Content uploaded by Paul Thibodeau
All content in this area was uploaded by Paul Thibodeau on Aug 17, 2015
Content may be subject to copyright.
When a bad metaphor may not be a victimless crime:
The role of metaphor in social policy
Paul Thibodeau (firstname.lastname@example.org)
James L. McClelland (email@example.com)
Lera Boroditsky (firstname.lastname@example.org)
Stanford University, Department of Psychology
Jordan Hall (Bldg 420), 450 Serra Mall, Stanford, CA 94305 USA
Metaphors are pervasive in our discussions of abstract and
complex ideas (Lakoff & Johnson, 1980), and have been
shown to be instrumental in problem solving and building
new conceptual structure (e.g., Gentner & Gentner, 1983;
Nersessian, 1992; Boroditsky, 2000). In this paper we look at
the role of metaphor in framing social issues. Our language
for discussing war, crime, politics, healthcare, and the
economy is suffused with metaphor (Schön, 1993; Lakoff,
2002). Does the way we reason about such important issues
as crime, war or the economy depend on the metaphors we
use to talk about these topics? Might changing metaphors lead
us to different conceptions and in turn different social
policies? In this paper we focused on the domain of crime
and asked whether two different metaphorical systems we
have for talking about crime can lead people to different ways
of approaching and reasoning about it. We find that framing
the issue of crime metaphorically as a predator yielded
systematically different suggestions for solving the crime
problem than when crime was described as a virus. We then
present a connectionist model that explores the mechanistic
underpinnings of the role of metaphor.
Keywords: Metaphor, analogy, connectionism, social policy
We use a variety of metaphors when discussing crime. In
some cases, crime and criminals are described as predators,
preying on the public, lurking in the shadows, stalking their
victims. As William James put it, “Man… is simply the
most formidable of all beasts of prey, and, indeed, the only
one that preys systematically on its own species” (p. 846,
1904). The police in this set of metaphors are meant to hunt
down, lay traps and attempt to catch or capture the
criminals, so as to lock them away.
In other cases, crime is described as a disease or
epidemic. It infects cities and plagues neighborhoods. On
this framing, the job of police is centered on diagnosing and
treating the root cause of the problem, stopping the spread
of the infection, identifying risk factors to prevent future
outbreaks, and restoring the health of the community. Public
health researchers have explicitly proposed that treating
crime as a disease will help us find the cure (Guerrero &
Concha-Eastman, 2001). A violence prevention program
operated by an epidemiologist in Chicago takes this
metaphor to heart, treating crime according to the same
regimen used for contagious diseases like AIDS and
tuberculoses (Kotlowitz, 2008).
In some cases, scholars have even cast bad metaphors as a
societal danger. George Kelling, a criminal justice scholar,
describes a case in which a serial rapist attacked 11 girls
over a 15-month period before being captured by the police
(Kelling, 1991). The police later revealed that they had
deliberately withheld information from the public that could
have prevented at least 8 of the attacks, because it might
have compromised the traps they had laid for the suspect.
The girls, Kelling argues “were victims… not only of a
rapist, but of a metaphor” (p. 1, 1991). The police in this
view were too entrenched in the role of hunting down and
catching the criminal, and neglected their responsibility to
inoculate the community against further harm.
In this paper we empirically test whether metaphors
indeed structure how people reason about social issues like
crime. We also present a computational model that explores
the mechanisms through which metaphors may shape
The Current Study
We focus on two common frameworks for talking about
crime: “crime as a predator” and “crime as a virus.” Both
are used frequently and productively in discourse about
crime. However, if we take these metaphors seriously, they
offer very different implications for how societies should
deal with crime. For example, to deal with a dangerous
predator on the loose, one might try to hunt, trap or cage the
animal. If crime is like a predator, then the best way to deal
with crime is to catch and imprison as many criminals as
possible. Solutions to the crime problem might include
increasing the police force, harsher enforcement of laws,
longer prison terms, and so on. If crime is a disease, the set
of implications is rather different. To treat a disease, one
might attempt to diagnose and treat the root cause of the
problem, and one would also aim to restore the organism’s
immune system so that it is not susceptible to future
infections1. If crime is a disease, then really dealing with
crime involves not only treating the symptoms, but getting
1 There are two somewhat different metaphorical frameworks that
treat crime as an illness. In one, the community or population is
seen as an organism, and crime is a disease that is developing
inside that organism (e.g., “"Violent crime is a cancer that eats
away at the very heart of society."). In another, the community is
seen as individual agents and crime is a contagious disease that can
be passed on from one person to another forming an epidemic. In
this paper the stimuli did not strongly distinguish between these
two metaphors, but doing so would be an interesting extension of
this work, as the two metaphors suggest somewhat different
implications for treating crime.
to the root cause of the problem, and restoring the health of
the community so that it is no longer susceptible to future
While these analyses of the metaphors seem plausible,
what we don’t know is whether such metaphors in fact have
any psychological weight. Does casually encountering one
metaphor or another in discourse about crime actually lead
regular English speakers to come to different conceptions of
the crime problem? Would people unwittingly come up
with different types of solutions for the crime problem when
exposed to one metaphor versus another?
The experiment was designed to explore whether simply
embedding a common metaphor in an otherwise neutral
report about crime can systematically influence people’s
approach to solving the crime problem. In the task,
participants read a report about crime in a fictional city and
then answered questions about the city. The report
contained mostly crime-relevant statistics, and also two
brief instances of either the crime as predator metaphor or
the crime as virus metaphor. After reading the report,
participants answered questions relating to crime in the city.
Critically, in one of these questions, participants were asked
to propose a solution to the crime problem.
If metaphors in fact have psychological weight, then
being exposed to different metaphors for crime may lead
people to propose different solutions to the city’s crime
problem. For example, people exposed to the crime as a
predator metaphor might propose toughening law
enforcement, while people exposed to the crime as disease
metaphor might think about dealing with problems in the
community and improving the social environment to
prevent future crime. Of course, it is also possible that such
metaphors are simply ornamental flourishes of language,
and do not influence how people conceive of important
social issues like crime.
Participants Four-hundred sixty-three students participated
in the study as part of a course requirement – 104 from
Stanford University and 359 from the University of
California–Merced. The same patterns were found in both
samples, so we report pooled data. Gathering data from the
two populations allowed us to get a somewhat broader
cross-section of the general population. This seemed
important since people’s conceptions of social issues like
crime are likely to differ as a function of factors like
socioeconomic status and personal experience.
Materials The survey was included in a larger packet of
questionnaires that were unrelated to this study. Participants
filled out the packet individually in a quiet room. Our
survey consisted of a single page which included a short
paragraph about crime in the fictional city of Addison and
some follow-up questions. The paragraph mostly contained
crime statistics, which were the same in the two conditions.
The two conditions differed only in two sentences in the
paragraph which were used to embed either the crime-as-
predator metaphor or the crime-as-disease metaphor. Half
of the subjects were given the crime-as-predator version and
half the crime-as-disease version. The report read:
Crime is a (wild beast preying on / virus infecting)
the city of Addison. The crime rate in the once
peaceful city has steadily increased over the past
three years. In fact, these days it seems that crime
is (lurking in / plaguing) every neighborhood. In
2004, 46,177 crimes were reported compared to
more than 55,000 reported in 2007. The rise in
violent crime is particularly alarming. In 2004,
there were 330 murders in the city, in 2007, there
were over 500.
Below this report, subjects were asked to briefly describe
the crime situation in Addison (to make sure they read and
understood the story), and were then instructed to answer
the following two questions: 1. In your opinion what does
Addison need to do to reduce crime? 2A) What aspect of the
report was most influential in your decision? 2B) Please
underline the part of the report that was most influential in
Fifty-two participants were given version A of question
2, and 411 participants were given version B. This
question was aimed at discovering if participants
explicitly noticed or made use of the metaphor, and we
wanted to allow participants different opportunities to
The solutions participants proposed to the crime problem in
Addison differed depending on the metaphorical frame used
in the crime report. Interestingly, most participants were not
consciously aware of this influence. In response to the
second question (which part of the report was most
influential in your decision?), only 3% of the participants
reported that the metaphorical framing influenced their
Participants’ proposed solutions to the crime problem in
Addison were coded into two categories: 1) social
environment and 2) law enforcement / punishment.
Responses were categorized as “social environment” if they
suggested a social reform (e.g., healthcare or educational or
welfare programs) or investigating the cause of the problem
(e.g., “look for the root cause”). Responses were
categorized as “law enforcement / punishment” if they
suggested restructuring the police force (e.g., hiring more
officers, retraining officers, calling in the national guard) or
modifying the penal structure (e.g., instituting harsher
penalties, building more jails).
Each participant’s response was coded as either 1 point
for the social environment category, 1 point for the
enforcement/punishment category, or split .5 points for each
category if both types of solutions were proposed.
Responses were coded by a blind coder. Of all responses,
9% did not fit into either category. In nearly every case this
was because the response lacked a suggestion (e.g., “I don’t
know”, “I need more information”, “It should be
Results are shown in Figure 1. Overall, participants were
more likely to suggest an enforcement/punishment solution
than a social environment solution (74% enforcement, 23%
social environment), χ2 = 98.12, p < .001. However,
participants given the crime-as-virus metaphorical framing
were more likely to suggest social reform (31%) than
participants given the crime-as-predator framing (20%), χ2
= 6.88, p < .01.
Figure 1: Proportion of proposed solutions broken down by
The results suggest that metaphors can influence how
people conceptualize and in turn hope to solve important
social issues. It appears that even casually encountering one
metaphor or another in discourse about crime can lead
people to unwittingly propose different types of solutions
for the crime problem. Importantly, it appears that the
metaphors had a subconscious effect on people’s reasoning.
Very few of our participants thought that the metaphors
influenced their crime-reducing suggestion.
To further explore a possible mechanism for the effects
observed in this experiment, we created a computational
model, using a connectionist architecture. Our goal was to
explore how connectionist architectures can model people’s
reasoning in highly structured and relational domains such
as metaphor and analogy, often seen as an area of weakness
for connectionist architectures (Fodor & Pylyshyn, 1988;
Gentner & Markman, 1993). Recent work by Rogers and
McClelland (R&M, 2008) suggests that connectionist
networks may indeed be sensitive to the types of relational
similarity that defines analogy and metaphor (Gentner,
To explore metaphors in a connectionist model we
considered three domains (crime, predator, and virus). By
gathering descriptions of the three domains we were able to
create a set of 36 representative phrases (11 for predator, 11
for virus, and 14 for crime). Each phrase is characterized by
a “domain-label” (crime, predator, or virus) as the first
object of the phrase, a relation (e.g., “caused_by”, “infects”,
“capture_with”) as the means of associating objects, and a
“completion” (e.g., “poverty”, “people”, and “trap”) as the
second object of the phrase. For example, one phrase might
This list of 36 phrases included 17 “relations” and 19
“completions”. Each relation and completion fell into one
of four categories - crime-specific (e.g., "arrest"), predator-
specific (e.g., "trap"), virus-specific (e.g., "infect"), or
shared (e.g., "cause").
There were four domain-specific relations that reflected
suggestions for solving each problem (“solutions units”):
these included “treat_by” in the virus domain,
“exterminate_by” in the predator domain, and “reform_by”
and “fight_by” in the crime domain. These relations
reflected the way that people suggested solving each
problem – e.g., “solve crime by investigating/reforming the
education system” or “solve crime by fighting it with more
police.” Figure 2 illustrates the set of relations and
completions in each domain as well as how they were linked
to domains in training. The circles highlight completions
that were activated as a function of the domain-specific
We trained the model by presenting one of the relations
and the corresponding domain-label as inputs and checking
the completions. Back-propagation of error was used to
adjust the network so that eventually the pairing of a
domain-label and relation activated all of the relevant
completions. In the test phase, we presented the model a
domain-label and relation that belonged to different domains
– pairings that the model had not seen in training (e.g.,
"crime" and "exterminate_by").
The virus and predator domains both shared some
structural similarity with the crime domain. If the model
was able to capture this structural similarity in learning, then
it should be able to interpret a structurally similar but
crossed pairing. If not, it might activate some unsystematic
set of irrelevant completions.
Network architecture. The network’s architecture is
shown in Figure 3. It is modeled on R&M’s feed-forward
network and is trained through back propagation. There are
two input layers (labeled "domain-labels" and "relations"),
three hidden layers, and one output layer (labeled
"completions"). Each of the input layers projects to an
associated set of hidden units - one is the learned
representation of the “domain-labels” and the other is the
learned representation of "relations." These two sets of
hidden units project forward into the third set of hidden
units - the learned representation of the item and relation
units combined. This hidden layer feeds forward to the
output layer – the “completions”.
Note that the two input layers and the output layer are
divided into distinct domains. This separation is what
allows us to explore metaphor processing. That is, in the
training phase, the model learns aspects of the three
domains separately. In the test phase, previously unpaired
relations and domains are crossed.
Figure 3: Network architecture
Creating concepts We sought to create ecologically valid
representations of our target domains, so we gathered
descriptions of the three situations. Forty-one participants
from Foothill College completed an online survey in
exchange for course credit. The survey asked participants to
describe, in at least ten sentences, several situations. Three
of the situations were: 1) worsening crime, 2) a wild beast
on the loose, and 3) a spreading virus. These descriptions
formed the basis of the training patterns that were presented
to the model.
We selected a set of relations by reading through the
descriptions and collecting lists of the verbs used to describe
the situations. For each domain, we identified a set of the
most commonly used verbs – often combining counts of
verbs that shared a similar meaning (e.g., “capture” and
“catch”). We then read through the stories again to identify
completions. We created a list of the subjects and objects of
the sentences that were commonly associated with the
selected verbs. This list was similarly condensed.
Structural Similarity Previous research has demonstrated
that structural similarity is a critical feature of a
comprehensible metaphor (e.g., Gentner & Clement, 1988).
The situation descriptions revealed some critical structural
features, which mirrored the findings of the experiment.
Descriptions of the virus situation often emphasized the
underlying aspects of the city that enabled the virus to
spread (e.g., investigating the water supply and food sources
as well as the ways in which people interact socially).
Descriptions of the predator situation, on the other hand,
often emphasized targeted responses (e.g., hunting down
and caging the predator). Descriptions of the crime
situation fit two schemas. In one, participants focused on
the reforming underlying causes of crime – e.g., social
welfare. In the other, participants focused on fighting crime
head on – e.g., hiring police officers and building jails.
One relation in the virus set (“treat_by”), one relation in
the predator set (“exterminate_by”), and two relations in the
crime set (“reform_by” and “fight_by”) were critical for
highlighting structural similarity. These relations reflect the
participants’ consensus on how to solve each situation. The
“treat_by” relation is associated with the causal elements of
a virus. The “exterminate_by” relation is associated with
targeted and immediate responses to a predator. The
“reform_by” relation is associated with the causal elements
of crime and the “fight_by” relation is associated with the
targeted and immediate responses to crime.
Network training The training patterns paired a single
domain-label with a single relation and all appropriate
completions. In some cases, completions were linked to a
domain via a shared relation; in other cases they were linked
via a domain-specific relation. Figure 2 illustrates the set of
relation-completion pairings within each domain (note that
there were not two distinct representations of crime).
The connection weights in the network were initialized
with very small random values. There were 19 training
patterns. All patterns were presented to the model an equal
number of times in random order. Back-propagation was
used to do online learning after each pattern. The network
was trained for 30,000 epochs without momentum with a
learning rate of .05 at which point 99% of the output units
were activated to within .1 of their target values.
Network testing In the testing phase, we explored whether
the model had learned and could utilize the relational
structure of the three target domains. Two test patterns
illustrate the critical pairings: 1) Crime + Treat and 2) Crime
These label–relation pairings were never seen by the
network in the training phase. If the model could harness
the relational structure of the domains, we would expect the
underlying causes of crime activated when crime was paired
with “treat_by” (as in the lower left quadrant of Figure 2).
We would expect the immediate responses of crime
activated when crime was paired with “exterminate_by” (as
in the lower right quadrant of Figure 2).
We ran three simulations with different initial starting
weights, and the resulting activation in the completions was
to ensure that the results do not reflect chance findings from
a particular set of starting weights. Figure 4 illustrates the
average level of activation in completions in response to the
two test inputs.
Figure 4: Average activation level of completions when
crime was paired with “treat_by” (left) and
“exterminate_by” (right) as inputs. No other completion
averaged an activation level greater than .1.
The results reveal that pairing “crime” with “treat_by”,
which were unpaired in the learning phase, activated the
underlying causes of crime and not the immediate targeted
responses to crime or causes of a virus. Pairing “crime”
and “exterminate_by”, which were also unpaired in the
learning phase, activated the immediate targeted responses
to crime and not the underlying causes of crime or the
immediate responses to a predator.
A closer look at the learned and distributed
representations of the relation-units reveals how the
metaphoric influence works. The model naturally
represents “treat_by” and “reform_by” as well as
“exterminate_by” and “fight_by” similarly because the two
pairs of relations serve similar roles in their specific
domains and because there are a scarce number of
representation units for the relations. That is, both the
“treat_by” and “reform_by” relations activate completions
that are associated with “cause”. Therefore, the model can
represent the two domains with very similar patterns of
representation. The model does not confuse the two
representations because it activates the completions only
after integrating the information from the domain-label
stream of input (i.e., in the subsequent representation layer).
The similarity between the “exterminate_by” and
“fight_by” relations is slightly more difficult to explain
because there are no shared relation units that could underlie
the mapping. Instead, as in the case of R&M’s simulation,
the network succeeds because the two domains are nearly
identical structurally (identical except for the additional
crime-specific relation unit, “reform”). Figure 5 illustrates
the learned and distributed representational similarity
between the solution units.
Figure 5: Learned, distributed representational similarity
among relations that serve a structurally similar purpose.
The columns represent activation levels in the eight units
that learn a representation for relations.
The model presented here learned and exploited the
relational similarity between the three target domains in a
way that was consistent with the results from the
Experiment. Specifically, the model learned to represent the
three domains by learning both the low level particulars of
each domain as well as high-level structural similarity. This
model builds on earlier simulations by Hinton's "family
tree" simulations (1989) as well as R&M's (2008).
Our model is different from previous computational
models of analogy/metaphor (e.g., Falkenhainer et al., 1989;
Eliasmith & Thagard, 2001; Hummel & Holyoak, 1997) in
that the conceptual representations are entirely learned and
distributed. In fact, it is precisely because the
representations of the relation units are distributed that
cross-domain inputs yield appropriate outputs.
Limitations and future directions It should be noted that
the model presented here does not fully reflect the
experimental paradigm. Participants in the experiment are
not asked to, e.g., “treat crime.” Future work will attempt to
design a model in which the representation of one domain
(e.g., virus) is activated and remains activated while inputs
from a distinct domain (e.g., crime) are presented to the
model. The model presented here is intended as a first step
in closely studying the mechanisms that underlie metaphor
processing from a connectionist perspective.
In this paper we presented one empirical demonstration of
the role of metaphors in shaping people’s thinking about an
important social issue, and a computational model which
explored a possible mechanism for such effects. In the
experiment, participants’ suggestions for solving a crime
problem were systematically influenced by a metaphoric
frame. When crime was compared to a virus, participants
were more likely to suggest reforming the social
environment of the infected community. When crime was
compared to a predator, participants were more likely to
suggest attacking the problem head on – hiring more police
officers and building jails.
In the simulation we explored how neural networks are
able to learn and use structural similarity to successfully
interpret cross-domain relations. When domain-specific
relations activated structurally similar completions, the
model naturally learned to represent them in a similar way.
The experiment and simulation confirm recent speculation
by policy makers, academics, and journalists who have
suggested that the metaphors we use to talk about important
issues shapes how we think about the issues and even how
we approach solving them. This research suggests that it is
vital that we choose good metaphors to frame social issues.
Far from being mere rhetorical flourishes, metaphors may
have a profound influence on how we act to deal with
important societal issues.
Boroditsky, L. (2000). Metaphoric Structuring:
Understanding time through spatial metaphors. Cognition,
Eliasmith, C. & Thagard, P. (2001). Integrating structure
and meaning: a distributed model of analogical mapping.
Cognitive Science, 25, 245-286.
Falkenhainer, B., Forbus, K. D., & Gentner, D. (1989).
The Structure Mapping Engine: Algorithms and
examples. Artificial Intelligence, 41, 1-63.
Fodor, J. & Pylyshyn, Z. (1988). Connectionism and
cognitive architecture: A critical analysis. Cognition, 28,
Gentner, D. (1983). Structure mapping: a theoretical
framework for analogy. Cognitive Science, 7, 155-170.
Gentner, D., & Clement, C. (1988). Evidence for relational
selectivity in the interpretation of analogy and metaphor.
In G. H. Bower (Ed.), The psychology of learning and
motivation: Advances in research and theory (Vol. 22, pp.
307-358). New York: Academic Press.
Gentner, D., & Gentner, D. (1983). Flowing waters and
teeming crowds: mental models of electricity. In Gentner,
D., & Stevens, A. (ed.) Mental Models. Hillsdale, NJ:
Lawrence Erlbaum Associates, Publishers.
Gentner, D., and Markman, A. (1993). Analogy-watershed
or waterloo? Structural alignment and the development of
connectionist models of analogy. In S.J. Hanson, J.D.
Cowan, & C.L. Giles (Eds.), Advances in Neural
Information Processing Systems 5 (pp. 855-862). San
Mateo, CA: Morgan Kaufmann Publishers.
Guerrero, R., & Concha-Eastman, A. (2001). An
Epidemiological Approach for the Prevention of Urban
Violence: The Case of Cali, Colombia. World Health and
Hinton, G. (1989). Learning Distributed Representations of
Concepts. In R. G. M. Morris (Ed.), Parallel distributed
processing: Implications for psychology and
neurobiology (Chap. 3, pp. 46–61). Oxford: Clarendon
Hummel, J.E., & Holyoak, K.J. (1997). Distributed
Representations of Structure: A Theory of Analogical
Access and Mapping. Psychological Review, 104, 427-
James, W. (1904). Remarks at the peace banquet. The
Atlantic Monthly, 94, 845-847.
Kelling, G. (1991). Crime and metaphor: Toward a new
concept of policing. City Journal, 1.
Kotlowitz, A. (May 4, 2008). Is urban violence a virus? New
York Times Magazine, 52-59.
Lakoff, G., & Johnson, M. (1980). Metaphors We Live By.
Chicago: Univ. of Chicago Press.
Lakoff, G. (2002). Moral Politics: How Liberals and
Conservatives Think. Chicago: University of Chicago
Nersessian, N. (1992). How do scientists think? Capturing
the dynamics of conceptual change in science. In Giere,
(ed.). Minneapolis: Univ. of MN Press.
Rogers, T. McClelland. (2008). Précis of Semantic
Cognition: A Parallel Distributed Processing Approach.
Behavioral and Brain Sciences, 31, 689-749.
Schön, D. (1993). Generative metaphor: A perspective on
problem-setting in social policy. In Ortony, A. (ed.)
Metaphor and Thought. Cambridge, UK: Univ. Press.