A Knowledge Representation Approach to Understanding Metaphors 
E. Judith Weiner 
Computer and Information Sciences 
Computer Activity Building 
Temple University 
Philadelphia, PA 19122 
This study represents an exploration of the phenomenon of non-literal language 
("metaphors") and an approach that lends itself to computational modeling. Ortony's 
theories of the way in which salience and asymmetry function in human metaphor process- 
ing are explored and expanded on the basis of numerous examples. A number of factors 
appear to be interacting in the metaphor comprehension process. In addition to salience 
and asymmetry, of major importance are incongruity, hyperbolicity, inexpressibility, 
prototypicality, and probable value range. Central to the model is a knowledge representa- 
tion system incorporating these factors and allowing for the manner in which they interact. 
A version of KL-ONE (with small revisions) is used for this purpose. 
1. Introduction 
One can hardly fail to notice the flurry of intellectual 
activity that currently surrounds the understanding of 
the use of figurative language. The interest is multi- 
disciplinary - linguistics, psychology, philosophy, edu- 
cation, to name a few of the more active disciplines. 
The reason, which anyone writing on the subject has- 
tens to point out, is that the observation of natural 
speech demonstrates clearly that it is rarely confined 
to the strictly literal. Figurative language is not mere- 
ly an ornament of the poet but abounds in the every- 
day speech of everyday people and as such is a legiti- 
mate area of inquiry for researchers - in any discipline 
- who are concerned with understanding natural lan- 
guage. The interest in metaphor in computer under- 
standing of natural languages stems from this same 
source. It is well understood that people, when con- 
versing with machines, can no more be constrained to 
literal language than they can be expected to be long 
contented, within the confines of a synthetic language. 
2. Scope of the Study 
The heading "figurative language" comprises the tradi- 
tional figures of speech know as synechdoche, metony- 
my, hyperbole, personification, irony, etc., as well as 
the more common metaphor and simile. I am going to 
focus here on these latter two in order to narrow my 
view in the hope of achieving some depth and also 
because of a belief that the other figures may operate 
under similar principles. Except where noted, I will 
use the term "metaphor" in referring to both similes 
and metaphors. Please note that this does not imply 
that I am taking the position that metaphors and sim- 
iles are the same; in fact, there is some evidence that 
they function differently from one another. At the 
least, it seems possible that the distinction between 
these two is more than the traditional one of implicit- 
ness versus explicitness since there are instances of 
metaphors that sound strange when "transformed" 
into similes and vice versa. I therefore am using the 
term "metaphor" in a very loose way to cover the area 
metaphors and similes have in common (for example, 
the similarity in the figurative reads of John is an 
animal and John is like an animal), without pausing at 
this time to delve into its exact nature and ignoring for 
the moment the apparent differences. 
To start, I will work only with isolated sentences of 
the form 
(1) A is (like) B. 
In sentences of this form, A is commonly referred to 
as the "topic", the B term as the "vehicle". That 
which they have in common is called the "ground". In 
a sentence like 
Copyright 1984 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted 
provided that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the 
first page. To copy otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/84/010001 - 14503.00 
Computational Linguistics, Volume 10, Number 1, January-March 1984 1 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
(2) Billboards are like warts 
then, the topic is billboards, the vehicle warts and the 
ground ugliness and (perhaps) prominence. 
In restricting this study to sentences of the form 
(1), my motives here again are to constrain the un- 
wieldiness of the subject. Of course this rules out a 
large body of possible metaphorical utterances of other 
forms. Many of these, however, if confined to one 
sentence, could be restated in the form of (1) with no 
significant loss of meaning. I will not discuss them 
here. I will, however, have something to say later 
about the larger linguistic context (discourse.) 
The typed word, the presumed form of input of 
natural language to a computer until such time as actu- 
al speech understanding systems develop sufficiently, 
imposes limitations of its own on the scope of any 
language processing system. The most obvious is, of 
course, that variation in intonation of the input is lim- 
ited to its most "neutral" pattern; prosodic features 
must largely be ignored. (A certain amount of empha- 
sis or contrastive stress may be obtained by underlin- 
ing, but the study of this should be considered sepa- 
rately.) Another, perhaps more relevant consideration, 
is the use of a space to separate parts of what must be 
considered a single lexical item, e.g., blind alley. His- 
torically, this was undoubtedly a metaphor (and a 
candidate for this study); today it is most probably 
interpreted as a single unit. Although most native 
speakers of English would classify it as an idiom 
through an awareness of the written form and the fact 
that even in the spoken version the component parts 
are clearly recognized, there should be no attempt to 
componentially process such forms. 
3. Salience 
One of the most useful notions for modeling meta- 
phoric understanding is that of salience (Ortony 
1979a), which Ortony takes to mean an estimation of 
"prominence of a particular attribute with respect to a 
concept to which it does or could apply." (p. 162.) 
He later speaks of "predicates" rather than attributes" 
(1979b, p. 191): "A predicate can be attributed to, or 
predicated of, something. It can represent knowledge, 
a belief, or an attitude about or toward something." I 
too prefer the flexibility of "predicate" and shall fol- 
low Ortony in the use of this term. 
The notion of salience makes use of the apparent 
fact that metaphorical statements are asymmetric: 
(2) Billboards are like warts 
means something different from 
(3) Warts are like billboards 
Ortony's explanation is that in isolated sentences of 
the form 
(1) A is (like) B 
those predicates that have high salience in B and low 
salience in A are the ones being considered in the 
metaphor. 1 The effect is one of raising the salience of 
these predicates in A. Thus, sentence (2) is generally 
understood as meaning that billboards are ugly, where- 
as in (3), those predicates that have high salience for 
billboards (but low salience in warts) - for example, 
prominence - are attributed to warts (that is, the sali- 
ence is raised). An additional requirement is that 
there be high salient predicates of B that cannot apply 
to A. 2 
My working definition of salience includes the as- 
sumption of a taxonomic structure of concepts with 
the most general at the top and the most specific at 
the bottom. Figure 1 provides an illustration of a 
simple taxonomy. 
I define a salient predicate of a concept as one that 
implies inherent prominence (for example, saturation 
of color, largeness of size, etc.) or else is definitional 
in that it entails a concept's separation from others in 
the hierarchy (for example, the dog's domesticity sepa- 
rates it from the wolf). My notion of salience paral- 
lels that of Tversky (1977) in that intensity and diag- 
nosticity are the critical factors, but it makes the addi- 
tion of hierarchical organization to facilitate 
diagnosticity. 3 Context, both linguistic and extralin- 
guistic, is of course a major contributing factor too, 
but it is outside the scope of this study. The way in 
which these factors interrelate is a fertile area for psy- 
chological research. 
4. Prototypicality 
Another valuable contribution coming out of cognitive 
psychology is prototype theory (Rosch 1973, Rosch 
and Mervis 1975), which holds that a concept may 
belong to a category even if it is somewhat atypical in 
terms of the predicates usually (typically) associated 
with members of that category. A chicken is a bird 
even though it can't really fly. Here, bird refers to 
some prototype from which chicken represents a de- 
parture. 
In terms of metaphors, there is much value in in- 
eluding prototype theory in any model. For example, 
(4) Mary's cheeks are like apples 
10rtony acknowledges that the att.!butes may be similar, not 
identical, in the vehicle and topic (1979a, p. 167). While I am in 
agreement with Ortony, I will, for the purposes of this paper, make 
the assumption of predicate identity. 
2 In testing this hypothesis experimentally, Gentner (1980) 
showed that salience did not appear to be a relevant mechanism in 
metaphor processing. This seems to me, however, to be partially a 
result of how salience was measured and the need for a clearer 
analysis of how metaphoric interpretation proceeds. Salience, 
properly defined, may provide a necessary but not sufficient expla- 
nation. 
3 Conklin and McDonald (1982) have used salience as a means 
of solving the selection problem in natural language generation 
using KL-ONE as the representation language. 
2 Computational Linguistics, Volume 10, Number 1, January-March 1984 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
Figure I. 
would probably mean to most people that Mary's 
cheeks are round and red. A different interpretation 
would be obtained if the concept of round, red apple 
were replaced by a withered, rotten one or even, for 
that matter, by a green one. For communication to 
take place among people in a speech community, some 
sort of prototypicality considerations are essential. 
There is another way in which prototypicality might 
figure in a discussion of metaphors. It seems probable 
that the B term in 
(1) A is like B 
represents the epitome of the predicate(s) that are true 
of A and of interest in a given utterance, that is, B is 
the prototypical representative of these predicates. 
(Tversky (1977) observed that the B term is the more 
prototypical of the two in literal sentences.) If the 
sentence reverses A and B, then A becomes the proto- 
type of (probably) different predicates. The vehicle of 
choice should be one in which the cluster of predicates 
is (ideally) uniquely appropriate, prototypical, and 
therefore also salient. For example, 
(5) A hose is like a snake 
Snakes are typically, even classically, the ultimate in 
long, narrow, coiledness; these characteristics can be 
thought of as distinguishing snakes from other mem- 
bers of the category ANIMALS. Sentence (5) draws 
the reader's attention to these (perhaps) slightly less 
salient qualities of a hose. 
I have intentionally limited myself to sentences 
taken out of the discourse context. One of the bene- 
fits of doing this is that there is considerable context 
within the sentence itself that can influence its inter- 
pretation. Consider the following pair of sentences 
(6) My cat's tail is like a carrot 
(7) John's hair is like a carrot 
Without adding any context, it is unlikely that a per- 
son would miss the fact that the relevant salient predi- 
cate of (6) is shape (and perhaps color) and that of 
(7) is color. Our knowledge of prototypical cats with 
prototypical cat tail shapes and colors and prototypical 
hair shapes and colors leads us to the right conclusion. 
If no additional information is available, then it 
wouldn't be likely that John was wearing a pony tail. 
On the other hand 
(8) John's nose is like a carrot 
again, taken out of context, would indicate a comment 
about shape. From these examples it should be clear 
that prototypicality considerations are relevant to both 
topic and vehicle. 
5. Prototypicality and Possibility 
How then, does this relate to prototype theory? The 
relationship appears to be that, in the prototypical tail, 
nose, or hair, certain predicates are probable. These 
are the ones most likely to match those salient in B. 
Thus in order to process metaphors, it is necessary to 
know, in addition to the nature of the prototype, a 
range of probable values for a given predicate. This 
would facilitate Ortony's determination (1979a, p. 
173) of "whether any gross incompatibility would 
result by applying the predicate in question to the 
Computational Linguistics, Volume 10, Number 1, January-March 1984 3 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
concept." This range can help determine whether the 
statement is literally true or not. For example, 
(9) John's hands are like ice. 
If a range of possible temperatures were built into the 
representation for human hands, it would be known 
that John's hands could not possibly be literally as 
cold as ice (there could not be an actual equivalence 
of temperature in hands and in ice). 4 The figurative 
interpretation would thus prevail. 
Prototypicality and range of possible values appear 
to operate throughout language as a whole; they are 
not confined to figurative usage. Labov (1973) re- 
ported on an experiment on the denotation of the 
word cup. He gave subjects pictures of cuplike objects 
to identify and observed differing percentages of the 
use of cup as form and function were varied. He con- 
cluded the existence of an invariant core 
(corresponding to my use of prototype) as well as a 
range of deviations through which recognition still 
occurred, albeit at lower percentages. The inclusion of 
these elements in the knowledge representation of a 
system for understanding natural language is therefore 
broadly motivated. 
6. Metaphors as Hyperboles 
All metaphors are hyperbolic in a sense. They seem to 
say: the predicates A shares with B are in A so ex- 
treme that they can only be expressed by relating them 
to some object in which they are epitomized, that is, 
B. In 
(9) John's hands are like ice, 
the exaggeration is apparent. This is evident because 
of the range of possible temperature values known to 
be actually attributable to human hands. The sentence 
(7) John's hair is like a carrot. 
is less metaphoric (consequently more literal) in this 
sense. It is possible that hair could be the same color 
as the prototypical carrot, but the probability is low. 
Consequently, the following sequence does not seem 
absurd: 
(10) John's hair is like a carrot. 
Is it really that color? 
whereas 
(11) John's hands are like ice. 
Are they really that cold? 
would lead one to think the response peculiar at the 
least. On some scale of metaphoricity then, (10) is 
less metaphoric than (11). Since a hose may in fact 
be as long, narrow, coiled as a snake, 
4 If John is assumed to be living and context does not indi- 
cate the possibility of frostbite conditions. 
(5) A hose is like a snake 
is the least metaphoric of the examples given. This 
supports Ortony's claim that high salient predicates of 
A matched with high salient predicates of B make for 
a literal statement. A response of Is it really that long, 
narrow, coiled? could easily be followed by an unquali- 
fied response of "'Yes. '" 
Ortony (1975) denies the possibility of the ground 
consisting merely of a single predicate. "People sim- 
ply do not use metaphors to transfer one characteristic, 
even if it is a distinctive one, when there is a ready 
literal way of making the point." (p. 50.) Sentence 
(9), however, provides a fairly good counterexample to 
this claim. Here, the hyperbolic nature of the meta- 
phor rather than the size of the ground provides the 
incentive for its use. 
7. Taxonomic Structure and Incongruity 
The conclusion which should be drawn from this dis- 
cussion is that all of the above factors must be brought 
to bear in an analysis of metaphor understanding. The 
result of using these measures will be an isolation of 
those predicates of B that are true of A and the estab- 
lishment of a relative degree of metaphoricity within 
the sentence context. 
Given a corpus of sentences of the form 
(1) A is (like) B, 
some will appear to be literal similarity statements; 
others will appear to be metaphors. 
(12) John is like his father. 
(13) John is like a snake. 
(14) John is like a black box. 
In (12), the sentence appears to be a literal compari- 
son. John shares certain characteristics with his fa- 
ther. John and his father are already known to be 
similar on the basis of the fact that they are members 
of the same superordinate category (males) or are 
already known to participate in a relationship to one 
another (father-son). There is no element of surprise 
or incongruency in statement (12). As Ortony has 
proposed, high salient predicates of B are also high 
salient predicates of A. This is Ortony's criterion for 
a literal similarity statement as opposed to a metaphor. 
I think, however, that viewing this phenomenon from 
the perspective of category membership, relationship 
and consequent incongruity will shed more light on its 
computation representation. 
There do seem to be metaphorical statements in 
which there is matching of high salient predicates in 
both the vehicle and the topic. If sentence 
(8) John's nose is like a carrot 
were uttered by one of John's friends to another, it 
would not represent new information. It would proba- 
4 Computational Linguistics, Volume 10, Number 1, January-March 1984 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
bly produce laughter because of the hierarchical incon- 
gruity but beyond this should have no more of an ef- 
fect than a literal paraphrase of the same sentence. In 
terms of those predicates that can be shared, those 
peculiar to the vehicle's hierarchical position - for 
example, in (8), a carrot's being a root crop - seem 
the least likely candidates. 
The position in a taxonomic hierarchy is important 
in another way. Consider 
(15) Penguins are like wolves. 
(16) Dogs are like wolves. 
(16) is a similarity statement; (15), assuming an inter- 
pretation can be found, is a metaphor. Figure 1 pro- 
vides a possible explanation: In terms of this diagram, 
dogs and wolves are siblings but penguins and wolves 
are not. The latter relationship is more distant, the 
sentence more metaphorical. Although distance met- 
rics are notoriously difficult to pin down, fairly clear 
cases like this one indicate that members of categories 
at a level of abstraction close to that of Rosch's basic 
level categories (1973) and with a shared superordi- 
nate can be considered in some way closer than those 
without a shared one. Empirical research on other 
categorical relationships may provide additional sup- 
port for this approach. Although less is known about 
the relationships in the examples that follow, since 
they no longer deal with a basic level category and its 
immediate superordinate, future research may shed 
some light on this phenomenon as well. 
For example, that incongruity may be a factor of 
"goodness" of metaphor seems to be illustrated by 
(13) and (14). Metaphors may be judged by the ap- 
parent unlikeliness of an A B juxtaposition. This of 
course is under the provision that the metaphor is 
understandable to the hearer. Thus (14), if under- 
stood to mean that John is somehow unknowable, is 
better than (13) in which John is thought to be 
sneaky. The difference may be that in (13) A and B 
are of a shared superordinate category (ANIMALS), 
whereas in (14) they are categorically more remote. 
Incongruity appears to be the reason that the best 
metaphors often produce a smile by the hearer as they 
are comprehended. ("How unlikely yet how apt" may 
be the attendant thought.) So there is a connection 
between metaphor and humor, or the intelligence and 
wit of the speaker who first utters a good, novel meta- 
phor. 
8. Number of Shared Predicates 
In addition to incongruity, another attribute of meta- 
phoric quality is the number of shared characteristics 
under incongruous circumstances. The more the better 
so long as incongruity is maintained: 
(17) Jane's eyes are like stars 
Although (17) is somewhat hackneyed, its survival and 
wide use may be a result of the fact that it is good in 
the sense just described, that is, incongruity plus mul- 
tiplicity of shared predicates (twinkliness, brightness, 
beauty, clarity, etc.). 
Some writers have talked of the magic of meta- 
phors, the idea that the whole is equal to more than 
the sum of its parts (Verbrugge 1980). While there 
may be other factors involved than those I have men- 
tioned above, I believe ultimately that metaphors can 
at least theoretically be accounted for formally and 
without appeal to the supernatural. 
Figure 1. 
Computational Linguistics, Volume 10, Number 1, January-March 1984 5 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
9. Expressing the Inexpressible 
Why do people use metaphors? Inexpressibility in 
literal terms seems to provide part of the answer. 5 
Hyperbole as discussed above is one way that this is 
overcome. When an extreme in terms of predicates is 
unavailable to the speaker, he or she may search other 
domains (categories) for something that epitomizes the 
desired quality. Metaphor makes the difficult to ex- 
press possible to express. 
This provides an explanation of the fact that human 
emotions (love, hate, etc.) are often described meta- 
phorically, that is, the less well-understood in terms of 
the more thoroughly understood. It also helps explain 
the observation that more abstract concepts with few 
if any high salient predicates are used for A while 
more concrete concepts (with predicates of high sali- 
ence) are used for B. There are many examples of 
metaphors for love, friendship, etc. in which B is a 
concrete concept having usable salient predicates. 
Metaphors appear commonly with regard to people 
to express something about their more abstract person- 
al characteristics (personality, character, value) rather 
than the more easily stated physical attributes. 6 Thus, 
when searching for the meaning of such metaphors, all 
other things being equal, one should generally exhaust 
those first. When one says, 
(18) You are my sunshine, 
one is not attributing yellowness to the addressee but 
rather those characteristics that are both more abstract 
and possible to attribute to a human - that is, warmth, 
brightness, cheerfulness, etc. If such characteristics 
are unavailable or not salient in the B term, then oth- 
er, physical ones are used in the interpretation. 
(19) Jane is a string bean, 
for example, makes a statement about Jane's shape. 
It should be noted at this point that I am not mak- 
ing any claims about the historical primacy of physical 
over more abstract predicates (as do Lakoff and John- 
son 1980). I am not assuming (nor am I contradict- 
ing) the possibility that this is true. The process by 
which it came to pass that we can now say "John is a 
bright person" or "That is a bright light" is not the 
issue. It does seem reasonable that the abstract was at 
one time a metaphor based on the physical, but I don't 
know what the historical evidence is for this. My 
analysis of metaphor is strictly a synchronic one. I am 
50rtony's inexpressibility thesis (1975) deals with transfer- 
ring from vehicle to topic "characteristics which are unnameable." 
6 Carbonell (1981) has proposed an invariance hierarchy for 
explaining this phenomenon in which physical descriptors occupy a 
relatively low position. An interpretation is obtained by searching 
downward through the hierarchy and stopping When knowledge 
common to A and B is encountered. 
saying that this is how the language at any given his- 
torical stage can be perceived as operating. 
10. A Word about Context 
In extending the context of a metaphorical sentence to 
include the surrounding linguistic environment, Ortony 
(1979a) has suggested that the effect of the linguistic 
environment is to raise the salience of certain predi- 
cates. To analyze this notion, consider the following 
pair: 
(20a) Look how the highway curves 
(20b) It's just like a snake. 
(20a) can be thought of as raising the salience of cur- 
viness in that sentence. Compare this pair to 
(21) This highway is like a snake, 
that is, the isolated sentence. The difference between 
(20) and (21) is that in (20) the discourse phenome- 
non of focus (Grosz 1977, 1981) is operating. Atten- 
tion is focused (and salience consequently raised) to 
the snakelike curviness of the highway. Note however 
that the other attributes that have high salience in (20) 
can also be applied metaphorically. It is just that one 
particularly snakelike attribute is highlighted in the 
pair. 
The following linguistic environment is also of im- 
portance and 
(22) Look at how he eats. 
Isn't John a pig! 
(23) Isn't John a pig! 
Look at how he eats. 
Metaphors in literature, especially poetry, represent 
a possible limit to which a computational model might 
aspire since poets are experts in the novel use of lan- 
guage and in explicating human experiences. But it is 
true that these sources should not be overlooked just 
because they seem to present difficult problems; a 
language understander should at least theoretically be 
able to understand poetic metaphors. This poem by 
Emily Bront~ 7 is metaphorically fairly straightforward 
as poems go and a good illustration of how metaphors 
in context can function. 
Love and Friendship 
Love is like the wild rose-briar, 
Friendship like the holly tree- 
The holly is dark when the rose-briar blooms 
But which will bloom most constantly? 
The wild rose-briar is sweet in spring, 
Its summer blossoms scent the air; 
Yet wait till winter comes again 
7 The Mentor Book of Major British Poets, edited by Oscar 
Williams, 1963. 
6 Computational Linguistics, Volume 10, Number 1, January-March 1984 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
And who will call the wild-briar fair? 
Then scorn the silly rose-wreath now 
And deck thee with the holly's sheen, 
That when December blights they brow 
He still may leave thy garland green. 
The first two lines represent two similes of the sort 
that I have been discussing; in each case an abstract 
concept is juxtaposed with a concrete one. The re- 
mainder of the poem proceeds to describe the ground 
of the similes. Certainly the ground consists of some 
of the low salient predicates of wild rose-briars and 
holly bushes. This seems to expand Ortony's thesis 
that high salient predicates of B are the only ones to 
be considered. Those of low salience are also eligible. 
Including these predicates in the context of the poem 
serves to raise the salience of some predicates the 
reader may not have even had in his/her conceptual 
representation prior to reading the poem. Love and 
wild rose-briars do make an incongruous twosome as 
do friendship and holly trees. The success of the met- 
aphor rests on this and on the size of the ground. The 
best metaphors are those presented by the best poets, 
those in which a vaguely understood experience is 
clarified through the predicates, salient or otherwise, 
of the B terms. Ortony's compactness thesis (1975), 
which allows metaphors to cause the transfer of fea- 
tures or characteristics as a "chunk" from vehicle to 
topic, does not account for this type of metaphoric 
discourse. 
11. The Semantic Net Approach 
From the above discussion it is obvious that some sort 
of conceptual representation underlies human meta- 
phor understanding; at some level people know the 
predicates of concepts and presumably something 
about their organization in terms of a generality hier- 
archy. Our "knowledge" undoubtedly includes, in 
addition to what might be labeled general knowledge, 
the values and beliefs of our speech community. Al- 
though the computational representation of such a 
base is a formidable task, the purpose of this paper is 
to delineate its nature and boundaries. It is assumed 
that the implementation of a base, in actuality, is a 
separable task. 
KL-ONE is one of a number of extant knowledge 
representation languages that allow Concepts to be 
arranged in a generality hierarchy with the characteris- 
tic that properties of the more general Concepts are 
inheritable by the more specific ones. It furthermore 
provides for the Concept to be represented as a struc- 
tured object, allowing one, in effect, to get inside the 
Concept and to see its relationship to other Concepts. 
In this discussion, I will assume that the features of 
KL-ONE are available for use in metaphor processing. 
The KL-ONE entities most relevant to this study are 
Concepts (diagrammatically represented by ellipses) 
and Roles (represented by encircled squares). 8 The 
structured Concept is the primary representational 
entity. A Role is internal to a Concept; it can be a 
part (for example, a hand is a Role of the Concept 
BODY) or what is commonly called an attribute (for 
example, a PERSON - the Concept - has habits - a 
Role). I will use Roles to represent predicates as they 
have been described above. The hierarchical classifi- 
cation aspect of KL-ONE allows lower, less general, 
Concepts to inherit structured description from those 
in an ancestral relationship to them. 
At the topmost level are the most general Con- 
cepts, called Generic Concepts (GCs). As one pro- 
ceeds downward through the network, one encounters 
more and more specific GCs. At the lowest level is 
knowledge about an individual, called an Individual 
Concept (IC) (see Figure 2). ANIMAL (a GC of a 
higher level) passes down to PERSON (a GC of a low- 
er level) all of its Roles and the interrelationships 
among them. So if ANIMALS have noses then so do 
PEOPLE. PERSON in turn passes to JOHN (an IC, 
shaded in Figure 2) those Roles PERSON got from 
ANIMAL as well as those unique to PERSON. 9 The 
Value and Satisfies links indicate the relationship be- 
tween a Role of an IC (called an IRole) and its parent 
role. 
Since the kind of knowledge necessary for meta- 
phor processing must include the beliefs and cultural 
values of the members of the community for which the 
system is being designed, in the representation of a 
prototype, some stereotyping is inevitable. That 
snakes are frightening and perhaps evil creatures is a 
commonly held opinion, although of course this is not 
true of all snakes. Metaphors seem to tap these kinds 
of generalizations in their insistence on prototypes, 
and they do seem necessary for understanding to take 
place. 
(24) The whip lay coiled on the ground like a 
snake. 10 
8 The local internal structure of every Conoept is made up of 
Roles and RoleSet Relations (RSRs). This discussion will not 
describe the functioning of RSRs but will focus on Roles since they 
are adequate to support this theory at its present stage of develop- 
ment. Furthermore, RSRs are less clearly understood at the current 
state of KL-ONE's design than are Roles. (I use the term "Roles" 
instead of "RoleSets" for discursive simplicity.) Those readers who 
have an interest in a more detailed description of the knowledge 
representation language KL-ONE should consult Brachman (1978, 
1979) and Schmolze and Brachman (1982). 
9 A Concept description in KL-ONE represents the intension of 
the Concept and there is a clear distinction drawn between the 
intension and its extension (the Concept in the real world). There 
is also a distinction drawn between definitions and assertions. My 
examples will deal only with definitions. I will not discuss issues 
related to the definability of Concepts here. For a treatment of 
this, see Cohen (1982b). 
~0 My thanks to David Weiner (personal communication) for 
this example. 
Computational Linguistics, Volume 10, Number 1, January-March 1984 7 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
ANIMAL NOSE NAME 
DOG PERSON NAME SPEECH 
O0 
~---~VALUE ENGLISH 
Figure 2. 
That the more affective (here, negative) aspects of 
peoples' impressions of snakes are of importance, and 
not simply shape or the more physical characteristics, 
seems evident by making a comparison 
(25) The whip lay coiled on the ground like a strand 
of spaghetti. 
The fact that a snake might strike and inflict harm 
from a coiled position (even though many types of 
snakes would not) shows how generalizations seem to 
operate. The menacing nature of the whip is high- 
lighted in (24), but in (25) that aspect of whips is 
much less important. In fact, it can be considered to 
be "negated" by (25). 
It should be noted at this point that one network 
alone may not suffice in processing metaphors or, for 
that matter, in processing other language phenomena. 
The existence of sublanguages is generally accepted 
within the field of linguistics: there are technical sub- 
languages for technical fields, for example. The repre- 
sentation must parallel and support this phenomenon. 
An individual's style of speech also changes according 
to the social setting; a person surely has more than 
one style. At the least, there is careful speech and 
casual speech. In lecturing to a class, one would use a 
more careful variety than in chatting with one's 
friends. I propose that there also exist sub-knowledge 
networks to support different styles of speech. Again, 
at the least there would be a careful and a casual vari- 
ety. Formal situations would favor the careful; infor- 
mal situations the casual. The other could then be a 
reasonable second choice. 
As an example, a veterinarian would have one rep- 
resentation of the animal kingdom for use on the job 
and one for home use. The careful (or more techni- 
cal) one might look like an expanded and detailed 
version of Figure 2. The sentence John is an animal 
might receive one interpretation (the literal one) at 
work where, due to the nature of the representation 
(and of course the context), the metaphorical interpre- 
tation is less likely. At home, the metaphorical one 
might prevail. Underlying it could be a representation 
like Figure 3. (Notice that in addition to the NAME of 
the Role (of no computational interest) there is the 
pointer labeled V/R This stands for Value/Restriction 
and provides information about the fillers of a Role. 
V/Rs must be other Concepts. NAME and V/R are 
two facets that Roles can be thought of as having.) 
The common person may well consider him or herself 
to be different from the animals (witness the 
creationist-evolutionist debates). Then John is an 
animal would no longer fall into a generality, specifici- 
ty situation and, although PEOPLE and ANIMALS are 
in a sibling relationship to one another, the relation- 
ship is between members of categories far more ab- 
stract than basic level categories. Thus, the incongrui- 
ty would make the metaphorical interpretation more 
likely. 11 
11 1 am grateful to Robert Dietz and Loretta Hirsch (personal 
communication) for the example that led me to this refinement. 
8 Computational Linguistics, Volume 10, Number 1, January-March 1984 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
WILD 
ANIMATE 
LIVING 
THINGS 
RATIONAL 
BEHAVIOR|N~ 
NAME ANIMALS 
PEOPLE ~ NAME 
BEHAVIOR 
JOHN 
Figure 3. 
With metaphors as with other forms of user input, 
contradictions with the knowledge base may occur. 
The idea of a user contradicting the knowledge of the 
system raises the issue of the relationship between 
these two. The core of knowledge originally stored in 
the base should be thought of as having been created 
by an expert to be used by a layman. Consequently, 
information supplied by the user should have a differ- 
ent status from that of the original designer of the 
system. However, for communication with the system 
to take place, discourse elements must at least tempo- 
rarily be integrated into the knowledge base. 12 
The kind of contradictions most common in meta- 
phors will be those involving a change in salience. For 
example, the system may know that John is an attrac- 
tive guy with sloppy eating habits. Both may be con- 
sidered of equal salience in this case. If the user ut- 
ters 
(27) John is a pig, 
then the salience of his eating habits for this user has 
been elevated beyond that of his otherwise pleasant 
demeanor. (This is referred to as predicate promotion 
in Ortony 1979b.) On the other hand, 
12 Others have provided for discourse phenomena in repre- 
senting knowledge. For example, see sections on semantic knowl- 
edge and discourse knowledge in Walker (1978) for a discussion of 
the discourse component of the SRI speech understanding system. 
The relationship of context to non-literal language is explored in 
Ortony, Schallert, Reynolds, and Antos (1978). 
(28) John is a doll 
would have the opposite effect. John's being an at- 
tractive guy is a more important characteristic of John 
for this user. In the somewhat unlikely event that 
both (27) and (28) were uttered by the same user, the 
salience of both would be elevated. 
How then can a representation system like KL-ONE 
be used? It should first be noted that in the interest 
of prototypicality considerations I will follow Cohen 
(1982b) in allowing V/Rs on Roles of a Concept to 
include an exclusive disjunction of possible values, 
weighted by typicality. The lower Concepts can then 
restrict these to the appropriate ones. These Concepts 
can in turn also be ordered by typicality. In Figure 4 
then, RED is a more typical color for an APPLE than 
GREEN and a DELICIOUS apple is a more typical ap- 
ple than a GRANNY SMITH. These rankings are indi- 
cated by the symbol ">". This representation of 
APPLE, having as it does only one Role, is of course 
highly simplified. Here, color is restricted to RED or 
GREEN, where RED and GREEN are other Concepts in 
the network. 
If the sentence 
(29) Jane's cheeks are like apples 
is to be understood, it is necessary to have a Concept 
of the prototypical apple. Among the members of the 
community that would understand this, surely the col- 
or of prototypical apples is red (see Figure 4 - RED > 
GREEN and DELICIOUS > GRANNY SMITH). 
Computational Linguistics, Volume 10, Number 1, January-March 1984 9 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
(30) Jane's cheeks are like Granny Smith apples 
would certainly have a different interpretation. 
In Figure 5, notice that for the concept HAND, 
temperature is not terribly salient. 13 I have also intro- 
duced a range of values as a possible value restriction, 
here 3-6. So this means that hands prototypicality 
range from hot to cold. In considering the Concept 
ICE (Figure 6), notice that it has a temperature of 7, 
that is, extremely cold. Temperature is highly salient 
for ICE(= 1). 
Figure 6 also illustrates the relationship between a 
higher Concept, SOLIDS, and a lower one, ICE. The 
Roles of SOLIDS are inherited by ICE. In some repre- 
sentations they are inherited intact, but here the Re- 
stricts link causes a restriction of the fillers of the Role 
in question. Thus, TEMPERATURE is restricted to 7; 
TEXTURE is restricted to HARD/SMOOTH. 
13 I have provided for salience to assume values between zero 
(least. salient) and one (most salient). The algorithm for the com- 
putation of these values awaits further empirical results. At pres- 
ent, they represent an estimate based on my intuitions. 
.So 
REPRESENTATION OF HAND 
(1) 
\[ T,~,,~,~TIIP~ ~ (2) 
............... i (3) \ / (4) 
//, ,, .~ -Vtc. (7) 
C_D 
TEMPERATURE: 
BEYOND LINGUISTIC DESCRIPTION (HOT) 
EXTREMELY HOT 
HOT 
LUKEWARM 
COOL 
COLD 
EXTREMELY COLD 
BEYOND LINGUISTIC DESCRIPTION (COLD) 
REPRESENTATION OF TEMPERATURE 
Figure 5. 
10 Computational Linguistics, Volume 10, Number 1, January-March 1984 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
RED 
APPLE 
GREEN 
GRANNY 
SMITH DELICIOUS 
V/R 
Figure 4. 
INORGANIC 
SUBSTANCES 
TEMPERATURE TEXTURE 
TEMPERATURE 
SOLIDS NAME TEXTURE 
7 ~1 Ig----4 ICE HARD/SMOOTH 
1.0 
Figure 6. .85 
Computational Linguistics, Volume 10, Number 1, January-March 1984 11 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
What then would be the result of the user input 
(9) John's hands are like ice. 
Looking at a portion of John's network integrated with 
the one for HAND (Figure 7), sentence (9) would 
cause one to look at the Roles for John's hands, see 
that they do not Restrict the Roles of the Generic 
Concept HAND. Therefore, they are thought by the 
system to be prototypical hands. Temperature is the 
most salient feature of ICE for which John has a Role. 
This must be the Role indicated by (9). The user has 
made a hyperbolic statement. He or she has said that 
John's hand temperature is inexpressible by reference 
to normal hand temperatures. Thus, to understand the 
sentence, it is necessary to observe that ICE has a 
much more extreme value for temperature and it is of 
the highest salience (= 1). As a result, for the purpos- 
es of this discourse, the salience of John's hand temp- 
erature is given the value of 1, implying that for the 
speaker John's hands are one of his most salient fea- 
tures. This information may be useful in interpreting 
the discourse that follows the sentence. 
The incongruity that must be present for metaphors 
to work can also be seen by referring to these dia- 
grams, particularly with regard to their hierarchical 
nature. It is clear that a metaphorical statement is 
possible between members of an inheritance relation- 
ship 
(31) John isaperson 
(32) Ice is a solid 
since these are actual statements of that relationship. 
A deeper analysis of human classification devices 
promises to yield further constraints on pairs that can 
relate metaphorically to one another. For example, 
looking back at 
(12) John is like his father, 
since both John and his father share the same immedi- 
ate superordinate category, male, (12) cannot be 
metaphoric. This may also explain why 
(33) Encyclopedias are like gold mines 
is metaphorical but 
(34) Encyclopedias are like dictionaries 
seems to be a similarity statement. (Examples from 
Ortony, Reynolds, and Arter 1978.) 
The algorithm implied here can be expressed as 
follows: 
1. If the topic is an individual constant (IC), establish 
restrictions (using the Restricts link), if any, on the 
Role in question (for example, JOHN's HANDS). If 
% 
.50 
Figure 7. 
NAME 
12 Computational Linguistics, Volume 10, Number 1, January-March 1984 
E. Judith Weiner A Knowledge Representation Approach to Understanding Metaphors 
there are restrictions, note these; otherwide, note 
inherited V/Rs. 
2. Establish those salient predicates for the vehicle for 
which the topic also has a Role (for example, 
TEMPERATURE for ICE and JOHN's HANDS). 
3. If the V/Rs for these Roles are extreme in the vehi- 
cle but not in the topic, the utterance is hyperbolic. 
If, in addition, the vehicle and topic are in the 
proper relationship to one another with respect to 
the taxonomy, the utterance is metaphorical. Giv- 
en that these conditions hold, raise the salience of 
the relevant Roles of the topic. 
We have begun the computer implementation of 
this algorithm using NIL on a VAX 11/780. We have 
implemented enough of the features of KL-ONE to 
allow us to build a prototype knowledge base (in prog- 
ress) that will be rich enough to permit experimenta- 
tion using input consisting of novel metaphors. We 
intend to exercise the system with the goal of estab- 
lishing the correctness or need for refinement of the 
algorithm. 
12. Conclusions 
To summarize, I have demonstrated how metaphor 
comprehension can proceed on the foundation of sali- 
ence, with the following modifications to the theories 
of Ortony: High salient predicates of the A term can 
be those at issue in a metaphorical (as opposed to 
strictly literal) interpretation and low salient features 
of the B term likewise are of concern. Because of 
other factors that serve to motivate the use of meta- 
phors (incongruity, hyperbole, inexpressibility), meta- 
phors are not always compact, nor are they prohibited 
from being used for a single predicate. 
Prototype theory applies in two ways: the B term is 
generally chosen as prototypical of certain predicates; 
the real-world representative of the B term is a proto- 
typical member of its class. 
In my utilization of KL-ONE, I have allowed for a 
range of possible values in the value restriction facet 
and introduced salience as a role facet. I have also 
demonstrated the need for sub-knowledge networks in 
dealing with metaphors and other natural language 
issues. 
13. The Large Scope of Things 
Although this approach to natural language and 
knowledge representation has been from the point of 
view of metaphors, it seems clear that at least some of 
the factors operating in metaphor understanding oper- 
ate in literal language as well. There is undoubtedly a 
relationship between salience raising in metaphors and 
the resultant effect on discourse and focus of atten- 
tion. In fact, salience raising undoubtedly contributes 
to focusing. I have only dealt lightly with discourse 
problems and recognize these as crucial to all language 
understanding, whether the language be literal or fig- 
urative. 
It should be noted that the approach I have taken 
ignores the possibility of considering metaphors as 
analogies. Since many are not analogies, I will save 
that for future work. That they exist is clear: 
(35) Giraffes are like skyscrapers 
is somewhat more complicated to understand 
(computationally speaking) than many of the ones I 
have used in this discussion because it involves rela- 
tionships among Roles and not Roles simply. 
(Giraffes are the tallest animals, skyscrapers are the 
tallest buildings.) Also, clearly, some analogies and 
some metaphors are instructional (Ortony 1975). 
(37) The structure of an atom is like the structure of 
the solar system. 
These appear to be used in building the representation 
of a new concept (here, atoms). I have dealt here 
only with representations of existing concepts. 14 
In addition, I have chosen to develop a method that 
could be applied to handling novel metaphors as op- 
posed to those recognized by Lakoff and Johnson 
(1980) as general metaphors. A system could be 
made more efficient by utilizing a technique such as 
Carbonell (1981) has described for recognizing the 
latter whenever they occur and incorporating my pro- 
posals elsewhere. 
Finally, as in the case of literal language, it is es- 
sential to study recordings of natural speech to see 
what people actually say. In metaphors as elsewhere, 
there will be many surprises. 
Acknowledgments 
I'd like to thank Ralph Weischedel, Michael Freeman, 
and Genevieve Berry-Rogghe for reading and com- 
menting on an earlier draft of this paper. 

References 
Anderson, R.C. and Ortony, A. 1975 On Putting Applies into 
Bottles - A Problem of Polysemy. Cognitive Psychology 7: 
167-180. 
Brachman, R.J. 1977 A Structural Paradigm for Representing 
Knowledge. Ph.D. thesis. Harvard University, Cambridge, 
Massachusetts. 
Brachman, R.J. 1978 Theoretical Studies in Natural Language 
Understanding. Annual Report, 1 May 1977 to 30 April 1978; 
BBN Report No. 3888 (September). Bolt Beranek and New- 
man, Inc., Cambridge, Massachusetts. 
Brachman, R.J. 1979 On the Epistemological Status of Semantic 
Networks. In: Findler, N.V., Ed., Associative Networks: Repre- 
sentation and Use of Knowledge by Computers. Academic Press, 
New York, New York: 3-50. 
Carbonell, J.G. 1980 Metaphor - a Key to Extensible Semantic 
Analysis. In Proceedings of the 18th Annual Meeting for the 
Association for Computational Linguistics and Parasession on 
14 Ortony (1979b, p. 199) recognizes the distinction in terms 
of "predicate promotion" versus "predicate introduction". 
Topics in Interactive Discourse, June 19-22, 1980. University of 
Pennsylvania, Philadelphia, Pennsylvania: 17-21. 
Carbonell, J.G. 1982 Metaphor: An Inescapable Phenomenon in 
Natural Language Comprehension. In: Lehnert, W. and Ringle, 
U., Eds., Strategies for Natural Language Processing. Lawrence 
Erlbaum Associates, Hillsdale, New Jersey: 415-434. 
Cohen, B. 1982a A Knowledge Representation Model of Proto- 
type Theory. In: Schmolze, J.G. and Brachman, R.J., Eds., 
Proceedings of the 1981 KL-ONE Workshop. BBN Report No. 
4842 (June). Bolt Beranek and Newman, Inc., Cambridge, 
Massachusetts: 124-127. 
Cohen, B. 1982b Understanding Natural Kinds. Ph.D. Thesis. 
Stanford University, Stanford, California. 
Conklin, E.J. and McDonald, D.D. 1982 Salience: The Key to the 
Selection Problem in Natural Language Generation. In Proceed- 
ings of the 20th Annual Meeting of the Association for Computa- 
tional Linguistics, 16-18 June 1982. University of Toronto, 
Toronto, Canada: 129-135. 
Gentner, D. 1980 Studies of Metaphor and Complex Analogies. 
Paper presented at the APA Symposium on Metaphor as Proc- 
ess. 
Gentner, D. 1981 Are Scientific Analogies Metaphors? In: Miall, 
D.S., Ed., Metaphor: Problems and Perspectives. Harvester Press, 
Ltd., Brighton, Sussex, England. 
Grosz, B.J. 1977 The Representation and Use of Focus in Dialog 
Understanding. Ph.D. dissertation. University of California at 
Berkeley, Berkeley, California. 
Grosz, B.J. 1981 Focusing and Description in Natural Language 
Dialogues. In: Joshi, A.; Webber, B.; and Sag, I., Eds., Ele- 
ments of Discourse Understanding. Cambridge University Press, 
New York, New York. 
Haynes, F. 1975 Metaphor as Interactive. Educational Theory 
25(3): 272-277. 
Labov, W. 1973 The Boundaries of Words and their Meanings. 
In: Bailey, C.J.N. and Shuy, R.W., Eds., New Ways of Analyzing 
Variation in English. Georgetown University Press, Washington, 
D.C.: 340-373. 
Lakoff, G. and Johnson, M. 1980 Metaphors We Live By. Universi- 
ty of Chicago Press, Chicago, Illinois. 
Ortony, A. 1975 Why Metaphors Are Necessary and Not Just 
Nice. Educational Theory 25(1): 45-53. 
Ortony, A. 1979a Beyond Literal Similarity. Psychological Review 
86(3): 161-179. 
Ortony, A. 1979b The Role of Similarity in Similes and Meta- 
phors. In: Ortony, A., Ed., Metaphor and Thought. Cambridge 
University Press, New York, New York. 
Ortony, A.; Reynolds, R.E.; and Arter, J.A. 1978 Metaphor: 
Theoretical and Empirical Research. Psychological Bulletin 
(September) :919-943. 
Ortony, A.; Schallert, D.L.; Reynolds, R.E.; and Antos, S.J. 1978 
Interpreting Metaphors and Idioms: Some Effects of Context on 
Comprehension. Journal of Verbal Learning and Verbal Behavior 
17: 465-477. 
Rosch, E.H. 1973 On the Internal Structure of Perceptual and 
Semantic Categories. In: Moore, T.E., Ed., Cognitive Develop- 
ment and the Acquisition of Language. Academic Press, New 
York, New York: 111-144. 
Rosch, E. and Mervis, C.B. 1975 Family Resemblances: Studies in 
the Internal Structure of Categories. Cognitive Psychology 7: 
573-605. 
Rumelhart, D.E. and Ortony, A. 1977 The Representation of 
Knowledge in Memory. In: Anderson, R.C.; Spiro, R.J.; and 
Montague, W.E., Eds., Schooling and the Acquisition of 
Knowledge. Lawrence Erlbaum Associates, Hillsdale, New 
Jersey: 99-135. 
Schomlze, J.G. and Brachman, R.J. 1982 Summary of the KL- 
ONE Language. In: Schmolze, J.G. and Brachman, R.J., Eds., 
Proceedings of the 1981 KL-ONE Workshop. BBN Report No. 
4842 (June). Bolt Beranek and Newman, Inc., Cambridge, 
Massachusetts: 233-260. 
Tversky, A. 1977 Features of Similarity. Psychological Review 
84(4) 327-352. 
Verbrugge, R.R. 1980 Transformations in Knowing: A Realist 
View of Metaphor. In: Honeck, R.P. and Hoffman, R.R., Eds., 
Cognition and Figurative Language. Lawrence Erlbaum Associ- 
ates, Hillsdale, New Jersey: 87-125. 
Walker, D.E., Ed. 1978 Understanding Spoken Language. Elsevier 
North-Holland, Inc., New York, New York. 
Weiner, E.J. 1982 Using KL-ONE to Understand Metaphors. In: 
Schmolze, J.G. and Brachman, R.J., Eds., Proceedings of the 
1981 KL-ONE Workshop. BBN Report No. 4842 (June). Bolt 
Beranek and Newman, Inc., Cambridge, Massachusetts: 194- 
197. 
Weiner, E.J. 1983 The Role of the Knowledge Network in the 
Computational Processing of Non-literal Language. In Proceed- 
ings of the International Conference on Data Bases in the Humani- 
ties and Social Sciences. (in press) 
Williams, O., Ed. 1963 The Mentor Book of Major British Poets. 
The New American Library, Inc., New York, New York. 
