The Primitive ACTs of Conceptual Dependency 
Roger C. Schank 
Yale University 
New Haven CT 
How should a proposition appear in 
memory? It simply cannot be stored in a form 
that bears a strong relationship to the 
original form of the input in natural 
language. Since it is possible to say .the 
same thing in any number of different ways, 
it is unreasonable to suppose that people 
are constantly checking to see if a 
proposition that they have stored one way in 
memory is the same as another they have 
stored somewhere else in a different form. 
What is far more likely is that people have 
developed a kind of canonical form. No two 
propositions in memory can be equivalent in 
this view, since the canonical form for 
their meaning forces them to be represented 
in the same way and thus stored and treated 
in the same way by memory. 
A canonical representational scheme 
cannot have natural language words as its 
elements. Words can be ambiguous and a 
meaning representation system must be 
unambiguous. Furthermore, words can be 
synonymous and overlap in meaning. By 
disallowing synonymous canonical forms of 
propositions we have also disallowed the use 
of words in the representation. 
The consequences of this are consistent 
with what we know about people. In memory 
experiments people consistently exhibit 
recognition confusions because of their 
remembrance of the "gist" of a sentence 
rather than the words that they actually saw 
or heard. 
Conceptual Dependency has as its basic 
premise that the ACTION is the basis of any 
proposition that is not descriptive of a 
static piece of the world. 
All propositions that describe events 
are made up of CONCEPTUALIZATIONS. A 
conceptualization is composed of an ACTOR, 
an ACTION, and a set of cases that are 
dependent on that ACTION. 
The main insight of Conceptual 
Dependency theory is that there must be a 
canonical form for meaning representations. 
The types of conceptual roles that there are 
in this canonical form do not correspond to 
syntactic roles. Rather, words are broken 
down into their conceptual parts and then 
placed in a meaning representation with 
respect to their conceptual role rather than 
the syntactic role that was chosen for them 
as a result of some, perhaps random, lexical 
choices. For example, the word "danger" has 
a nominal form, a verbal form (endanger), an 
adjectival form (dangerous), and an 
adverbial form (dangerously). The meaning 
of the concept underlying "danger" is the 
same no matter what lexical form is selected 
to realize that concept. Basically, 
"danger" means that something can possibly 
happen that will result in a negative effect 
34 
in somebody's physical state. "Danger" is 
represented conceptually as a 
conceptualization linked to a state 
description. Certain elements of the 
conceptualization are unknown, but can be 
filled in by the words surrounding "danger'. 
The state description is partially known. 
It says that somebody's physical state may 
be negatively changed. 
So we see that "endanger" is no more an 
action than it is an actor. Rather, 
"endanger" refers to a partial conceptual 
structure. The word "endanger" also sets up 
an expectation that the syntactic subject of 
that verb will contain the item that will 
help fill in the missing action that was 
"dangerous'. Thus, in the sentence "The 
bees endangered Bill," the "bees" are the 
actors of a conceptualization that is 
causally related to the state description 
that the concept "Bill" will undergo a 
negative change. The action is still 
unknown, but now can be inferred. The 
memory is consulted for actions that bees 
are known to perform that endanger humans. 
The simplest explanation is preferred and 
the best candidate for the action is 
"sting'. The object of that action is 
"Bill'. The conceptual representation of 
"The bee's endangered Bill" says that "the 
bees may do an action (possibly "sting') 
which will cause Bill to become hurt." 
We said before that words do not appear 
in conceptualizations. Following that point 
of view it is obvious that "sting" cannot 
really satisfy our needs here. The idea of 
"sting" is correct, but it must be broken 
down into its component parts. We define an 
action as something an actor can do to an 
object. We define an actor as an animate 
object, and an object as any concrete 
physical object. "Sting" still works under 
these constraints, but couldn't we just have 
well used the word "bite'? Although these 
two words are slightly different in their 
strict interpretations, to the average user, 
these words are synonyms. We are left with 
the problem of describing the concept 
underlying them in order to satisfy the 
requirements that we have set out above. 
In order to handle similarities and 
overlaps in meaning as used by the man in 
the street we have developed the concept of 
the PRIMITIVE ACTION. Primitive actions are 
intended to be the building blocks out of 
which the meanings of verbs and abstract and 
complex nouns are constructed. The 
primitives are not category names but rather 
elements that can be used in many varied 
combinations to express the meaning of what 
underlies a given word. 
Conceptual Dependency representations 
use only twelve primitive acts. These acts 
were arrived at by noticing structural 
similarities that existed when sentences 
were put into an actor-action-object 
framework. Using this set of acts, a larger 
set of states and a set of conceptual roles, 
it is possible to express a large amount, if 
not all, of the meanings expressible in 
natural language. We will now describe the 
most important of the primitive ACTs and 
I 
I 
I 
I 
I 
I 
I 
I 
briefly mention the others. 
ATRANS The transfer of an abstract 
relationship such as possession, 
ownership or control. ATRANS 
requires an actor, object and 
recipient. 
ex. 
give=ATRANS an object from the 
actor to the recipient. 
take=ATRANS an object from 
someone to the actor. 
buy=two ATRANS actions, each 
causing the other. An 
ATRANS of money from the 
first actor to the second 
actor. An ATRANS of an 
object from the second to 
the first. 
PTRANS The transfer of physical location 
of an object. PTRANS requires an 
actor, object and direction. 
ex. 
go=PTRANS of an actor by an 
actor to a location. 
fly=PTRANS of an actor to a 
location. Any ACT can have 
another act as its 
instrument. For "fly" that 
instrument is PROPEL. 
PROPEL The application of a physical 
force to an object. If movement 
takes place as a result of a 
PROPEL, a PTRANS has taken place 
too if it can be determined that 
that PROPEL was intended by the 
actor. Direction, object and actor 
are required. 
ex. 
push=PROPEL an object in a 
direction by an actor. 
pull=PROPEL an object in a 
direction (towards actor) by 
an actor. 
kick:PROPEL an object in a 
direction by an actor using 
the instrument MOVE object 
foot. 
MTRANS The transfer of information 
between animals or within an 
animal. For the purpose of dealing 
with the words people use to 
discuss memory, we partition memory 
into three pieces: the CP (the 
conscious processor9. Something 
being in CP means that it is being 
thought of at that moment. The IM 
(the intermediate memory) where 
current context is kept and 
contextually relevant information 
is stored while in use. The LTM 
(long term memory) where informaton 
is stored. The various sense 
organs serve as the origins in an 
MTRANS. Recipient, object and 
actor are required, although here 
these cases refer to the above 
mental locations as recipients and 
mental information as possible 
objects. 
ex. 
remember:MTRANS information from 
LTM to IM 
35 
forget:MTRANS cannot be 
accomplished on some 
information from LTM 
see:MTRANS information to IM 
from the eye 
tell=an actor MTRANS 
information to another 
actor 
read=MTRANS information from 
a book to IM by means of 
ATTEND eye to the book. 
MBUILD The construction within an animal 
of new informaton from old 
information. MBUILDing takes place 
within IM and receives its inputs 
from the CP and from IM. It 
transforms them into a new idea and 
places that idea in the CP. 
ex. 
describe=MBUILD a future action 
or a new state. 
imagine:MBUILD a new idea that 
is hypothetical 
consider=MBUILD from a mentioned 
input into an 
unspecified output 
answer=MBUILD a specific output 
to somebody else's input 
that was just received 
in the CP. 
INGEST The taking in of an object by an 
animal into the inner workings of 
that animal. INGEST requires an 
actor, object, and direction. 
ex..left margin 16 
eat=INGEST a solid into the body 
through the mouth 
breathe=INGEST air into the body through 
the mouth 
shoot up=INGEST a narcotic into the body 
through a vein by PROPEL narcotic 
through needle. 
smoke:INGEST smoke from a cigarette into 
the mouth by INGEST air through the 
cigarette. 
All conceptualizations require 
conceptualizations that are instruments for 
them. The remainder of the primitive ACTs 
are primarily the ACTs of instrumental 
conceptualizatons. They are: 
GRASP To grasp an object. 
ATTEND To focus a sense organ to an object. 
SPEAK To make a noise. 
MOVE To move a bodypart. 
EXPEL To push something out of the body. 
We have listed above only one sense of 
each verb that we have chosen to describe. 
The power of the primitive ACTs is that they 
can point up similarities between words 
{such as the use of ATRANS for "give', 
"take', and "buy" while also highlighting 
their differences. In addition they serve 
as a vehicle for disambiguating ambiguous 
words. Thus, "smoke" as in "the barn is 
smoking- would not have INGEST in it but 
would use a state description involving 
"smoke" and "heat'. Similarly, "take" as in 
"take an aspirin" would be INGEST rather 
than ATRANS. A program that makes these 
disambiguations and assigns the correct 
conceptual structure to an input sentence 
was written by Riesbeck (1974). The program 
uses knowledge about how words fit together 
in English and some general knowledge about 
the physical world to make its distinctions. 
Upon encountering "take" for example, the 
program asks if there is an object around 
which is medication. It also must find out 
if the sentential object is really " an 
action, as in "take a beating." If it is we 
have neither ATRANS or INGEST but a causal 
structure. This series of tests helps to 
determine the correct conceptual sense for a 
word. 
The primitive ACTs are useful for 
organizing the inference process. Rather 
than stating that if you see something, then 
you know it and if you hear something then 
you know it and if you read something then 
you know it and so on, we simply state that 
whenever an MTRANS exists, a likely 
inference is that the MTRANSed information 
is in the mental location LTM (our 
representation for "know'). This is a 
tremendous savings of time and space. For 
example, any number of forces can cause 
different kinds of state changes of damage, 
location and the like. By listing the 
criteria that must be met for these 
inferences once under PROPEL, rather than 
under each individual verb that might call 
them up, we organize our knoweldge of facts 
and events in a very useful way. 
(Organizing the inferences under the 
ACTs also serves to define the ACTs 
themselves. The theoretical decision for 
what constitutes a primitive ACT is based on 
whether the proposed new ACT carries with it 
a set of inferences not already accounted 
for by an ACT that is already in use. 
Similarly, a primitive ACT is dropped when 
we find that its inferences are already 
handled by another ACT.) 
We have recently added a new primitive 
ACT called PLAN. PLAN is intended to 
account for an individual's ability to 
decide on a step by step course of action 
that leads to a goal. PLAN takes as input a 
goal and a set of possible plans that are 
known to be ways to achieve that goal. PLAN 
decides among them and produces as output a 
sequential chain of conceptualizations that 
the actor intends to act out in order to 
achieve the desired goal. PLAN takes MBUILD 
as its instrument where the items MBUILDed 
are the particular decisions that make up 
the selection of a given plan. 
The general form of PLAN is with an 
object that takes as input a goal and the 
plans associated with that goal and produces 
as output the chain of acts that the actor 
intends to perform to achieve the goal. 
MBUILD is now redefined as having an 
object that takes multiple inputs dealing 
with the current facts and the logical 
consequences and beliefs associated with 
those facts and produces a new thought which 
36 
may or may not be an intended future action. 
ACTI,ACT2,ACT3... 
actor <=> PLAN <- 
GOAL 
ACTS and PLANBOXES 
CONCEPTUALIZATION 
actor <=> MBUILD <- 
FACT 
FACT 
BELIEF 
RESULTS OF FACTS 
The difference in these two ACTs with 
respect to their realizations in ENGLISH is 
as follows: 
John decided to go to the Bahamas. 
MBUILD new thought. 
John developed a plan to go to the 
Bahamas. 
PLAN unknown set of ACTs. 
GOAL that is input to ACT 
PLAN is known 
John knows how to get the money. 
PLAN with known goal. 
John figured out who the murderer is. 
MBUILD new thought 
John did that to get Mary to like him. 
PLAN 
John intended to hurt Mary. 
PLAN (ambiguous if no ACT was done.) 
John intends to hurt Mary. 
Ambiguous (not known if he has 
figured out how) 
The differences with respect to inference 
are: 
MBUILD We can infer that the facts were 
there to make the decision and question 
the applicability of those facts. 
PLAN We can infer that the steps to a goal 
were thought about before an ACT was 
taken. 
We MBUILD goals. We PLAN ways to get 
those goals to become reality. 
The idea of primitive ACTs has met with 
a general agreement that such a thing is in 
principle reasonable on the one hand, and a 
general uneasiness with the set that 
Conceptual Dependency proposes on the other. 
Do we really need such a small set? Must we 
break down all words into these primitives 
each and every time? How do you arrive at 
the correct set? What about words like 
"drive" or "dance" or "hunt'? 
The answer to these questions is not 
simple. We never sought to create an 
extremely small set. We just happended on 
that set in the course of attempting to find 
out what the entities such as "understand3" 
and "believe1" that we were creating were. 
The set that we came up with has so far been 
adequate to handle a myriad of domains that 
we have tried. We expect that over time a 
different set than we are currently using 
will emerge. However, our expectation is 
that the order of magnitude is correct. It 
should be possible to handle most worlds 
with a very small set. 
I 
I 
I 
I 
I 
I 
I 
I 
! 
I 
I 
I 
I 
I 
I 
! 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
The question of whether we must break 
down everything we see in terms of primitive 
ACTs seems to be the one that excites the 
most controversy. Wouldn't it be simpler to 
use English words where we can and break 
down into primitives only when we have to? 
Yes, it would be simpler to do that if we 
had a data base that had in it "John drives 
a red car" and we were asked "Who drives a 
red car?" Researchers who work with highly 
restricted data bases probably have no need 
for primitive ACTs. But real understanding 
systems must be capable of handling qustions 
such as "Who was that I saw moving down the 
street on those fancy wheels??" or even just 
"Who parked that car over there?" Life is 
rarely as simple as naive question answering 
systems would like it to be. Real 
understanding implies knowing what the 
meaning of everything you hear is, before 
you attempt to process it in memory. The 
truth is that it just isn't any harder to 
map input sentences into conceptualizations 
involving the primitive ACTs than to do a 
syntactic parse. (See the paper by Riesbeck 
in this volume.) 
But what about words such as "drive" 
and "dance'? The answer is that these words 
(and many others) have no simple meaning in 
Conceptual Dependency terms. They are what 
we used to call SDO, meaning they were names 
of prescribed sequences of 
conceptualizations, any subset of which 
might not be meant when the word was being 
used. These action sequences can be handled 
using the primitive ACTs. The problem is in 
describing just what "drive" does mean. The 
answer is that it has a basic primitive 
associated with it (PTRANS by PROPEL here) 
but it also has what we now (see Schank and 
Abelson \[1975\]) call a script. This script 
carries along with it all the mundane 
information that is known about "drive'. 
In creating the primitive ACTs we made 
no claim that these were the last primitive 
entities that we would be creating. We are 
now working on a set of primitives plans and 
planboxes (described by me in another paper 
in this volume). These higher level 
abstract entities serve to organize 
information about people's goals and plans. 
They work on top of the conceptualizations 
that are received as input. Knowledge 
organizing primitives are every bit as 
important as event organizing primitives. 
It is important to realize why we are 
creating these abstract entities in memory. 
Basically, the reason for creating them is 
organization. A primitive ACT in memory is 
useful for organizing inferences around. 
Whenever that ACT appears in a 
representation, we know that a certain body 
of knowledge applicable to that ACT is going 
to be relevant. Primitive or abstract 
constructionns in a memory are there 
precisely because a given body of knowledge 
is frequently accessed as a unit whenever 
that abstract entity is accessed. 
Because we do not remember the original 
inputs, the abstract entities that we create 
to represent them must serve a dual role. 
They must both accurately represent the 
37 
meaning of the input as well as call up 
relevant bodies of knowledge that will be 
applicable as data upon which to base 
inferences. This dual responsibility of the 
conceptual representation means that the 
abstract elements within that representation 
must be selected with respect to how they 
affect the other elements of the 
representation as well as whether they serve 
to organize information in an efficient 
manner. Since the primitive ACTs must be 
more carefully chosen, there are serious 
questions as to what can constitute an 
acceptable and workable set. 
These considerations do not hold in the 
situation where abstract entities are 
created and the input information is 
retained. Thus, in situations where 
conceptual representations are to be 
evaluated for goals that are implicitly 
expressed within them, we are simply 
evaluating the input conceptualizations. 
They will not be replaced by whatever new 
constructions are invented. Thus, the 
selection of these higher level abstractions 
depends mainly on whether we can find a use 
for them. 
So, we are saying that there are many 
possible kinds of primitives. Primitive 
actions underlie the words of a language. 
Operating on top of those primitives, and 
therfore affected by them, are a set of 
primitive knowledge organization pieces. 
Getting at these is what understanding 
understanding is all about. 
