Generating Apt Metaphor Ideas for Pictorial Advertisements
Ping Xiao, Josep Blat
Department of Information and Communication Technologies
Pompeu Fabra University
C./ Tànger, 122-140 Barcelona 08018 Spain
{ping.xiao, josep.blat}@upf.edu
Abstract
Pictorial metaphor is a popular way of expression in
creative advertising. It attributes certain desirable quality
to the advertised product. We adopt a general twostage
computational approach in order to generate apt
metaphor ideas for pictorial advertisements. The first
stage looks for concepts which have high imageability
and the selling premise as one of their prototypical
properties. The second stage evaluates the aptness of
the candidate vehicles (found in the first stage) in regard
to four metrics, including affect polarity, salience,
secondary attributes and similarity with tenor. These
four metrics are conceived based on the general characteristics
of metaphor and its specialty in advertisements.
We developed a knowledge extraction method for the
first stage and utilized an affect lexicon and two semantic
relatedness measures to implement the aptness metrics
of the second stage. The capacity of our computer
program is demonstrated in a task of reproducing the
pictorial metaphor ideas used in three real advertisements.
All the three original metaphors were replicated,
as well as a few other vehicles recommended, which,
we consider, would make effective advertisements as
well.
Introduction
A pictorial advertisement is a short discourse about the
advertised product, service or idea (all referred to as ‘product’
afterwards). Its core message, namely the selling premise,
is a proposition that attributes certain desirable quality
to the product (Maes and Schilperoord 2008). A single
proposition can be expressed virtually in an unlimited
number of ways, among which some are more effective
than the others. The ‘how to say’ of an ad is conventionally
called the ‘idea’. ‘Pictorial metaphor’ is the most popular
way of expression in creative advertising (Goldenberg,
Mazursky and Solomon 1999). A pictorial metaphor involves
two dimensions, ‘structural’ and ‘conceptual’
(Forceville 1996; Phillips and McQuarrie 2004; Maes and
Schilperoord 2008). The structural dimension concerns
how visual elements are arranged in a 2D space. The conceptual
dimension deals with the semantics of the visual
elements and how they together construct a coherent message.
We see that the operations in the structural and conceptual
dimensions are quite different issues. In any of
these two dimensions, computational creativity is not a
trivial issue. In this paper, we are focusing on only one
dimension, the conceptual one.
The conceptual dimension of pictorial metaphors is not
very different from verbal metaphors (Foss 2005). A metaphor
involves two concepts, namely ‘tenor’ and ‘vehicle’.
The best acknowledged effect of metaphor is highlighting
certain aspect of the tenor or introducing some new information
about the tenor. Numerous theories have been proposed
to account for how metaphors work. The interaction
view is the dominant view of metaphor, which we also
follow. It was heralded by Richards (1936) and further
developed by Black (1962). According to Black, the principal
and subsidiary subjects of metaphor are regarded as
two systems of “associated commonplaces” (commonsense
knowledge about the tenor and vehicle). Metaphor works
by applying the system of associated commonplaces of the
subsidiary subject to the principal subject, “to construct a
corresponding system of implications about the principal
subject”. Any associated commonplaces of the principal
subject that conform the system of associated commonplaces
of the subsidiary subject will be emphasized, and
any that does not will be suppressed. In addition, our view
of the subsidiary subject is also altered.
Besides theories, more concrete models have been proposed,
mainly the salience imbalance model (Ortony
1979), the domain interaction model (Tourangeau and
Sternberg 1982), the structure mapping model (Gentner
1983; Gentner and Clement 1988), the class inclusion
model (Gluckberg and Keysar 1990, 1993) and the conceptual
scaffolding and sapper model (Veale and Keane
1992; Veale, O’Donoghue and Keane 1995). Furthermore,
these models suggest what make good metaphors,
i.e. metaphor aptness, which is defined as “the extent to
which a comparison captures important features of the topic”
(Chiappe and Kennedy 1999).
In the rest of this paper, we first specify the problem of
generating apt metaphor ideas for pictorial advertisements.
Then, the relevant computational approaches in the literature
are reviewed. Next, we introduce our approach to the
stated problem and the details of our implementation. Subsequently,
an experiment with the aim of reproducing three
pictorial metaphors used in real advertisements and the
Proceedings of the Fourth International Conference on Computational Creativity 2013 8
results generated by our computer program are demonstrated.
In the end, we conclude the work presented in this
paper and give suggestion about future work.
Problem Statement
The whole range of non-literal comparison, from mereappearance
to analogy (in the terms of Gentner and Markman
(1997)), is featured in pictorial advertisements. But,
analogies are rare. What appear most frequently are metaphors
with the mapping of a few attributes or relations.
This type of pictorial metaphors is the target of this paper.
To generate pictorial metaphors for advertisements, our
specific problem is searching for concepts (vehicles), given
the product (tenor), its selling premise (the property concept)
and some other constraints specified in an advertising
brief. The metaphor vehicles generated have to be easy to
visualize and able to establish or strengthen the connection
between the product and the selling premise.
There are two notes specific to advertisements that we
would like to mention. One is about the tenor of metaphor.
In pictorial ads, not only the product, but also “the internal
components of the product and the objects that interact
with it” are often used as tenors (Goldenberg, Mazursky
and Solomon 1999). The other note is about the selling
premise. Metaphors in advertisements are more relevant to
communicating intangible, abstract qualities than talking
about concrete product facts (Phillips and McQuarrie
2009). Therefore, we are primarily considering abstract
selling premises in this paper. In the next section, we review
the computational approaches to metaphor generation
that are related to the problem just stated.
Computational Approaches to Metaphor
Generation
Abe, Sakamoto and Nakagawa (2006) employed a threelayer
feedforward neural network to transform adjectivemodified
nouns, e.g. ‘young, innocent, and fine character’
into ‘A like B’ style metaphors, e.g. ‘the character is like a
child’. The nodes of the input layer correspond to a noun
and three adjectives. The nodes of the hidden layer correspond
to the latent semantic classes obtained by a probabilistic
latent semantic indexing method (Kameya and Sato
2005). A semantic class refers to a set of semantically related
words. Activation of the input layer is transferred to
the semantic classes (and the words in each class) of the
hidden layer. In the output layer, the words that receive
most activation (from different semantic classes) become
metaphor vehicles. In effect, this method outputs concepts
that are the intermediates between the semantic classes to
which the input noun and three adjectives are strongly associated.
If they are associated to different semantic
classes, this method produces irrelevant and hard to visualize
vehicles.
A variation of the above model was created by Terai and
Nakagawa (2009), who made use of a recurrent neural
network to explicitly implement feature interaction. It differs
with the previous model at the input layer, where each
feature node has bidirectional edge with every other feature
node. The performance of these two models was compared
in an experiment of generating metaphors for two tenors.
The model with feature interaction produced better results.
Besides, Terai and Nakagawa (2010) proposed a method
of evaluating the aptness of metaphor vehicles generated
by the aforementioned two computational models. A candidate
vehicle is judged based on the semantic similarity
between the corresponding generated metaphor and the
input expression. A candidate vehicle is more apt when the
meaning of the corresponding metaphor is closer to the
input expression. The semantic similarity is calculated
based on the same language model used in the metaphor
generation process. The proposed aptness measure was
tested in an experiment of generating metaphors for one
input expression, which demonstrated that it improved the
aptness of generated metaphors.
Veale and Hao (2007) created a system called Sardonicus
which can both understand and generate propertyattribution
metaphors. Sardonicus takes advantage of a
knowledge base of entities (nouns) and their most salient
properties (adjectives). This knowledge base is acquired
from the web using linguistic patterns like ‘as ADJ as *’
and ‘as * as a/an NOUN’. To generate metaphors, Sardonicus
searches the knowledge base for nouns that are associated
with the intended property. The aptness of the found
nouns is assessed according to the category inclusion
theory, i.e. “only those noun categories that can meaningfully
include the tenor as a member should be considered
as potential vehicles”. For each found noun, a query in the
format ‘vehicle-like tenor’ is sent through a search engine.
If there are more than zero results returned, the noun is
considered an apt vehicle. Otherwise, it is considered not
apt or extremely novel.
The above reviewed effort of generating metaphor converges
at a two-stage approach. These two stages are:
 Stage 1: Search for concepts that are salient in the
property to be highlighted
 Stage 2: Evaluate the aptness of the found concepts
as metaphor vehicles
This two-stage approach of metaphor generation is adopted
by us. We provide methods of searching and evaluating
metaphor vehicles, which are different from the literature.
In addition, special consideration is given to the aptness of
metaphor in the advertising context.
An Approach of Generating Apt Metaphor
Ideas for Pictorial Advertisements
We adopt a general two-stage computational approach of
metaphor generation (as introduced in the last section) to
generate apt metaphor ideas for pictorial advertisements.
At the first stage, we look for concepts which have high
Imageability (Paivio, Yuille and Madigan 1968; Toglia,
and Battig 1978) and the selling premise as one of their
prototypical properties. At the second stage, we evaluate
the aptness of the candidate vehicles using four metrics,
including affect polarity, salience, secondary attributes and
Proceedings of the Fourth International Conference on Computational Creativity 2013 9
similarity with tenor. Vehicles that are validated by all the
four metrics are considered apt for a specific advertising
task. In the following sections, we explain the rationale of
our approach and its computational details.
Stage 1: Search Candidate Metaphor Vehicles
To find entities which have the selling premise as one of
their prototypical properties, our strategy is searching for
concepts that are strong semantic associations of the selling
premise. One note to mention is that the concepts soughtafter
do not need to be the ‘absolute’ associations, because
the meaning of a metaphor, i.e. which aspect of the tenor
and vehicle becomes prominent, does not only depend on
the vehicle, but on the interaction between the tenor and
vehicle. In the past, we developed an automatic knowledge
extraction system, namely VRAC (Visual Representations
for Abstract Concepts), for providing concepts of physical
entities to represent abstract concepts (Xiao and Blat
2011). Here we give a brief introduction of this work.
We look for semantic associations in three knowledge
bases, includingword association databases (Kiss,
Armstrong, Milroy and Piper 1973; Nelson, McEvoy and
Schreiber 1998), a commonsense knowledge base called
ConceptNet (Liu and Singh 2004) and Roget’s Thesaurus
(Roget 1852). The reason for using three of them is that we
want to take use of the sum of their capacity, in terms of
both the vocabulary and the type of content. The nature of
these three knowledge bases ensures that the retrieved concepts
have close association with the selling premise.
Vehicles of pictorial metaphors should have high imageability,
in order to be easily visualized in advertisements.
Imageability refers to how easy a piece of text elicits
mental image of its referent. It is usually measured in
psychological experiments. The available data about word
imageability, at the scale of thousands, does not satisfy our
need of handling arbitrary words and phrases. As imageability
is highly correlated with word concreteness, we developed
a method of estimating concreteness using the
ontological relations in WordNet (Fellbaum 1998), as an
approximation of imageability.
To evaluate the capacity of VRAC, we collected thirtyeight
distinct visual representations of six abstract concepts
used in past successful advertisements. These abstract concepts
have varied parts of speech and word usage frequency.
We checked if these visual representations were included
in the concepts output by VRAC, with the corresponding
abstract concept as input. On average, VRAC
achieved a hit rate of 57.8%. The concepts suggested by
VRAC are mostly single objects. It lacks the concepts of
scenes or emergent cultural symbols, which also play a role
in mass visual communication.
Stage 2: Evaluate the Aptness of Candidate
Vehicles
The aptness of the candidate vehicles generated in Stage 1
is evaluated based on four metrics, including affect polarity,
salience, secondary attributes and similarity with tenor.
Affect Polarity Most of the time, concepts with negative
emotions are avoided in advertising (Kohli and Labahn,
1997; Amos, Holmes and Strutton 2008). Even in provocative
advertisements, negative concepts are deployed with
extreme caution (De Pelsmacker and Van Den Bergh 1996;
Vézina and Paul 1997; Andersson, Hedelin, Nilsson and
Welander 2004). In fact, negative concepts are often discarded
at the first place (Kohli and Labahn 1997). Therefore,
we separate candidate vehicles having negative implication
from the ones having positive or neutral implication.
For this purpose, affective lexicons, which provide affect
polarity values of concepts, come in handy. We decided to
use SentiWordNet 3.0 (Baccianella, Esuli and Sebastiani
2010), due to its big coverage (56,200 entries) and fine
grained values. It provides both the positive and negative
valences, which are real values ranging from 0.0 to 1.0. If a
candidate vehicle is found in SemtiWordNet 3.0, its affect
polarity is calculated by subtracting the negative valence
from the positive valence. The candidate vehicles which
are not included in SemtiWordNet 3.0 are considered being
emotionally neutral.
Salience Salience refers to how strongly a symbol evokes
certain meaning in humans’ mind. The candidate vehicles
found by VRAC have varying association strength with the
selling premise, from very strong to less. The vehicle of a
metaphor has to be more salient in the intended property
than the tenor (Ortony 1979; Glucksberg and Keysar
1990). We interpret salience as a kind of semantic relatedness
(Budanitsky and Hirst 2006), which reflects how far
two concepts are in the conceptual space of a society. We
calculate the semantic relatedness between each candidate
vehicle and the selling premise, and between the product
and the selling premise. Candidate vehicles that are more
remote from the selling premise than the product are discarded.
We will talk more about semantic relatedness and
the specific measures we used in a later section.
Secondary Attributes Metaphors that capture the appropriate
number of relevant features are considered especially
apt (Glucksberg and Keysar 1990, 1993; Chiappe and
Kennedy 1999). Phillips (1997) found that strong implicatures
as well as weak implicatures were drawn from pictorial
advertisements. Strong implicatures correspond to
the selling premise of an ad, while we use ‘secondary
attributes’ for referring to the weak implicatures. We have
not seen literature on the salience of the secondary
attributes in metaphor vehicles. We think the candidate
vehicles should, at least, not contradict the secondary
attributes prescribed to a product. For this end, we use a
semantic relatedness measure to filter candidate vehicles
that are very distant from the secondary attributes. This is
‘soft’ filtering, in contrast to the ‘hard’ filtering used in the
previous two metrics, i.e. affect polarity and salience, in
the sense that the current criterion might need be tighten in
order to ensure the aptness of generated metaphors.
We compare the above approach with an alternative,
which is using both the selling premise and the secondary
attributes to search for candidate vehicles. This alternative
Proceedings of the Fourth International Conference on Computational Creativity 2013 10
method indeed looks for concepts that are salient in all
these properties. This is possible, but rare. Most of the
time, no result will be returned. On the other hand, there is
a natural distinction of priority in the attributes (for a product)
desired by advertisers (recall the strong and weak implicatures
just mentioned). To represent this distinction,
weighting of attributes is necessary.
The computational model proposed by Terai and Nakagawa
(2009) also uses multiple features to generate metaphors.
The weights of the edges connecting the feature
nodes in the input layer vary with the tenor. Specifically,
the weight of an edge equals to the correlation coefficient
between the two features respecting the tenor. The calculation
is based on a statistic language model built on a Japanese
corpus (Kameya and Sato 2005), which means the
weighting of features (of a tenor) is intended to be near
reality. However, this idea does not suit advertising, because
the features attributed to a product are much more
arbitrary. Very often, a product is not thought possessing
those features before the appearance of an advertisement.
Similarity with Tenor Good metaphors are those whose
tenor and vehicle are not too different yet not too similar to
each other (Aristotle 1924; Tourangeau and Sternberg
1981; Marschark, Kats and Paivio 1983). For this reason,
we calculate the semantic relatedness between the product
and each candidate vehicle. Firstly, candidate vehicles
which have zero or negative semantic relatedness values
are discarded, because they are considered too dissimilar to
the product. Then, the candidate vehicles with positive
relatedness values are sorted in the descending order of
relatedness. Among this series of values, we look for values
that are noticeably different from the next value, i.e.
turning points. Turning points divide relatedness values
into groups. We use the discrete gradient to measure the
change of value, and take the value with the biggest change
as the turning point. Candidate vehicles with their relatedness
value bigger than or equal to the turning point are
abandoned, for being too similar to the tenor. Figure 1
shows the sorted relatedness values between the candidate
vehicles and the tenor ‘child’ in the ad of the National Museum
of Science and Technology. The turning point in this
graph corresponds to the concept ‘head’.
Semantic Relatedness Measures In general, semantic
relatedness is measured through distance metrics in certain
materialized conceptual space, such as knowledge bases
and raw text. A number of semantic relatedness measures
have been proposed. Each measure has its own merits and
weakness. We employed two different measures in the
current work, including PMI-IR (Pointwise Mutual Information
and Information Retrieval) (Turney 2001) and LSA
through Random Indexing (Kanerva, Kristofersson and
Holst 2000). PMI-IR is used to compute salience, because
we found it gives more accurate results than other available
measures when dealing with concept pairs of high semantic
relatedness. The relatedness between the selling premise
and candidate vehicles is deemed high. Therefore, we use
Similarity between Candidate Vehicles and 'Child'
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
0,16
half
head
car
uptake
Einstein
loaf
button
headpiece
mankind
alien
sage
brainpan
highbrow
chess
owl
reader
serpent
cerebrum
professor
Candidate Vehicles
Similarity with 'Child'
Figure 1: Similarity between candidate vehicles and 'Child'
PMI-IR to give a delicate ordering of their association
strength. LSA is employed for the metrics of secondary
attributes and similarity with tenor. The motivation behind
this choice is to capitalize on LSA’s ability of ‘indirect
inference’ (Landauer and Dumais 1997), i.e. discovering
connection between terms which do not co-occur. Recall
that candidate vehicles are assumed to have strong association
with the selling premise, but not necessarily the secondary
attributes. In most cases, the association between a
candidate vehicle and a secondary attribute is not high.
Thus, we need a measure which is sensitive to the lowrange
semantic relatedness. LSA has demonstrated capacity
in this respect (Waltinger, Cramer and Wandmacher
2009). For LSA, values close to 1.0 indicate very similar
concepts, while values close to 0.0 and under 0.0 indicate
very dissimilar concepts. In our computer program, we
utilize the implementation of Random Indexing provided
by the Semantic Vectors package1
. Two-hundred term vectors
are acquired from the LSA process for computing semantic
relatedness. In the present work, both PMI-IR and
LSA are based on the Wikipedia corpus, an online encyclopedia
of millions of articles. We obtained the English
Wikipedia dumps, offered by the Wikimedia Foundation2
on October 10th, 2011. The compressed version of this
resource is about seven gigabytes.
An Example
We intend to evaluate our approach of generating apt metaphor
ideas for pictorial advertisements based on checking
whether this approach can reproduce the pictorial metaphors
used in past successful advertisements. We have
been collecting a number of real ads and the information
about the product, selling premise, secondary attributes,
and the tenor and vehicle of metaphor in these ads. Nonetheless,
it is a tedious process.
 1 http://code.google.com/p/semanticvectors/
2 http://download.wikipedia.org/
Proceedings of the Fourth International Conference on Computational Creativity 2013 11
Rank Vehicle Rank Vehicle
1 IQ 19 reader
2 Mensa 20 child
3 brain 21 sage
4 computer 22 serpent
5 cerebrum 23 owl
6 alien 24 car
7 mankind 25 whale
8 highbrow 26 horse
9 Einstein 27 pig
10 head 38 half
11 professor 29 needle
12 dolphin 30 button
13 chess 31 table
14 lecturer 32 uptake
15 geek 33 storey
16 headpiece 34 loaf
17 newspaper 35 brainpan
18 atheist 36 latitudinarian
In this paper, we use the information of three real ads to
show what our computer program generates. These three
ads are for the Volvo S80 car, The Economist newspaper
and the National Museum of Science and Technology in
Stockholm respectively. Each of them has a pictorial metaphor
as its center of expression. All the three ads have the
same selling premise: 'intelligence'. However, three different
vehicles are used, including ‘chess’, ‘brain’ and ‘Einstein’
respectively. The selection of these particular ads
aims at testing whether our aptness metrics are able to differentiate
different tenors.
Table 1 summarizes the three aspects of the three ads,
including product, secondary attributes and the tenor of
metaphor. For both of the car and newspaper ads, the tenors
of metaphor are the products. For the museum ad, the
tenor is the target consumer, children.
We found the secondary attributes of the Volvo S80 car
in its product introduction3
. For the other two ads, the
Economist newspaper and the National Museum of
Science and Technology, we have not found any secondary
attributes specified. Instead, their subject matter is used to
distinguish them from the products of the same categories
Furthermore, we think it is more accurate to use the
Boolean operations ‘AND’ and ‘OR’ in describing the relation
between multiple secondary attributes. As consequence,
candidate vehicles have to be reasonably related to
both attributes at the two sides of AND; at least one of the
two attributes connected by OR.
Product Secondary Attributes Tenor
car
4 elegance AND luxury AND
sophisticated car
newspaper
5 international politics OR
business news newspaper
museum6 science OR technology child
Table 1: Information about the three real ads
For the concept 'intelligence', VRAC provides eightyseven
candidate vehicles, including single words and
phrases. We keep the single-word concepts and extract the
core concept of a phrase, in order to reduce the complexity
of calculating the aptness metrics at the later stage. An
example of the core concept of a phrase is the word ‘owl’
in the phrase ‘wise as an owl’. The core concepts are extracted
automatically based on syntactic rules. This process
introduces noise, i.e. concepts not related to 'intelligence',
such as 'needle' of the phrase 'sharp as a needle' and 'button'
of the phrase 'bright as a button'. In total, there are thirtyfour
candidate vehicles of single words. All the three metaphor
vehicles used in the three real ads are included.
 3 http://www.volvocars.com/us/all-cars/volvo-s80/pages/5-
things.aspx, retrieved on April 1st, 2012. 4 http://adsoftheworld.com/media/print/volvo_s80_iq
5 http://adsoftheworld.com/media/print/the_economist_brain
6
http://adsoftheworld.com/media/print/the_national_museum_of_
science_and_technology_little_einstein
As to affect polarity, the majority of the candidate vehicles,
thirty out of thirty-four, are emotionally neutral.
Besides, ‘highbrow’ is marked as positive, while ‘geek’
and ‘serpent’ as negative.
The ranking of candidate vehicles by its salience in the
selling premise is shown in Table 2. The semantic relatedness
calculated by PMI-IR correctly captured the main
trend of salience. ‘IQ’, ‘Mensa’ and ‘brain’ are ranked top,
while ‘needle’, ‘button’ and ‘table’, which are the noise
introduced by the core concept extraction method, are
ranked very low. The positions of the products are marked
in italic. Only candidate vehicles having higher salience
than a product are seen as valid. For instance, ‘horse’,
ranked the twenty-sixth, is not selected for the Volvo S80
car ad, since car is judged as more intelligent than horse by
PMI-IR. On the other hand, all the metaphor vehicles used
in the original ads, i.e. chess, brain and Einstein, have
higher rankings than the corresponding tenors, which supports
Ortony’s salience imbalance theory.
Table 2: Candidate vehicles sorted in the descending order of
salience
Table 3 shows how candidate vehicles are filtered by the
secondary attributes of products, where candidate vehicles
that are not contradictory to the secondary attributes are
presented. Table 4 shows the candidate vehicles that are
not too different yet not too similar with the tenors of the
three ads respectively. For both results, the metaphor vehicles
used in the original ads survived the filtering, which
gives support to the domain interaction theory proposed by
Tourangeau and Sternberg. Nevertheless, there is also flaw
in the results produced by the LSA-IR measure. For instance,
regarding the fourth column of Table 3, we suspect
Proceedings of the Fourth International Conference on Computational Creativity 2013 12
‘brain’ should not have nothing to do with ‘science’ and
consulted several other semantic relatedness measures,
which confirmed our skepticism.
Product car newspaper museum
Secondary
Attributes
elegance AND
luxury AND
sophisticated
international
politics OR
business news
science OR
technology
Candidate
Vehicle
chess
half
geek
IQ
brain
computer
cerebrum
mankind
highbrow
head
professor
dolphin
chess
lecturer
geek
headpiece
atheist
reader
sage
owl
car
whale
horse
half
needle
button
table
uptake
storey
brainpan
IQ
Mensa
computer
cerebrum
alien
mankind
highbrow
Einstein
head
professor
chess
lecturer
headpiece
atheist
reader
sage
owl
whale
half
needle
button
table
storey
loaf
brainpan
Table 3: Candidate vehicles NOT contradictory to the secondary
attributes of the three products respectively
Tenor car newspaper
child
(museum)
Candidate
Vehicle
pig
storey
mankind
uptake
button
half
serpent
whale
lecturer
chess
latitudinarian
sage
professor
alien
horse
IQ
professor
loaf
whale
table
atheist
geek
mankind
brainpan
head
Mensa
button
dolphin
brain
sage
pig
headpiece
uptake
storey
car
uptake
Einstein
loaf
button
headpiece
mankind
alien
sage
brainpan
highbrow
chess
owl
reader
serpent
cerebrum
professor
Table 4: Candidate vehicles that are not too different yet not too
similar with the tenors of the three ads respectively
We show in Table 5 the metaphor vehicles suggested by
our computer program for each of the three ads after applying
all the four aptness metrics. For all the three ads, the
vehicles used in the original ads are included in the vehicles
suggested by our computer program, as marked in
italic. For the Volvo S80 car ad, the original metaphor vehicle
is the only one recommended by our program. For the
other two ads, our program also proposed other five and
eight vehicles respectively. Considering that there are thirty
four candidate vehicles input to the second stage, we
think the four aptness metrics together did an acceptable
job.
Regarding the generated vehicles other than the one used
in the original ad: are they equally effective? We will have
a closer look at the metaphor vehicles generated for the ad
of the National Museum of Science and Technology, since
it has the most suggested vehicles. It is easy to spot a semantic
cluster among these eight vehicles. Five out of eight
are humans or human-like entities bearing high intellect,
including ‘Einstein’, ‘mankind’, ‘alien’, ‘highbrow’ and
‘professor’. ‘Einstein’, as the most prototypical within this
cluster, fits best this specific advertising task. Besides, other
vehicles in this cluster are also highly relevant to a setting
like museum for people, especially children, to increase
knowledge and encounter inspiration. They may be
optimal for other advertising tasks with slightly different
focus. The only exception is ‘mankind’, which is a very
general concept. As to the rest of the suggested metaphor
vehicles, certain ‘headpiece’ is possibly kind of symbol of
intelligence; playing ‘chess’ shows someone is intelligent,
and ‘cerebrum’ is strongly associated with intelligence. It
is not difficult to imagine a picture of juxtaposing a headpiece
and a child, a child playing chess or a child whose
cerebrum is emphasized, all of which would be effective to
associate a child with intelligence. However, strictly speaking,
they are not metaphors.
On the other hand, the existence of candidate vehicles
other than the ones used in the original ads may suggest,
firstly, our implementation of the four aptness metrics may
not sufficiently reduce inapt vehicles. Secondly, more metrics,
representing other factors that affect metaphor aptness,
may be necessary.
Ad Tenor Vehicle
Volvo S80 car car chess
The Economist newspaper newspaper
professor
mankind
head
dolphin
brain
headpiece
National Museum of Science
and Technology child
Einstein
headpiece
mankind
alien
highbrow
Proceedings of the Fourth International Conference on Computational Creativity 2013 13
chess
cerebrum
professor
Table 5: Metaphor vehicles considered apt for the three ads respectively
Conclusions
In the work presented in this paper, we adopted a general
two-stage computational approach to generate apt metaphor
ideas for pictorial advertisements. The first stage
looks for concepts which have high imageability and the
selling premise as one of their prototypical properties. The
second stage evaluates the aptness of the candidate vehicles
(found in the first stage) with regard to four aspects,
including affect polarity, salience, secondary attributes and
similarity with tenor. These four metric are conceived
based on the general characteristics of metaphor and its
specialty in advertising. For the first stage, we developed
an automatic knowledge extraction method to find concepts
of physical entities which are strongly associated
with the selling premise. For the second stage, we utilized
an affect lexicon and two semantic relatedness measures to
implement the four aptness metrics. The capacity of our
computer program is demonstrated in a task of reproducing
the pictorial metaphors used in three real advertisements.
All the three original metaphors were replicated, as well as
a few other vehicles recommended, which, we consider,
would make effective advertisements, though less optimal.
In short, our approach and implementation are promising
in generating diverse and apt pictorial metaphors for advertisements.
On the other hand, to have a more critical view of our
approach and implementation, larger scale evaluation is in
need. Continuing the evaluation design introduced in this
paper, more examples of pictorial metaphors used in real
advertisements have to be collected and annotated. This
corpus would not only contribute to building our metaphor
generator, but also be an asset for the research on metaphor
and creativity in general.
Moreover, the results provided by our aptness metrics
support both the salience imbalance theory and the domain
interaction theory.
Future Work
We intend to compute more ways of expression appeared
in pictorial advertisements. Firstly, our current implementation
can be readily adapted to generate visual puns. In a
pun, the product (or something associated to it) also has the
meaning of the selling premise. An example is an existing
ad which uses the picture of an owl to convey the message
‘zoo is a place to learn and gain wisdom’. As we all know,
owl is both a member of the zoo and a symbol of wisdom.
Secondly, we found some other fields of study are very
relevant to computing advertising expression, such as the
research and computational modeling of humor (Raskin
1985; Attardo and Raskin 1991; Ritchie 2001; Binsted,
Bergen, Coulson, Nijholt, Stock, Strapparava, Ritchie, Manurung,
Pain, Waller and O’Mara, 2006). Finally, we are
especially interested in investigating hyperbole. Hyperbole
has nearly universal presence in advertisements, but its
theoretic construction and computational modeling are
minimal. There exist some ad-hoc approaches: for instance,
we can find the exaggeration of the selling proposition
by the AlsoSee relation in WordNet; or, we should
first think about a cognitive or linguistic model of hyperbole
instead.
<references_biblio/>
References
Abe, K., Sakamoto, K., and Nakagawa, M. 2006. A computational
model of metaphor generation process. In Proceedings of the
28th Annual Meeting of the Cognitive Science Society, 937–942.
Amos, C., Holmes, G., and Strutton, D. 2008. Exploring the relationship
between celebrity endorser effects and advertising effectiveness:
A quantitative synthesis of effect size. International
Journal of Advertising 27(2):209–234.
Andersson, S., Hedelin, A., Nilsson, A., and Welander, C. 2004.
Violent advertising in fashion marketing. Journal of Fashion
Marketing and Management 8:96-112.
Aristotle. 1924. The Art of Rhetoric. Translated by W. Rhys Roberts.
The Internet Classics Archive, accessed October 1, 2012.
http://classics.mit.edu//Aristotle/rhetoric.html.
Attardo, S., and Raskin, V. 1991. Script theory revis(it)ed: Joke
similarity and joke representation model. Humor: International
Journal of Humor Research 4(3-4):293-347.
Baccianella, S., Esuli, A., and Sebastiani, F. 2010. SentiWordNet
3.0: An enhanced lexical resource for sentiment analysis and
opinion mining. In Proceedings of the 7th Conference on Language
Resources and Evaluation (LREC’10), Valletta, MT.
Binsted, K., Bergen, B., Coulson, S., Nijholt, A., Stock, O.,
Strapparava, C., Ritchie, G., Manurung, R., Pain, H., Waller, A.,
and O’Mara, D. 2006. Computational humor. IEEE Intelligent
Systems March-April.
Black, M. 1962. Models and metaphors. NY: Cornell University
Press.
Budanitsky, A., and Hirst, G. 2006. Evaluating WordNet-based
measures of semantic distance. Journal of Computational Linguistics
32(1):13–47.
Chiappe, D., and Kennedy, J. 1999. Aptness predicts preference
for metaphors or similes, as well as recall bias. Psychonomic
Bulletin and Review 6:668–676.
De Pelsmacker, P., and Van Den Bergh, J. 1996. The communication
effects of provocation in print advertising. International
Journal of Advertising 15(3):203–22.
Fellbaum, C. D., ed. 1998. WordNet: An Electronic Lexical Database.
Cambridge: MIT Press.
Forceville, C. 1996. Pictorial Metaphor in Advertising. London:
Routledge.
Foss, S. K. 2005. Theory of Visual Rhetoric. In Smith, K., Moriarty,
S., Barbatsis, G., and Kenney, K., eds., Hand- book of
Visual Communication: Theory, Methods, and Media. Mahwah,
New Jersey: Lawrence Erlbaum. 141-152.
Gentner, D. 1983. Structure-mapping: A theoretical framework
for analogy. Cognitive Science 7:155-170.
Proceedings of the Fourth International Conference on Computational Creativity 2013 14
Gentner, D., and Clements, C. 1988. Evidence for relational selectivity
in the interpretation of analogy and metaphor. In Bower,
G., ed., The Psychology of Learning and Motivation, Vol. 22.
Orlando, FL: Academic Press.
Gentner, D., and Markman, A. B. 1997. Structure mapping in
analogy and similarity. American Psychologist 52:45-56.
Glucksberg, S., and Keysar, B. 1990. Understanding Metaphorical
comparisons: beyond similarity. Psychological Review 97:3-
18.
Glucksberg, S., and Keysar, B. 1993. How metaphors work. In
Ortony, A. ed., Metaphor and Thought, 2nd ed. Cambridge:
Cambridge University Press. 401- 424.
Goldenberg, J., Mazursky, D., and Solomon, S. 1999. Creativity
templates: towards identifying the fundamental schemes of quality
advertisements. Marketing Science 18(3):333-351.
Kameya, Y., and Sato, T. 2005. Computation of probabilistic
relationship between concepts and their attributes using a statistical
analysis of Japanese corpora. In Proceedings of Symposium
on Large-Scale Knowledge Resources, 65-68.
Kanerva, P., Kristofersson, J., and Holst, A. 2000. Random indexing
of text samples for latent semantic analysis. In Proceedings of
the 22nd Annual Conference of the Cognitive Science Society
(CogSci'00), Erlbaum.
Kiss, G. R., Armstrong, C., Milroy, R., and Piper, J. 1973. An
associative thesaurus of English and its computer analysis. In
Aitken, A. J., Bailey, R. W., and Hamilton-Smith, N. eds., The
Computer and Literary Studies, Edinburgh: University Press.
153-165.
Kohli, C., and Labahn, D. W. 1997. Creating effective brand
names: A study of the naming process. Journal of Advertising
Research 37(1):67–75.
Landauer, T. K., and Dumais, S. T. 1997. A solution to Plato's
problem: The latent semantic analysis theory of the acquisition,
induction, and representation of knowledge. Psychological Review
104:211-240.
Liu, H., and Singh, P. 2004. ConceptNet: A practical commonsense
reasoning toolkit. BT Technology Journal 22(4):211-26.
Maes, A., and Schilperoord, J. 2008. Classifying visual rhetoric:
Conceptual and structural heuristics. In McQuarrie, E. F., and
Phillips, B. J., eds., Go Figure: New Directions in Advertising
Rhetoric. Armonk, New York: Sharpe. 227-257.
Marschark, M., Katz, A. N., and Paivio, A. Dimensions of metaphor,
Journal of Psycholinguistic Research 12(1):17.
Nelson, D. L., McEvoy, C. L., and Schreiber, T. A. 1998. The
university of south Florida word association, rhyme, and word
fragment norms. University of South Florida Website, accessed
January 13, 2012. http://www.usf.edu/FreeAssociation/.
Ortony, A. 1979. Beyond literal similarity. Psychological Review
86:161-180.
Paivio, A., Yuille, J. C., and Madigan, S. 1968. Concreteness,
imagery, and meaningfulness values of 925 nouns. Journal of
Experimental Psychology 76:1-25.
Phillips, B. J. 1997. Thinking into it: Consumer interpretation of
complex advertising images. Journal of Advertising 26(2):77–87.
Phillips, B. J., and McQuarrie, E. F. 2004. Beyond visual metaphor:
A new typology of visual rhetoric in advertising. Marketing
Theory 4(1/2):111-134.
Phillips, B. J., and McQuarrie, E. F. 2009. Impact of advertising
metaphor on consumer beliefs: delineating the contribution of
comparison versus deviation factors. Journal of Advertising
38(1):49-61.
Raskin, V. 1985. Semantic Mechanisms of Humor. Dordrecht,
Boston, Lancaster: D. Reidel Publishing Company.
Richards, I. A. 1936. The Philosophy of Rhetoric. London: Oxford
University Press.
Ritchie, G. D. 2001. Current directions in computational humour.
Artificial Intelligence Review 16(2):119-135.
Roget, P. M. 1852. Roget’s Thesaurus of English Words and
Phrases. Harlow, Essex, England: Longman Group Ltd.
Terai, A., and Nakagawa, M. 2009. A neural network model of
metaphor generation with dynamic interaction. In Alippi, C.,
Polycarpou, M., Panayiotou, C., Ellinas, G., eds., ICANN 2009,
Part I. LNCS, 5768, Springer, Heidelberg . 779–788.
Terai, A., and Nakagawa, M. 2010. A computational system of
metaphor generation with evaluation mechanism. In Diamantaras,
K., Duch, W., Iliadis, L. S., eds., ICANN 2010, Part II. LNCS,
6353, Springer, Heidelberg. 142–147.
Toglia, M. P., and Battig, W. F. 1978. Handbook of Semantic
Word Norms. Hillsdale, NJ: Erlbaum.
Tourangeau, R., and Sternberg, R. J. 1981. Aptness in metaphor.
Cognitive Psychology 13: 27-55.
Tourangeau, R., and Sternberg, R. J. 1982. Understanding and
appreciating metaphors. Cognition 11: 203-244.
Turney, P. D. 2001. Mining the Web for synonyms: PMI-IR versus
LSA on TOEFL. In Proceedings of the Twelfth European
Conference on Machine Learning, 491-502. Berlin: SpringerVerlag.
Veale, T., and Keane, M. T. 1992. Conceptual Scaffolding: A
spatially-founded meaning representation for metaphor comprehension.
Computational Intelligence 8(3):494-519.
Veale, T., O Donoghue, D., and Keane, M. T. 1995. Epistemological
issues in metaphor comprehension: A comparative analysis of
three models of metaphor interpretation. In Proceedings of
ICLC’95, the Fourth Conference of The International Cognitive
Linguistics Association, Albuquerque NM.
Veale, T., and Hao, Y. 2007. Comprehending and generating apt
metaphors: A Web-driven, case-based approach to figurative
language. In Proceedings of the Twenty-Second AAAI Conference
on Artificial Intelligence, Vancouver, Canada.
Vézina, R., and Paul, O. 1997. Provocation in advertising: A conceptualization
and an empirical assessment. International Journal
of Research in Marketing 14(2):177-192.
Waltinger, U., Cramer, I., and Wandmacher, T. 2009. From social
networks to distributional properties: A comparative study on
computing semantic relatedness. In Proceedings of the ThirtyFirst
Annual Meeting of the Cognitive Science Society, CogSci
2009, 3016- 3021. Cognitive Science Society, Amsterdam, Netherlands.
Xiao, P., and Blat, J. 2012. Image the imageless: Search for pictorial
representations of abstract concepts. In Proceedings of the
Seventh International Conference on Design Principles and Practices,
Chiba, Japan.
Proceedings of the Fourth International Conference on Computational Creativity 2013 15