PRIMITIVES AND WORDS 
Yorick Wilks 
Istituto per Gli Studi 
Semantici e Cognitivi 
Castagnola, Switzerland 
We may usefully distinguish between 
internal and external questions when 
discussing the use of primitives for 
representing natural language content and 
doing related semantic computations. Here I 
shall give a few examples of internal 
questions; go on to explain why I shall turn 
immediately to external questions; and 
finally discuss two of the latter: the 
justification of primitives in general, and 
the distinction, if any, between primitives 
and words. 
What I mean by "internal questions" 
about primitives are detailed considerations 
about what semantic primitives to choose, or 
how to insert them into larger structures in 
particular cases so as to represent some 
complex concept or conceptual relation, etc. 
These are questions that can only arise when 
the general notion of semantic primitive has 
already been accepted. It is not possible 
to discuss such internal questions while one 
is at the same time answering external 
questions, such as the justification of 
semantic primitives i__Dn general. 
There are fairly straightforward 
internal criteria for the selection and 
maintenance of a primitive vocabulary: the 
vocabulary should not be obviously 
redundant, with two primitives covering the 
same, or nearly the same, range of meaning. 
If one could show of anyone's suggested set 
of primitives that this was so, it would 
follow that he did not have a good set. 
Secondly, a primitive vocabulary should not 
be obviously oriented towards a particular 
subject area, if it is at the same time 
claimed to be a general set. So, for 
example, if, in a proffered set of primitive 
actions, we saw a majority of primitives 
concerned with human bodily actions such as 
moving, expelling, ingesting, etc., we might 
well wonder how such a system would cope 
with the expression of general actions such 
as "divide", "separate", "specify", 
"undertake", "delay", etc. 
Now there would be a drawback in 
centering a contribution, for a meeting as 
general as this one, on only internal 
questions. To my knowledge there are only 
two effective systems (in the AI approach to 
natural language understanding*) that make 
use of primitives, namely Schank's (1973a) 
and my own (Wilks 1972), and therefore any 
such concentration would be something of a 
*It will be clear here that I refer only to 
AI systems. There were, of course, working 
primitive systems in the earlier days of 
Computational Linguistics, of which the best 
known was perhaps Gardin's (1965), and in 
Generative Linguistics there is the well 
known work of Fodor, Katz, Postal, Lakoof, 
McCawley, etc. 
38 
private fight. External\[ questions, however, 
are of wider interest and it is my 
contention that, if we follow certain 
external questions through, we can see that 
other workers in natural language 
understanding are also using primitives, 
though they may not be aware of the fact. 
One issue bridges the gap between 
internal and external questions in an 
interesting way, and will serve me here as a 
new point of departure. A point of 
difference between Schank's view and my own 
has always been over the appearance in 
semantic representations of what appeared to 
be simply the surface words of the language. 
So, for example, in a representation of 
"John shot Mary" by Schank (1973b) there 
will normally appear the English word "gun". 
Whereas, in my representations, all one sees 
are structurings of the 80 primitives, and 
"gun" is not one of them. I have suggested 
that Schank's diagrams are therefore of 
mixed type, as between primitives and words. 
However, I now believe this criticism of 
mine to have been badly put, and to rest 
upon a confusion, namely that there is some 
clear distinction between primitives and 
non-primitives (i.e. surface vocabulary). 
Historically, what happened in Schank's 
system is clear: he wanted a surface-free 
semantic representation without words in it, 
and has gradually achieved this, most 
recently by the elimination of many 
noun-words in favour of Fodor and Katz type 
primitive lists, as developed by Weber. I 
am proposing now in my own system to reverse 
the process, as it were, and starting from 
an inventory of only primitives (about 54 in 
1967, now grown to 80, see Wilks (1972)) to 
begin to insert nonprimitives, i.e. words, 
into the formula structures that define word 
senses, but with one important proviso that 
I shall now explain. 
Let us consider the English action 
"fire at", as a single unit, and its 
expression in terms of more primitive 
concepts. I have expressed it until now 
(1975a) as the formula, or tree structure, 
below: 
(*HUM SUBJ) (STRIK GOAL) (, CAUSE) /\ 
(THING MOVE) 
The structure need not detain us except 
to note that the right-most primitive CAUSE 
is the head, or principal primitive, of the 
action, and the rightmost THING that is 
caused-to-move is, of course, the bullet, 
while the leftmost THING, that is the 
INSTRument, is the gun. 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
i 
I 
I 
I 
I 
I 
I 
I 
What I now propose to do is to insert 
English words into the formula: in 
particular the words "gun" and "bullet", as 
follows (in list form): 
((*HUM SUBJ) ((*ANI OBJE) ((STRIK GOAL) 
((gun INST) ((bullet MOVE) CAUSE))))) 
but to do this if and only if there are also 
formulas elsewhere in the system defining 
the meanings of "gun" and "bullet" Thus, 
in the decomposition of a formula, two kinds 
of entities may be encountered, primitives 
and "words", and if an atom is found that is 
not a primitive then it must be a word and 
so a formula must exist elsewhere for it. 
Now, there is a danger of circularity in all 
this: in that, looking for the formula 
defining "gun" as that atom appearing in the 
formula for "fire at", we might find a 
formula that told us (in primitives) that a 
gun was an object used by human beings for 
firing at something. True, but since the 
head (rightmost element) of the formula for 
"gun" will be THING, this re-entrant method 
of formula construction cannot give less 
information than formulas consisting only of 
primitives, as mine have until now. 
Now, does this method of re-entrant 
formulas -- re-entrant in the sense of 
mentioning formulas inside each other (as 
that for "fire at" now mentions that for 
"gun" and "bullet") -- force me to withdraw 
the "mixed-type" criticism I made in the 
past of Schank's descriptions. Well, yes 
and no. No, in that Schank's use of, say, 
the word "gun" in the conceptual dependency 
representation for "shoot" was not 
re-entrant in the sense I have defined it. 
There was, if I understand him, no formula 
elsewhere for "gun" in his example: there 
was the English lexeme and no more. Hence, 
if a program parsing his system searched in 
a text for "gun" as the instrument of 
"shoot", and found "bow and arrow", it would 
be helpless because it did not find exactly 
what it was looking for -- it would have no 
descriptive formula for the sense of "gun" 
to help it know when it had found roughly 
the same sort of thing. Just as in "John 
shot her with a colt", it would have no 
information with which to separate the horse 
and gun senses of "colt": it would either 
find the sought English lexeme or, as in 
these cases, not. 
But then again, on the other hand, the 
criticism i__ss changed because the claim that 
there should not i__nn princiDle be mixed type 
(word and primitive) semantic descriptions 
is implicitly withdrawn by the above 
proposal for re-entrant formulas. Now there 
need be no serious "theoretical" 
considerations involved in such a proposal. 
It can be seen as simply a notational 
convenience: in that "gun" in a formula for 
"shoot" is now just a shorthand form for the 
formula for "gun" existing elsewhere in the 
system. This makes the formulas easier to 
read for a human user by avoiding the 
insertion of too much repetitive material in 
terms of primitives into the body of the 
formula itself. 
39 
Yet, there is a more fundamental point 
buried in all this, very much an external 
question, and one which led me to mis-state 
my criticism of mixed type descriptions in 
the past. Let me now go into this a little: 
though it is not, as I shall show, intended 
to lead to conclusions that should give any 
comfort, either to those who have used mixed 
type descriptions in the past or to those 
who wish to avoid primitive descriptions 
altogether. 
My basic error was to defend the use of 
primitives by implicitly assuming that they 
were in some way essentially different from 
natural language words, and therefore, that 
structures of primitives were in some way 
different from structures of natural 
language words, which is to say, assuming 
that a semantic description in terms of 
primitives was something other than a 
reduced micro-language, with all the normal 
weaknesses and vaguenesses of a natural 
language. I now see more clearly that that 
is not and cannot be so. Or, to put the 
message of this position paper in a slogan 
form, there is no escape from a natural 
language and certainly not via the primitive 
route. 
Let me take first the .iustification 
aspect of this. Those of us who have used 
primitives for some time have all I think 
made use of the innocent notational 
convention of writing primitives in upper 
case letters thus: MAN, CAUSE, TRANS, etc. 
And I would not wish to change this; it is 
useful for indicating whether we are, at any 
given moment, describing a structure in the 
primitive or the surface language. The 
trouble is that such a usage does inevitably 
carry the suggestion that the upper case 
entities are something other than (over and 
above, or "deeper" than, but anyway 
different from) the ordinary words they look 
like. And this, of course, is error, pure 
and simple. There have certainly been 
claims from some quarters that these 
entities are quite other: that CAUSE has 
almost a halo around it, and represents, or 
refers to, some other kind of entity in the 
brain or mind directly. 
At its crudest, this is just the old 
referentialist fallacy moved up a level, as 
it were. The low level fallacy, still alive 
and well, is that the meanings of words are 
physical objects. "Chair", the story goes, 
signifies by referring to things like the 
one I am sitting on, and so therefore do 
"mind", "action", "friendship" and 
"cunning", though perhaps in a slightly more 
roundabout way. I do not want to discuss 
this view here, but only to point out that 
the view of primitives under discussion 
holds that primitives, like words, have 
their meaning/significance in the same sort 
of way, but in their cases by referring to 
certain ill-defined mental entities. I have 
argued in detail elsewhere (Wilks 1974) that 
this could not conceivably be so, or be 
known even if it were so. 
But this comic-strip philosophy is not 
the heart of the matter: what this view of 
primitives does, in real terms, is to lead 
to research that attempts to justify 
particular sets of primitives directly in 
some way, and these days that usually means 
psychologically*. But what I am arguing is 
that, if formulas, templates, conceptual 
dependency structures, etc. are simply 
usages of a language of primitives, then no 
direct .justification of the vocabulary makes 
sense, and certainly not any justifications 
of a correct vocabulary, any more than it 
would make sense to try to establish the 
correct vocabulary of English or any other 
natural language. It follows from this bhat 
there can be a variety of primitive 
languages for semantic descriptions, no one 
necessarily better or worse than any other, 
any more than my vocabulary is better or 
worse than yours if I know 100 English words 
you don't, and you know 101 that I don't. 
In the case of each primitive vocabulary, 
the only ultimate test will be the success 
or failure of linguistic computations that 
make use of it. 
Now there are limits to this sweeping 
parallel between primitive and natural 
languages. Clearly, and as I mentioned 
earlier, a primitive vocabulary should not 
have synonymous primitives, whereas in a 
natural language it is often a point of 
aesthetic pride to have as large a range of 
semi-synonyms as possible: "cemetery" and 
"graveyard" in English, for example. Again, 
there is at least one interesting piece of 
indirect justification of a primitive 
vocabulary, namely the project at Systems 
Development Corporation that put the whole 
of Webster's Third International Dictionary 
onto tape and counted the rank frequency 
list of words used in the definition of 
other words. That was, ignoring very 
frequent words like "a" and "the", down to 
the 80th rank order, pretty close to my own 
list of primitives, and naturally I was 
pleased. That is what I, like anyone else, 
would have hoped for; since primitives are 
in fact used to define the senses of other 
words, we should be happy to be close to the 
list used unconsciously by the makers of a 
large and efficient dictionary. 
Nevertheless, and in spite of these two 
caveats, it seems to me that a primitive 
vocabulary is nothing other than a small 
natural language, and therefore not open to 
methods of justification unavailable to any 
other natural language, and that it is a 
mistake to pretend otherwise. To sum up, if 
words of a natural language like English are 
not justified nor gain significance by their 
direct reference to things, but only by 
their function within the overall language, 
then we may expect precisely the same to be 
*I cannot see that the notion of the 
"psychological justification of primitives" 
makes sense, though I would be happy to be 
shown. I do not of course refer here to 
work like (Johnson-Laird 1974) showing that 
humans seem not to store surface language. 
Such results are quite consistent with a 
"primitive hypothesis", but do not support 
it, since they are also consistent with the 
hypothesis that human semantic 
representation is not linguistic at all, but 
consists of, say, binary numbers! 
40 
true of a language of primitives, and no 
amount of writing in upper case letters is 
going to change that fact. 
Now, if I take this to heart, that 
there is no difference at bottom between 
primitives and other words, then I find that 
another criticism I have made implicitly or 
explicitly in the past must also be 
completely recast. This concerns the 
initial representation of sentences in 
PLANNER-type formalisms, where it seems not 
unnatural to represent "John is at the 
station" as (AT JOHN STATION). 
I have been tempted to criticise such 
representations, to myself at least, on the 
grounds that they were simply the English 
words of the sentence (or something very 
like them) rearranged in some plausible way, 
and that therefore nothing had been shown or 
structured. But, as I said, if I take to 
heart the point that primitive vocabularies 
are not essentially different from those 
comprising more obviously surface words, 
then I cannot maintain that criticism in the 
same way. Why shouldn't "John", "at" and 
"station" be somebody's primitive 
vocabulary? 
Now the structural criticism of this 
method of coding up sentences is unaffected. 
That is to say, little is revealed by such a 
method of expression, unless done in a very 
systematic manner*, and it normally relies 
over much on our intuitive appreciation of 
the structure of the original sentence, in 
such a way as to be not really the 
structuring of an example but a mere 
reprojection of the example itself. To see 
this, one only has to think of what it would 
be like to express the first sentence of 
this paragraph by such a method, the one 
beginning "Now the structural 
criticism...etc.". 
However, the point at issue here is not 
this structural one but that of the status 
Of the items in the description: as I raised 
the question, but did not answer it, why 
should those English words not be declared 
to be part of the primitive vocabulary? For, 
if as I have just argued no serious 
distinction of type can be maintained 
between words and other primitives, what 
could be wrong with that? 
Well, it is easy to see what is wrong, 
given that one accepts one other principle: 
namely, that one s system, whatever it is, 
should be extensible in a non-trivial 
manner. What weare now discussing is the 
fallacy of the map, to adapt a philosophical 
cliche. We have a system claiming to 
represent the structure of natural language 
but which in fact represents it in the way a 
map would if its scale was 
one-mile-to-one-mile. There would be 
something wrong** with such a map, that much 
is clear, and similarly there is something 
wrong with a system which is only extensible 
on the same scale as what it represents: it 
adds to its primitives (ordinary words in 
*The only attempt I know to do this sort of 
thing systematically is (Sandewall 1972). 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
1 
I 
I 
I 
I 
this case) at the same rate as it adds to 
the sentences covered. I am not trying to 
smuggle back any of the distinction between 
primitives and non-primitive words that I 
have abjured, but am only pointing out that, 
if one chooses surface words as one s 
primitives, there is nothing theoretically 
wrong, but just the practical (and in view 
insuperable) difficulties of (I) inability 
to state significant semantic 
generalizations, and (2) the inability to 
extend one s coverage of the language in 
anything other than a mile-to-mile manner. 
It is for this reason that it still seems 
possible, to me, to give up believing in any 
difference in principle between primitives 
and other words but still advocate strongly 
the use of a sensible selection of words as 
a reduced, or primitive, sublanguage for 
semantic expression. 
One final point, which is not argument 
but merely the drawing of a bead on a more 
distant target. It is the case that the use 
of a PLANNER, or Predicate Calculus, type of 
representation in "semi-English" that I have 
just discussed is intimately associated with 
the view that the study of reasoning is 
dissociable, or decouplable, from the study 
of the semantic structure of natural 
language, and can be pursued in isolation. 
In (Wilks 1975b) I have raised a number of 
doubts about that view, but here I want to 
add another which follows directly from the 
argument of this paper. 
If it is true, as I have argued here, 
that there is no escape from a natural 
language into some other realm, and that a 
language of representation is just another 
natural language whether of primitives or of 
"semi-English", then it follows that there 
is no special extra-terrestrial sphere for 
the examination of reasoning*, but only 
translations into another natural language. 
Hence there is no reasoning about natural 
language separate from natural language, and 
all we can do is to choose the language in 
which we prefer to model the reasoning and 
over which we prefer to compute. Thus, to 
compress things somewhat, we have the choice 
between computing about reasoning in a 
primitive-like language, or one reducible to 
it by the "re-entrant" method I described, 
or in one like PLANNER semi-English in which 
little is made explicit, and which would 
require another system to make its internal 
relationships explicit for any but the most 
trivial examples. 
**One way out I have not discussed here is 
for someone to argue that, for the 
representation of language, the 
mile-to-a-mile scale is not a fallacy, 
because no significant linguistic 
generalizations are possible. 
*I am not denying here that some 
non-linguistic forms may explicate our 
reasoning about, say, position in space or 
numerical relationships; nor am I denying 
that there is reasoning in the sphere of 
vision, which is also possessed by dogs, and 
cannot therefore be linguistic. Granted all 
that, I am arguing that the sort of 
reasoning required to understand the 
argument of this paper, or of a standard 
newspaper editorial, must (a) be linguistic 
in nature, but (b) will not be explicated in 
an interesting and non-circular manner by a system using ony semi-English. 
It may turn out that it is more 
sensible to say that language understanding 
depends on reasoning, rather than 
vice-versa. Everyone in A.I. seems to 
believe it without question, and I have done 
no more here than raise a few small doubts 
that it might, after all, turn out to be the 
other way round. 
REFERENCES 
Gardin, J., SYNTOL, in Artandi (ed) Rutgers 
Series on Systems for the Intellectual 
Organization of Information, New Jersey, 
1965. 
Johnson-Laird, P., "Memory for Words", 
Nature, 1974. 
Sandewali, E., "PCF-2, A first-order 
calculus for expressing conceptual 
information", Dept. Computer Science, 
Uppsala Univ., 1972. 
Schank, R., "The Fourteen Primitive Actions 
and their Inferences", Stanford Univ., AI 
Lab. Memo No. 183, 1973a. 
Schank, R., "Identification of 
Conceptualizations underlying Natural 
Language", in Schank and Colby (eds), 
ComPuter Models of Thought and Language, 
San Francisco, 1973b. 
Wilks, Y., Grammar, Meanin~ and the Machine 
Analysis of Language, London, 1972. 
Wilks, Y., "One Small Head", Foundations of 
Language, 1974. 
Wilks, Y., "An intelligent analyser and 
understander for English", Comm. ACM, 
1975a. 
Wilks, Y., "Natural Language Systems within 
the AI Paradigm", Stanford Univ. AI Lab, 
Memo No. 337, 1975b. 
41 
