Coulson, S. & Fauconnier, G. 1999. Fake Guns and Stone Lions: Conceptual Blending and Privative Adjectives. In B. Fox, D. Jurafsky, & L. Michaelis (Eds.) Cognition and Function in Language. Palo Alto, CA: CSLI.
Fake Guns and Stone Lions: Conceptual Blending and Privative Adjectives
University of Arizona
Thus mental spaces are used to represent each of the
domains, while linked elments represent the identity of the ball in Trashcan
Basketball with the crumpled paper in the Trash domain. Moreover, the idea
of trashcan basketball doesn't come out of nowhere, but by analogy to the
real game of basketball played with a leather ball and a ten-foot basket.
Mental space connections can be based on identity, similarity, or analogy.
Moreover, when a mental space connection exists between two elements, the
Access Principle allows speakers to refer to an element in one mental space
with a word more appropriate for a linked element in a different space
Because a proper understanding of the game of trashcan
basketball involves the simultaneous apprehension of the relationship between
knowledge recruited from established domains as well as aspects of the
novel concept, it is represented in a conceptual integration network
(Fauconnier & Turner, in press). A conceptual integration network is
a network of mental spaces structured with frames the speaker constructs
from contextual information and background knowledge.The prototypical integration
network is comprised of four mental spaces, one for each of the input domains,
one for the blended domain, and a generic space which represents abstract
properties that apply to structure in all of the spaces.
For example, Figure 1 represents the conceptual integration network for the concept of a trashcan basketball. In this network, there is an input space structured by a model built from knowledge about basketball, an input space with cognitive models of trashcans, and a blended space with a cognitive model that combines information from the two inputs. Although the generic space in this example is not elaborated in Figure 1, it is included in the network because abstract, schematic information derived from the blending process can be appealed to in extensions of the concept. In this example, the generic space involves throwing an object into a container and suggests the initial cross-space mappings between the Trash and Basketball domains.
Figure 1.Conceptual Integration Network for Trashcan Basketball
Importantly, trashcan basketball does not result from
an isolated mapping between a wad of paper and a basketball. Rather, it
arises from a whole system of correspondences which speakers can establish
between the various domains. Further, the emergent features of the combined
concept trashcan basketball can be seen as arising in part from abstract
processes of conceptual projection from one domain to another -- understanding
a wad of paper as a basketball. However, it also arises out of concrete
interaction with the trashcan basketball in the course of the game.Because
the physics of interacting with a piece of paper are different from those
of interacting with a leather ball, participants in the game will naturally
discover the differences in the objective features of shooting a regular
versus a trashcan basketball.
Meaning construction in conceptual blending is thus
both flexible and knowledge-rich. Frames recruited for integration in a
network can vary from context to context, as can the nature of the integration.
The content of a particular conceptual blend depends on the amount and
type of knowledge available about the input domains. In the absence of
particular knowledge about a domain speakers rely on default models of
prototypical instances. However, speakers can also construct models which
are highly idiosyncratic in response to situational demands.
Completion is pattern completion which occurs
when structure projected from the inputs matches information in long-term
memory. For example, in trashcan basketball, if one student shoots and
the other attempts to defend the goal, pattern completion could result
in evoking the frame for a one-on-one basketball game. Completion is closely
related to elaboration, a process which involves performance and/or
mental simulation of the event in the blend. For example, one might employ
elaboration in order to understand the concept of moon rock basketball,
basketball played on the moon with moon rocks. The activation of novel
structure can either be done by computation, or, as in the trashcan basketball
examples, may rely extensively on interaction with the environment as construed
with existent blended models.
In this section we have gone over some of the introductory
concepts in conceptual blending theory. We suggest that the processes of
conceptual blending are rooted in speakers' imaginative capacities. In
particular, blending processes rely heavily on cross-space mapping abilities
which enable speakers to forge links between elements whose objective properties
can differ quite substantially. We return to conceptual blending in section
4 where we discuss how conceptual integration networks can be used to represent
the meaning evoked by privative constructions. However, in the next section
we turn to the sense generation model, a leading model of concept combination
in cognitive science.
One feature of Franks' account is his acknowledgment
of the importance of perspectival relativity in classification judgments
and how it serves to limit the generalizability of those statements in
particular ways. As in mental space theory, perspectival relativity implies
that we can adopt various perspectives on the same discourse referent.
This allows us to understand a particular apple as a dessert, a cricket
ball, or a still life, depending on what perspective we have adopted. In
the sense generation model, perspectival relativity is handled by pegs,
or discourse referents, and alecs, perspective-relative instantiations
Moreover, as in conceptual blending, meanings are
built up in context based on information available from long-term memory.
In-context meanings, or senses, consist of attribute-value structures
(AVS). Senses of privative constructions are built up in a similar way
as senses for other sorts of compounds: by recruiting attribute-value structures
for the modifier and the head, and performing operations on the representations
in order to yield the resultant compound sense. The core of the sense generation
model is its three processes for sense generation: unification, priority
union, and metonymic type coercion.
Unification is a monotonic operation which
simply summates the two sets of features. Consequently, unification can
only occur when there is no conflict between the values of shared attributes
of the component concepts. The second process,
priority union, is
used in cases where the two concepts have conflicting values for a given
attribute. When combining concepts A and B to form AB, the combined concept
inherits all nonshared attributes and non-conflicting attribute-value pairs
directly. In cases where A and B have a conflicting value for a given attribute,
the combined concept inherits the value of the modifier. Thus priority
union is unification with overrides.
More relevant to privative constructions, metonymic
type coercion (MTC) is a process by which the modifier can alter the
sense of the head noun. There are two types of metonymic type coercion,
with rebuttal and MTC with undercutting. In both sorts of MTC,
the attribute-value structure of the head is divided into central and diagnostic
attributes. Privatives work as operators on the attribute-value structures
of the head. Negative privatives such as 'fake' in 'fake gun,' or 'false'
in 'false eyelashes,' trigger MTC with rebuttal. In this process, the central
features of the head are negated while its diagnostic features are combined
with those of the modifier via priority union. Equivocating privatives
such as 'alleged' in 'alleged bomber' are dealt with via MTC with undercutting.
In this process, the diagnostic attributes are inherited while the central
attributes of the head are left unspecified.
An additional feature of the model, known as implicit
attachment, can modify explicitly evoked senses by unification with
contextually evoked concepts. If there is an informational demand for further
specification, the attribute-value structure of an implicitly attached
noun can be unified with explicit senses in the compound. For example,
if you are in the park and someone says, 'stone lion,' the AVS's for 'stone'
and 'lion' could be unified with the AVS for an implicitly evoked noun
such as statue.
The sense generation model also provides two mechanisms
for context-sensitivity. The first is the context-sensitivity of the initial
activation of the AVS for 'gun'. Contextual and communicative constraints
can constrain the initial activation of values in the AVS. Moreover,
if there is a need for further specification in order to individuate the
referent, the sense generated for 'fake gun' can be unified with the AVS
for a situationally appropriate noun via implicit attachment. This can
result in different AVS matrices for a 'fake gun' which is a toy,
and a 'fake gun' which is a replica. Overall, the sense generation
model provides a rigorous way of accounting for similarities and differences
between real and fake guns.
On the other hand, imagine that someone buys the defunct
gun at a garage sale with the express intention of using it in a robbery.
Perhaps he wants people to think it's a gun, but doesn't want to actually
hurt anyone. In this scenario, our rusty gun probably would be deemed a
'fake gun,' albeit an atypical one. The importance of the faker's plans
and intentions in this example is not readily handled by the sense generation
model. This is because it does not incorporate a crucial component of being
fake: the faker's intention to create a discrepancy between his own beliefs
about the fake object and those of his victim or victims. Moreover, a successful
fake implies that the audience reacts to a fake gun in the same way they
might react to a real gun.
We suggest that the real/fake distinction lies beyond
the attributes of the gun itself. Rather, the character of a fake gun will
depend on the faker's motivation, the scenario in which the gun is to function,
and the knowledge of the prospective victim. For example, in a prototypical
scenario where a burglar holds up a store with a fake gun, the important
thing is for the object to look like a gun. Other diagnostic features such
as weight are far less important. However, if the victim is a gun collector,
and the con-artist wants to sell him a fake 19th-century gun,
a successful fake will likely have most of the diagnostic features, and
most of the central ones as well.
In fact, the best fake 19th-century gun
is probably a real gun which was manufactured just after the turn of the
20th century. At this point an advocate for classical (and 'quasi-classical')
approaches to concept combination might object that the modification of
'gun' by '19th -century' will require adding Manufactured-in-the-19th-
century(+) to the set of central attributes. However, while this objection
explains why our fake gun was not manufactured in the 19th century,
it leaves the issue of why it can fire bullets unanswered. Further,
we can just as well imagine a parallel scenario in which the fake gun is
lacking in the diagnostic features of a gun.
For example, imagine that we are attempting to sell
a fake Colt 45 to a collector. Although the object has been mangled beyond
recognition, we have cleverly inscribed the trademark onto what might have
once been the gun's handle. Knowing that collectors consider this inscription
to be the hallmark of authenticity, we tell our victim that the gun was
damaged when it was confiscated by the FBI and show him the trademark.
This mangled mass of metal lacks any gun features except for one (presumably
central) one. Thus it would seem possible to have fake guns which possess
central attributes of guns,
as well as ones which lack diagnostic attributes.
We suggest speakers treat 'fake' as a space builder
that prompts a mapping between an actual scenario in which the actor employs
the fake gun, and a counterfactual scenario in which his audience reacts
as if it were a real gun. For example, in Figure 2 we sketch a blend in
which the actor uses a plastic gun to rob his victim. The object in the
blended space inherits the property of being plastic from the space that
represents the actor's knowledge. Moreover, it inherits the property of
being a gun from the victim's belief space. The establishment of this sort
of a mapping goes beyond the mere presence or absence of central versus
diagnostic features, and relies on the speaker's ability to coordinate
frames in actual and counterfactual spaces via cross-space mappings.
In the default scenario that Franks evokes to demonstrate
the sense generation model, the mappings between the fake gun model and
its counterpart in the gun domain are similarity mappings between features
in the object frames. However, being fake need not entail similarity between
the fake object and its counterfactual counterpart. The important thing
is that its properties might induce the victims to believe that the counterfactual
scenario obtains. In the right context, a single feature might produce
the desired belief. For example, a pipe in the back of the neck, a hairbrush
in the raincoat, or a balloon popping in a room full of hostages with their
heads to the floor might all serve as fake guns. In 1997, in a testament
to the adage that truth is stranger than fiction, Carlos Diaz was sentenced
to 18 years in prison for stealing $20 and a watch while armed with a zuchinni
concealed in his jacket.
Determination of what features are relevant requires
the construction of a model of the actor's actions and the victim's perception
of those actions as illustrated in Figure 3. Mappings are constrained by
the understanding that in the faker's mind there is a causal connection
between his own actions and the victim's beliefs. Of course, the victim
actually be deceived for the fake object to be a fake.However,
the characteristics of fake objects arise because of the way in which the
intent to deceive is central to the concept of fake. Conceptual combination
in each case is driven by the way in which theory of mind determines cross-space
mappings between the actor's intention and the victim's would-be belief.
Diagnostic features of guns, both fake and real, result from apprehension
of the more extensive set of mappings between the relevant spaces.
In conceptual blending, the ascription of features
is only a side effect of coordinating the representational structure in
the various spaces. Consequently, projection, representation, and rerepresentation
can all make objective features of the gun largely irrelevant. For example,
suppose we're comparing two gun diagrams in order to determine which one
is the fake. While neither of the drawings has the central attributes of
a gun, we might single out one of them because (in the drawing) the barrel
is solid rather than hollow. Although diagrams bear little objective resemblance
to guns, conventions for diagrammatic representation allow us to map between
relevant aspects of the diagram and our knowledge of guns. Further, socially
defined activities such as play present similar problems for a feature-based
For example, if I'm playing cops and robbers, I might stick my thumb in the air and point my index finger at my playmate saying, 'Give me the money, or I'll shoot!' After he complies I might laugh saying, 'Hah! It's only a fake!'In this case, whether I had a real gun or not, bullets wouldn't come out of my finger. Moreover, in such a context, the difference between a real gun and a fake gun might well involve whether my playmate pretends to die after being shot with the gun. Because conceptual blending theory relies on the establishment of mappings based on similarity, identity, and analogy, it predicts the use of these relational counterparts as well as the similarity-based ones.We elaborate on the importance of relational counterparts in the next section.
Figure 3.Actor's Intention in 'Fake Gun'
The NI lexical concept is then the concept
associated with the head noun of a phrase that individuates the referent.
For example, if a discourse or text states that, A girl was sketching
a stone lion in the park, the instantiation of stone lion would
likely be as a statue of a lion, and not an ornament in the shape of a
lion. The NI lexical concept selected would then be statue, and not ornament.
This is one constraint on NI selection: In addition to being situationally
appropriate, an AVS description of the instantiation must be subsumed by
the sense AVS of the phrase that it instantiates; in this case, the AVS
for a statue of a lion would be subsumed by one for stone lion.
Although the sense generation
model provides a clever mechanism to provide the meanings for functional
privatives in which the head is physically constructed from the modifier
(as in 'stone lion'), it is compromised by the requirement that the implicitly
attached lexical concept be subsumed by the phrase it instantiates. In
the sentence, 'A girl was sketching a stone lion in the park,' Franks discusses
the reading where 'stone lion' refers to the statue in the park. However,
in this context 'stone lion' might just as well be interpreted as referring
to the girl's drawing. A sketched stone lion is not a lion any more than
a stone lion is a lion (from the reality perspective), but unless the lexical
concept for 'sketch' is subsumed by that for 'stone lion' it cannot serve
as an implicitly attached concept.
For the sense generation
model to work, then, the lexical concepts for stone and lion must be such
that they can subsume senses of 'statue,' 'ornament,' and, apparently,
'sketch.' However, this suggests the lexical concept for 'stone lion' is
a veritable Pandora's Box, containing all the AVS's which 'stone lion'
might might possibly assume in context. As noted above, 'stone lion' could
refer to a pictorial representation of a stone lion such as a sketch, a
painting, or a photograph. In a dramatic production (perhaps a sequel to
'stone lion' might refer to a plastic prop which represents a statue in
the play. Or, if the statue is played by a human actor, 'stone lion' would
refer to a person representing herself as a statue representing a lion.
In fact, the flexibility of speakers is such that unless the lexical concept
of for each word contains an infinite set of AVS's, the sense generation
model will be unable to account for all possible instances of word usage.
Further, a general blending mechanism provides a better
account of the way in which features accommodate (in Langacker's 1987 sense)
to the different domains in which they occur. For example, the sense in
which a trashcan basketball is 'round' differs from the sense in which
a leather basketball is 'round.' In conceptual blending, to say that both
balls are 'round' means that the element set up to represent each ball
can be integrated with a frame for round. The jagged character
of the roundness of a trashcan basketball is an emergent property
which can be perceived by participants, and conceived by
people who can imagine a crumpled up piece of paper. In blending theory,
this sort of composition need not entail the ascription of objective properties.
Rather, it relies on the speaker's ability to exploit relational aspects
of cognitive models to establish mappings.
Moreover, 'stone lion' can also assume any number
of metaphorical meanings that the sense generation model is entirely unable
to handle. A 'stone lion' might be a very stoic lion, or a lion which stands
very still. Similarly, if the family cat assumed a frozen pose in the course
of hunting a robin, it might be referred to as a 'stone lion.' Finally,
'stone lion' could be used to refer to objects with almost no connection
to real lions, such as a deposed dictator in exile, or an idea which seemed
forbidding at one time, but was never instantiated.
While the connection between 'stone lion' and these
latter two concepts is not straightforward, neither is it random. Because
mappings in mental space theory can be based on identity, similarity, and
analogy, it affords a unified treatment of stone lions which are representations
of lions, representations of stone lions, metaphorical stone lions, and
even representations of metaphorical stone lions, as in a painting of the
family cat stalking his prey.
Moreover, while most of the above examples are privative,
the conflicting features of 'stone' and 'lion' do not seem to preclude
non-privative treatment of 'stone lion.' For example, in an exhibit at
the zoo, we might use 'stone lion' to refer to the lion asleep on a stone,
the lion standing by the big stone, the lion who was once asleep on the
stone, a lion with a stone around his neck, a lion playing with a stone,
and so on. These are examples of a phenomenon Clark (1983) has referred
to as the contextuality of intrpretation.
In conceptual blending theory, the contextuality of
language use is attributed to creative mechanisms that can build new models
in response to background and local information. Because the overt language
of a nominal compound such as 'stone lion' provides minimal clues to how
the integration of input frames is to proceed, the language user is forced
to rely on contextual information and background knowledge. An important
aspect of conceptual blending theory, absent from the attribute-value listing
employed in the sense generation model, is the use of structured representations
which facilitate cross-domain mappings. Because the representations evoked
in mental spaces are hierarchically structured such that causal and relational
information is readily available, they enable the establishment of mappings
that help constain the projection of structure to the blend (see also Wisniewski
& Gentner, 1991).
For humans, representation is not just an internal
mental activity, but a physical and a social one. Sometimes part of being
a gun is the builder's intention to make a gun. Similarly, sometimes being
a gun is the result of stipulating that an object is a gun. Because human
activity involves the projection of partial structure from domain to domain,
there is far more counterpart structure in the world than we typically
realize. Like our two students playing trashcan basketball, this counterpart
structure is not just recognized but actively imposed.
Further, tracking this counterpart structure does
not involve the apprehension of objective features, but the integration
of their perception with contextually relevant frames. Consequently, an
adequate account of meaning construction needs a mechanism (such as mental
spaces) which can represent the complex embeddings due to drama, pictorial
representation, pretense, beliefs, and so on. Moreover, it points to a
non-traditional locus of conceptual productivity. Rather than the algorithmic
combination of discrete concepts (monotonic and nonmonotonic combinations
alike), our observations point to the importance of the human ability to
accommodate frames at various levels of abstraction to suit varying contextual