on the assumption t h a t one's knowledge is fairly coneistent for similar

o b j e c t s .

4 . 3 Functional Inferences

F m c tional inferences are ca n

in

t h e dialogues we collected

6

(Ccllins, Wainock, and Passafiume ) . Examples 1, 3, and 4 in

Figure 3 i l l u s t r a t e the three different ways m have s e e n people

use functional knowledge: in quasi calculations, in analogies,

Functional knowledge, which includes knowledge about func- t i o n a l determinants and their i n t g r a c t i o n s ,

i e

learned, j u s t as

is

f a c t u a l knowledge, and therefore

is

stored in SCHOLAR'S d a t a base under concept8 such as climate or a g r i c u l t u r a l products. We w u P d argue t the representation of f u n c t i o n a l knowledge

should be in a f o m that different procedures can use. One

problem is to find a why to represent such knowledge in SCHOLAR so t h a t it can be more or less precise, and s t i l l be accessible to d i f f e r e n t s a r s m ~ n e s that. infer anmers to q u e s t i o n s or t h a t describe the functional relation to students.

FmctPanaB c a l c u l a t i o n s can be used Pa both. a p o s i t i v e and negative m y . One simple positive f u n c t i o n now in SCHOLAR

calculates the climate of a place if the information is not

s t w e d , Based ow the major functional dete nawts of climate,

which are l a t i t u d e m d a l t i t u d e , SCMOUR will i n f e r *ether t h e

climate is t r o p i c a l , sub-*apical, temperate, or cold/palar. A

n e g a t i v e use 0% c a l c u l a t i o n based on t h e a g r i c u l t u r a l products

e m ~ t i ~ n

is s h o w in

me

first pa9$ of t h e tutoro s aaamr in

Example 1. The f u n c t i o n a l determinants of agricultural products

Pwelude the c l $ m a & e , s o i l , and r a i n f a l l , The t u t o r p i e k e d the

l a c k of r a i n as a basis f o r a t e n t a t i v e wNo.w N q a t i v e cabcula-

t i a n s do n o t require a s precise knawledge as positive calculations. They often only require t h a t one or two of the f u n c t i o n a l

Like funcf i o n a l c a l c u l a t i o n e , functional analogies can be

m e i t i v e

or

negative. Example 3 shows the tutor making a positive

functional analogy, again

with

the agricultural products function. Phere he thought of a region, western Texas, t h a t matched the

2haeo En terns of climate a d r a i n f a l l , the functiokaf B e t e m i n -

s n t s of c a t t l e raising, Since he knew that western W x a s was

c a t t l e country he i n f e r r e d t h a t t h e Chaco might be as well. A

negative f nnc t i o n a l analogy might have occurred if the student had a s k e d whether the Chiieo produced rubber. S i n c e the zon jungle and Xndonesfa produce rubber, the tutor could have s a i d

W N ~ W on t h e basis of the m i s m t c h between the Chaea and those

r e g i o n s , w i t h respect to climate and r a i d f a l l .

A positive and _{n e g a t i v e} analogy subroutine has been

implemented in SCHOLAR. It is a f a l l b a c k strategy to be used icf them is not enough i n f o t f s n stared c a l c u l a t e a e

f u n c t i o n a l relationship. For a functional analogy it is only

necessary to know t h e functionally relevant a l t r f i u t e s and Uaeir

relative importance. Then SCHOLAR looks to see if it knows any

similar objects where the property in question is in f a c t stored.

I t tries to find a match or a mismatch by comparing t h e g i o e n object and the similar object w i t h respect to their v a l u e s on

t h e Eunctiollally relevant a t t r i b u t e s . P e o p l e frequently uae

The laet example in Figure 3 sh&s the use o f a f u n c t i o n a l r e l a t i o n to anewer a @Whyw question. T h e population density of

a place depends on an indefinite set o f functional determinants: climate, soil, and rainfall are major o n e s but d i s t a n c e from the

sea, the par'icular c o n t i n e n t , presence o f valuable minerals, all c o n t r i b u t e in different ways. The tutor p i c k e d one determinant

t had a value inapprgpriate f o r a large p o p u l a t i o n density and gave t h a t as a reason. By contrast a geographer c o u l d

probably m i t e a mole treatise on why t h e Chaco has a l o w

pspulation d m s i t y , W h a t we a s p i r e POP SCHOLAR to do 1s w h t the

tumr

d i d p t h a t is, to pick o n e or t w o of the

rnarmr

d e t e m i n a n t a

w i t 2 1 a p p r o p r i a t e v a l u e s and g i v e those as a reason.

Q,4 Xndubtive Inferences

We m e n u o n i n d u c t i v e i n f e r e n c e s here o n l y because t h e y a r e

a

major c l a s s of human i n f e r e n c e . We have n o t y e t t r i e d to progrm *em in I A R since they occur m s t l y i n s t a r i n g

r a t h e r than retrieving i n f o m a t i o n , The generalization and d f s - c r i d n a t i o n processes underlyfwg i n d u e t i on have been dasenssed

4.5

The inferential procdaserr described can combine

in

a ~ r i e t y

of

ways. For inehance, contradictions can combine w i t h deductive

inferences. SCBOWLR will anmer a question like "Is the A t l a n t i c orange?" w i t h wNo, it

is

blue, because it finds blue f e atorad with the S U P E R . , ocean. A l s o one fanctfanal inference may c a l l

another. If the agricultural products function needs a value for

bhe climate of some region, it could call the climate function to compute it.

A more important way t h a t inferences combine s h o w s up when d i f f e r e n t strategies reach independent canclusiona about the same question. A good example is Example

I.

in F i g u r e 3 . Them a

negative functional i n f e r e n c e , w k t h an implicit lack-of-knowledge ihferance,

f i e s t

led to a tentative * N o u m a w e r , but *en a

proximity inference produced a p o s s i b l e 'Yeem a n s w e r , and so the tutor backed o f f his earlier ""No." men several i n f e r e n c e s co-ine to yield the s m e conclusion, a e y increase the c e r t a i n t y of

answer, and when they produce oppasite e o n e l u s l o n s , a e y d e r e s s e

the certainty,

There are a n er of sources of m c e r a i n t y in i n f e r e n t i a l

procedures. Uncertainty can derive from the s i z e sf Ule diffemrence between I-tags &n Ule lack-of-knowledge inference, it can derive

d a t e m i n a n t s , andl as m discussed

earlier,

it can derive

f m m

Uae

degree of certainty about the info t f o n storad. Theae s o u c e s

of uncertainty may be c i n e d ta produce an o v e r d l uncertainty

9

(see for example tling T h i s overall uncertainty ia important so t h a t long, tenuous chains of reasoning are not pursued to their pointless end, and so t h a t the degree o f u n c e r t a i n t y in the a n s w e r c a n be S a d i c a t e d to the student,

5, Copcluaions

what we have t r i e d Ca show in t h i s paper i s t h e fuzzy, ill- def Fned, mcer-in n a t u r e of much of h n h o w l e d g e and thinking.

We want = H O U R ta be just a a f u z x y - t h i n k i n g a s we are.

6,

This paper was s t a r t e d by Jaim R. Carbonell who died

quddsnly F & a a r y 2, 1973, I have completed it as best I c o u l d following h i s o u t l i n e , I w a n t to thank E l e a n o r H, Wamock wha

helped me w i t h the e d i t i n g and D a n i e l 6 , B o b r a w and M s s Quillian w h o have contributed many ideas to our work. The programing of

v a r i o u s s u b r o u t i n e s described in the paper was done by Nelleke A i e l l o , Jaime G. Carbonell, Susan k. Greesser, Mark L o Miller.

T h i s research was supparted

in

part by the Office o f Naval

-search, XmPo t i o n Systsma, uader Contract No, NO00f4-70-C-0264, and a l s o in part by Ule Office sf Naval Research, Permnnel and Training, under Contract No, M00014-7%-C-0228, and by the Air Fame Systams Command, Electronic S y s t m s Division, under

J. D. Becker, "A Model for the Encoding of E x p e r i a t i a l I n f o t i o n , *

in

R. _Schank & K. Colby ( B d s . ) ,

,

Freeman., S a F r a n c i ~ c o

J. R. Carbonell, * ~ i x e d - I n i a t i a t i w Man-Computer

Instructional Dialogues. * P h . L Thesis, M.I,T,, Dept. of

Electrical Engineering (June 1970).

J. R. C a r b o n e l l . *A. I. in C . A . I . : An Artificial-Intelligence

Approach to Computer-Ass is-d I n s t m c t ion" IEEE T r a n s . on

I V o l , MMS-11, No. 4 (Dece&er 1970).

J, R. Carbonell, *Artificial Intelligence and Large

I n t e r a c t i v e Man-Computer Systems " Proceedings of 1971 IEEE

Systems, M a n , and Cybernetics Conference, Anaheim, C a l i f o r n i a

A, M, Collins, J, R. C a r k n e l l , and E, H. Warnoek, "Semmtic

I n f e r e n t i a l Processing by Computer" in J. Rose (Ed.) Advances

Wrdaw 6 Breach, Mndon ( 1 9 9 4 ) .

A. M. C o l l i n s , E. H, Warnock, and J. J Q Passafiume, "Analysis and S y n t h e s i s of Tutorial Dialoguesn in G. B o w e r (Ed. )

.

-

The

T Q A. M.

Collins and M.

R e

QuiL'lian,

"How to

Make

a

Language

Usera

in

E. Tulving

and

W e

Donaldeon

(Eds.

)

Academic

Press, New York

(1972).

8. C.

Pillmre,

"The Case for Casew

in

Bach

and

Hams ( ~ d a . )

,

M o l t , Mnehart, and

Winston, N. Y e (1968).

9. R e Kling, "FUZZY P NER: Computing Inexactness

in

a

Procedural

Problem

Salving Language.

"

U n i w r s i t y of

Wisconsin Technical Report NO. 168 (February 1973).

ridge University

Press,

Cambridqe, England (1968)

.

.LA o D. R e Olson, wLanguage and Thought: Aspects of a Cognitive

Theory of SemanticsY I

Vole

77, No. 4 ( J u l y 1 9 7 0 1 .

12. M. R. Q u i l l i a n , 'The Teachable Language Comprehender: A Simulation

Program

and Theory of Languagew CACM,

-

Vol.

12,

No. 8 (August 1969).

13. B.

Raphael,

*SIR: A Computer Program f o r Semantic Information

Retrievaln in

M. L.

Minsky

(Ed. ) Semantic

Information

Proceers-

6ng7 M.1.T. Press, Cambridge, Mass. (1968).