What Makes Something "Ad Hoe" 
by Roger C. Sehank 
Yale University 
Department of Computer Science 
New Haven, Conn. 06520 
Only one of the questions posed before this 
session really inspires me to take pen in hand. 
"How general are various formalisms? Are they 
really ad hoc solutions to relatively narrow 
domains?" 
That is not exactly my favorite question. I 
find the thought of having to address it palatable 
only if I can delude myself into believing that 
this is the last time I shall have to deal with 
it. So, proceeding on the basis of that 
delusional belief, I shall begin. 
Ad Hoeness, I have come to believe, is a 
disease that all new theories in the three fields 
in which I consider myself well-versed, namely 
linguistics, psychology and Artificial 
Intelligence, contract at conception, sort of like 
original sin. This would not be so bad if it were 
a disease for which there were a cure, but alas 
there is none. 
We are all familiar with the phrase "beauty 
is in the eye of the beholder." In this case we 
have an instance of "the disease is in the eye of 
the beholder" which of course explains whY the 
cure is so elusive. The beholder rarely wants to 
do anything about it. To discuss this more 
subjectively, let's take a neutral case. Before 
doing so, we shall have to point out what a case 
can be expected to look like. A case of "ad 
hocness" usually fits the form (or should I say 
the "ad hoe" form) 
Theory X is called "ad hoc" by group with 
rival theory Y 
The research described in this paper was supported 
by the Advanced Research Projects Agency of the 
Department of Defense and monitored by the Office 
of Naval Research under contract N00014-75-C-IIII. 
To get to our neutral case, we shall start our 
discussion where X is Conceptual Dependency and Y 
is Transformational Generative Grammar. Before I 
begin, I should note that there are conditions on 
X and Y relative to each other, namely that X must 
be a theory that has been conceived at a date 
later than Y was conceived. Furthermore Y should 
have been dominating some academic field which X 
is seeking to invade. 
What makes a theory X assailable by Y as ad 
hoc? There are a n~nber of criteria: 
1 - X must explain a phenomenon that Y chose 
to ignore and that Y would rather go on 
ignoring since Y could not possibly explain 
it. 
2 - X must be fundamentally at variance with 
y, so that if X were right Y would be 
necessarily wrong. 
3 - X must use different criteria of judgment 
of how a phenomena should be explained than Y 
does • 
The following rules are used for the 
strategy to be followed in labelling an X as 
ad hoc: 
1 - Since X will undoubtedly show how its 
theory explains a given particular 
phenomenon, accuse X's theory of only working 
in that case. This will put the burden of 
proof for generality on X rather than Y and 
also has the desirable effect of putting X in 
the position of not being able to prove 
anything with out proving everything. 
2 - Choose a phenomenon to explain in which 
it is virtually impossible to explain 
everything, thus giving game and set to Y. 
Consider our hypothetical ease where 
Conceptual Dependency is X and Tranformational 
Grammar is Y. An examination of the literature 
will show that criteria I through 3 as well as the 
two available strategies have been used by the 
Transformationalists- In various articles and 
public performances charges of "ad hotness" have 
been raised against Conceptual Dependency. We are 
told that our structures only work for the 
examples we discuss; that we have "no principled 
way of going from a sentence to a 
conceptualization" (Dresher and Hornstein (1976)) 
or that "Schank provides no demonstration that his 
scheme is more than a collection of heuristics 
that happen to work on a specific class of 
examples" (Weizenbaum (1976)). (If the reader is 
wondering how Weizenba~ got to be a 
transformationalist in my view, he need only read 
Weizenbaum's further remarks extolling Chomsky as 
having met the criteria that he claims I have not 
met • ) 
To what extent are these charges valid? To 
not knowing if one can extract a conceptualization 
from any sentence (and its corroborating charge of 
not proving that there exists a right CD diagram 
for any sentence) I plead guilty. But of course, 
I would be less than completely honest if I did 
not also note that there does not exist any theory 
or theorist who would not also have to plead 
guilty. Have the transformationalists shown us 
that they have some principled way of extracting 
conceptualizations from sentences or determining 
the correct representation for any sentence? 
Unless they are keeping their solution as a secret 
plan not to be revealed until after the election, 
I would have to imagine that the answer to this is 
that they do not have a solution to the problem. 
So clearly, they are no more or less ad hoc than 
we are. (Of course I might note here that we do 
have programs that suggest that we can do a large 
class of examples and show that our parsers are at 
least the beginning of some set of principles that 
work, but I won't). 
What about Weizenba~'s attack? Perhaps it 
is all heuristics. To this charge I plead no 
contest. It might be that, in the end, we will 
have built a working program that solves the 
entire natural language problem and it will be 
easily labelled as a grand set of heuristics. 
Won't that be terrible! To quote Dresher and 
Hornstein again, "Not only has work in AI not yet 
made any contribution to a scientific theory of 
language, there is no reason to believe that 
(Al)...will ever lead to such theories". 
And what will they say after success has been 
achieved and the ultimate natural language system 
has been designed? The same thing of course. 
Chomsky himself (personal communication) has 
claimed that such an achievement would be no more 
interesting than the achievement of the 16th 
century clockmakers. 
I mention all this in the hope of pointing 
out that it is not just me and my theories that 
are damned by criticisms of ad hocness. We are 
all damned by them. Our ultimate success would 
not be even recognized much less applauded by 
those who criticize our solutions as ad hoc. 
Suppose every domain we worked on required yet 
another ad hoc solution. This might well be the 
case after all. What would we lose if this 
happened? Nothing at all. That's what artificial 
intelligence is all about. AI is the designing 
and testing of theories about human understanding 
capabilities. There is, at the moment, no reason 
to believe that people solve puzzles the way they 
read newspapers or that they play chess the way 
they answer questions. Of course, we all hope 
that there exist some general mechanisms that 
solve all these problems in some neat way. We 
hope this in large part because we are lazy. We 
would not like to have to work on each problem 
individually. We also hope this because we 
believe our intuitions when they tell us how 
reading a newspaper is a lot like watching a soap 
opera. A word of caution is necessary here. 
Beware of your intuitions. As a child you learned 
how to do each of these things separately and were 
pained to deal with each one of them. Of course, 
we do expect there to be some general principles 
that apply across domains. But if these 
principles are affix - hopping or trace - deletion 
we are all in trouble. 
Part II 
~aving said all this, now let me tell you 
what I actually believe. I do not believe that 
any of our theories are ad hoc. Just because CD 
needed to be modified by causal chaining rules, 
and those by scripts and those by plans and goals 
and themes, and those by triangles, does not mean 
that what we are doing is ad hoc. We are no more 
ad hoc in hypothesizing our primitive elements 
than chemists were in hypothesizing theirs. I do 
not know what the ultimate result will be. How 
many elements make up the correct number, or what 
other kinds of formalisms will need to be added to 
those listed above is still unknown. 
I do know how AI does its research however. 
We build a program to do a small class of examples 
and when we are finished we rip it apart and build 
a bigger and better program to do larger examples. 
In so doing, ad hoe entities (oftimes called 
kludges) cannot survive. If a formalism does not 
keep handling more data it is either abandoned or 
moved down to a special purpose role within a 
larger program. 
Well, in ten years of research by my research 
group what has survived? After ten years and 
probably a hundred different kinds of programs, 
Conceptual Dependency is still with us. It still 
works for us. I challenge any other theory that 
has been programmed to say the same! Is it ad 
hoc? I leave that as an exercise for the reader. 
PART III 
Just to give the reader a feel for the nature of 
ad hoc thinking in AI that I believe to be worth 
espousing, I will now consider a problem that I 
have recently been working on. We have had a 
problem in representing certain kinds of political 
concepts in our old representation, Since we have 
been very concerned with the problem of newspaper 
story understanding it is very important that we 
be able to handle such concepts in a clean 
representation that will facilitate computer 
under stand ing. 
The problem we are attempting to solve can be 
illustrated by looking at a recent New York Times 
headline: "Catawba Indians land claim supported." 
The problem here is to be able to represent what 
"land claim" and "supported" mean. We know that a 
land claim is more than what we might use to 
represent it in Conceptual Dependency. 
Something like "Indians MTRANS land be 
possessed by Indians" is possibly true, but it 
misses the point. A "land claim" is in a sense a 
petition to a higher authority to resolve a 
dispute between two parties. That is, the Indians 
are saying to the U.S. Government, "this land is 
ours". It may not be possible to infer the 
particulars of this land claim. Indians have been 
known to take the land by force, to file doc~nents 
in government offices, to complain to newsmen and 
so on. The important point here is that we really 
need not know• and in most cases a reader would 
not bother to worry about, exactly which method 
has been selected. Rather. a reader feels that he 
understands such a sentence when he has been able 
to identify the relationships and aims of the 
parties involved. 
A program must recognize that a "land claim" 
is a type of petition to a higher authority to 
resolve a dispute about land ownership. We do not 
know who presently owns the land• but we know 
enough about ownership of property to infer that 
there is probably a counter petition of some sort. 
We also know about petitions to authority. They 
usually get resolved by the authority. In this 
case then, "supported" refers to the decision of 
the authority in the case. 
This information can be represented 
graphically by a kind of triangle (example I); 
AUTHORITY 
Catawba Indians 
(c) 
Other 
(a) 
In this triangle (a) represents the dispute 
between the Indians and the owners of the land, 
) represents the appeal to authority to resolve 
the dispute made by the Indians• and (c) 
represents the authority's decision. 
Triangles of this sort have use in 
representing any type of dispute. For example, in 
(2) and (3) such triangles can also be 
constructed : 
(2) Burma appeals to UN to settle border dispute 
with Thailand. 
UN 
Burma ~ Thailand 
(3) John complained to Bill's mother that Bill hit 
him. 
Bill" s Mother 
Jo hn Bill 
Of course• these triangles just suggest the 
basic relationships involved. In order to add 
substance to the bare bones of the triangles we 
shall have to deal with some representational 
issues that are being glossed over here. The 
important point at this juncture is that there is 
an essential similarity across (I)• (2) and (3)• 
that the similarity must be represented in some 
way, and that that similarity can be exploited for 
use in an understanding system. 
The first representational problem we 
encounter in trying to make explicit much of what 
is implicit in the triangle representation is that 
we will need to design a new set of ACTs to take 
care of the various relationships. 
In the primitive ACTs of Conceptual Dependency 
we have a system that represents physical actions 
by using a small set of basic actions that can 
combine in various ways to describe detailed or 
complex actions that underlie seemingly simple 
verbs and nouns. The primitive ACTs do not 
account for intentionality and goals underlying 
physical action. To account for such things we 
devised a complex apparatus discussed in Schank 
and Abelson (1977). If we wish to account for 
social events, we will need a system of basic 
social ACTs to represent the social actions that 
comprise the events. I term these "basic social 
ACTs" rather than primitive ACTs because in the 
end most social ACTs have some physical 
manifestation. Often their physical manifestation 
is uninteresting however. For example a 
government decision may be MTRANS-ed in a variety 
of ways. The manner of the MTRANS (written• 
announced in a speech, etc) is often not 
significant with respect to the overall social 
effect of the action. Furthermore the MTRANS 
itself is only slightly interesting. The standard 
inferences from MTRANS apply, but there are some 
highly significant inferences that need to be made 
that are not obviously available. 
For example, the most significant inference 
to be made from an authority's decision is that 
simply by virtue of that decision something has 
actually happened. That is • a government 
authorization is a truly performative ACT. Thus, if 
the government says some property is mine, or that 
a man is a criminal, then it is so by virtue of 
their saying it. Similarly other authority 
figures have the same power. A professor can say 
a thesis is finished and a student has a Ph.D. 
and these things are the case by virtue of his 
saying it. 
Not all authority's decisions are like this 
to be sure. Sometimes an authority gives an order 
and that order must be carried out for the 
decision to have effect. Frequently these orders 
come about as a result of a governmental decision 
or authorization. If the government says the land 
belongs to the Catawba Indians, then it does, but 
they may have to send in the National Guard to get 
the original owner off the property. 
What I am proposing then is two basic social 
ACTs - AUTHORIZE (abbreviated AUTH) and ORDER. 
AUTH is something only an authority can do. (This 
is a bit circular actually since if you actually 
can AUTH then that defines you as an authority.) 
i0 
In a sense then, an authority is one who when he 
acts like he is doing an AUTH (that is he does the 
physical ACTs that ordinarily correspond to an 
AUTH) in fact causes some things to happen as a 
result of the AUTH that were supposed to be the 
results of the AUTH. In other words, you cannot 
really tell if an AUTH has taken place until it 
becomes clear that the person doing the AUTH can 
back up his AUTH in some way.) The object of the 
AUTH is the authorization or new state of the 
world. AUTH takes a recipient, namely the 
relevant parties in the dispute. 
ORDER is a frequent inference of AUTH. The 
government can AUTH the army to fight a war, but 
that doesn't, simply by virtue of the statement, 
imply that they are fighting it, A subsequent 
ORDER is required that carries with it the 
implicit punishments that are relevant in carrying 
out an order. 
Why can't we do these things with CD 
primitives we now have? What is the advantage of 
these new ACTs? To answer these questions, we 
need to look at the purpose of a primitive ACT. 
It is possible to represent ORDER in CD for 
example. The verb 'order" means to MTRANS to 
someone that they must do a particular action 
or face some (usually implicit) consequence. 
Thus, implicit in the verb "order', but explicit 
in the CD representation for 'order', is the idea 
that if the required ACT is not performed, then 
someone will possibly do something to harm the 
recipient of the order in some way. This implied 
punishment is a part of the concept 'order" but is 
it necessary that we think of it each time that we 
understand an "order" to have taken place? 
The same question can be asked with respect 
to 'authorize'. We understand what authorization 
or governmental decision is, but we need not 
access all that information each time we 
understand the word. Consider the problem of 
explaining the meaning of these words to a child 
for example. It is very difficult to explain them 
precisely because they are so complicated at the 
level of physical primitive AC~s, Yet these ideas 
are really not complicated at all at a social 
level of ACTs. Such simple concepts such as ORDER 
and AUTHORIZE form the basis of the organization 
of societies. What is complex at one level is 
simple at another. This idea of nested levels of 
complexity, each with their own set of primitives, 
is a very important one for the representation of 
information in artificial intelligence. By 
choosing a good set of primitives we can 
effectively organize what we need to know. Thus, 
ORDER and AUTHORIZE have inferences that come from 
them just as the physical primitive ACTs do. The 
main difference is that these basic social ACTs 
are not primitive in the same sense. They can be 
broken down, but we would rarely choose to do so. 
The use of these new basic ACTs is much like 
the use of the original primitive ACTs. We can 
predict what will fill slots reasonably in a 
conceptualization and make inferences about slot 
fillers and consequent inferences as we would any 
conceptualization. Thus we represent sentences 
such as the following using AUTH: 
(4) The Supreme Court decided segregation 
illegal. 
O R people of 
S.C.<=>AUTH<---segregation<---IU. S. 
US<~-->ORDER< ~- punishment 
is 
(5) The cop gave the speeder a ticket. 
0 driver drive --->driver 
Cop<=>AUTN< --- ~ V $ <---1 
$TRIAL ATRANS ---<gov t. 
&DEFENDANT 
money 
? 
govt • 
In (4) we have chosen to ignore representing 
"segregation" for the moment, since it is 
obviously complex. Supreme Court decisions are 
AUTHs. They also carry with them (as do most 
AUTHs) an implicit ORDER for 'punishment" if 
certain circumstances are not met. The 
straightforwerd inference from (4) then is that 
someone practicing segregation can expect to be 
punished. 
Policemen are authorities also. In (5) the 
ticket is a written manifestation of an AUTH that 
either puts the driver in a DEFENDANT role in a 
$TRIAL script or forces him to pay a fine. The 
instrument of the AUTH is the actual PTRANS of the 
ticket (left out here). The important point here 
is that we could represent (5) using PTRANS only. 
However, what we would be describing is the 
physical ACT itself when it is the social ACT that 
is significant here. (When I was young there was 
much talk of bad kids getting "JD cards". I never 
understood what was so horrible about that. 
Couldn't they just throw them away?) The social 
significance of an ACT must be represented if it 
is understood. 
Now that we have presented these two ACTs 
let's return to our triangle: 
AUTH 
~followed by a possible ORDER) 
We have named one side of the triangle. The other 
sides represent ACTs as well. The complete 
triangle is as follows: A 
PETITION/ ~AU~H 
/ ~ (ORDER) 
DIS PUTE 
The ACT PETITION represents an individual or 
group's act of requesting AUTH" s from an 
authority. Thus a "civil suit" is a PETITION to 
the courts using some legal scripts. A protest 
demonstration is a PETITION to unstated 
Ii 
authorities using some demonstration script. The 
point here is that we cannot do away with the 
scripts that describe the actual physical 
manifestations of these events. However, the 
scripts are instr~ents of the social ACT involved 
- PETITION. The most important inference from 
PETITION is, of course, that an AUTH is expected 
that will resolve the issue that is the object of 
the PETITION. 
The issue that is the object of the PETITION 
is the DISPUTE itself. DISPUTE takes two actors 
(one of whom may be quite passive). The object of 
the DISPUTE is the issue involved. DISPUTE takes 
no recipient as it is not an inherently directed 
ACT. It is the ACT of PETITION that directs it to 
a particular authority who can AUTH something that 
will resolve it. 
We are now ready to deal with sentence (I) 
(Catawba Indians Land Claim Supported). The 
representation using the new social ACTs is: 
o 
Indians<=>DISPUTE<--(OWN (land) <=>?) 
other 
--->US o 
Indians<=>PETITlON<-- (OWN(land) <=> ?) < -- I 
---<Indians 
--->Indians 
o I or other 
U. S. <=>AUTH<--(OWN(land)<=>Indians)l 
---<US 
Since this representation is not as easy to 
write as the triangular one, we shall continue to 
use triangles in the remainder of the paper. Thus 
(I) is: 
U.S. Gov't. 
Indians 
OWNS(land)<=>Indians 
Other 
OWN(land)<=>? 
We will leave out the arrows and the ACTS for 
diagrammatic purposes, but the above triangle 
should be understood as containing all the 
information given in the CD diagram for (I). 
(Actually~the triangles contain more information.) 
Triangles provide us with a method for 
representing the social significance of actions. 
As with any other representation scheme, the 
advantage of the symbols we create can only be in 
the new symbols or actions that they spawn. That 
is, it is the inferences that come from the 
triangles that are of key importance. When we 
created the original primitive ACTs we said that 
PROPEL was no more than the set of inferences that 
it fired off. The same is true here, so we must 
ask what these inferences are. 
The first thing we can recognize about 
potential inferences here is that they will come 
in two varieties. The first are the inferences 
that are fired off from the new social ACTs that 
we have created. The second kind are those that 
come from the triangles themselves. That is, 
there should be patterns of triangles that are 
recognizable for the triangles they spawn as well 
as a set of inferences that come from the fact 
that certain triangles exist. 
As examples of this let us consider again 
sentence (2): 
(2) Burma appeals to UN to settle border dispute 
with Thailand. 
Since the representation of (2) involves a 
PETITION we can employ the inference rules that 
are fired by PETITION. Some of these are: 
a. For every PETITION we can expect a 
corresponding AUTH. 
b. For every PETITION there was probably a 
DISPUTE that gave rise to it. 
These rules lead us to the inferences available 
from AUTH and DISPUTE. Of course, inferences from 
inferences have a lower probability of truth, so 
for (2) the inferences below would be somewhat 
less certain. 
c. An AUTH can cause a DISPUTE to end. 
d. An AUTH can cause a PETITION to a higher 
authority from the party unfavorably affected by 
the AUTH. 
e. An unfavorable AUTH can cause a rebellion; or 
lack of acceptance of the validity of the AUTH. 
This can give rise to ORDERs to effect the AUTH in 
the case of individuals versus governments or wars 
in the case of governmental conflicts. 
f. An AUTH causes a new state of the world to 
exist, often ending an old state in conflict with 
the new state. 
g. A DISPUTE can cause one party to PETITION. 
h. A DISPUTE can cause a PROPEL to cause damage 
to occur for individuals, or a WAR triangle to be 
initiated for countries. 
There are, of course, a great many more of these 
kinds of inferences than we are listing here. The 
above list is mostly intended to give the flavor 
of basic social ACT inferences. It is important 
to note that the social ACTs give rise to 
inferences at both of the other levels of 
representation besides those at the same level of 
representation. That is, given a social ACT we 
may be able to infer another social ACT, a new 
primitive ACT, or a new triangular representation. 
Thus, for (2) we have two representations to 
start with: one is at the standard CD level and 
uses MTRANS; the other is at the social level and 
uses PETITION. Both of these representations 
would be available as output from the parser. 
12 
The MTRANS representation would fire off 
inferences about the methods of communication 
possibly used - that the UN now knows about the 
problem and so on. 
The PETITION representation would fire off 
inferences about the expected AUTH from the UN. 
Since we know how the UN does its AUTHs, this 
would fire off a UN script of some kind that dealt 
with voting and debate. PETITION would also cause 
DISPUTE to be inferred which would cause 
inferences about the kind of methods possibly 
employed by the quarreling countrles~ both in 
creating the DISPUTE and escalating it. 
The existence of the PETITION-AUTH-DISPUTE 
triangle would fire off an inference that the 
country kind of triangle existed. Thus, a new 
triangle that was lopsided showing possible 
aggression from Thailand towards Burma would be 
created. This triangle would in turn fire off 
inferences about attempts to RESOLVE the DISPUTE 
(one of which was (3) itself) and weuld predict an 
escalation towards the WAR triangle with its 
normal inferences if a RESOLVE did not take place. 
Although the above is rather sketchy, the 
point should be clear. We need additional 
representational mechanisms to handle the many 
levels at which statements can be interpreted. 
Triangles provide us with a new set of inference 
rules providing more power to the understanding 
system. Are they ad hoc? Of course they are. My 
point is simply that such ad hoc mechanisms will 
either solve the problem or help us create a more 
general solution that will solve the problem. The 
proRram that we are writing that uses triangles is 
also ad hoe. Is is a kludge? No. If it were it 
wouldn't be worth a thing. But, here again, if 
the program we write can handle many examples as 
we rewrite it because of what we have learned from 
it, then it will have been werthwhile. 
The program below reads newspaper headlines 
in English and generates, by use of triangles and 
the inferences available from triangles, a 
paraphrase of the input. This English para- 
phrase is generated by the program. 
******************************************* 
TRIANGLE analyzer loaded. 
INPUT SENTENCE: 
(CATAWBA INDIAN LAND CLAIMS SUPPORTED) 
(PARSE II) : CON4 
~panding token: CON4 = 
((CON ((ACTOR (*PP* CLASS (#GROUP) 
CFEATURE (*AMERINDIAN*) TYPE 
(*ETHNIC*) NAME (CATAWBA) TOK NPI) 
<=> (*PETITION*) OBJECT ((ACTOR 
(*PP* CLASS (#REGION) TOK NP2 REL CON1) 
IS (*OWN* VAL NPI)) TOK CON1) 
FROM NPI TO (*PP* CLASS (#INSTITUTION) 
MEM *COURT* TOK NP3)) TOK CON2) 
IR ((ACTOR NP3 <=> (*AUTH*) OBJECT CON1 
RECIPI NPI RECIP2 GAP1 FROM 
GAP2) TOK CON3)) TOK CON4) 
The Catawba Indians asked a Federal 
Court to rule that they own the land. 
The Catawba Indians requested a Federal 
Court to rule that the land is owned by 
t hem. 
The Catawba Indians appealed to a Federal 
Court. 
The Catawba Indians asked a Federal Court 
to rule that they own the land and it 
decreeed that the land is owned by them. 
\[ Generating inferences from CON4 \] 
>(TELL-STORY) 
The Catawba Indians and the other parties 
disagreed over the ownership of the land. 
The Catawba Indians requested a Federal 
Court to rule that they own the land. 
A Federal Court decided that the land is 
owned by the Catawba Indians. 
The other parties will probably appeal the 
decision. 
The other parties might use force against 
the Catawba Indians to assert that they 
own the land. 
***** 
This program was written by Jaime Carbonell and 
Stephen Slade. 
References 
Dresher, B.E. and Hornstein, N., \[1976\], On some 
supposed contributions of artificial intelligence 
to the scientific study of language, Cognition, 
4(1976) 321-398. 
Schank, R.C. and Abelson, R.P., \[1977\], Scripts, 
Plans, Goals and Understanding: An Inquiry into 
Human Knowledge Structures, Lawrence Erlbaum 
Associates, Hillsdale, New Jersey. 
Weizenbaum, J., \[1976\], Computer Power and Human 
Reasoning, W.H. Freeman and Company, San 
Francisco. 
13 
