On primitives, prototypes, and other semantic anomalies 
Terry Winograd 
Stanford University 
Over the past few years, there have been a number of 
papers arguing the relative merits of primitives and proto- 
types as representations for the meaning of natural 
language. Much of the discussion has been both pug- 
nacious and confused, with each author setting up one or 
another straw-man to knock down. Much of the confusion 
has resulted from a lack of agreement as to what it would 
mean for a system to use primitives or prototypes. There 
are several different dimensions along which semantic 
folxnalisms vary, and many of the arguments have blurred 
these into a single distinction. 
In this paper, I propose a framework within which to 
compare a variety of semantic formalisms which have 
been proposed in linguistics and artificial intelligence. The 
paper lays out three dimensions (called ontological, logical, 
and relational), describing the relevant options along each 
and the implications of making alternative choices in the 
design of a formalism. It does not attempt to demonstrate 
that one or another alternative is right, but instead tries to 
clearly state the advantages and disadvantages of each in a 
non-partisan way. It is more in the style of a text-book 
than of a research paper. Its contribution will, I hope, be 
in dissolving some non-issues which have occupied 
previous discussion, and in focussing attention on the real 
distinctions between alternative proposals. My own 
prejudices are set forth in Winograd (1976) and Bobrow 
and Winograd (1977). In addition to citing primary 
sources, I will make particular reference to the discussion 
by Wilks (1977) since it is recent and sets out a number of 
the same issues. 
The ontological dimension 
The formalisms we want to compare are all based on the 
use of symbol structures to represent meaning. There are 
deep philosophical questions as to how much of meaning 
can be captured in a formal system, but such questions are 
outside the scope of this paper. We will take it for granted 
that meaning is to be characterized in terms of structured 
relationships between discrete symbols. The first question, 
then, is just what these symbols are. There are three basic 
positions which have been taken: 
LINGUISTIC. In many older accounts of meaning, the only 
entities which take part in the formal structure are the 
entities of language: words, morphemes, phrases, and sen- 
tences. The dictionary is an account of meaning within 
this tradition. The meaning of a word is expressed in 
terms of structures made up of other words, without any 
direct appeal to concepts which lie outside the language. 
PSYCHOLOGICAL. Most current work in AI and psycho- 
linguistics assumes that the entities which are manipulated 
in the formal theory represent some sort of concepts which 
underlie language use, but are' not themselves part of the 
language. These concepts have psychological reality, in 
that they correspond to functional components in the 
memory and language activity of a person. Words and 
sentences are seen as corresponding to structures of under- 
lying concepts. A psycholinguistic theory includes an ac- 
count of the processes by which language is translated into 
conceptual structures, and generated from them. In the 
case of AI systems (such as the conceptual dependency 
formalism of Schank (1972)), the commitment to PSYCHO- 
LOGICAL entities is a global assumption which plays little 
role in the methodology of the work. In the case of 
psychological experimentation (for example, much of the 
work described by Clark and Clark (1977)), it is a hypo- 
thesis to be tested explicitly. Some theoretical psycho- 
logists (such as Miller and Johnson-Laird (1976) and 
Fodor (1975)) have characterized it is a private "language 
of thought". 
TItEORETICAL. A more cautious stance is taken by most 
theorists who work within the generative linguistics para- 
digm. They argue that the symbols of their formal seman- 
tic theories need not correspond to functional psycho- 
logical entities. The symbols and structures play a role 
similar to that of postulated theoretical entities in physics, 
such as neutrinos and probability waves. A system based 
on them is justified in terms of its resulting overall 
simplicity and ability to account for the observable 
phenomena, not by finding psychological correlates for its 
individual terms. This view shares with the psychological 
view the notion of lexical decomposition. Words and sen- 
tences of the language correspond to structures built up of 
non-linguistic symbols. 
25 
There has been a certain amount of confusion within 
both syntactic and semantic theory about whether there is 
any psychological reality to the formal constructs postu- 
lated by linguists. In the 60's, experiments w.re carried 
out (e.g., Miller, 1962) looking for psychological correlates 
of transformations, with generally negative results. Chom- 
sky has repeatedly reiterated his official stance that the 
validity of transformational theory is not based on any 
assumption as to whether transformations play a functional 
role in language comprehension or production. Similarly, 
as Wilks (1977) points out, Katz's view of semantic mar- 
kers shifted from PSYCHOLOGICAL (in Katz and Fodor, 
1965) to THEORETICAL (in Katz, 1972). 
In doing AI research, the issue can be finessed. In 
building a program, one must develop a set of symbolic 
structures which are used functionally--they play a direct 
role in the memory and reasoning of the system. In this 
sense they are purely psychological (the psychology of the 
computer program, not of a person). When the program is 
viewed as a 'theory of human language use', two routes 
can be taken. If strong psychological equivalence is 
claimed, there is an assumption that the internal organ- 
ization and objects of the program correspond to the 
organization and objects in the mind of a human language 
user. An alternative position of weak psychological equiv- 
alence is similar to that of the generative linguists. The 
program as a whole is justified by its ability to match 
human performance, but no claims are made about the 
ways in which its organization maps onto psychological 
phenomena. Since programs can be built without con- 
fronting this issue, there has been a tendency by AI 
researchers to handwave about it, taking whichever 
viewpoint seems most advantageous in a given discussion. 
Begging the fundamental question of 
semantics 
A persistent cause of misunderstanding in arguments 
about semantics has been a lack of agreement over what a 
'semantic theory' should achieve. From a philosophical 
stanclpomt, Uae issue centers around what meaning is. The 
fundamental question is that of the relationship between 
symbols (words) and a world about which they speak. 
From an AI standpoint, the question is operational--how 
can a symbolic system be organized which accounts for the 
phenomena of language use. As pointed out by Fodor 
(1978), no answer to the second question, no matter how 
clever or elegant, is an answer to the first. In creating a 
system which accepts text, answers questions, or enters 
into a dialog, we have not created a theory of semantics, 
we have created another class of objects for which such a 
theory is needed. 
This observation applies regardless of which of the 
three choices is taken along the ontological dimension. In 
taking words as the formal objects, we leave the semantic 
problem completely unaddressed. In relying on psycho- 
logical entities, we transform the question into the equally 
difficult one "How are concepts related to the world which 
they are concepts about?". Similarly, with theoretical 
objects we beg the question by pushing it into a different 
domain. As many people have argued, (e.g. Lewis (1972) 
in discussing Katz and Fodor's theory of semantic 
markers), translating English into 'Markerese' doesn't 
illuminate the fundamental nature of meaning any more 
than translating it into French. 
Wilks (1977) describes several papers which argue for 
the necessity of a semantic theory along the general lines 
of Tarski and recent work in model-theoretic semantics for 
formal languages. He characterizes them as criticisms of 
semantic primitives and argues that they are based on 
weak 'escape arguments'. He is correct in concluding that 
the concerns of these authors are orthogonal to the specific 
technical debate about primitives, but wrong in assuming 
that they are arguments in the same domain at all. In 
creating formal systems for representing and manipulating 
structures corresponding to meaning, we are not forced to 
answer the fundamental question of what meaning is. As 
Wilks points out, this question has been asked for 
thousands of years, and technical progress does not seem 
to depend on clearing it up. 
There are valid doubts about whether adequate 
semantic formalisms (in the AI/operational sense) can be 
developed without more careful thought about the basic 
questions. In particular, our unexamined assumptions 
about the nature of meaning can lead us down paths in the 
problems we choose to look at, which may in the long run 
conceal other more fruitful paths. However, this sort of 
question has not been addressed in current AI work, and 
for the purposes of setting up a clear framework for 
understanding that work, we will continue to ignore it. A 
characterization of a semantic formalism in terms of the 
dimensions of this paper has nothing to say about the 
fundamental nature of semantics. 
The logical dimension 
As implied in the previous section, we are primarily 
concerned with the operational implications of different 
formalisms--the ways in which they can be used in 
language comprehension and production. Each symbol or 
structure of symbols plays a role in reasoning processes 
which underlie language activities, and there are a number 
different approaches to dealing with them. There are three 
basically different views of the logical status of the 
individual concepts (or words): 
ABSTRACTION. The tradition drawn from logic and 
linguistics is to view the elements of a semantic formalism 
as logical abstractions--predicates and constants within a 
logical system. The meaning of a word is a structure of 
semantic elements which express the logical truth 
conditions determining its applicability. For example, if 
we analyze one sense of "bachelor" as having the semantic 
components HUMAN, MALE, and UNMARRIED, it is 
implied that any object to which that sense of the word 
could be properly applied will fit the truth conditions 
corresponding to those terms. If "kill" is analyzed as a 
structure of the form CAUSE(X, DIE(Y)), then we can 
safely deduce from the fact that "A killed B" that, among 
other things, B died. 
26 
There are many old and unsettled debates about the 
status of such knowledge as analytic or synthetic. The 
issue here is not that distinction, but the status of the 
semantic analysis as leading to logical consequences which 
can be drawn from the the application of a given word. 
PROTOTYPE. One of the currently fashionable trends in 
AI is the development of languages and systems based on 
some kind of frame or prototype representation. The basic 
motivation comes from the observation that much of what 
we know about the world is not in the form of simple 
logical statements, but in knowledge about what is typical 
or expected. If we represent the meaning of "buy" and 
"sell" in terms of a COMMERCIAL-TRANSACTION scenario 
which includes the transfer of money, we also want to be 
able to apply it to cases which involve the e~_change of 
valued objects other than money. However, we do not 
want to do this by creating an abstraction (e.g. the 
exchanged object is a VALUED-OBJECT) and thereby lose 
the information that it is usually money. 
Many papers have been written on the advantages and 
problems of including prototypical infoixnation as a 
fundamental part of a semantic representation. Formally, 
such systems are distinct from those based on logical 
abstraction only if issues of computational order and 
resources are taken into account (See Winograd (1976), for 
a discussion of these issues). However, it is important not 
to focus too narrowly on form rather than use: there is a 
clear difference in approach between the adherents of the 
alternate views. Some systems (such as Schank's (1972) 
system of primitives) are clearly based on prototypes even 
though they may not appear as such in the formal 
characterization. The inferences they draw from semantic 
decomposition are based on typical expectation, rather 
than logical certainty. 
Prototype-based systems have often gone along with a 
psychological view of the status of the symbols they use. 
Some of the motivation has come from psycholinguistic 
experiments which indicate that in many cases people are 
uncertain about the applicability of words to 'borderline 
cases', although they have a clear notion of the 'proto- 
typical case'. This applies to areas of the vocabulary as 
varied as color terms (Berlin and Kay, 1969) and simple 
nouns such as "cup", "glass", and "bowl" (Labov, 1973). 
The implication is that the semantic representation of 
words is organized around a set of 'most typical' cases 
rather than around a checklist of logical criteria which 
must be met for the word to be applied. 
EXEMPLAR. Extending the prototype notion one step 
further, some psychologists have suggested that our 
understanding of words is based on having exemplars 
which are drawn from experience. Rather than having a 
semantic prototype for "fruit", we may have an exemplary 
fiuit (e.g. a red apple) and understandthe use of the word 
by comparison to what we know about this apple. The 
line between prototypes and exemplars is not sharp, but 
there is a difference in emphasis. Prototypes emphasize 
the presence of information which is typical to the class of 
objects described by a word, while exemplars emphasize 
the ability to reason by comparing one specific object to 
another specific object, which may have its own peculi- 
arities which are not general to the class. 
Although there has been some discussion of reasoning 
by analogy (e.g. Moore and Newell, 1973), no system I 
know of has really made use of exemplars in a substantial 
way. There are many difficult issues surrounding the 
selection of the 'important' or 'invariant' aspects of the 
exemplar in a specific context. Critics of AI (e.g. Dreyfus, 
1972) see this as being impossible to adequately represent 
in a formal system. Whether this turns out to be ulti- 
mately true or not, we are far from having explored the 
potential for such reasoning within AI programs. 
What is a primitive? 
Before going on to the third dimension--the way in which 
the symbols within a semantic formalism are inter- 
related--it is useful to examine the notion of primitive 
which plays a central role in arguments on semantics. In 
understanding the properties of semantic primitives, it is 
helpfid to look at two other domains where primitives 
have played an important role: chemistry and math- 
ematics. Much of the thinking and discussion about 
primitives draws on conscious or unconscious comparisons 
with these two domains, often without recognition that 
they differ in some critical ways. 
Chemistry. One exemplar of a system based on primitives 
is the analysis of physical substances as structures made up 
of elements. There are atomic elements (note how much 
of the abstract vocabulary comes from this exemplar), and 
well-defined rules for the ways they can be combined into 
structures. Every substance, no matter how complex, can 
be analyzed as a compound of these primitive elements. 
The set of elements is experimentally determined and 
dealt with as a fact of nature--no two chemists would 
imagine postulating different sets of elements in their 
theories. Similarly, the structural analysis of a substance is 
not a matter of theoretical choice, but can be determined 
empirically. 
Mathematics. One of the methodological advances in the 
foundations of mathematics at the beginning of this 
century was the understanding of how complex mathema- 
tical systems could be constructed in a systematic way 
from small sets of primitive concepts. Beginning with a 
primitive basis (such as the notions of set, inclusion, and 
the null set), one can define complex constructions, and 
use these in still further definitions to build up ever- 
widening circles of complexity. In doing this, each new 
term is defined in terms of previous terms and simple rules 
of composition. The meaning of a complex term like 
"abelian group" or "divisor field" can be reduced step by 
step to primitives through these definitions. The choice of 
primitives is not determined by the domain to be covered. 
For any field of mathematics, there are alternative 
axiomatizations which take different things as primitive, 
and define others in terms of them. Even with the same 
set of primitives, there are alternative ways of defining 
27 
higher order concepts. For example, there are different 
ways of embedding the real numbers in the rational 
numbers for which it is quite difficult to prove 
equivalence. 
These two examples illustrate some typical features of 
primitives listed below (the terms used here are somewhat 
expanded from those in Wilks, 1977). Not every system 
based on primitives exhibits all of them, but they form a 
part of our understanding of what it is to be 'primitive': 
1. Finitude. A system contains a relatively small closed set 
of primitives. As it is applied to a wider range of things 
(substances, mathematical constructs, vocabulary items), 
the set of primitives remains fixed. The number of 
primitives should be substantially smaller than the number 
of things which can be reduced to combinations of 
primitives. 
2. Comprehensiveness. The set of primitives covers the 
range of phenomena. Every entity of interest can be 
expressed as a structure of primitives. For example, a 
chemist would be upset by a new substance which was not 
built of the available elements, and a mathematician would 
reject a new definition which was not in terms of the 
primitives of his or her axiomatization. 
3. Completeness. A description of an entity in terms of 
primitives is sufficient for generating all of the information 
about the entity, There are no 'hidden propeldes'. This 
does not mean that the information must be explicit--a set 
of mathematical definitions does not provide all of the 
theorems, but it does provide a basis for proving all those 
which could be proved. In the case of substances, this 
criterion does not apply. Information other than the 
chemical structure (for example energy, phase, crystalline 
structure, etc.) is needed for determining the properties of 
a substance. 
4. Independence. Primitives should not be definable in 
terms of one another. This is clear in the case of chemical 
elements, and in mathematics it provides a strong metric 
for judging axiomatizations. There is a high value placed 
on reducing the primitives to an absolutely minimal set. 
5. Canonicality. The analysis of an entity as a structure of 
primitives should be unique and unambiguous. Chemists 
agree on the structure of a compound as a unique formula. 
Within a particular axiomatization of a mathematical 
system, there is one and only one way a term such as 
"integer" is defined in terms of the primitives. 
6. Irreducibility. The meaning of a primitive cannot be 
expanded within the same level of theory. There are many 
issues here as to what a 'level of theory' is, but the 
application is clear in chemistry. The primitive elements 
can indeed be described as composite structures made up 
of even more primitive sub-atomic particles. But in doing 
so, we move from chemistry to atomic physics. For the 
purposes of doing normal chemistry, it is more useful to 
treat them as primitives. It is important to recognize that 
'primitivity' is always relative to an overall choice of the 
scope of the theory. 
In comparing the various forms of semantic primitives, we 
will look at the ways in which they match these criteria. 
The relational dimension 
The notion of primitive makes sense only within a system 
of interrelated terms. The basic idea of composition from 
primitives is only one of several possible ways of organ- 
izing such sets of relationships: 
PRIMITIVES. The most straightforward use of semantic 
primitives would be a system in which the full meaning of 
any word or phrase could be expressed as a structure 
whose components are chosen from a small set of primi- 
tives, combined according to a well-defined set of rules. 
No existing system is pure in this sense, as discussed 
below. 
MUTUAL. Another approach is to have a web of mutually 
related elements, with no primitive set on which to 
'bottom out'. A standard dictionary describes word mean- 
ing in this way. Words are defined using other words 
which are defined using others, and so on, inevitably 
leading to circularity. A mutually related system of terms 
can be either DEFINITIONAL or DESCRIPTIVE. In a 
DEFINITIONAL system, each item is defined by giving a 
structure made up of other items. The definition is 
complete, in that no information which is available from 
the term itself is lost by replacing it with the definition. In 
a DESCRIPTIVE system, each term is described by 
structures of other terms, but these do not necessarily 
capture its full meaning. Althotigh the dictionary is 
normally thought of as being DEFINITIONAL, this is the 
case only for very precise technical terms. For most of the 
common vocabulary, the 'dictionary definition' is a quite 
partial account of the meaning of the word. 
DISTINGUISHED. In systems based on mutual relations, it 
will often be the case that some terms tend be be used in 
definitions or descriptions much more often than others. 
There may be small finite distinguished subsystems of terms 
which form a standardized basis for a large number of 
descriptions. These terms need not be primitive in the 
senses discussed above--they may be further reducible, 
definable in terms of each other, and may provide only a 
partial coverage of the meanings to be expressed. 
However, there are organizational (and computational) 
advantages to granting them a privileged status in the way 
other definitions and descriptions are built up. In fact, 
most of the argument in favor of semantic primitives for 
AI systems has been (as we will see below) argument in 
favor of having one or more preferred subsystems within a 
mutually related system. 
Some examples 
The following table summarizes the dimensions and 
choices described above. In this section, we will use it to 
characterize a number of existing formalisms. 
28 
Ontological • Logical Relational 
iinguistic ~ (Abstractionl~ /Primitives ,Definitiona l sychological| ~ Prototype | ~lMutual 
heoretical / ~Exemplar / ~Distinguished~ tDescriptive~ 
Dimensions of choice in a semantic formalism 
The traditional dictionary. The traditional dictionary is 
clearly LINGUISTIC, based primarily on ABSTRACTION, 
and MUTUAL relationships. It varies between being DEF- 
INITIONAL and DESCRIPTIVE, and at times does include 
some PROTOTYPE information. The popular view of the 
dictionary tends to ignore the PROTOTYPE and DESCRIP- 
TIVE aspects. 
Theories from generative linguistics. Semantic theories 
within the Chomskian tradition of generative linguistics 
tend to be THEORETICAL, based on ABSTRACTION and 
PRIMITIVES. Katz and Fodor (1964), Jackendoff (1976), 
and Leech (1969) all fit these categories. There is an 
occasional hint of PSYCHOLOGICAL relevance, but it does 
not play a major role in the methodology. Within the 
school of 'generative semantics', there are many approa- 
ches. Much of FiUmore's (1974, 1975) work is an exam- 
ination of how PROTOTYPE and EXEMPLAR systems can 
provide insights which do not fit neatly into ABSTRAC- 
TION. Some of the earlier work on 'underlying verbs' 
takes a more LINGUISTIC turn, in which the underlying 
components are seen as closely related to actual lexical 
items. 
Semantics based on formal logic. Much of the work on the 
semantics of natural language has been closely related to 
work on the semantics of formal languages. This includes 
the classical work on issues like reference, and more recent 
attempts to view English as a formal language, as 
developed in Montague grammar. On the first two 
dimensions, this work is clearly THEORETICAL and 
ABSTRACFION based. On the third, the relationship 
between the symbols used for semantic representation 
carries over that of an underlying logical system. From 
the point of view of the semantic theory (the relationship 
between words and underlying entities), each p:edicate or 
constant is a PRIMITIVE. The fact that these are related by 
theorems, definitions, etc. within the logical system is 
independent of the semantic formalism in the same sense 
that the representation of elements in terms of sub-atomic 
particles is independent of ordinary chemistry. The clarity 
of this distinction (between the semantic roles and the 
• reasoning rules) is one of the advantages of this style of 
work, not shared by most AI programs, which use data 
structures and procedures which make no clear distinction. 
Conceptual Dependency. Schank has been one of the most 
insistent advocates of primitives, and his early (1972) work 
was clearly PSYCHOLOGICAL based on PRIMITIVES. As 
mentioned above, his attention to 'typical' inferences 
places it closer to PROTOTYPE than to ABSTRACTION. In 
trying to expand his theory beyond the set of simple 
actions for which it was initially developed, he has 
gradually shifted away from a strong PRIMITIVES based 
view, and has been one of the major developers of systems 
based on DISTINGUISHED subsystems. Schank and 
Abelson (1977), provide subsystems for actions, scales 
reflecting a person's state, causes, scripts, goals, plans, goal 
outcomes, interpersonal themes, and life themes. Their 
students have carried out the same kind of activity in other 
areas, such as the uses and classification of physical ob- 
jects. In all of this work, the emphasis is on finding a 
plausible and useful set of terms, rather than on justifying 
their primitive status. Most of the arguments are based on 
the pragmatics of doing language comprehension and 
reasoning within the system. 
KRL. KRL provides a language for representation within 
computer systems. As such, it is neutral between a 
PSYCHOLOGICAL and THEORETICAL stance, but the 
authors lean heavily towards the PSYCHOLOGICAL in 
developing their formalism. It is clearly based on 
PROTOTYPES, and much of the discussion (see Bobrow 
and Winograd, 1977) centers around this aspect. It is 
based on a MUTUAL DESCRIPTIVE set of relationships. 
DISTINGUISHED subsystems have been developed within 
specific applications (see Bobrow, Winograd, et. al., 1977), 
but these have not been a part of the basic formalism. 
Preference Semantics. Wilks' system of 'preference 
semantics' is one of the hardest to understand, since he 
seems to combine many different (and often incompatible) 
views. He insists that his system is based on PRIMITIVES, 
but it has few of the characteristics described above. In 
fact, his discussion argues strongly for the possibility of a 
MUTUAL DEFINITIONAL system, and he provides an 
interesting set of DISTINGUISHED subsystems (1977, 
Appendix A). In stating that "primitives are to be found 
in all natural language understanding systems" (1977, p. 
19) he seems to be using the term 'primitive' to cover any 
formal symbol used in a semantic system. He argues 
against the PSYCHOLOGICAL basis, but alternates between 
the other two possibilities along the ontological dimension. 
He is LINGUISTIC in stating that his formalism is con- 
sistent with the view that "Every semantic primitive can 
appear as a surface word in a natural language", and 
THEORETICAL in arguing that the primitives are part of an 
interlingual "primitive language" which is a "useful 
organizing hypothesis" which has no independent justi- 
fication in psychological terms, and "has no correct 
vocabulary, any more than English has". His formulas 
generally contain only ABSTRACTION information in their 
structure, but have PROTOTYPE information (or in his 
terms, 'preferences') in the assignment of types of objects 
to the nodes. 
OWL. The OWL representation is much closer to a 
LINGUISTIC base than any of the others listed here. It is 
described as a system of 'concepts', but its developers 
(Szolovits, Hawkinson, and Martin, 1977) have paid a 
good deal of attention to the way that natural language 
words and collocations can be preserved in the repre- 
sentation. It has a MUTUAL DESCRIPTIVE organization, 
which focuses on ABSTRACTION sorts of information, 
29 
although the semantics of the reasoning process are not 
clearly enough specified to distinguish between this and 
other choices on the logical dimension. The term 
'exemplar' is used in OWL to refer to sub-classes of a 
larger class, a concept related to but not the same as the 
one described above. 
Semantic networks. There are many versions of semantic 
networks, and it is hard to say anything which applies 
across the board. The majority have been argued on 
PSYCHOLOGICAL grounds, have focussed on ABSTRAC- 
TION information, although with some PROTOTYPE, and 
have been a web of MUTUAL DESCRIPTION. The network 
notation is well suited to MUTUAL (as opposed to 
PRIMITIVE), but is general enough to be used for almost 
anything. 
Properties of semantic systems 
The purpose of the classification given above is to provide 
a basis for comparing the merits and problems of 
alternative formalisms. Rather than arguing whether 
primitives are right or wrong, we will examine some 
desirable properties for semantic systems and see what 
they imply for the choices to be made along the three 
dimensions. This paper cannot hope to cover the full range 
of important issues, but as examples we will consider the 
following properties: 
The ability to state significant generalizations 
Criteria for deciding on a set of semantic entities 
Coverage of relevant semantic phenomena 
Canonicality and its effects on memory form 
Possibilities for dealing with extended meaning and 
metaphor 
The ability to state significant generalizations. The raison 
d'etre of a semantic theory is the desire to find regularities 
in the way language conveys meaning. Rather than 
enumerating the relationships among every possible set of 
texts, we can assign formal semantic structures to texts in a 
regular way, and systematically describe relationships 
between these structures. The theory is interesting to the 
extent that the formal semantic system allows us to find 
regularities and state broader generalizations than we 
could at the surface level. 
There are many possible views as to what kinds of 
generalizations are most interesting. Linguists look for 
generalizations which predict the judgements of native 
speakers as to whether sentences are well-formed. Some, 
like Jackendoff (1976) also look for generalizations as to 
the entailment relations between sentences. AI work, such 
as that of Rieger (1975) emphasizes inferential general- 
izations--that certain inferences will be made whenever a 
given underlying semantic structure appears. _-\I systems 
in general are based on 'reasoning' programs which make 
use of semantic representations to do reasoning which is 
independent of the specific linguistic form in which the 
knowledge was stated. 
In some discussions of primitives, it is implied that it is 
necessary to have a system based on primitives in order to 
make significant generalizations. It should be clear from 
the discussion above that this is a confusion of categories. 
Any system of formal semantics is based on generalization. 
The specific choice to base it on primitive decomposition 
may lead to a different set of generalizations, but not a 
necessarily better one. 
Criteria for deciding on a set of semantic entities. The 
main factor influencing the choice and justification of 
semantic entities within a formalism is the choice along the 
ontological dimension. Those who take a LINGUISTIC 
position need make no choice--the words of the language 
are themselves the entities of the semantic theory. There 
is work to be done in determining the relations between 
them, but the set of entities is given from the beginning. 
Those who take a THEORETICAL stance are flee to create 
semantic entities at will, but must justify them by demon- 
strating that the set chosen leads to generalizations and 
simplifications which are not shared by alternative sets. In 
the generative grammar tradition, a good deal of attention 
is given to finding a highly valued set. Through careful 
work, one can construct tests in the form of sentences 
whose acceptability would be predicted by one possible 
set, and not by another. Simplicity of stating the semantic 
theory is used to choose between sets with equal coverage. 
In the AI tradition, the selection of entities is more 
intuitive and less careful. A system as a whole is claimed 
to 'work', and there is little precise evaluation of which 
aspects of the formalism were critical, and what might be 
done with alternatives. In this context, there are only 
vague intuitions and heuristics to guide the choice of 
entities and their relationships. Wilks accepts this, in 
noting that "no direct justification of the vocabulary \[of 
primitives\] makes any sense." 
The m~,st interesting problems arise if the formalism is 
intended as a PSYCHOLOGICAL theory. In this case, the 
determination of a set of semantic entities is an empirical 
question. There is an implicit claim that there are 
functional equivalents to the elements of the semantic 
theory within the psychological activities of compre- 
hending and generating language. It is possible to invent 
experiments which can choose between alternative theories 
according to the detailed predictions they make about 
human performance. Some of the distinctions above (such 
as that between ABSTRACTION, PROTOTYPE and EXEM- 
PLAR) grew out of experiments of this type. However, 
there is a large gap between the isolated examples handled 
in experiments and the kind of coverage needed in a 
comprehensive semantic formalism. Those people in AI 
who have built large-scale systems have not looked to 
detailed psychological justifications, even though they 
often informally describe their formalism as a 
psychological theory. When Schank (1972) calls his 
formalism 'conceptual dependency', or Jackendoff desc- 
ribes his system as using 'cognitive primitives' the appeal 
to psychology is suggestive, not of direct relevance to the 
methodologies they follow. 
30 
i 
Within a PSYCHOLOGICAL viewpoint, there are many 
further issues as to the generality of the postulated 
semantic entities. Are they idiosyncratic, or shared by all 
competent speakers of a language? Are they language- 
specific, or do they represent a more basic experiential 
knowledge which cuts across cultures and languages? If 
they are not language-specific, then are they innate or 
learned? There has been some interesting work done on 
these questions in very specific semantic domains such as 
the lexicon for describing colors, but once we move 
outside of these limited domains, most of what can be 
said is anecdotal or purely speculative. 
Coverage of relevant semantic phenomena. In developing 
a comprehensive semantic theory, there are many aspects 
of meaning which must be taken into account. A 
formalism which is developed for one aspect of meaning 
(for example, the hierarchical relationships between the 
classes named by common nouns) may be inadequate or 
completely irrelevant for others (for example, the ways in 
which participants are related to events). In some cases, a 
general approach cuts across several aspects. Much of the 
discussion of primitives and prototypes above can be 
applied both to classification (for example, Schank's (1972) 
classification (:,f acts vs. LakofFs (1977) 'gestalts') and to 
the case relationships between participants and an act 
(Fillmore's (1968) notion of a primitive set of cases vs. the 
Bobrow and Winograd (1977) notions of hierarchies of 
prototypes with named 'slots'). 
Existing semantic formalisms are all partial, and many 
of the arguments in the literature are of the "I can do 
something you can't do" style. It is clear, for example, 
that PRIMITIVES are not well suited for handling the broad 
vocabulary of nouns and verbs describing the objects and 
actions of our world in all their variety. As Wilks says, 
"No representation in primitives could be expected to 
distinguish by its structure hammer, mallet, and axe.'" 
Formalisms based on ABSTRACTION are problematic when 
we attempt to deal with lexical fields where there are no 
clear criteria for whether a word applies. This includes the 
naming of simple objects, such as "cup" and "bowl" 
(Labov, 1973), as well as the more obvious areas of 
metaphor. On the other hand, alternatives, such as 
PROTOTYPE systems based on MUTUAL relations have 
been far less developed in the details of the generalizations 
they allow, and the specification of how they would deal 
with any specific semantic domains. 
It is clear that no formalism at this point has a claim to 
"Anything you can do, I can do better." Intuitions as to 
which aspects of language are most central play the 
leading role in determining which of the competing 
theories seems most promising. 
Canonical form and its effects on memory and reasoning. 
In early work on semantic primitives, there was a good 
deal of debate about the advantages provided by a 
canonical form for the representation of meaning. Two 
words or sentences with the same meaning have identical 
semantic representations in a formalism based on 
canonical form. In other formalisms, they may have equi- 
valent representations (anything inferrable from one would 
be inferred from the other) which nevertheless differ in 
form. Typically, PRIMITIVE systems tend to support a 
canonical form, while MUTUAL organizations do not. 
However, DISTINGUISHED subsystems can be used to cre- 
ate a canonical form for their particular aspect of meaning 
in a system which does not depend on primitives. By 
choosing to always expand into the terms of this subsystem 
in the same way, all of the properties of canc fical form 
apply. 
In evaluating the benefits of canonical form, it is 
important to take into account the procedural aspects. In 
its simplest usage, each piece of input text is converted 
immediately to canonical form and stored that way. 
Inferences are based on the elements of this expanded 
form, and memory search depends on finding the form 
corresponding to the query as a subset of what is stored. 
In a more sophisticated use, the canonical form is available 
for potential expansion, but memory can include unex- 
panded structures built up out of a vocabulary of non- 
primitive semantic entities. Expansion is done only when 
needed for a specific task such as matching a new input to 
previous knowledge in answering a question. The 
advantages and disadvantages of canonical form are 
somewhat different for these two organizations. The 
primary ones can be summarized: 
1. Absence of ambiguity and vagueness. This property 
applies to the canonical form after expansion. It is a 
global property of systems based on expansion at 
input--since meanings are expanded into canonical struc- 
tures of primitives at the time they are analyzed, there is 
no remaining uncertainty about their meaning. This is 
viewed as an advantage by those who emphasize the use of 
the formalism in abstract reasoning, and as a disadvantage 
by those (like Martin, 1976) who emphasize the impor- 
tance of context and interpretation in using knowledge. 
Martin argues that a semantic representation for natural 
language must share its ability to represent imprecise 
meanings. 
2. Reasoning activity at input time. The process of 
expansion to canonical form can be used as a procedural 
driver for carrying out inference. Much of the work on 
conceptual dependency makes use of this organization. 
The advantage is a unitorm way o! tnggenng stanOarcl 
inferences. The disadvantages come from the problems of 
triggering too much----of drawing inferences far below the 
level of detail relevant to the particular context because 
the canonical form demands expansion to that level. 
3. Uniqueness for indexing and search. A canonical form 
can be stored and indexed in a uniform way which makes 
it possible to use straightforward algorithms for memory 
search and consistency checking. These have the 
advantages and disadvantages of most uniform procedures 
for dealing with complex structures--they are easy to write 
and understand, but they suffer from combinatorially 
explosive inefficiency and tend to bog down for all but 
tiny toy bodies of knowledge. One of the fundamental 
technical differences among existing systems is in whether 
they emphasize uniformity (as in most logic-based systems, 
and in early versions of conceptual dependency) or the 
31 
provision of explicit tools for controlling memory search 
and inference (as in KRL). 
4. Association of inference rules with primitive elements. 
In a system which is expected to expand meanings into 
canonical form (either at input time or in the process of 
reasoning), inference rules can be associated with the most 
general primitives (e.g. GO, used in a sense which covers 
all sorts of change, as in Jackendoff (1976)). In a system 
which does not expand to a common base, the same 
inference might have to be repeated in a number of places. 
The disadvantage arises in the case where an inference is 
associated with a higher-level meaning (such as "flee" 
having implications not shared by other instances of 
going). In a fully canonical system, it is necessary to 
recognize the particular combination of primitives which 
triggers the inference. In systems like that of Rieger 
(1975), there are discrimination nets, used to sort out the 
appropriate inferences from the expanded forms. This 
again leads to a combinatorial problem which becomes 
untenable in all but the smallest systems. Like the other 
issues, this one is complicated by the ability to build 
systems which partake of canonical expansion to some 
degree, either by expanding only along certain dimensions, 
or by operating with a mixture of expanded forms and 
non-primitive-based forms from which they were derived. 
Possibilities for dealing with extended meaning and 
metaphor. A recurring theme in discussions of semantics 
is that of metaphor. Any reahstic view of language must 
take into account the fact that words are used in ways 
which defy simple analytic characterization of their 
meaning. There are explicitly poetic metaphors, conven- 
tional metaphors ("His ideas were beyond me", "Carter 
named three _nain targets in his war on inflation"), and a 
wide range of cases in which meanings are extended 
beyond their prototypical application. For example, if we 
define "spend" in terms of a commercial transaction, then 
it must be extended to deal with "I spent a week in 
Boston." In general, formal semantic theories have not 
gone very far in dealing with these problems. Those who 
base systems on PROTOTYPE or EXEMPLAR reasoning 
argue that this is an important step towards dealing with 
the fuzzier aspects of language. However, the compu- 
tational details needed to make the power of such systems 
clear have not been filled in. They either stick to trivial 
cases (as in Moore and Newell, 1973), or operate in ways 
which do not depend on going beyond standard logical 
meaning. This area remains one of the most tantalizing 
and difficult for future research. 
REFERENCES 
Berlin, B. and P. Kay, Basic color terms: their universality and 
evolution, Berkeley: Univ. of California Press, 1969. 
Bobrow, D.G. and T. Winograd, An ove~iew of KRL, a Knowledge 
Representation Language, Cognitive Science 1:1 (January, 1977), 3-46 
Bobrow, D.G., T. Winograd, and the KRL Research Group, 
Experience with KRL-0: One cycle of a knowledge representation 
language, Proceedings of the Fifth International Joint Conference on 
Artificial Intelligence (August, 1977), 213-222. 
Clark, H.H., and E.V. Clark, Psychology of Language: An. 
h+troduction to Psycholinguistics, New York: Harcourt Brace, 1977. 
Dreyfus, H. L., What computers can't do: a critique of artificial reason, 
New York: Harper & Row, 1972. 
Fillmore, C., The case for case, In Bach and Harms (Eds.), Universals 
in Linguistic Theory, Chicago: Holt, 1968, 1-90. 
Fillmore, C., The future of Semantics, Berkeley Studies in Syntax 
and Semantics I, Dept. of Linguistics, Univ. of California Berkeley, 
1974. 
Fillmore, C., An Alternative to Checklist Theories of Meaning, 
Proceedings of the First Annual Meeting of the Berkeley Linguistics 
Society, Cogen et al. (Eds.), University of California, Berkeley, 1975. 
Fodor, J.A., The Language of Thought, New York: Cromwell, 1975. 
Fodor. J.A., Methodological solipsism as a research strategy in 
psychology, unpublished draft, 1978. 
Jackendoff, R., Toward an explanatory semantic representation, Linguistic Inquiry 
7:1 (Winter, 1976) 89-150. 
Katz, J.J,, Semantic Theory, New York: Harper and Row, 1972. 
Katz, J.J., and J.A. Fodor, The Structure of a Semantic Theory, in J. 
Fodor and J. Katz, (eds.) The Structure of Language, Prentice Hall, 1964. 
Labov, W., The boundaries of words and their meanings, in C-J. N. 
Bailey and Roger Shuy (eds.), New Ways of Analyzing Variation in English, Georgetown Univ., 1973. 
Lakoff, G., Linguistic Gestalts, Proceedings of the Chicago Linguistic Society (CLS 13). 1977, 236-287. 
Leech, G., Towards a semantic description of English, London: Longman, 1969. 
Lewis, D., General semantics, in Davidson and Harman (eds.), 
Semantics of Natural Language, Dordrecht: Reidel , 1972. 
Martin, W.A., A theory of English grammar, unpublished notes, MIT, 1976. 
Miller, G.A., Some psychological studies of grammar, American Psychologist 17 (1962), 748-762. 
Miller, G.A., and P.N. Johnson-Laird, Language and Perception, 
Cambridge: Harvard University Press, 1976. 
Moore J., and NeweU. A., How can MERLIN understand?, In 
Gregg (Ed.), Knowledge and Cognition, Baltimore, Md.: Lawrence Erlbaum Associates, 1973. 
Rieger, C., Conceptual memory and inference, in R.C. Schank, 
Conceptual Information Processing, Amsterdam: North Holland, 1975, 157-288. 
Schank, R. C., Conceptual dependency: A theory of natural 
language understanding, Cognitive Psychology, 1972, 552-631. 
Schank, R.C. and R.P. Abelson,, Scripts Plans Goals and 
Understanding, Hillsdale: Lawrence Erlbaum Associates, 1977. 
Szolovits, P., L.B. Hawkinson, and W.A. Martin, An Overview of 
OWL, an language for knowledge representation, M.I.T. LCS-TM-86, 1977, 
Wilks, Y., Good and bad arguments about semantic primitives, 
D.A.I. Research Report No. 42, University of Essex, May 1977. 
Winograd, T., Towards a Procedural Understanding of Semantics, 
Revue lnternationale de Philosophie, 1976 fasc. 3-4 (117-118). 
32 
