THE REASONER AND THE INFERENCER 
DON'T TALK MUCH TO EACH OTHER 
Robert P. Abelson 
Yale University 
In this paper, I wish to point out an 
important problem which has largely escaped 
notice in the field(s) of natural language 
processing~ Curiously, it is implicit in 
the topic title of this session, "Reasoning 
and Inference"~ The fact that two terms 
were used in the title instead of one 
implies a meaningful distinction between two 
different types of intelligent processing~ 
The dictionary definitions of reasoning 
and inference are not much help in making a 
sharp distinction, but in actual use in 
artificial intelligence work these are two 
quite different traditions, each laying 
claim to one of the terms~ Reasoning refers 
to the processing of abstract propositions 
to reach abstract conclusions~ Reasoning is 
formal; it is based on the use of logical or 
mathematical rules, or special devices in a 
particular problem-solving domain~ 
Inferencing, on the other hand, refers to 
the processing of concrete facts to reach 
concrete conclusions~ It is informal and 
"common-sensical" (whatever we quite mean by 
that)~ It includes the curren~ vein of work 
on "frames" (Minsky, 1974; Winograd, 1975) 
or "scripts" (Schank & Abelson, 1975) or 
"conceptual overlays" (Rieger, 1975a) in 
which input patterns are matched to big 
memory structures from which further 
expectations are read out~ 
It is conceivable that one of these two 
approaches is "right" and the other "wrong". 
It is also conceivable that the two 
approaches are really the same, if only we 
were smart enough to see how~ Much more 
likely, mechanisms of (at least!) two rather 
different general types can coexist within 
the same intelligent system, each taking 
over the processing burden in its 
appropriate contexts~ 
The question I want to raise concerns 
th~ interface between reasonin~ and 
in~erencing~ I will initially explore 
examples of this interface in the human 
mind, and then comment with respect to 
artificial intelligence. My very tentative 
conclusions will be that psychological 
reasoning and inference processes are 
relatively insulated from one another; that 
this has unfortunate, or at any rate 
peculiar consequences for human thinking; 
that it is tempting to want to design 
artificial intelligence systems wherein the 
reasoner and the inferencer talk to each 
other more, although it is very unclear how 
best to do this; thus this problem may 
constitute a major agenda item for 
artificial intelligence in the next decade. 
M~xin~ abstract aD~ concrete information 
In the course of human affairs, there 
often seem to be cases where lip service is 
paid to abstract values (say, racial 
integration) but behavior is by contrast 
responsive to concrete concerns (say, being 
3. 
fearful for one's own children). As 
political scientist Robert Dahl puts it, 
"lit is\] a common tendency ... of 
mankind ~ to qualify universals in 
application while leaving them intact in 
rhetorics" On the face of it, this contrast 
appears hypocritical~ More charitably, 
however, it may in some cases represent a 
natural difficulty in applying abstract 
principles to concrete cases~ Recent 
experimental studies have demonstrated such 
a difficulty very strikingly. 
Several studies by Kahneman and Tversky 
(1973) have dealt with the use of abstract 
"base rate" information~ Subjects are given 
a personal background description of an 
individual and are asked to rate the 
likelihood that he chose one or another of 
two careers~ For example: 
Jack is a 45-year-old man~ He is 
married and-has four children~ He is 
generally conservative, careful, and 
ambitious. He shows no interest in 
political and social issues and spends 
most of his free time on his many 
hobbies which include home carpentry, 
sailing, and mathematical puzzles. 
This background biography is written so as 
to suggest strongly the stereotypic 
inference that one occupation is much more 
likely than the other (above, engineer). 
Now suppose that base rate information 
is introduced. Some of the subjects are 
told that the given biography was randomly 
sampled from a population of individuals 
with a fixed ratio of the two career 
choices, say 70% lawyers and 30% engineers. 
Rationally, this base rate information ought 
to make the more densely represented 
occupation more likely to be the correct 
characterization of any single individual. 
Yet it turns out that subjects given base 
rate information make occupational 
predictions no different from subjects given 
no base rate information or contrary base 
rate information. That subjects might not 
understand the meaning of the base rate 
information cannot be the explanation of 
this irrational result. A separate group of 
subjects given only base rate information 
and no biography correctly applies the 
statistical odds from the population 
proportions to the single case. 
The explanation for this effect seems 
to be that the concrete impression from the 
biography subjectively wipes out the 
abstract statistical information. It is as 
though the inferential program scanning the 
biography as a single case has no way to 
borrow from or communicate with the 
reasoning program which scanned the base 
rate information. This explanation gains 
credence from the further finding that the 
biography doesn't necessarily have to be 
strongly stereotypic to wipe out the 
influence of base-rate information. 
At ~a recent conference at 
Carnegie-Mellon on "Cognition and Social 
Behavior", papers by Abelson (1975a), 
Slovic, Fischhoff, and Lichtenstein (1975) 
and Nisbett, Borgida, Crandall, and Reed 
(1975) discussed several related phenomena. 
Nisbett et al. gave a number of striking 
experimental findings, and the following 
powerful anecdotal example: 
"Suppose you wish to buy a new car and 
have deczded that ... you want to 
purchase ... either a Volvo or a Saab. As a 
prudent and sensible buyer, you go to 
Consumer Reports, which informs you that the 
consensus of their experts is that the Volvo 
is mechanically superior, and the consensus 
of the readership is that the Volvo has the 
better repair record. Armed with this 
information, you decide to go and strike a 
bargain with the Volvo dealer before the 
week is out. In the interim, however, you 
go to a cocktail party where you announce 
this intention to an acquaintance~ He 
reacts with disbelief and alarm: "A Volvo! 
You've got to be kidding. My brother-in-law 
had a Volvo. First, that fancy fuel 
injection computer thing went out~ 250 
bucks. Next he started having trouble with 
the rear end~ Had to replace it. Then the 
transmission and the clutch~ Finally he 
sold it in three years for junk~ The 
logical status of this information is that 
of the N of several hundred Volvo-owning 
Consumec Reports readers has been increased 
by one, and the mean frequency-of-repair 
record shifted up by an iota on three or 
four dimensions~ But anyone who maintains 
that he would reduce the encounter to such a 
net informational effect is either 
disingenuous .or lacking in the most 
elemental self-knowledge~" 
If we accept this example as indicative 
of a general phenomenon, it may seem to 
imply that people do not use statistical 
information properly~ However, that's not 
quite it, because if the protagonist of the 
story had not gone to the cocktail party, he 
would have used the Consumer Report~ 
information (properly) to buy a Volvo~ 
Rather, like the experiment with the 
biographies, it appears that there is an 
inability to translate information from one 
mode to the other~ If you have an 
abstraction (a good statistical repair 
record) without direct episodes to 
instantiate it, and then in another context 
you are given an episode implying but not by 
itself proving the contrary abstraction, it 
is not so easy to put these pieces of 
information together. You cannot 
comfortably add the new episode to the 
instance set stored under the old 
abstraction; it's the only instance, and it 
doesn't support the abstraction. There is 
no commensurability of information, either 
at a concrete or at an abstract level. How 
would a cleverly designed artificial 
knowledge system cope with such a case? I do 
not know. 
Consider another psychological 
phenomenon involving more reasoning than the 
above example, but still pitting abstract 
against concrete information. In the lingo 
of social psychology's ,,attribution theory" 
(Jones, et al., 1972), both ,,consensus" 
information and ,,distinctiveness" 
information are useful in assigning causal 
responsibility for an event~ Given a 
statement about an actor's response to a 
stimulus ("Mary ran away from the dog"), it 
is possible to locate the cause either 
primarily in the actor ("Mary was fearful") 
or in the stimulus ("The dog was 
frightening"). From simple reasoning 
(setting aside complicating factors) it 
follows that if the given actor responds 
non-distinctively to other similar stimuli 
("Mary is afraid of other dogs") and there 
is non-consensus from other actors toward 
this stimulus ("No one else is afraid of 
this dog") then it is something about Mary 
which has caused the event. On the other 
hand, if the given actor responds 
distinctively to other stimuli ("Mary is not 
afraid of other dogs"), and there is 
consensus from other actors toward this 
stimulus ("Everyone else is afraid of this 
dog"), then it is something about the dog 
which has caused the event~ 
In a formal logical sense, consensus 
and distinctiveness information are 
symmetric and equivalent~ However, it has 
been found in questionnaires posing events 
and asking for attributions (McArthur, 1972) 
that distinctiveness information (what this 
actor does with other stimuli) is more 
powerful than consensus information (what 
other actors do with this stimulus)~ In 
fact, consensus information is curiously 
weak in a number of contexts~ 
Miller, Gillen, Schenker, and Radlove 
(1973) asked college students to read the 
procedure section of the classic Milgram 
(1963) study of obedience, in which subjects 
were found willing to administer 
overwhelming doses of electric shock to 
another person~ Half the subjects were 
given the actual data of the Milgram study 
showing that all subjects administered 
substantial shocks and a clear majority 
continued throughout the experiment, even 
beyond the point of apparent danger to the 
victim's life~ The other half of the 
subjects were left with the naive--and 
typical--expectation that such behavior 
would be quite rare~ Then all subjects were 
asked to rate two individuals, both of whom 
had administered maximal shock, on eleven 
evaluatively loaded trait dimensions such as 
warmth, aggressiveness, etc~ 
Social psychologists have learned to 
interpret the apparently vicious behavior of 
Milgram's subjects as being due to the 
subtly very powerful situational pressures 
toward obedience, pressures with universal 
effect transcending personal values~ If 
everybody does something bad, it is not as 
logical to fault the given individual who 
does it as it would be if he were one of 
only a few individuals who did it~ 
Nevertheless, logic does not prevail in this 
judgment~ The Miller et al~ subjects given 
the true consensus information judge the two 
shock-givers in virtually equivalently 
wicked terms as do subjects not given this 
information~ Somehow subjects are not able 
to adjust their concretely moralistic 
interpretations to take account of the base 
rate of the behavior in question~ The value 
inference is impervious to statistical 
4. 
reasoning. 
It may seem that these examples are 
special in some way, and that perhaps the 
reasoner and the inferencer do converse in 
other contexts. I want to turn now to 
another, rather different context which the 
artificial intelligence worker will find 
somewhat more familiar, but where the same 
puzzle arises. 
How abstraetly should a planner plan? 
One of the vexing issues in artifieial 
intelligenee which everyone is aware of but 
not eager to discuss, is the lack of 
generality of problem contents. There is a 
kind of taeit agreement that it is OK for 
everyone to define his own arena or table 
top or eoneeptual eorpus, and do his thing 
within that context without worrying yet 
about extrapolation. 
I can understand the arguments for this 
permissive attitude--the necessity to start 
somewhere, the danger of doing nothing by 
worrying about everything, and so on--yet 
somehow it still bothers me. Thus I note 
that Terry Winograd's (1972) tour de foree 
with the "blocks world" is applauded for its 
cleverness, but no one to my knowledge has 
tried to generalize any of the,mechanisms to 
other problems. 
Yet other problems arise which seem 
somehow very closely related. For example, 
Charniak's (1975) paper at this conference 
raises questions about the "supermarket 
frame"~ The shopper must get items to the 
checkout counter with the sometime 
instrumentality of a shopping cart. It 
seems clear that there is an analogy between 
the planned transport of a grocery item to a 
checklist counter, enabled by the prior 
placement of that item in a mobile cart, and 
the planned transfer of a block to an empty 
spot on a table top, enabled by the prior 
grasping of that block by a mobile robot 
arm. To be sure, there are some differences 
in side constraints (Shrdlu's arm can only 
hold one thing at a time, while the cart can 
hold many), but it is the similarities which 
interest us, even at this very. simple 
analogic level~ 
Is there any way to take advantage of 
such similarities in the design of a more 
abstract problem understander? Yes, in my 
view. In a recent paper (Abelson, 1975b), I 
attempted to develop a set of intention 
primitives as building blocks for plans~ 
Each primitive is an act package causing a 
state change. Each state change helps 
enable some later action, such that there is 
a chain or lattice of steps from the initial 
states to the goal state the actor desires. 
These act primitives we call 
"deltacts". They ~re distinguished by the 
state they change, and are notated by the 
symbol ~ preceding a state name. There are 
nine deltacts in my system, designed 
(hopefully) such that they could be applied 
to almost any natural world content~ (Some 
of these deltacts make cameo appearances in 
a Sehank (1975) paper in this volume, 
5. 
embellished with his idea of planboxes, and 
with a new deltact added~ He and I have 
been. trying to work out ways to graft the 
deltacts into an expanded Conceptual 
Dependency formalism)~ For purposes of the 
present paper, it is not necessary to 
explain the entire set of deltacts, but 
merely to focus on a portion of the 
machinery needed for the robot-arm and 
shopping basket cases~ 
Consider first the abstract goal (or 
subgoal) that an object X be located at a 
certain place Y. In our abstract system, 
the achievement of this goal requires the 
deltact ~PROX, a change in the proximity 
relations of an object. This deltact has 
five arguments: the actor A, the object X, 
the object's starting place Z, the final 
place Y, and the means M~ In turn, this 
deltact has a set of enabling states, which 
may have to be achieved by other deltacts. 
The Abelson (1975b) paper considers the 
general case where the~PROX may be achieved 
via a "carrier system" (say, an airplane) 
run by agents other than the main actor. If 
we ignore this unnecessary complication 
(along with one other) here, the enabling 
states for ~PROX may be listed as follows, 
where "I" denotes an instrumental device 
(such as a shopping cart) used for means M~ 
a) PROX(A,Z)--the actor must be at the 
starting point; 
b) PROX(X,Z)--the object is at the same 
starting point; 
c) HAVE(I,A)--the actor must have 
instrumental device; the 
d) UNIT(X,I)--the object must be in a 
"unit relation" with the instrumental 
device; 
e) OKFOR(I,M)--the device must be in good 
condition for the transportation means~ 
These enablements are very general~ 
They must be satisfied no matter what the 
content~ Perhaps the level of generality is 
too high, so that some of the states such as 
OKFOR may have a feeling of kluge about them 
(as Rieger (1975b) wonders), but this is not 
a problem for other states such as UNIT 
which have rather clean properties. (There 
are two variants of UNIT: nested ("in", or 
"on"), and joined ("with"). Various nice 
axioms characterize objects in UNIT 
relation, for example, common fate from 
~PROX; transitivity of nesting, etc~) 
Concretizing these conditions to the 
two applications: 
Robot arm moving block: The arm 
must be at the place where the 
block is; the arm must have its 
"hand" (satisfied by definition); 
the object must be put in unit 
relation with the hand; the hand 
must not be damaged (satisfied by 
definition. 
Shoppgr moving an item: The shopper 
must be at the place where the item 
is; the shopper must have (say) a 
cart; the item must be put in unit 
relation with the cart; the cart 
must be all right for whatever it 
is that carts do (roll)~ 
In placing these two applications in 
parallel under a common abstract rubric, we 
clearly see their similarities and 
differences. The robot problem is 
simplified because its hand is always there, 
undamaged~ To satisfy the remaining two 
states enabling the desired ~PROX, it is 
necessary to get the arm to where the block 
is (another ~PROX, or in Winograd's terms, 
MOVE), and to join the arm to the block (a 
~UNIT in our terms, or GRASP in the 
original)~ The supermarket problem is one 
more complex because the shopper may not 
have a cart, or may have a damaged one~ If 
he doesn't have one, he must execute a 
~HAVE, and if it is damaged, he might 
execute a ~OKFOR (fix it) or a ~HAVE of 
another~ Additionally, he must get to where 
the item is (~PROX) and nest the item in the 
cart (~UNIT)~ 
Each of these prior deltacts has its 
own abstract list of enablements~ For a 
~UNIT, these include a couple of OKFORs 
which for grasping are not so klugey: for 
the robot, they map into the conditions that 
the arm be empty and the block have a clear 
top. For a ~HAVE (which can be achieved 
either by taking the object- or having 
someone give it to you), the enablements 
include other PROX and UNIT conditions, etc~ 
With a well-designed system of deltacts 
(which mine probably is not, because of 
various loopholes), it ought to be possible 
to have a planner (or an understander of the 
plans of others) reason appropriately at a 
level well above the specific set of content 
bindings of a particular problem frame or 
script~ Apart from the academic theoretical 
interest of such an abstract system, would 
it ever be useful to human or artificial 
intelligence working in a given practical 
context? 
Well, suppose something went wrong in 
the execution of the usual plans in the 
context~ Suppose, for example, that 
Shrdlu's hand were damaged so that it 
couldn't grasp the smallest blocks--could it 
ever move them? Or suppose that the shopper 
absolutely couldn't find a cart~ 
In these cases, one could imagine the 
inferencer (say, of the shopper) contacting 
the reasoner for help, "Do something! Get 
through your such-thats and Use a theorem or 
something~ I'm in trouble here without a 
cart~" And the reasoner might say, "Well 
let's see, you ask for an instrumental 
device such that you can have it and it's OK 
for forming unit connections with a number 
of these--what did you call them?--grocery 
items, in order to move them a short 
distance? Well, how big and heavy are these 
grocery items, and how soon do you want an 
answer?" And three minutes later the 
6. 
reasoner would come back and say, "Since 
it's an emergency, I'll be more specific 
than I usually like to be~ I recommend 
something flat, with a big surface area, 
that you already have and that could be 
pushed or pulled by hand~ Have you got 
anything on your person?" 
"Well, I have this overcoat~" 
"Oh, good~ Yes, use your overcoat to pull 
the grocery items along the floor to 
the check-out counter~" 
"What?" (Horrified) "But the overcoat will 
get filthy!" 
"Sorry~ That's not my department~ That's 
a low-level inference~ I thought you 
wanted high-level reasonings" 
An alternative fantasy is that the 
reasoner might be so busy working out 
abstract puzzles with its abstract 
mechanisms (say, running through a UNIT 
exercise on the Towers of Hanoi problem) 
that it wouldn't be interested in troubling 
with a silly applied problem~ "Go away and 
don't bother me! I'm thinking about the 
Towers of Hanoi~ What do I know from 
groceries?l" 
Reprise 
The argument I have stated as a devil's 
advocate goes back at least to William 
James, writing in 1890 of the several 
"worlds" of the mind: 
"Every object we think of gets at last 
referred to one world or another~It 
settles into our belief as a 
common-sense object, a scientific 
object, an abstract object, a 
mythological object, an object of 
someone's mistaken conception, or a 
madman's ~ object, and it reaches this 
state~often only after being hustled 
and bandied about amongst other 
objects until it finds some which will 
tolerate its presence and stand in 
relations to it which nothing 
contradicts~ The molecules and 
ether-waves of the scientific world, 
for example, simply kick the object's 
warmth and color out, they refuse to 
have any relations with them~\[The\] 
world of classic myth takes up the 
winged horse;.~the world of abstract 
truth, the proposition that justice is 
kingly, though no actual king be just~ 
The various worlds themselves, 
however, appear (as aforesaid) to most 
men's minds in no very definitely 
conceived relation to each other, and 
our attention, when it turns to one, 
is apt to drop the others for the time 
being out of its account~ 
Propositions concerning the different 
worlds are made from "different points 
of view'; and in this more or less 
chaotic state the consciousness of 
most thinkers remains to the end~ 
Each world whilst it is attended t_oo is 
real after its own fashion; only the 
reality lapses with the attention." 
(James, 1950, p~ 293)~ 
Is James right? If not, wherein? If so, 
is that how artificial intelligence--which 
possibly has design options not available to 
the human mind--would like to keep it? 
References 
Abelson, R~P~ Script processes in attitude 
formation and decision-making~ Presented 
at the Eleventh Carnegie Symposium on 
Cognition, Carnegie-Mellon University, 
April, 1975~ (a) 
Abelson, R~P~ Concepts for representing 
mundane reality in plans~ In D~ Bobrow 
and At Collins (Eds~), Representation 
and Understanding~ New York: Academic 
Press, 1975. (b) 
Charniak, Et Organization and inference in 
a frame-like system of common knowledge~ 
In Theoretical Issues in Natural Language 
Processing~ Proceedings of conference at 
the Massachusetts Institute of 
Technology, June, 1975t 
James, W~ (1890) Principles of Psychology. 
Vol~ 2~ New York: Dover ed'ition, 1950. 
Jones, E~E~, Kanouse, D~E~, Kelley, H.H~, 
Nisbett, RtE~, Valins, St, & Weiner, B~ 
Attribution: Perceiving the Causes of 
Hehavior. General Learning Press, 1972. 
Kahneman, D~, & Tversky, A~ On the 
psychology of prediction~ Psychological 
Review, 1973, 80, 237-251~ 
McArthur, LtA~ The how and what of why: 
Some determinants and consequences of 
causal attribution~ Journal of 
Personality and Social Psychology, 1972, 
22, 171-193t 
Milgram, S. Behavioral study of obedience. 
Journal of Abnormal and Social 
Psychology, 1963, 67, 371-378~ 
Miller, A~G~, Gillen, B., Schenker, C., & 
Radlove, S~ Perception of obedience to 
authority~ Proceedings. 81s.t Annual 
Conventioq of the American Psychological 
Associatiqn, 1973, 127-128~ 
Minsky, M~ A framework for representing 
knowledge~ MIT AI Memo 306~ June, 1974~ 
Nisbett R.E~, Borgida, E~, Crandall, R~, & 
Reed, H~ Popular induction: Information 
is not necessarily informative. 
Presented at the Eleventh Carnegie 
Symposium on Cognition, Carnegie-Mellon 
University, April, 1975. 
Rieger, C~ Conceptual overlays: A mechanism 
for the interpretation of sentence 
meanings in context. To appear in the 
Proceedings of the 4th International 
Joint Conference on Artificial 
Intelligence~ Tbilisi, 1975. (a)~ 
7. 
Rieger, C~ The commonsense algorithm as a 
basis for computer models of human 
memory, inference, belief, and contextual 
language comprehension~ In Theoretical 
Issues in Natura\] Language P__roqessing~ 
Proceedings of conference at the 
Massachusetts Institute of Techn~logy, 
June, 1975~ (b) 
Schank, R.C. Using knowledge to understand~ 
In Theoretical Issues in Natural Language 
processing~ Proceedings of conference at 
the Massachusetts Institute of 
Technology, June, 1975. 
Schank, R~C~, & Abelson, R.P~ Scripts, 
plans, and knowledge~ Prepared for 
presentation at the 4th International 
Joint Conference on Artificial 
Intelligence. Tbilisi, 1975. 
Slovic, P~, Fischhoff, B~, & Lichtenstein, 
S. Cognitive processes and societal risk 
taking~ Presented at the Eleventh 
Carnegie Symposium on Cognition, 
Carnegie-Mellon University, April, 1975~ 
Winograd, 
Language. 
1972~ 
Tt 9nderstanding Natural 
New York: Academic Press, 
Winograd, T~ Frame representations and the 
declarative/procedural controversy~ In 
D~ Bobrow & A. Collins (Eds~), 
~epresentation and Understanding. New 
York: Academic Press, 1975~ 
