Once More, With Feeling!
Using Creative Affective Metaphors to Express Information Needs
Tony Veale
Web Science & Technology Division, KAIST / School of Computer Science and Informatics, UCD
Korean Advanced Institute of Science & Technology, South Korea / University College Dublin, Ireland.
Tony.Veale@gmail.com
Abstract
Creative metaphors abound in language because they
facilitate communication that is memorable, effective and
elastic. Such metaphors allow a speaker to be maximally
suggestive while being minimally committed to any single
interpretation, so they can both supply and elicit information
in a conversation. Yet, though metaphors are often used to
articulate affective viewpoints and information needs in
everyday language, they are rarely used in information
retrieval (IR) queries. IR fails to distinguish between
creative and uncreative uses of words, since it typically
treats words as literal mentions rather than suggestive
allusions. We show here how a computational model of
affective comprehension and generation allows IR users to
express their information needs with creative metaphors that
concisely allude to a dense body of assertions. The key to
this approach is a lexicon of stereotypical concepts and their
affective properties. We show how such a lexicon is
harvested from the open web and from local web n-grams.
Creative Truths
Picasso famously claimed that “art is a lie that tells the
truth.” Fittingly, this artful contradiction suggests a
compelling reason for why speakers are so wont to use
artfully suggestive forms of creative language – such as
metaphor and irony – when less ambiguous and more
direct forms are available. While literal language commits
a speaker to a tightly fixed meaning, and offers little scope
to the listener to contribute to the joint construction of
meaning, creative language suggests a looser but
potentially richer meaning that is amenable to collaborative
elaboration by each participant in a conversation.
A metaphor X is Y establishes a conceptual pact between
speaker and listener (Brennan & Clark, 1996), one that
says ‘let us agree to speak of X using the language and
norms of Y’ (Hanks, 2006). Suppose a speaker asserts that
“X is a snake”. Here, the stereotype “snake” conveys the
speaker’s negative stance toward X, and suggests a range
of talking points for X, such as that X is charming and
clever but also dangerous, and is not to be trusted (Veale
& Hao, 2008). A listener may now respond by elaborating
the metaphor, even when disagreeing with the basic
conceit, as in “I agree that X can be charming, but I see no
reason to distrust him”. Successive elaboration thus allows
a speaker and listener to arrive at a mutually acceptable
construal of a metaphorical “snake” in the context of X.
Metaphors achieve a balance of suggestiveness and
concision through the use of dense descriptors, familiar
terms like “snake” that evoke a rich variety of stereotypical
properties and behaviors (Fishelov, 1992). Though every
concept has the potential to be used creatively, casual
metaphors tend to draw their dense descriptors from a large
pool of familiar stereotypes shared by all speakers of a
language (Taylor, 1954). A richer, more conceptual model
of the lexicon is needed to allow any creative uses of
stereotypes to be inferred as needed in context. We will
show here how a large lexicon of stereotypes is mined
from the web, and how stereotypical representations can be
used selectively and creatively, to highlight relevant
aspects of a given target concept in a specific metaphor.
Because so many familiar stereotypes have polarizing
qualities – think of the endearing and not-so-endearing
qualities of babies, for instance – metaphors are ideal
vehicles for conveying an affective stance toward a topic.
Even stereotypes that are not used figuratively, as in the
claim “Steve Jobs was a great leader”, are likely to elicit
metaphors in response, such as “yes, a true pioneer” or
“what an artist!”, or even “but he could be such a tyrant!”.
Proper-names can also be used as evocative stereotypes, as
when Steve Jobs is compared to the fictional inventor Tony
Stark, or Apple is compared to Scientology, or Google to
Microsoft. We use stereotypes effortlessly, and their
exploitations are common currency in everyday language.
Information retrieval, however, is a language-driven
application where the currency of metaphor has little or no
exchange value, not least because IR fails to discriminate
literal from non-literal language (Veale 2004, 2011, 2012).
Speakers use metaphor to provide and elicit information in
Proceedings of the Fourth International Conference on Computational Creativity 2013 16
casual conversation, but IR reduces any metaphoric query
to literal keywords and key-phrases, which are matched
near-identically in texts (Salton, 1968; Van Rijsbergen
1979). Yet everyday language shows that metaphor is an
ideal form for expressing our information needs. A query
like “Steve Jobs as a good leader” can be viewed by an IR
system as a request to consider all the ways in which
leaders are stereotypically good, and to then consider all
the metaphors that are typically used to convey these
viewpoints. The IR staple of query expansion (Vernimb,
1977; Vorhees, 1994,1998; Navigli & Velardi, 2003; Xu &
Croft, 1996) can be made both affect-driven and metaphoraware.
In this paper we show how an affective stereotypebased
lexicon can both comprehend and generate affective
metaphors that capture or shape a user’s feelings, and show
how this capability can lead to more creative forms of IR.
Related Work and Ideas
Metaphor has been studied within computer science for
four decades, yet it remains at the periphery of NLP
research. The reasons for this marginalization are, for the
most part, pragmatic ones, since metaphors can be as
varied and challenging as human creativity will allow. The
greatest success has been achieved by focusing on
conventional metaphors (e.g., Martin, 1990; Mason, 2004),
or on very specific domains of usage, such as figurative
descriptions of mental states (e.g., Barden, 2006).
From the earliest computational forays, it has been
recognized that metaphor is essentially a problem of
knowledge representation. Semantic representations are
typically designed for well-behaved mappings of words to
meanings – what Hanks (2006) calls norms – but metaphor
requires a system of soft preferences rather than hard (and
brittle) constraints. Wilks (1978) thus proposed his
preference semantics model, which Fass (1991,1997)
extended into a collative semantics. In contrast, Way
(1990) argues that metaphor requires a dynamic concept
hierarchy that can stretch to meet the norm-bending
demands of figurative ideation, though her approach lacks
computational substance.
More recently, some success has been obtained with
statistical approaches that side-step the problems of
knowledge representation, by working instead with implied
or latent representations that are derived from word
distributions. Turney and Littman (2005) show how a
statistical model of relational similarity can be constructed
from web texts for retrieving the correct answer to
proportional analogies, of the kind used in SAT tests. No
hand-coded knowledge is employed, yet Turney and
Littman’s system achieves an average human grade on a
set of 376 real SAT analogies.
Shutova (2010) annotates verbal metaphors in corpora
(such as “to stir excitement”, where “stir” is used
metaphorically) with the corresponding conceptual
metaphors identified by Lakoff and Johnson (1980).
Statistical clustering techniques are then used to generalize
from the annotated exemplars, allowing the system to
recognize and retrieve other metaphors in the same vein
(e.g. “he swallowed his anger”). These clusters can also be
analyzed to identify literal paraphrases for a metaphor
(such as “to provoke excitement” or “suppress anger”).
Shutova’s approach is noteworthy for operating with
Lakoff & Johnson’s inventory of conceptual metaphors
without using an explicit knowledge representation.
Hanks (2006) argues that metaphors exploit
distributional norms: to understand a metaphor, one must
first recognize the norm that is exploited. Common norms
in language are the preferred semantic arguments of verbs,
as well as idioms, clichés and other multi-word
expressions. Veale and Hao (2007a) suggest that
stereotypes are conceptual norms that are found in many
figurative expressions, and note that stereotypes and
similes enjoy a symbiotic relationship that has some
obvious computational advantages. Similes use stereotypes
to illustrate the qualities ascribed to a topic, while
stereotypes are often promulgated via proverbial similes
(Taylor, 1954). Veale and Hao (2007a) show how
stereotypical knowledge can be acquired by harvesting
“Hearst” patterns of the form “as P as C” (e.g. “as smooth
as silk”) from the web (Hearst, 1992). They show in
(2007b) how this body of stereotypes can be used in a webbased
model of metaphor generation and comprehension.
Veale (2011) employs stereotypes as the basis of a new
creative information retrieval paradigm, by introducing a
variety of non-literal wildcards in the vein of Mihalcea
(2002). In this system, @Noun matches any adjective that
denotes a stereotypical property of Noun (so e.g. @knife
matches sharp, cold, etc.) while @Adj matches any noun
for which Adj is stereotypical (e.g. @sharp matches sword,
laser, razor, etc.). In addition, ?Adj matches any property
or behavior that co-occurs with, and reinforces, the
property denoted by Adj; thus, ?hot matches humid, sultry
and spicy. Likewise, ?Noun matches any noun that denotes
a pragmatic neighbor of Noun, where two words are
neighbors if they are seen to be clustered in the same adhoc
set (Hanks, 2005), such as “lawyers and doctors” or
“pirates and thieves”. The knowledge needed for @ is
obtained by mining text from the open web, while that for
? is obtained by mining ad-hoc sets from Google n-grams.
There are a number of shortcomings to this approach.
For one, Veale (2011) does not adequately model the
affective profile of either stereotypes or their properties.
For another, the stereotype lexicon is static, and focuses
primarily on adjectival properties (like sharp and hot). It
thus lacks knowledge of everyday verbal behaviors like
cutting, crying, swaggering, etc. So we build here on the
work of Veale (2011) in several important ways.
First, we enrich and enlarge the stereotype lexicon, to
include more stereotypes and behaviors. We determine an
affective polarity for each property or behavior and for
Proceedings of the Fourth International Conference on Computational Creativity 2013 17
each stereotype, and show how polarized +/- viewpoints on
a topic can be calculated on the fly. We show how proxy
representations for ad-hoc proper-named stereotypes (like
Microsoft) can be constructed on demand. Finally, we
show how metaphors are mined from the Google n-grams,
to allow the system to understand novel metaphors (like
Google is another Microsoft or Apple is a cult) as well as
to generate plausible metaphors for users’ affective
information needs (e.g., Steve Jobs was a great leader,
Google is too powerful, etc.).
Once more, with feeling!
If a property or behavior P is stereotypical of a concept
C, we should expect to frequently observe P in instances of
C. In linguistic terms, we can expect to see collocations of
“P” and “C” in a resource like the Google n-grams (Brants
and Franz, 2006). Consider these 3-grams for “cowboy”
(numbers in parentheses are Google database frequencies).
a lonesome cowboy 432
a mounted cowboy 122
a grizzled cowboy 74
a swaggering cowboy 68
N-gram patterns of the above form allow us to find
frequent ascriptions of a quality to a noun-concept, but
frequently observed qualities are not always noteworthy
qualities (e.g., see Almuhareb and Poesio, 2004,2005).
However, if we also observe these qualities in similes –
such as "swaggering like a cowboy” or “as grizzled as a
cowboy” – this suggests that speakers see these as typical
enough to anchor a figurative comparison. So for each
hypothesis P is stereotypical of C that we derive from the
Google n-grams, we generate the corresponding simile
form: we use the “like” form for verbal behaviors such as
“swaggering”, and the “as-as” form for adjectival
properties such as “lonesome”. We then dispatch each
simile as a phrasal query to Google: a hypothesis is
validated if the corresponding simile is found on the web.
This mining process gives us over 200,000 validated
hypotheses for our stereotype lexicon. We now filter these
hypotheses manually, to ensure that the contents of the
lexicon are of the highest quality (investing just weeks of
labor produces a very reliable resource; see Veale 2012 for
more detail). We obtain rich descriptions for commonplace
ideas, such as the dense descriptor Baby, whose 163 highly
salient qualities – a set denoted typical(Baby) – includes
crying, drooling and guileless. After this manual phase, the
stereotype lexicon maps 9,479 stereotypes to a set of 7,898
properties / behaviors, to yield more than 75,000 pairings.
Determining Nuanced Affect
To understand the affective uses of a property or behavior,
we employ the intuition that those which reinforce each
other in a single description (e.g. “as lush and green as a
jungle” or “as hot and humid as a sauna”) are more likely
to have the same affect than those which do not. To
construct a support graph of mutually reinforcing
properties, we gather all Google 3-grams in which a pair of
stereotypical properties or behaviors X and Y are linked
via coordination, as in “hot and spicy” or “kicking and
screaming”. A bidirectional link between X and Y is added
to the graph if one or more stereotypes in the lexicon
contain both X and Y. If this is not so, we consider whether
both descriptors ever reinforce each other in web similes,
by posing the web query “as X and Y as”. If this query has
a non-zero hit set, we still add a link between X and Y.
Next, we build a reference set -R of typically negative
words, and a disjoint set +R of typically positive words.
Given a few seed members for -R (such as sad, evil,
monster, etc.) and a few seed members for +R (such as
happy, wonderful, hero, etc.), we use the ? operator of
Veale (2011) to successively expand this set by suggesting
neighboring words of the same affect (e.g., “sad and
pathetic”, “happy and healthy”). After three iterations in
this fashion, we populate +R and -R with approx. 2000
words each. If we can anchor enough nodes in the graph
with + or – labels, we can interpolate a nuanced positive /
negative score for all nodes in the graph. Let N(p) denote
the set of neighboring terms to a property or behavior p in
the support graph. Now, we define:
 (1) N+(p) = N(p) ∩ +R
 (2) N-(p) = N(p) ∩ -R
We assign positive / negative affect scores to p as follows:
 (3) pos(p) = |N+(p)|
|N+(p) ∪ N-(p)|
 (4) neg(p) = 1 - pos(p)
Thus, pos(p) estimates the probability that p is used in a
positive context, while neg(p) estimates the probability that
p is used in a negative context. The X and Y 3-grams
approximate these contexts for us.
Now, if a term S denotes a stereotypical idea that is
described in the lexicon with the set of typical properties
and behaviors denoted typical(S), then:
 (5) pos(S) = Σp∈typical(S)
pos(p)
 |typical(S)|
 (6) neg(S) = 1 - pos(S)
So we simply calculate the mean affect of the properties
and behaviors of s, as represented in the lexicon via
typical(s). Note that (5) and (6) are simply gross defaults.
Proceedings of the Fourth International Conference on Computational Creativity 2013 18
One can always use (3) and (4) to separate the elements of
typical(s) into those which are more negative than positive
(a negative spin on s) and those which are more positive
than negative (a positive spin on s). Thus, we define:
 (7) posTypical(S) = {p ∈ typical(S) | pos(p) > neg(p)}
 (8) negTypical(S) = {p ∈ typical(S) | neg(p) > pos(p)}
For instance, the positive stereotype of Baby contains the
qualities such as smiling, adorable and cute, while the
negative stereotype contains qualities such as crying,
wailing and sniveling. As we’ll see next, this ability to
affectively “spin” a stereotype is key to automatically
generating affective metaphors on demand.
Generating Affective Metaphors, N-gram style
The Google n-grams is also a rich source of copula
metaphors of the form Target is Source, such as
“politicians are crooks”, “Apple is a cult”, “racism is a
disease” and “Steve Jobs is a god”. Let src(T) denote the
set of stereotypes that are commonly used to describe T,
where commonality is defined as the presence of the
corresponding copula metaphor in the Google n-grams. To
also find metaphors for proper-named entities like “Bill
Gates”, we analyse n-grams of the form stereotype First
[Middle] Last, such as “tyrant Adolf Hitler”. For example:
src(racism) = {problem, disease, joke, sin, poison,
crime, ideology, weapon}
src(Hitler) = {monster, criminal, tyrant, idiot, madman,
vegetarian, racist, …}
We do not try to discriminate literal from non-literal
assertions, nor indeed do we try to define literality at all.
Rather, we assume each putative metaphor offers a
potentially useful perspective on a topic T.
Let srcTypical(T) denote the aggregation of all
properties ascribable to T via metaphors in src(T):
 (9) srcTypical (T) = M∈src(T)
typical(M)
We can also use the posTypical and negTypical variants of
(7) and (8) to focus only on metaphors that place a positive
or negatve spin on a topic T. In effect, (9) provides a
feature representation for topic T as viewed through the
creative lens of metaphor. This is useful when the source S
in the metaphor T is S is not a stereotype in the lexicon, as
happens when one describes Rasputin as Karl Rove, or
Apple as Scientology. When the set typical(S) is empty,
srcTypical(S) may not be, so srcTypical(S) can act as a
proxy representation for S in these cases.
The properties and behaviors that are salient to the
interpretation of T is S are given by:
(10) salient (T,S) = [srcTypical(T) ∪ typical(T)]
 ∩
 [srcTypical(S) ∪ typical(S)]
In the context of T is S, the metaphorical stereotype M ∈
src(S)∪src(T)∪{S} is an apt vehicle for T if:
(11) apt(M, T,S) = |salient(T,S) ∩ typical(M)| > 0
and the degree to which M is apt for T is given by:
(12) aptness(M,T,S) = |salient(T, S) ∩ typical(M)|
 |typical(M)|
We can now construct an interpretation for T is S by
considering the stereotypes in src(T) that are apt for T in
the context of T is S, and by also considering the
stereotypes that are commonly used to describe S that are
also potentially apt for T:
(13) interpretation(T, S)
 = {M ∈ src(S)∪src(T)∪{S} | apt(M, T, S)}
In effect, the interpretation of the creative metaphor T is S
is itself a set of more conventional metaphors that are apt
for T and which expand upon S. The elements {Mi
} of
interpretation(T, S) can be sorted by aptness(Mi T,S) to
produce a ranked list of interpretations (M1 … Mn). For a
given interpretation M, the salient features of M are thus:
(14) salient(M, T,S) = typical(M) ∩ salient (T,S)
So if T is S is a creative IR query – to find documents in
which T is viewed as S – then interpretation(T, S) is an
expansion of T is S that includes the common metaphors
that are consistent with T viewed as S. In turn, for any
viewpoint Mi in interpretation(T, S), then salient(Mi
, T, S)
is an expansion of Mi that includes all of the qualities that
T is likely to exhibit when it behaves like Mi
.
A Worked Example: Metaphor Generation for IR
Consider the creative query “Google is Microsoft”, which
expresses a user’s need to find documents in which Google
exhibits qualities typically associated with Microsoft. Now,
both Google and Microsoft are complex concepts, so there
are many ways in which they can be considered similar or
dissimilar, whether in a good or a bad light. However, we
can expect the most salient aspects of Microsoft to be those
that underpin our common metaphors for Microsoft, i.e.,
the stereotypes in src(Microsoft). These metaphors will
provide the talking points for an interpretation.
The Google n-grams yield up the following metaphors,
57 for Microsoft and 50 for Google:
∪
Proceedings of the Fourth International Conference on Computational Creativity 2013 19
 src(Microsoft) = {king, master, threat, bully, giant,
leader, monopoly, dinosaur …}
 src(Google) = {king, engine, threat, brand, giant,
leader, celebrity, religion …}
So the following qualities are aggregrated for each:
 srcTypical(Microsoft) = {trusted, menacing, ruling,
threatening, overbearing,
admired, commanding, …}
 srcTypical(Google) = {trusted, admired, reigning,
lurking, crowned, shining,
ruling, determined, …}
Now, the salient qualities highlighted by the metaphor,
namely salient(Google, Microsoft), are:
{celebrated, menacing, trusted, challenging, established,
threatening, admired, respected, …}
Finally, interpretation(Google,Microsoft) contains:
{king, criminal, master, leader, bully, threatening, giant,
threat, monopoly, pioneer, dinosaur, …}
Let’s focus on the expansion “Google is king”, since
according to (12), aptness(king, Google, Microsoft) = 0.48
and this is the highest ranked element of the interpretation.
Now, salient(king, Google, Microsoft) contains:
{celebrated, revered, admired, respected, ruling,
arrogant, commanding, overbearing, reigning, …}
Note that these properties / behaviours are already implicit
in our consensus perception of Google, insofar as they are
highly salient aspects of the stereotypical concepts to
which Google is frequently compared on the web. These
properties / behaviours can now be used to perform query
expansion for the query term “Google”, to find documents
where the system believes Google is acting like Microsoft.
The metaphor “Google is Microsoft” is diffuse and
lacks an affective stance. So let’s consider instead the
metaphor “Google is -Microsoft”, where - is used to
impart a negative spin (and where + can likewise impart a
positive spin). In this case, negTypical is used in place of
typical in (9) and (10), so that:
srcTypical(-Microsoft) =
{menacing, threatening, twisted, raging, feared,
sinister, lurking, domineering, overbearing, …}
and
salient(Google, -Microsoft) =
{menacing, bullying, roaring, dreaded…}
Now, interpretation(Google, -Microsoft) becomes:
{criminal, giant, threat, bully, evil, victim, devil, …}
In contrast, interpretation(Google, +Microsoft) is:
{king, master, leader, pioneer, classic, partner, …}
More focus is achieved with this query in the form of a
simile: “Google is as -powerful as Microsoft”. For explicit
similes, we need to focus on just a sub-set of salient
properties, as in this varient of (10):
{p ∈ salient(Google, Microsoft) | p ∈ N-(powerful)}
In this case, the final interpretation becomes:
{bully, threat, giant, devil, monopoly, dinosaur, …}
A few simple concepts can thus yield a wide range of
options for the creative IR user who is willing to build
queries around affective metaphors and similes.
Empirical Evaluation
The affective stereotype lexicon is the cornerstone of the
current approach, and must reliably assign meaningful
polarity scores both to properties and to the stereotypes
that exemplify them. Our affect model is simple in that it
relies principally on +/- affect, but as demonstrated above,
users can articulate their own expressive moods to suit
their needs: for Stereotypical example, one can express
disdain for too much power with the term -powerful, or
express admiration for guile with +cunning and +devious.
The Effect of Affect: Stereotypes and Properties
Note that the polarity scores assigned to a property p in (3)
and (4) do not rely on any prior classification of p, such as
whether p is in +R or -R. That is, +R and -R are not used
as training data, and (3) and (4) receive no error feedback.
Of course, we expect that for p ∈ +R that pos(p) > neg(p),
and for p ∈ -R that neg(p) > pos(p), but (3) and (4) do not
iterate until this is so. Measuring the extent to which these
simple intuitions are validated thus offers a good
evaluation of our graph-based affect mechanism.
Just five properties in +R (approx. 0.4% of the 1,314
properties in +R) are given a positivity of less than 0.5
using (3), leading those words to be misclassified as more
negative than positive. The misclassified property words
are: evanescent, giggling, licking, devotional and fraternal.
Just twenty-six properties in -R (approx. 1.9% of the
1,385 properties in -R) are assigned a negativity of less
than 0.5 via (4), leading these to be misclassified as more
positive than negative. The misclassified words are: cocky,
dense, demanding, urgent, acute, unavoidable, critical,
startling, gaudy, decadent, biting, controversial, peculiar,
disinterested, strict, visceral, feared, opinionated,
humbling, subdued, impetuous, shooting, acerbic,
heartrending, ineluctable and groveling.
Proceedings of the Fourth International Conference on Computational Creativity 2013 20
Because +R and -R have been populated with words
that have been chosen for their perceived +/- slants, this
result is hardly surprising. Nonetheless, it does validate the
key intuitions that underpin (3) and (4) – that the affective
polarity of a property p can be reliably estimated as a
simple function of the affect of the co-descriptors with
which it is most commonly used in descriptive contexts.
The sets +R and -R are populated with adjectives, verbal
behaviors and nouns. +R contains 478 nouns denoting
positive stereotypes (such as saint and hero) while -R
contains 677 nouns denoting negative stereotypes (such as
tyrant and monster). When these reference stereotypes are
used to test the effectiveness of (5) and (6) – and thus,
indirectly, of (3) and (4) and of the stereotype lexicon itself
– 96.7% of the positive stereotype exemplars are correctly
assigned a mean positivity of more than 0.5 (so, pos(S) >
neg(S)) and 96.2% of the negative exemplars are correctly
assigned a mean negativity of more than 0.5 (so, neg(S) >
pos(S)). Though it may seem crude to assess the affect of a
stereotype as the mean of the affect of its properties, this
does appear to be a reliable measure of polar affect.
The Representational Adequacy of Metaphors
We have argued that metaphors can provide a collective
representation of a concept that has no other representation
in a system. But how good a proxy is src(S) or
srcTypical(S) for an S like Karl Rove or Microsoft? Can we
reliably estimate the +/- polarity of S as a function of
src(S)? We can estimate these from metaphors as follows:
(15) pos(S) = ΣM∈src(S)
pos(M)
 |src(S)|
(16) neg(S) = ΣM∈src(S)
neg(M)
 |src(S)|
Testing this estimator on the exemplar stereotypes in +R
and -R, the correct polarity (+ or -) is estimated 87.2% of
the time. Metaphors in the Google n-grams are thus
broadly consistent with our perceptions of whether a topic
is positively or negatively slanted.
When we consider all stereotypes S for which |src(S)| >
0 (there are 6,904 in the lexicon), srcTypical(S) covers, on
average, just 65.7% of the typical properties of S (that is,
of typical(S)). Nonetheless, this shortfall is precisely why
we use novel metaphors. Consider this variant of (9) which
captures the longer reach of these novel metaphors:
 (17) srcTypical2
(T) =
S ∈ src(T)
srcTypical(S)
Thus, srcTypical2
(T) denotes the set of qualities that are
ascribable to T via the expansive interpretation of all
metaphors T is S in the Google n-grams, since S can now
project onto T any element of srcTypical(S). Using macroaveraging
over all 6,904 cases where |src(S)| > 0, we find
that srcTypical2
(S) covers 99.2% of typical(S) on average.
A well-chosen metaphor enables us to emphasize almost
any quality of a topic T we might wish to highlight.
Affective Text Retrieval with Creative Metaphors
Suppose we have a database of texts {D1 … Dn} in which
each document Di offers a creative perspective on a topic
T. We might have texts that view politicians as crooks,
popes as kings, or hackers as heroes. So given a query +T,
can we retrieve only those texts that view T positively, and
given -T can we retrieve only the negative texts about T?
We first construct a database of artificial figurative
texts. For each stereotype S in the lexicon, and for each M
∈ src(S)∩(+R∪-R), we construct a text DSM in which S is
viewed as M. The title of document DSM is “S is M”,
while the body of DSM contains all the words in src(M).
DSM uses the typical language of M to talk about S. For
each DSM, we know whether DSM conveys a positive or
negative viewpoint on S, since M sits in either in +R or -R.
The affect lexicon contains 5,704 stereotypes S for
which src(S)∩(+R∪-R) is non-empty. On average, each of
these stereotypes is described in terms of 14 other
stereotypes (5.8 are negative and 8.2 are positive,
according to +R and -R) and we construct a representative
document for each of these viewpoints. We construct a set
of 79,856 artificial documents in total, to convey figurative
perspectives on 5,704 different stereotypical topics:
Table 1. Macro-Average P/R/F1 scores for affective retrieval of
+ and - viewpoints for 5,704 topics.
Macro Average
(5704 topics)
Positive
viewpoints
Negative
viewpoints
Precision .86 .93
Recall .95 .78
F-Score .90 .85
For each document retrieved for T, we estimate its polarity
as the mean of the polarity of the words it contains. Table 1
presents the results of this experiment, in which we attempt
to retrieve only the positive viewpoints for T with a query
+T, and only the negative viewpoints for T using -T. The
results are sufficiently encouraging to support the further
development of a creative text retrieval engine that is
capable of ranking documents by the affective figurative
perspective that they offer on a topic.
∪
Proceedings of the Fourth International Conference on Computational Creativity 2013 21
Concluding Thoughts: The Creative Web
Metaphor is a creative knowledge multiplier that allows us
to expand our knowledge of a topic T by using knowledge
of other ideas as a magnifying lens. We have presented
here a robust, stereotype-driven approach that embodies
this practical philosophy. Knowledge multiplication is
achieved using an expansionary approach, in which an
affective query is expanded to include all of the metaphors
that are commonly used to convey this affective viewpoint.
These viewpoints are expanded in turn to include all the
qualities that are typically implied by each. Such an
approach is ideally suited to a creative re-imagining of IR.
An implementation of these ideas is available for use
on the web. Named Metaphor Magnet, the system allows
users to enter queries of the form shown here (such as
Google is –Microsoft, Steve Jobs as Tony Stark, Rasputin
as Karl Rove, etc.). Each query is expanded into a set of
apt metaphors mined from the Google n-grams, and each
metaphor is expanded into a set of contextually apt
qualities. In turn, each quality is expanded into an IR query
that is used to retrieve relevant hits from Google. In effect,
the system – still an early prototype – allows users to
interface with a search engine like Google using metaphor
and other affective language forms. The system can
currently be accessed at this URL:
 http://boundinanutshell.com/metaphor-magnet
Metaphor Magnet is just one possible application of the
ideas presented here, which constitute not so much a
philosophical or linguistic theory of metaphor, but an
engineering-oriented toolkit of reusable concepts for
imbuing a wide range of text applications with a robust
competence in linguistic creativity. Human speakers do not
view metaphor as a problem but as a solution. It is time our
computational systems took a similarly constructive view
of this remarkably creative cognitive tool.
In this vein, Metaphor Magnet continues to evolve as a
creative web service. In addition to providing metaphors
on demand, the service now also provides a poetic framing
facility, whereby the space of possible interpretations for a
given metaphor is crystallized into a single poetic form.
More generally, poetry can be viewed as a means of
reducing information overload, by summarizing a complex
metaphor – or the set of texts retrieved using that metaphor
via creative IR – whose interpretation entails a rich space
of affective possibilities. A poem can thus be seen in
functional terms as both an information summarization tool
and as a visualization device. Metaphor Magnet adopts a
simple, meaning-driven approach to poetry generation:
given a topic T, a set of candidate metaphors with the
desired affective slant is generated. One metaphor is
chosen at random, and the elements of its interpretation are
sampled to produce different lines of the resulting poem.
Each element, and the sentiment it best evokes, is rendered
in natural language using one of a variety of poetic tropes.
For example, Metaphor Magnet produces the following
as a distillation of the space of feelings and associations
that arise from the interpretation of Marriage is a Prison:
The legalized regime of this marriage
My marriage is a tight prison
The most unitary federation scarcely organizes so much
Intimidate me with the official regulation of your prison
Let your close confines excite me
O Marriage, you disgust me with your undesirable security
Each time we dip into the space of possible interpretations,
a new poem is produced. One can use Metaphor Magnet to
sample the space at will, hopping from one interpretation
to the next, or from one poem to another. Here is an
alternate rendition of the same metaphor in poetic form:
The official slavery of this marriage
My marriage is a legitimate prison
No collective is more unitary, or organizes so much
Intimidate me with the official regulation of your prison
Let your sexual degradation charm me
O Marriage, you depress me with your dreary consecration
In the context of our earlier worked example, which
generated a space of metaphors to negatively describe
Microsoft’s perceived misuse of power, consider the
following, which distills the assertion Microsoft is a
Monopoly into an aggressive ode:
No Monopoly Is More Ruthless
Intimidate me with your imposing hegemony
No crime family is more badly organized,
or controls more ruthlessly
Haunt me with your centralized organization
Let your privileged security support me
O Microsoft, you oppress me with your corrupt reign
Poetry generation in Metaphor Magnet is a recent addition
to the service, and its workings are beyond the scope of the
current paper (though they may be observed in practice by
visiting the aforementioned URL). For details of a related
approach to poetry generation – one that also uses the
stereotype-bearing similes described in Veale (2012) – the
reader is invited to read Colton, Goodwin & Veale (2012).
Metaphor Magnet forms a key element in our vision of a
Creative Web, in which web services conveniently provide
creativity on tap to any third-party software application
that requests it. These services include ideation (e.g. via
metaphor generation & knowledge discovery), composition
(e.g. via analogy, bisocation & conceptual blending) and
framing (via poetry generation, joke & story generation,
etc.). Since CC does not distinguish itself through distinct
algorithms or representations, but through its unique goals
Proceedings of the Fourth International Conference on Computational Creativity 2013 22
and philosophy, such a pooling of services will not only
help the field achieve a much-needed critical mass, it will
facilitate a greater penetration of CC ideas and approaches
into the commercial software industry.
Acknowledgements
This research was supported by the WCU (World Class
University) program under the National Research
Foundation of Korea (Ministry of Education, Science and
Technology of Korea, Project no. R31-30007).
<references_biblio/>
References
Almuhareb, A. and Poesio, M. 2004. Attribute-Based and ValueBased
Clustering: An Evaluation. In Proc. of EMNLP 2004.
Barcelona.
Almuhareb, A. and Poesio, M. 2005. Concept Learning and
Categorization from the Web. In Proc. of the 27th Annual
meeting of the Cognitive Science Society.
Barnden, J. A. 2006. Artificial Intelligence, figurative language
and cognitive linguistics. In: G. Kristiansen, M. Achard, R.
Dirven, and F. J. Ruiz de Mendoza Ibanez (Eds.), Cognitive
Linguistics: Current Application and Future Perspectives,
431-459. Berlin: Mouton de Gruyter.
Brants, T. and Franz, A. 2006. Web 1T 5-gram Ver. 1. Linguistic
Data Consortium.
Brennan, S. E. and Clark, H. H. 1996. Conceptual Pacts and
Lexical Choice in Conversation. Journal of Experimental
Psychology: Learning, Memory and Cognition, 22(6):1482-93.
Colton, S., Goodwin, J. and Veale, T. 2012. Full-FACE Poetry
Generation.In Proc. of ICCC 2012, the 3rd International
Conference on Computational Creativity. Dublin, Ireland.
Fass, D. 1991. Met*: a method for discriminating metonymy and
metaphor by computer. Computational Linguistics 17(1):49-90.
Fass, D. 1997. Processing Metonymy and Metaphor.
Contemporary Studies in Cognitive Science & Technology.
New York: Ablex.
Fishelov, D. 1992. Poetic and Non-Poetic Simile: Structure,
Semantics, Rhetoric. Poetics Today, 14(1), 1-23.
Hanks, P. 2005. Similes and Sets: The English Preposition ‘like’.
In: Blatná, R. and Petkevic, V. (Eds.), Languages and
Linguistics: Festschrift for Fr. Cermak. Charles Univ., Prague.
Hanks, P. 2006. Metaphoricity is gradable. In: Anatol
Stefanowitsch and Stefan Th. Gries (Eds.), Corpus-Based
Approaches to Metaphor and Metonymy,. 17-35. Berlin:
Mouton de Gruyter.
Hearst, M. 1992. Automatic acquisition of hyponyms from large
text corpora. In Proc. of the 14th Int. Conf. on Computational
Linguistics, pp 539–545.
Martin, J. H. 1990. A Computational Model of Metaphor
Interpretation. New York: Academic Press.
Mason, Z. J. 2004. CorMet: A Computational, Corpus-Based
Conventional Metaphor Extraction System. Computational
Linguistics, 30(1):23-44.
Mihalcea, R. 2002. The Semantic Wildcard. In Proc. of the
LREC Workshop on Creating and Using Semantics for
Information Retrieval and Filtering. Spain, May 2002.
Navigli, R. and Velardi, P. 2003. An Analysis of Ontology-based
Query Expansion Strategies. Proc. of the workshop on
Adaptive Text Extraction and Mining (ATEM 2003), at
ECML, the 14th European Conf. on Machine Learning, 42–49
Salton, G. 1968. Automatic Information Organization and
Retrieval. New York: McGraw-Hill.
Shutova, E. 2010. Metaphor Identification Using Verb and Noun
Clustering. In the Proc. of the 23rd International Conference
on Computational Linguistics, 1001-1010.
Taylor, A. 1954. Proverbial Comparisons and Similes from
California. Folklore Studies 3. Berkeley: University of
California Press.
Turney, P.D. and Littman, M.L. 2005. Corpus-based learning of
analogies and semantic relations. Machine Learning 60(1-
3):251-278.
Van Rijsbergen, C. J. 1979. Information Retrieval. Oxford:
Butterworth-Heinemann.
Veale, T. 2004. The Challenge of Creative Information Retrieval.
Computational Linguistics and Intelligent Text Processing:
Lecture Notes in Computer Science, Vol. 2945/2004, 457-467.
Veale, T. and Hao, Y. 2007a. Making Lexical Ontologies
Functional and Context-Sensitive. In Proc. of the 46th Annual
Meeting of the Assoc. of Computational Linguistics.
Veale, T. and Hao, Y. 2007b. Comprehending and Generating
Apt Metaphors: A Web-driven, Case-based Approach to
Figurative Language. In Proc. of AAAI 2007, the 22nd AAAI
Conference on Artificial Intelligence. Vancouver, Canada.
Veale, T. and Hao, Y. 2008. Talking Points in Metaphor: A
concise, usage-based representation for figurative processing.
In Proceedings of ECAI’2008, the 18th European Conference
on Artificial Intelligence. Patras, Greece, July 2008.
Veale. T. 2011. Creative Language Retrieval: A Robust Hybrid of
Information Retrieval and Linguistic Creativity. In Proc. of
ACL’2011, the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Technologies.
Veale, T. 2012. Exploding the Creativity Myth: The
Computational Foundations of Linguistic Creativity. London:
Bloomsbury Academic.
Vernimb, C. 1977. Automatic Query Adjustment in Document
Retrieval. Information Processing & Mgmt, 13(6):339-53.
Voorhees, E. M. 1994. Query Expansion Using Lexical-Semantic
Relations. In the proc. of SIGIR 94, the 17th International
Conference on Research and Development in Information
Retrieval. Berlin: Springer-Verlag, 61-69.
Voorhees, E. M. 1998. Using WordNet for text retrieval.
WordNet, An electronic lexical database, 285–303. MIT Press.
Way, E. C. 1991. Knowledge Representation and Metaphor.
Studies in Cognitive systems. Holland: Kluwer.
Wilks, Y. 1978. Making Preferences More Active, Artificial
Intelligence 11.
Xu, J. and Croft, B. W. 1996. Query expansion using local and
global document analysis. In Proc. of the 19th annual
international ACM SIGIR conference on Research and
development in information retrieval.
Proceedings of the Fourth International Conference on Computational Creativity 2013 23