kmerican Journal of Computational Linguir tics 
Microfiche 3 
Jaime R. Carbonell andLAl'lan M. Collins 
Bolt Besanek and Newman Inc. 
Cambridge, Massachusetts 
01974 by the Association for Computations 1 Linguistics 
ABSTRACT 
This papar discu~see human semantic knowledge and proceesing 
in terms of the SCHOLAR system. In one major section we discuss 
the imprecision, the incompleteness, the open-endedness, and the 
uncertainty of psopl,eqs knowledge. In the other major section we 
diecuss strategies people use to make different types of deductive, 
negative, and functional inferences, and the way uncertainties 
combine in these inferences. 
Irapreeision can occur either in memory or in camunication. 
SCHOLAR can have precf ae values or fuzzy values stored, and its 
procedures can, to eome extent, deal with fuzzy questions when 
precise valtues are stored, and with precise queltions, when fuzzy 
values are stored. Embedding allows info tion to be specified 
in the data base to any level of detail or preci sfon. But SCHOLAR 
only camunicatea the nwst important info tion on any topic (as 
rneasured by importmce tags), unless more info tion is requested. 
1.t should also be possible by using importance tags to adjust what 
+jon SCHOLAR communicates, in accord with the sophistication 
and, interests of the listener. 
Inference atrategiee that are appropriate when the complete 
set of object attributes, or values, is known (i .em, in a closed 
world) do not apply when knowledge is incomplete e, in an 
open wrld) . There are a variety of uncert~in inf ercnces that 
people use to circumvent the hales in their knowledge, which are 
being progr ed in SCHOLAR. 
There is a set of transitive relations -- superordinate, 
superpart, s larity, praimity , subordinate, and subpart relations 
-- that people frequently use to make deductive inferences, 
Currently SCHO only handles superordinate inferences (em 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 
Rio is part of Brazil). 
~eductive inferences can be more or leas 
certain (similarity inferences are like suptordinate inferences. 
but leas certain) and can have restrictions on their use (only 
certain attributes transfer an superpart). 
~bn knowledge is incomplete, it is not safe to assume that 
something is not true just because it is not stored. 
Thus an in- 
ference is necessary to decide when to say 'Now and when to say 
'I donBt know,@ There is a eamp1Pcated set sf strategies in 
SCHOLAR to find vatious kinds of contradictions that people case 
to say *NO, " If a contradiction cannot be found, another nega8ive 
inference, called the Wlack-aF-knawledgea inference, f s tried* 
When enough is known abut an abjech it is possible to conclude 
that someaing is not true ut mat object ow the ground8 that 
if it were true, it muld be stared, 
another class of wcertain inferences depends an ill-def ined 
knovledge of functional datenainants , e .g. , that climate depends 
OR latitude and altitude. D%f ferent ways that pople use fmctional 
knowledge irnm1ve fmctianal ealeulatisms (a.g,, if a place has a 
particular latitude, it 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 sme chimte) , and ts answer Why 
questions (egg,, a place hg a particular elfmte because of its 
latitude and altitude), 
Different inf erenees can esIwBine in different ways, Somtims 
one strategy may call another strategy to hind an anewer, men 
different inferences independently reach the sme or differmt con-- 
elusions, aey coItnbiwe to increase or dsreass certainty, The pm- 
ing sf uncertain Bnfermces is wecessqw to make cawutars as 
clever and aa fuzzy-thinking as pmple. 
TABLE OF CONTENTS 
2. The Scholar System as an Environment to Study Natural 
~~ma~~i~~.~~.~.~..~~~~~~~~~~~~~~~.~~~~~~~~~.~..~.~~.~.~~~ 6 
3.1 Imprecision or Fuzziness..oa. ..ooOOoaI)DoOo~OoOoOOOoO~~ 
3.2 Incompleteness, Embedding, and R~levancy............ 12 
3-3 The Reference Problem and Context .................... 13 
3,4 Closed versus Open Worlds oOo~o~o~o~~O~DoDeOoo~d~OODo 15 
3.5 The True-False ~ichotorny and ~uantification ..,...... 17 
4, Natural Infeuen@es.,..o...,P0DoomoeOOo0OO~BOoOOOoo.~.o.o.0 18 
4.1 Deductive ~~f~~e~~e~..oOQ.oO.O~OoOO~~OO~oO~OOOOOO.Oo~~ 
4.2 Negative Infer~~~es..o..o.ooOoOoDooOOoOoOooOooaoO~o.~~ 
4-4 Inductive Inferenceso0.0.00D~000DoD.0..o~a~OOOoOOD~028 
4.5 Combining Inferences and Accumulating Uncertainty ... 29 
% , Introduction 
In this paper we will discuss how to eepresent and process 
infomatioil in a computer in ways that are natural to people. 
This does not mean doing away completely with representations and 
procedures which computers have traditionally used, but adding 
new representations and procedures which they have not used. 
People often store and communicate imprecise, incomplete, 
and unquantified info tion; they aften assert truth or falsity 
in relative terms; and they seldom seem to use rigorous logic in 
their inferential processes. Because of these conditions, people 
s,eem to have an almost infinite info tion processing capacity, 
with inference making and problem solving abilities more refined 
and far more flexible than any existing computer program. 
Now man we study aese hman eapabiLities in order to make 
our machines show similar perfbnnance? A combination of 
approaches is perhaps best. Observation of people ' s behavior, 
introspection, some experimentation, protocol analysf s, and 
synthesis of computer programs can all be vhluable technil~ues. A 
6 
recent paper (CalZins, mrnock and Passafime ) discusses a tech- 
nique for combining protocol analysis with program synthesis as 
applied to tutorial dialogues. The synthesis directs what to 
analyze, and the strategies observed in the analysis are evaluated 
by synthesis, in a kind of feedback Poop 
We have been using 
the SCHOLAR system in this way as a vehicle for experimentation 
with natural semnties. 
Before we discuss some of the major problems in natural 
semantics, we will briefly describe the SCHOLAR system, since it 
ie the enviroment for our research. A word of caution though: 
we are only trying to develop some insights, without attempting 
to be exhaustive. More questions wil.1 be raised than qnewers 
provided. There are many observable things people do that we do 
not how how to simulate, 
Semantics 
.- 
1x1 this section we will discuss, very briefly, same pertinent 
aspects of SCBO a mixed-initiative instructional system. More 
debiled dilseussions are pmvided in CarboneIlP 3p and Wamock 
14 
and Collins , . Several data bases currently exist: me 
is &out the geography of South America, anoeher about the ARPA 
network, and a third about a text-editing system called NLS. 
SCHOLAR s knowledge about any subject matter is in the form of a 
static semantic network of facts, concepts, and procedures. This 
12 
ia a modified and extended nemork a la Quillin 
and has a rich 
internal structure with a well-deQEined syntax. 
~ialogue with SCHOLAR take8 place $XI a aeubaet of ~nglish that 
is limited mainly by SCHOLAR% currently primitive syntactic 
capabi lities. In tutorial fashion, the system uses its semantic 
network to generate the material it presents, the questions it 
asks, and the corrections it makes. 
At any time SCHOLAR accepts 
questions and r ests from the student, and generates responses 
based on its semantic network, making calculations and f nferences 
of different types that it has been progaamed to handle. The 
dialogue is unanticipated, and is dependent on the student's 
rsspanses, questions, and requests. Figure 1 presents a sample 
protocol of an interaction with SCHOLAR: the person types after 
the asterisk and returns control to the system with a second 
asterisk, 
Figure 2 shows some excerpts of SCHOLAR'S semantic network. 
Properties, none of which are obligatory, can have as values 
(single words 
usually English words defined elsewhere in the 
network), n 
ers, different types of lirts, and other properties. 
Attributes are usually English words, but there is a set of 
special attributes for i rtant relations, like SUPERC (for 
superconcept or 
superardinate, e.g. , Lima is a city and a capf tal) , 
SUPERP (for superpart, e.g., Lima ia a part of Peru and South 
~merica), SUPERA (for superattribute, e,g,, fertile refers to soil 
and soil refers to topagraphy), APPLIEDFO (color applies to things, 
*PROBABLY SPANISH* 
"TELL ME SOFIETHING ABOUT PERU* 
PERU IS A COUNTRY, 
IT IS IN WESTERN SOUTH RSCCA w 
IT IS ON THE COAST OF THE PACIFIC, 
THE ITRL IS L a 
THE MOUNTAINS ARE W OUS . 
THE REGIONS ARE THE ZON mNGLE, THE ATaC DESERT, AND 
THE ANDES, 
NOW YOU CAN EITHER ANSmR MY QUESTION OR ASK A NEW ONE, 
*THE A OF ARGENTINA IS PIP(90,000 SQUARE MILES*. 
APPROX TELY TRUE, BUT A MOW EXACT VALUE I$ ~,200,000, 
THE EASTERN BORDERING COUNTRY OF PEhU IS ---- 
NO. 
THE SOUTHERN BORDEMNG COUNTRY OF PERU SfS CHILE, THE EASTERN 
BORDERING COUNTRY OF PERU IS BRAZIL, 
6 T 1s THE MOST mPOR'PCWT DIFFE CE BE EM PERU AND P 
THE PRf Y DIFFEmNCE 4115': 
PERU IS IN WESTERN SOUTH RICA. 
BUT PAmGWAY IS IN CENTRAL SOUTH RICA, 
Figure 11, A Sample Dialogue mtwean S&-IOUR and a Student. 
(Student lnputs are enclosed by asterisks.) 
PTAL 
SUPERC (r 0) CITY 
PLACE (X 01 
OP (I: 0) GOVElRNmNT 
ZED/m (T 4) COUNTRY STATE 
PLES (I 2) ($EOR BUEN MONTEVIDEO 
l3RAS ILIA GEORGECS(FOW 
BW(P"rA QUITO 
SAPaTfAGO ASUNCTON U/PAZ WASHINGTON) 
FERTILE 
CONTRA (I 0) BA 
SUPERA (I 0) SOfE 
PEW 
SUPERC (I 0) CO 
SUPEX (I 1 B) 
LOCATION (I 0) 
IN (I 0) 
~OUTN/MERICA (I. O)  STE EN 
ON (I 0) 
COAST (I 0) 
OF (I 0) PACIFIC 
IATITWDE (I 4) 
CE (I 0) -18 0 
LONG1 E (1 5) 
GE (1 0) -82 -68 
~RDERING/CO~TRIES (I: 1 ) 
NORT~~ (I 1) (SL COLOMBIA ECU 
EASTERN (I 11 IBWiZfL 
SOVTHEASTEHaM (1 1) BOLIVIA 
SOUTHERN (1 2) CHILE 
e-nTaE (I 1) EX~ 
CITIES (1 2) 
PR1PaCI:PAE (I 0) (,$L LIMA ClhEEWO AREQUIPA TRUJILLO 
CHICIAYO CUZCO) 
LIMA 
SUPEW (I 0) CfW CAPITAL 
SUPERC (I 1 B) PERU SOUTH/ 
LOCATION (I 0) 
IN (I 0) PERU 
Figure 2, Four Partial Entries from SCHOLAR'S Georgraphy Da'ta 
Base, 
and capit.1 to countrian and states) , CONTRA (for contradihflon, e .go 
barren contradicts fertile and democracy contradicts dictatorship) , 
8 
case-structure attributes like agent and instrument (see Fillmbre ) , 
and various oaerts, 
The entry for location under Peru in Figure. 2 illustrates an 
important aspect of SCWOIARBs semantic netm9k called 
Under the attribute location there is the value South 
plus several s&attributes ng which is hrdeefng countries. 
But mder bordering countries there are s-attributes like Sorthern 
and eastern, some af which have several values, E&eddfng 
describes the ability $a go dom as deep as nseessaq to describe 
a property in more or less detail. 
In the data base there are ale0 tags, such as the (X 0) after 
Pocation and the (1 1) after brdering csmtries, meae tags are 
eal led or tags (1-tags), and they vary 
from 0 to 6, The lower the tag, the mre bprtat the piece sf 
in fa tion is, The tags add up as you ga dam through Pmer 
e&edded Bevels. One of the ways SCHO uses 1-ags is to 
decide what +s relevant to say at any given ti-, 
In the rest of this paper, we will di~cuss how we are using 
SCHOLkR to cope: wi* gome of the problems in natural semantics. 
However, Lhero are still many nathral-aewantics problems we have 
not touched, 
3, 
In this section we discuss some aspects of natural semantic 
infomation and its relation to artf Ficial intelligence. 
3,f 
Imprecise language is an essential characteristic of human 
PO 
comnrun&cation. As Lyons says, *Far from being a defect as 
some philosophers have suggested, referential 'impreciseness' ... 
makes language a more efficient means of communication. " 
Talking 
ut a tall person or a blue-green object does not require 
precise specification of height 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 not 
how the figure. 
SCHOLAR currently stores areas and populations in n 
form, but it can respond to the fuzzy question "Is Montevideo 
large?" with a pertinent answer like: 'It is not one of the 
largest cities in South America, but it is the largest city in 
Uruguay, " Here SCWOLaR has found tm supeqarts, South meriea 
and Uruguay* and then campred Mntevidw to other cities in each 
with qespect to population. 
However, it is mrce co n for people to @tore values 
that are irarprecise or flfuzzy', what 2adeb19 calls 'linguistic' 
variables. This is the case with values like large', 'red', 'ho-t' , 
'richg, etc. It seems to us that one must be able to store 
either precise values or fuzzy values interchangeably. (In fact, 
SCHO has fuzzy values as well as preoiae values atoted, e.g., 
that the Brazilian Highlands has a large population.) Further- 
more, the procedures that act upon therre values must be flexible 
enough to deal with either. 
32 
Imprecise statements $re of ten motivated by incomplete 
specification. Since all specifications can be refined, they 
are essentially incomplete. We store what is necessary, and if 
we store more, we only comunieate what is pertinent. SCHOLAR 
does this through its I-tags. If it is asked 'Tell me about 
Perm," it only gims a few salient facts. 
Further specification can be added by refining existing 
values. lor axmple, instead of 'blue we can have 'Navy bluev, 
OB. 'quite dark MaPry blue" ,tee. Furbher specification can also 
be added by giving plow properties with attributes somenhat 
ortkogonal to previous ones. An example of this is 'tall man' 
veraua 'tall, heav mn wearing glasses0. Properties can be 
specified to any level of detail by embedding, an inherent quality 
of SCHO -type semantic networks. 
Somawhat related to incompleteness and relevancy is the 
eference problem (see Olsonl') 
Referring to a colleague, we 
nay 'define8 him as the father of Jack and Jill, or the author of 
that paper on self-referential otatments, or tihe tall thin fell- 
ufth glasses. We decide on some specification de.pending on the 
:ontext, including our assumptions about the person we are talking 
to. People usually specify only to the degree that is needed. 
In this sense, every partial specification is a 'definition'. 
The problem of context pemades matualb senam8;%es. 
Definitions and specifications, anaphoric references, what and 
how to answer, all depend on context. Furthe re, there usually 
eo-exist a range of contexts from overall context to short-tern 
running contexts. For example. at a given time, SCHOLAR may 
have the contexts South America, Argentina and Buenos Airesr each 
with some dwamically adjustable If fe. 
What is releva*. at any 
given the depends on this contextual hierarchy. 
A start toward making references specific to the listener 
is possible in a SCHOLAR-type system by using I-tags (see Collins, 
Warnock, and ~assaf iumeb) . 
The likelihood that another pereon 
will know about any concept ie raugkily pmwrtional to the 
importance of the concept, as measured by the 1-tags, with 
respect to the overall context. Therefore, it is possible to 
sstfmate .the sophistication of a permn based on the level of tags 
of the concepts he mentions in his conversation. This estimate 
then can influence the description one uses in referring to Some 
concept. For example, to an unsophisticated listener one might 
refer to the "capital of Argentina" rather than 'Buenos Aires, a 
because the I-tags for the concepts *capitalw and "Argentinaw are 
lamr than *ose for 'Buenos AiresFw as masured from a context 
such a8 geographyc 
In the fugure we want to have adjustable contexts in SCROLM, 
eo mat it can talk about ~e ARPA network, say, "from a communi- 
cations point of viewm to one person and *from a progr fng point 
a% viewm to another person. What this entails is a temporary 
alteration of the relative values of I-tags throughout the 
saantic nebork, Thwe concepts that are referred to under $Pae 
concept wcommicationU (such as message capacity bit-rate, ee . ) 
should be temporarily increased in importance wherever they occur 
dab base, for the person interested in communication. A 
corresponding chmge must be made for the person interested in 
ing OP any aaer concept or set of concepts. This kind 
of sensitivity to tihe interests and background of the person, and 
the kind of sensitivity (described above) to the saphistication 
sf Ule person may be the two major alments in the way people 
adapt what they say to the listener, 
3.4 
In some realm of discourse such as an airline resrsmatians 
17 15 
.ystam (Hoods ) , a blocks world (Winograd F, or a lunar rocks 
catalogue (Woods, Kaplan, and ~ash-~e~erl~) , there is a closed 
aet of objects, attributes, and rmluel~ to deal with. However, 
.ip most real world domains such as those faced by SIR (~a~hael"), 
2 
TLC (quillian'2) at SCHOLAR (Carbonell ) , there are open set8 of 
objects, attributes, and values. It turns out that the procedures 
and even the rulae of inferdnce that can be applied are different 
in closed and open worlds, 
The distinction between closed and open assts is one of 
exhaustiveness and not one of size, For ex le,, the set of 
states (e.g., Iowa) . which is a cYbeed set for most people, is 
probably larger than the set of cattle breeds (eager Aolatein), 
nfrich is an o set. However, &pa sets tmd to be lkrger in% 
general than closed sets. 
The distinction is important in a variety of ways. 
For 
example, if there are no basaleic rocks stored 191 a ehasad ata 
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 follm that the aur 
to the question "Are there any mlcandee in the U. S.7. A more 
appropriate answer is *I dont.t know.' Putthe re, it makes 
8en.e to ask *at the smallest block in a scene is or the rock 
wia least alminm c~ncmtRatfsn, But it makes no sense to ask 
what is the malleat city in Brazil or the leaat f 
in the U. S, It uould be an appmpriate strategy for deciding 
how many flights from Boston to Chitago are nonstop, to consider 
each flight and count how mny make O stops. But it would not 
be an appropriate strategy to consider each person smred in a 
1hPted hta base (such as h s bve), in orden to answer the 
question *How many people in the U. S. are over 30 years old?" 
Within open worlds there are cbsed sets, so that a question like 
'How many states are on the Pacific?' makes sense whereae 'How 
many cities are on the Pacific?" does not. SCAOLAR dais with 
thie by distinguishing exhaustive sets from n~n-exhaustive sets. 
We will discuss in Section 4 how SdlHOmR Begins to deal with 
open mrld semantics. We essmtial point here is that the well- 
defined pmcadums that are appropriate for a closed world shply 
do not carry over to an open world. Unfortunately, mat of h 
knowledge is open-ended, and w people have complex strategies for 
dealing with uncertainty and facing problears such as how to apply 
new attributes ar values Lo objects where they haven't applied in 
the paete 
33 
me two-valued logic that undatliee the propasitional 
calculus and related approachee to 
inference ca~ot encompass 
natural ssltwa cs. The tfouble arises because truth varies in 
degree, in the, in range, in certainty, and in point of view of 
the observer, when it is applied to real-world objects. 
We will 
briefly examine some of the implications of the multivalued nature 
or tgu%Pa for natural smaatica, 
olVc logic uses quantif icetion to distinguish between 
the universal and the particular, e.g., between "All men are 
mrtdW and amme men bve mrts, "" But *ere is no allowance 
mde for the degrees of truth as between say "Some nren have wartsR 
and .Some men have ears,* even though only a fraction have warts 
and a &t all haw ears, Pwple will infer that Mewbn had ears 
(given no info tion to the contrary as with Van Gogh) , but will 
not infer *at Nets&an had warts, The inference in *e fomer 
case treats the particular like the universal, because almost 
all men have ears. The more generally trw a statement is, the 
more certainty people assign to such an inference. There just are 
not many universal truths to be found out in the cold, cruel world, 
and so pa~ple make the Best of it, 
Degree of truth mries not only with respect ta fuzzy 
variables (see Section 3.1) and quantification. but alao in sther 
respects. The aky is blue, but not all the the. The yellow of 
a lemn Pa less variable  an the yellow sf corn, which sometimes 
brders OM white, IBsaton is cold in the winter, but it is not m%a 
cold from the paint of view of an Eskimo. Nixan told us that he 
didn't know about the cover-up of Watergate, but one is only 
=re or lees catain that he didn't know. What these examples are 
designed to show is eat people are uncertain about the truth of 
any propaition for a variety of reasons. Sometimes people seem 
to merge all the many sources of uncertainty together, but 
somethes they can distinguish different aspects sf aeir 
uncertainty with respect to a single proposition. 
SCMOLM does not wow have my means far representing 
uncertainty, but the natural way to add such info tian is in 
tags stored along with the I-tags. Just ae with I-tagsp U-tags 
can apply at all edded levels of ae data base. Beause we 
have ~tarted on prsgr ng uncertain in'ferences (discussed below), 
it has be~orrme desfraB%e ka represent the underlying uncertainty 
in cIatxa base as well., %n order ta evalute how certain any 
inference my be, 
4 , Natural Pfiferenees 
We classify hum senoantic inferences into Pour major types: 
deductive, negative, fuetionaE, and indueti= inferences, The 
varioue tnes are discussed in gomewhat greater deUP1 in mBlias 
5 
and Quillian7 and Collins, Carbonell, and Warnock 
We do 
not argue mat these describe all the inferential strategies that 
people use, but only some of the major varieties. 
The different 
strategies described are being implmented as subroutines in 
SCWO mile we think that people have a large eet of such 
strategies, the n er is probably less than one hundred. 
Therefore, despite the inelegance of such an approach, we do not 
regard it as an endless task to encompass the bag of inferential 
tricks a person uses, 
In Figure 3 we have included excerpts from tape-recorded 
dialogues between h n tutares and studmts to illustrate some 
of the more complicated strategies people use, and the ways they 
ca&ine togethere We will discuss examples individually 
be low, 
4.1 Deductive Inferences 
There are several transitive relations that people use 
frequently to infer that a property of one thing may be a property 
of the other. These include superordinate, superpart, similarity, 
proximity, s rdinate, and eubpart relations. 
Of the above types SCHO now hanaes only suprardinate 
and superpart infersnces, mich are the mast co n. For example, 
if asked *Does the Llanos have a rainy seas~n?~, 
SCHOLAR will 
There is some jungle in here (points to Venezuela) but 
this breaks into a savanna around the Orinoco, 
Oh right, that is where they grow the coffee up there? 
I don't think that the savanna is used for growing 
coffee. The trouble is the savanna has a rainy season 
and *u can't count on rain in general. But I don't 
know. This area around Sao Paula is caf fee regionr and 
it is sort of getting into the savanna region there. 
Are there any other areas where oil is found other than 
Venezuela? 
Not particularly. There is same ail offshore there but 
in general oil comes from Venezuela. Venezuela ie the 
only one that b making any money in oil. 
Is the Chac~ the cattle cauntq? I know the cattle 
country is down there, 
I think it's mre sheep country. It@,e like western Texas 
80 in same sense 1 guess itus cattle country. 
&nd the no~thern part oT Argentina has a large sort of 
semi-arid plain that extends into Paraguay. And that's 
a plains area hat is relatively unpopulated. 
Because it" pretty drya 
Figure 3. Tutor-Student Dialogue Excerpts 
first look under Llanos and failing to find the info 
there, will look under Llanosv SUPEX (for superordinate), which 
is savanna, and its SUPEW (far superpart), which is Venezuela 
and Colombia. A rainy season is a property of savannas and so 
the superordinate inference provies the answer. The superpart 
inference is less general because it is restricted to certain 
attltlbutes such as climate, language, and topography. One would 
not want to conclude that the capital of Massachusetts is 
Washington, D. C., just because Massachusetts is part of the 
United Shates, Because moat properties of a sdperordinate or 
superpart are only generally true, and hot universally true, 
exceptions must be stored to preclude an incorrect inference 
(~a~heel'~). 
similarity and proximity inferences parallel the eupemrdihate 
and superpart inferences, but they carry less certainty. An ex- 
ample of a person using a pmximity inference is sham in the 
latter part of the tutor's response in E 
The tutor first said that a savanna could not be used for growing 
caf fee, but then he backed off 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 amaller, then one will 
infer that a wallaby probably has a pouch. 
We plan to add 
similarity information to SCHOLAR in the near future, because it 
will also be useful in making functioaal analogies which are 
discussed below. The tecently added map facility (Harnock and 
~ollins'') which ties together visual and semantic representations, 
makes proximity inferences possible, but they are still a way off. 
Subordinate and subpart inf erencee follow a somewhat different 
pattern from the athers discussed. If asked whether South rfca 
produces any oil, a person will answer "YesM because Venezuela, 
which is part of South America, produces oil. But one does not 
want to conclude that South merica is hot because the mazon 
jungle is. We haven't warked out the details of the restrictions 
on these inferences as yet. 
There are other transitive relations that are used to mke 
deductive inferences but they are not as prevalent as the ones 
outlined here, 
Negative information, such as the fact that men do not have 
wheels, is not usually stared but rather inferred. In a closed 
world thia presents no prablemt it is reaeondle to assume that 
if something it? not stored, then it is not true. In factr early 
versions of SCHOLAR say *Now if asked "fs oil a product of Brazil?" 
just because oil isn't storad for Brazil. But in the real warld, 
the fact Ulat somthing is not stored does not necessarily mean 
t it is not true. People seem to have complex strategies for 
deciding when to say @Now and when to say @I don't know. ~4 We 
have recently been implementing these in SCHOLAR. 
One kind of negative inference now in SCHOLAR is a simple 
contradiction procedure. It relies on contradictory values 
stored with various concepts: for example, barren contradicts 
fertile, and democracy contradicts dictatorship. Suppose 
SCHOLAR ia asked 1s the Pampas barren?* It would find the soil 
of the Pampas 1s fertile, anti since fertile contradicts barren, 
it wuld say *No. The soil of the Pampas is fertile." 
There is an hpoNant class of contradictions that are not 
subsumed under the procedure ve. For example, conrtider the 
questio~ *Is Buenos Aires a city in Brazil?. The fact that 
Buenos Aires is not among the cities of Brazil is no reason to 
say "Norw because there are eitiea in Brazil, such as Cor 
which are not stored. But there are three facts that ether 
make a contradiction possible$ (1) Buenos Aires is located in 
Argentina, (2) cities only have one location, and 3) Argentina 
and Brazil are mutually exclusive. 
We can illustrate tne 
necessity for conditions (2) and (3) : (2) even though Portuguese 
is the language of Portugal, it is alsd the language of Brazil 
e, language can have mbre than one location]; (3) even though 
Sao Paulo is in South America, it is also in Brazil . South 
America and Brazil are not mutually exclusive). Making an 
incorrect negative inference about cities with more than one 
location (e.g. Kansas City) or different cities with 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 out and implemented to find different 
contradictions of this kind is fairly complex. 
Failure to find a contradiction leads to anather kind of 
negative inference people - use which we call the lack-of -knowledge 
5 
inference (Collins, Carbonell and Warnock . Ex le 2 of 
Figure 3 shows the tutor using this strategy. The baeiB of the 
tubrws inference is this: since he knows as muck abut other 
Sou* merican countries as he knows about Venezuela, it is a 
plausible but uncertain inference that. if other countries produced 
ail, he would how about Lt. (his conclusion was at least 
somavhat wrong, because there are in fact several other aountries 
in South Ame~ica that produce oil. though far those countries oil 
is not nearly so important a% it is for Venezuela.) 
Sach a strategy is currently being implemented in SCHO in 
the following way: If asked a question like *Is oil a product of 
Uruguay?" where no ail is stored, SCHOLAR can look for oil under 
aidlar object8 keg., Venezuela or Brazil) or objects with the 
sme SUPEX and SWEW, ff SCHOWR finds sik stored wia 
Venezuela (say with an 1 -tag of 3) and if it has enough 
tion stored abut Uruguay 
(up to an I-mg of 8, say) 
to know about oil if it were at all important, then it can infer 
that Uruguay probably has no I. The degree of certainty 
expressed in the answer ~hould depend on the difference in I-tags 
en the depth of what it knms ut Uruguay and the level at 
which oil is stored with sMlar ob jecta. If SCHOLAR can find no 
similar objects that 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 that one's knowledge is fairly coneistent for similar 
objects. 
4.3 Functional Inferences 
Fmc tional inferences are ca n in the dialogues we collected 
6 
(Ccllins, Wainock, and Passafiume ). Examples 1, 3, and 4 in 
Figure 3 illustrate the three different ways m have seen people 
use functional knowledge: in quasi calculations, in analogies, 
and in anmer to 'whyw questions. 
Functional knowledge, which includes knowledge about func- 
tional determinants and their intgractions, ie learned, just as 
is factual knowledge, and therefore is stored in SCHOLAR'S data 
base under concept8 such as climate or agricultural products. We 
wuPd argue t the representation of functional knowledge 
should be in a fom that different procedures can use. One 
problem is to find a why to represent such knowledge in SCHOLAR 
so that it can be more or less precise, and still be accessible 
to different sarsm~nes that. infer anmers to questions or that 
describe the functional relation to students. 
FmctPanaB calculations can be used Pa both. a positive and 
negative my. One simple positive function now in SCHOLAR 
calculates the climate of a place if the information is not 
stwed, Based ow the major functional dete nawts of climate, 
which are latitude md altitude, SCMOUR will infer *ether the 
climate is tropical, sub-*apical, temperate, or cold/palar. A 
negative use 0% calculation based on the agricultural products 
em~ti~n is show in me first pa9$ of the tutoro s aaamr in 
Example 1. The functional determinants of agricultural products 
Pwelude the cl$ma&e, soil, and rainfall, The tutor pieked the 
lack of rain as a basis for a tentative wNo.w Nqative cabcula- 
tians do not require as precise knawledge as positive calculations. 
They often only require that one or two of the functional 
deteminants have an inappropriate value. 
Like funcf ional calculatione, functional analogies can be 
meitive 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, that matched the 
2haeo En terns of climate ad rainfall, the functiokaf Betemin- 
snts of cattle raising, Since he knew that western Wxas was 
cattle country he inferred that the Chaco might be as well. A 
negative f nnc tional analogy might have occurred if the student 
had asked whether the Chiieo produced rubber. Since the zon 
jungle and Xndonesfa produce rubber, the tutor could have said 
WN~W on the basis of the mismtch between the Chaea and those 
regions, with respect to climate and raidfall. 
A positive and negative analogy subroutine has been 
implemented in SCHOLAR. It is a fallback strategy to be used 
icf them is not enough info tfsn stared calculate ae 
functional relationship. For a functional analogy it is only 
necessary to know the functionally relevant altrfiutes and Uaeir 
relative importance. Then SCHOLAR looks to see if it knows any 
similar objects where the property in question is in fact stored. 
It tries to find a match or a mismatch by comparing the gioen 
object and the similar object with respect to their values on 
the Eunctiollally relevant attributes. 
People frequently uae 
such analogical reasoning, probably because of the ill-defined 
nature of their knawledge about functional relations. 
The laet example in Figure 3 sh&s the use of a functional 
relation to anewer a @Whyw question. The population density of 
a place depends on an indefinite set of functional determinants: 
climate, soil, and rainfall are major ones but distance from the 
sea, the par'icular continent, presence of valuable minerals, all 
contribute in different ways. The tutor picked one determinant 
t had a value inapprgpriate for a large population density 
and gave that as a reason. By contrast a geographer could 
probably mite a mole treatise on why the Chaco has a low 
pspulation dmsity, What we aspire POP SCHOLAR to do 1s wht the 
tumr didp that is, to pick one or two of the rnarmr deteminanta 
wit21 appropriate values and give those as a reason. 
Q,4 Xndubtive Inferences 
We menuon inductive inferences here only because they are 
a major class of human inference. We have not yet tried to 
progrm *em in IAR since they occur mstly in staring 
rather than retrieving infomation, The generalization and dfs- 
cridnation processes underlyfwg indueti on have been dasenssed 
7' 
in detail elsewhere (~eeker'; WIRS~OP~~~; Collins and Quillian ) a 
4.5 
The inferential procdaserr described can combine in a ~riety 
of ways. 
For inehance, contradictions can combine with deductive 
inferences. SCBOWLR will anmer a question like "Is the Atlantic 
orange?" with wNo, it is blue, 
because it finds blue fe atorad 
with the SUPER., ocean. Also one fanctfanal inference may call 
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 that inferences combine shows up when 
different strategies reach independent canclusiona about the same 
question. A good example is Example I. in Figure 3. Them a 
negative functional inference, wkth an implicit lack-of-knowledge 
ihferance, fiest led to a tentative *Nou mawer, but *en a 
proximity inference produced a possible 'Yeem answer, and so the 
tutor backed off his earlier ""No." men several inferences co-ine 
to yield the sme conclusion, aey increase the certainty of 
answer, and when they produce oppasite eoneluslons, aey deresse 
the certainty, 
There are a n er of sources of mcerainty in inferential 
procedures. Uncertainty can derive from the size sf Ule diffemrence 
between I-tags &n Ule lack-of-knowledge inference, it can derive 
from the degree of match or mismatch in a functional analogy. it 
can derive from the degree of predictivcanees of the functional 
dateminants, andl as m discussed earlier, it can derive fmm Uae 
degree of certainty about the info tfon storad. Theae souces 
of uncertainty may be c ined ta produce an overdl uncertainty 
9 
(see for example tling This overall uncertainty ia important 
so that long, tenuous chains of reasoning are not pursued to their 
pointless end, and so that the degree of uncertainty in the answer 
can be Sadicated to the student, 
5, Copcluaions 
what we have tried Ca show in this paper is the fuzzy, ill- 
def Fned, mcer-in nature of much of h n howledge and thinking. 
We want =HOUR ta be just aa fuzxy-thinking as we are. 
6, 
This paper was started by Jaim R. Carbonell who died 
quddsnly F&aary 2, 1973, I have completed it as best I could 
following his outline, I want to thank Eleanor H, Wamock wha 
helped me with the editing and Daniel 6, Bobraw and Mss Quillian 
who have contributed many ideas to our work. The programing of 
various subroutines described in the paper was done by Nelleke 
Aiello, Jaime G. Carbonell, Susan k. Greesser, Mark Lo Miller. 
Joseph Jo PassaPbume, and Eleanor H, Warnock, AllIan M, Colbfns. 
This research was supparted in part by the Office of Naval 
-search, XmPo tion Systsma, uader Contract No, NO00f4-70-C-0264, 
and also 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 Systms Division, under 
Contract No. F19628-72-C-0163, 

References

J. D. Becker, "A Model for the Encoding of Experiatial 
Info tion,* in R. Schank & K. Colby (Bds.), Computer Models of Thought and Language, Freeman., Sa Franci~co 

J. R. Carbonell, *~ixed-Iniatiatiw Man-Computer 
Instructional Dialogues. * Ph.L Thesis, M.I,T,, Dept. of 
Electrical Engineering (June 1970). 

J. R. Carbonell. *A. I. in C.A.I.: An Artificial-Intelligence 
Approach to Computer-Ass is-d Instmct ion" IEEE Trans. on 
Man-Machine Systems. I Vol, MMS-11, No. 4 (Dece&er 1970). 

J, R. Carbonell, *Artificial Intelligence and Large 
Interactive Man-Computer Systems " Proceedings of 1971 IEEE 
Systems, Man, and Cybernetics Conference, Anaheim, California 

A, M, Collins, J, R. Carknell, and E, H. Warnoek, "Semmtic 
Inferential Processing by Computer" in J. Rose (Ed.) Advances 
in Cybernetics and Systems. Wrdaw 6 Breach, Mndon (1994). 

A. M. Collins, E. H, Warnock, and J. JQ Passafiume, "Analysis 
and Synthesis of Tutorial Dialoguesn in G. Bower (Ed. ) . The 
Psychology of Learning and Motivation. Val. 9, Academic Press 

A. M. Collins and M. Re QuiL'lian, "How to Make a Language 
Usera in E. Tulving and We Donaldeon (Eds. ) Organization of Memory
Academic Press, New York (1972). 

8. C. Pillmre, "The Case for Casew in Bach and Hams (~da.) , 
Universals in Linguistics Theory. Molt, Mnehart, and 
Winston, N. Ye (1968). 

9. Re Kling, "FUZZY P NER: Computing Inexactness in a 
Procedural Problem Salving Language. " 
Uniwrsity of 
Wisconsin Technical Report NO. 168 (February 1973). 

J. Lyons. Introduction to Theoretical Linguistics,
Cambridge University Press, Cambridqe, England (1968) . 

D. R. Olson, wLanguage and Thought: Aspects of a Cognitive 
Theory of SemanticsY I Vole 77, No. 4 (July 1970). 

12. M. R. Quillian, '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 for Semantic Information 
Retrievaln in M. L. Minsky (Ed. ) Semantic Information Proceers- 
6ng7 M.1.T. Press, Cambridge, Mass. (1968). 

14. E. H. Warnock and A. M. Collins, "Semantic ~etworks" ~olt 
Beranek and Ne n Inc. Report No. 2833 (May 1974). 

15. T. Winograd, Understanding Natural Language
Academic Press, New York. (1972). 

16. P. H. Winston, *Learning Structural Descriptions from 
Examples,* Project MAC TR-76, M.I.T. (September 1970). 

17, We A. Woods, vSemantics for a Question-Answering 
System, @ Aiken Computation Laboratory, Report No. NSF-19, 
Harvatd Univefsity (August 1967) . 

B. 8, W, A. Woods, R. M. Kaplan, and Be Nash-Webber, ''The Lunar 
Sciences Natural Language Infomation Systq,* Bo1.t Beranek 
and Ne n Inc., Report No. 2378 (June '1972). 

19. L. A. Zddeh, wOutlining of a New Approach to the Analysis of 
Complex Systems and Decision Processes" IEEE Trans. on Systems, Man, and Cybernetics, Vol. SMC-3, No. 1 (January 1973).
