“In reality there are as many religions as there are papers”
– First Steps Towards the Generation of Internet Memes
Diogo Costa, Hugo Gonc¸alo Oliveira, Alexandre Miguel Pinto
CISUC, Department of Informatics Engineering
University of Coimbra, Portugal
dcosta@student.dei.uc.pt, hroliv@dei.uc.pt, ampinto@dei.uc.pt
Abstract
We report on the first steps towards the automatic generation
of Internet memes starring public figures. Their images are
retrieved from the Web and combined with famous quotes,
altered according to recent information on the figures. Current
implementation, in Portuguese, exploits several computational
resources and aims to produce artifacts with coherent
text, image, and some humor value. A preliminary evaluation
survey confirmed a strong relation between generated memes
and present events. Results on humor were also positive.
Introduction
The term meme originally denotes an idea, behavior, or
style that spreads from person to person within a culture
(Dawkins, 1976; Blackmore, 2000). On the Internet domain,
memes became a popular and effective way of transmitting
an idea. They are a product of human creativity that
typically take the form of an image, often combined with a
short phrase. They tend to be funny, make people laugh, and
aim to be spread throughout the World Wide Web by sharing
and re-sharing in social media.
We present the first steps towards the development of
MemeGera, a system for the automatic generation of Internet
memes – or better, protomemes1 – starring public figures
(hereafter, characters). MemeGera uses famous quotes, altered
as follows: one word is replaced by another that is
semantically related to the character and its current information2.
These sentences, presented together with a character’s
image, should convey a simple and effective idea,
make sense for the character, even if only for a short period
of time after generation, and exhibit some novelty. To
deal with the latter, the system exploits fresh information on
the character, such as that in recent news or tweets. The
produced text+image combinations have thus a transient flavor
which, together with their humor potential, may qualify
them as “jokes du jour”. Long-term knowledge on the character,
from its Wikipedia page, is also explored, but so far
only used to favor fresh information.
1
The definition of meme implies social sharing, which will only
occur if people actually spread the protomeme. 2
As the title of the paper, a twist of Mahatma Gandhi’s quote:
“In reality there are as many religions as there are individuals”.
We see the generation of meme sentences as a kind of linguistic
creativity, a topic that covers tasks such as the generation
of: poetry (Toivanen, Gross, and Toivonen, 2014;
Gonc¸alo Oliveira and Cardoso, 2015); metaphors (Veale and
Hao, 2008); neologisms (Smith, Hintze, and Ventura, 2014);
or verbally-expressed humor (Binsted and Ritchie, 1994;
Valitutti et al., 2013). Given the funny aspect inherent to
memes, our work is probably closer to the latter. Yet, although
not essential, the character and its image also play
an important role in the success of our memes.
In the remaining of this paper, we provide some background
knowledge on the study of humor, together with
computational approaches to this topic. We then present the
automatic method for generating memes and list each of the
steps involved. The current implementation targeted Portuguese,
our native language, and is described right after. Although
the method may seem quite straightforward, our effort
involves the combination of several knowledge sources.
Before concluding, we describe an illustrative example and
report on the results of an online survey, which suggests that
we are heading in the right direction. All the memes used in
the survey are shown in the end of the paper, together with
information about their generation and evaluation.
Background and Related Work
This section addresses the topic of humor from a theoretical
point of view, followed by an enumeration of computational
approaches for humor generation and recognition.
Theoretical Study of Humor
Humor has been studied from a variety of perspectives ranging
from psychology and philosophy (Morreall (2013)), to
its sociological aspects in literature (Kuipers (2010)) and,
more recently, via the computational approach (e.g. Suslov
(1992); Ritchie (2014)). Theoretical accounts of humor
encompass the superiority theory, endorsed by Descartes,
where “our laughter expresses feelings of superiority over
other people or over a former state of ourselves”; the relief
theory, a hydraulic model proposed by Shaftesbury, and later
refined by Sigmund Freud, according to which laughter acts
as a mechanism for releasing accumulated nervous energy
built up from many possible emotionally-charged situations;
and the incongruity theory, proposed by Beattie and sponsored
by Kant, Schopenhauer, and Kierkegaard, among othProceedings
of the Sixth International Conference on Computational Creativity June 2015 300
ers, which claims “laughter is the perception of something
incongruous – something that violates our mental patterns
and expectations”, which is now the dominant theory.
Socioliterary studies (e.g. Kuipers (2010)) explore the
mechanisms through which humor is related to social
boundaries, and how it differs between groups; whereas
computational approaches address the building of formal
theories of humor (Ritchie (2014)), the synthesis of a sense
of humour via specific algorithms (Suslov (1992)), and the
generation of humorous text and jokes (Ritchie (2009)).
Humor expressed in Portuguese has also been studied
from a theoretical point of view. While presenting linguistic
mechanisms for achieving humor in this language, Tagnin
(2005) states that, since humor breaks conventionality in
language, understanding it is a sign of fluency.
Humor generation
The automatic generation of humor has been a research topic
for more than two decades. In early work by Binsted and
Ritchie (1994), a model, implemented under the name of
JAPE, was proposed for generating punning riddles. The
generated puns (e.g. What do you call a murderer that has
fiber? A cereal killer) took advantage of spelling or word
sense ambiguities. STANDUP (Manurung et al., 2008) follows
the lines of JAPE, but is more robust, user friendly, and
was developed with the purpose of allowing young children,
especially those with linguistic disabilities, to explore language
and improve their skills.
Given a concept and an attribute, HAHAcronym (Stock
and Strapparava, 2005) rewrites existing acronyms and generates
new ones with a humor intent. It relies on an incongruity
detector and generator that selects opposing domains
and opposing adjectives, while considering also rhythm and
rhymes. For instance, the acronym FBI may become Fantastic
Bureau of Intimidation. Or given the concept of ‘processor’
and the attribute ‘fast’, it generates the acronym OPEN
– Online Processor for Effervescent Net.
Valitutti et al. (2013) explored the generation of adult humor
based on the replacement of a word in a short message.
The word should introduce incongruity and lead to a humorous
interpretation, achieved by applying three constraints. It
must: (i) be of the same form as the original word, i.e. match
the part-of-speech and either rhyme or be orthographically
similar to the original word; (ii) convey a taboo meaning,
e.g. an insult or something related to sex; (iii) take place at
the end of the message and keep the coherence of the original
sentence. An example of an output is: I’ve sent you my
fart.. I mean ‘part’ not ‘fart’....
Besides English, there were attempts for generating puns
in Japanese (e.g. Sjobergh and Araki (2007a)). We are not ¨
aware of any work of this kind for Portuguese.
Humor recognition
In the scope of natural understanding, there has been work
on the automatic recognition of verbally-expressed humor.
Researchers typically focus on a specific kind of jokes, such
as knock-knock (Taylor and Mazlack, 2004) and That’s what
she said (Kiddon and Brun, 2011), or on a less specific kind
of humor but transmitted in bounded kinds of text, such as
single sentences (Mihalcea and Strapparava, 2006; Sjobergh ¨
and Araki, 2007b), or tweets (Barbieri and Saggion, 2014).
Humor recognition is generally seen as a text classification
problem and relies on a set of humor relevant features to
train a classifier, given their presence in humorous and nonhumorous
text. For instance, Barbieri and Saggion (2014)
exploit hashtags, such as #humuor or #irony, to collect positive
examples. Selected features generally include the occurrence
of antonymous or ambiguous words, alliteration, and
other words or expressions typically used in jokes, such as
slang or idiomatic expressions.
For Portuguese, the closest works to humor recognition
we are aware of include the automatic detection of
irony (Carvalho et al., 2009) or proverbs (Rassi, Baptista,
and Vale, 2014) in text.
Internet Memes
Internet memes are a current trend in social media. They
are typically a reusable combination of text and graphics.
Popular memes include Boromir from the Lord of the Rings
with the template “One does not simply X”, Morpheus from
the Matrix with “What if I told you Y”, or Batman slapping
Robin, with a personalized text in their speech balloons.
There is however a subtype of Internet memes related
to current events, where new images, text, or both, can be
used – if successful enough, they might be reused. Events
that triggered several memes include the football player Luis
Suarez biting his opponent in a World Cup 2014 match (e.g. ´
“If you can’t beat them, eat them”), or when the pop singer
Madonna fell on stage, while wearing a cape, during a performance
in the BritAwards 2015 ceremony (e.g. “56 years
old, still does her own stunts”, “Has a cape, can’t fly”).
While most memes show a break of conventionality (e.g.
unexpected situation, confusing interpretation, taboo meaning),
we address the previous subtype, which, as suggested
by the superiority theory, makes fun of the portrayed character.
In fact, the image is sometimes enough to make people
laugh (e.g. when it displays a funny person or situation).
We are not aware of any published work on the automatic
generation of Internet memes. Existing web services for
meme generation rely on the user input of both images and
text. There is work however on the automatic combination
of images and text, such as Grafik Dynamo (2005) and Why
Some Dolls Are Bad (2008), by Kate Armstrong3. In those
projects, a narrative is dynamically generated by combining
sequences of images, retrieved from social networks, with
speech balloons. The result is often non-sense.
Method
This section provides a high-level description of our proposed
method for meme generation. Specific details of its
current implementation are given in the next section.
Among other parameters, our algorithm for the generation
of memes (see figure 1) uses the name of a public figure,
our character, currently provided by the user. Informally,
it starts by retrieving n recent messages (e.g. tweets) mentioning
the character, from where the top-k frequent nouns
3
http://katearmstrong.com/
Proceedings of the Sixth International Conference on Computational Creativity June 2015 301
are collected. Then, it selects a random quote from a pool
of famous quotes, pairs it with one of the top-k nouns, and
generates a sentence, more precisely, an altered quote where
the last noun of the original quote is replaced by one of the
top-k nouns – similarly to Valitutti et al. (2013), replacing
the last noun will increase surprise and humor potential. After
repeating this process for a predefined number of times,
generated sentences are ranked by a dedicated scoring function.
The highest-ranked sentence is pasted on an image of
the character, automatically retrieved from the Web, and the
combination is finally returned as the generated meme.
The scoring function considers the humor value of the
sentence, the frequency of the replacement noun, and its
presence in a more stable long-term information source on
the character. Words without previous associations to the
character are considered novel and are thus favored in the
ranking. We may find some parallelism between this and
the work of Toivanen, Gross, and Toivonen (2014), where
novel associations in documents are identified by their overlap
with known associations from a background corpus.
Require:
charName:name of character
n : # of messages to retrieve
k : # of top frequent common nouns to consider
m : # of < quote,frequent noun> pairs to generate
1: procedure MEMEGERA
2: messages {msg : msg mentions charName}
: #messages = n
3: freqNouns top-k most frequent nouns in messages
4: quotes {quote : quote is a famous quote}
5: pairs {< quote, freqNoun > randomly generated}
: #pairs = m
6: maxEval 0
7: bestQuote ;
8: for each < quote, freqNoun >2 pairs do
9: nq replace last noun in quote with freqNoun
10: ne score(nq, freqNoun, charN ame)
11: if ne > maxEval then
12: maxEval ne
13: bestQuote nq
14: image get image of charName from the Web
15: resultingMeme paste bestQuote in image
16: return resultingMeme
Figure 1: Meme generation algorithm
Implementation
Although our method is language-independent, its current
implementation targets Portuguese. MemeGera was implemented
in Java and exploits several available resources, for
different purposes, including a classifier for Portuguese humor,
currently in development. We also describe the function
that currently ranks the generated sentences.
Tools and Resources
Famous quotes used in this work were acquired from the
Portuguese edition of Wikiquote4, a collaborative repository
of quotes, run by the Wikimedia Foundation. For the current
4
http://pt.wikiquote.org/
version of the system, we selected quotes from three wellknown
thinkers – Mahatma Gandhi, Aristotle and Confucius
– who were the authors of many quotes, most of them timeless
and generic enough for our purpose. We soon realized
that long quotes would not produce the desired effect, so we
only used quotes with up to 15 words, totaling 90.
We use the social network Twitter5 and Twitter4J6, a Java
API, to retrieve tweets mentioning the names of the selected
characters. While we could have used a news site or aggregator,
the choice of Twitter relied on the fact that its messages
are shorter, up-to-date, and mix different and less controlled
opinions. In recent years, Twitter has been widely exploited
by computer programs, not only for text mining, but
also in computational creativity research (e.g. Veale (2014);
Cook, Colton, and Gow (2014) or the recent PROSECCO
Code Camp7, focused on the development of creative Twitterbots).
Natural language processing is made by the OpenNLP
toolkit8 and its models trained for Portuguese tokenization
and part-of-speech tagging. Since the models were not
trained with tweets, a few annotation errors are expected.
But this is not severe because we end up using only words
in a morphological lexicon, LABEL-Lex9, in which we rely
to perform inflection, so that words agree with the sentence
they are put in. Also, we count the lemmas frequency in
the tweets, and not the words frequency. Lemmatization is
performed by LemPort (Rodrigues, Gonc¸alo Oliveira, and
Gomes, 2014), a Portuguese lemmatizer.
Nouns long-associated to famous people were collected
from the abstracts of their articles in the Portuguese
Wikipedia, retrieved directly from the DBPedia10 entries under
the category of Person.
Images of the meme characters are retrieved automatically
from Google Images11, at runtime. The first hit for
each character is always used.
The Mallet12 toolkit was used in the development of a humor
classifier for Portuguese, presented in the next section.
Given a positive and a negative dataset, Mallet automatically
converts input text to features, and learns a classifier, using
one of the algorithms available out-of-the-box.
Humor Classifier
We have recently started to work on a classifier for recognizing
humorous pieces of text, in Portuguese, currently
trained with the Mallet toolkit. The first step for its development
was the collection of examples of humorous and
non-humorous Portuguese documents, labeled respectively
as positive or negative. The selected datasets were then imported
to Mallet, which was used to train a classifier with the
5
https://twitter.com/
6
http://twitter4j.org/
7
http://codecampcc.dei.uc.pt/
8
https://opennlp.apache.org/
9
http://label.ist.utl.pt/pt/labellex_pt.php
10http://dbpedia.org/
11https://images.google.com
12http://mallet.cs.umass.edu/
Proceedings of the Sixth International Conference on Computational Creativity June 2015 302
best available learning algorithm. Instead of labeling the examples
manually, we collected them from selected sources
which we now present.
Positive Dataset: While it is rather easy to collect negative
examples, the same does not apply for humorous examples
in Portuguese. After searching in the Web, we were able to
find the following compilations of Portuguese jokes:
• B´ıblia de Anedotas13 (in English Bible of Jokes);
• O Sagrado Caderno das Piadas Secas14 (in English, The
Sacred Book of Dry Jokes).
To focus on shorter jokes, we discarded all with more than
25 words, and were left with 790 positive examples.
Negative Dataset: The non-humorous dataset should contain
text with a similar structure to the positive examples
but without a potential humor effect. We thus collected sentences
of similar length (25 words) from non-humorous
sources. Since many of the collected jokes have a questionanswer
structure, we included this kind of text as well. The
following resources were used:
• 304,211 sentences from the Portuguese Wikipedia, each
collected randomly from a different article;
• Text from Portuguese corpora available through the
AC/DC project (Santos and Bick, 2000)15:
– 81,478 sentences from CETEMPublico, a corpus with
editions of the Portuguese newspaper Publico (1991- ´
1998).
– 25,000 sentences from CONDIVport, a corpus of
sports newspapers, fashion and health magazines;
– 6,767 question-answer pairs from Museu Da Pessoa, a
corpus of interviews.
In the end, we had a total of 417,456 negative examples.
Validation: After importing the positive and negative
datasets, a classifier was trained with the Maximum Entropy
algorithm, selected after a 10-fold cross-validation, where it
yielded 99.8% accuracy. These numbers look promising, but
they were computed in a dataset with mostly negative examples.
Although the F1 for the negative class was 99.9%, it
was just 63.7% for the positive, with a recall of 49.4%.
We should stress that the classifier is still in an early stage
of development. In the future, instead of relying only in the
black-box text classification of Mallet, additional features
should be integrated, including a subset of those used by others
(Sjobergh and Araki, 2007b; Mihalcea and Strapparava, ¨
2006). Moreover, we are aware that we cannot expect much
of the current classifier, at least for the kind of sentences we
are generating. While it was trained with classic and timeless
jokes, understanding the generated sentences requires
not only general world knowledge, but further information
that may be valid only on a specific moment in time.
13http://rbep.cm-porto.pt/rbep/upload/
dnloads/BibliadeAnedotas.doc 14https://www.facebook.com/CadernoDasPiadas
15http://linguateca.pt/ACDC/
Figure 2: Meme generated for the pop singer Madonna. The
text translates to Keep your thoughts positive, because your
thoughts become your falls.
Ranking function
As referred earlier, MemeGera generates a set of m sentences
that combine a known quote with a noun f retrieved
from Twitter. Towards the selection of the most promising
generated sentences, these are currently ranked by the following
linear combination:
Score = humorP rob ⇤ ↵
+ wordF requency ⇤ "
+ notInW ikipedia ⇤ #
There, humorProb is the probability returned by the humor
classifier; wordFrequency is the number of tweets
where f occurs, divided by the total number of retrieved
tweets, n; and notInWikipedia is a binary function that is 1
if the word is not in the Wikipedia abstract of the character,
or 0 otherwise.
Results
We generated several memes with different configurations.
Although not enough experiments were performed to select
the best configuration, at a certain point, we started to
use fixed parameters, to have a base for comparison. In all
reported experiments, generation was based on 200 tweets
(n = 200), written in Portuguese (according to Twitter), and
using the top-5 frequent nouns (k = 5). The best sentence
was selected from a set of 20 (m = 20). The ranking function
used the weights: ↵ = 0.7, " = 0.25, # = 0.05.
Example
Figure 2 illustrates the output of MemeGera with a meme
generated on the 26th February 2015, the day after Madonna
fell on stage. The original quote, attributed to Mahatma
Gandhi, was Keep your thoughts positive, because your
thoughts become your words.
Since people were talking about the fall, the most frequent
nouns in tweets were: tombo (tumble), queda (downfall),
palco (stage), v´ıdeo (video) and madonno. The last one results
from an incorrect part-of-speech tag given to the proper
noun Madonna. There is no risk of using it though, because
Proceedings of the Sixth International Conference on Computational Creativity June 2015 303
Figure 3: Overall results of the performed evaluation.
it is not in the morphological lexicon. None of the top-5
nouns were in Madonna’s Wikipedia abstract.
In the end of this paper, we present more memes, together
with information that will help understanding why they were
generated, and their scores according to the online survey
where they were used.
Evaluation survey
In order to have a first appreciation of the results produced
by MemeGera, we made an online survey, answered by 41
human subjects, all Portuguese native speakers. The survey
had the title “Imagens com texto” (Images with text) and it
never mentioned the word meme, nor automatic generation.
The survey had 5 memes, for which the name of the
character was presented together with three questions, to be
answered according to a Likert scale: strongly agree (5),
partially agree (4), neutral (3), partially disagree (2) and
strongly disagree (1). The questions were:
1. The text is syntactically and semantically coherent (Does
it follow the grammar rules and makes sense?).
2. There is coherence in the combination of text, image and
the present time (We suggest to search for breaking news
about the character).
3. The combination of the text and the image produce a humorous
effect (Did it make you smile?).
The used memes were generated between the 25th February
2015 night and 26th morning. Their characters were
manually selected for being mentioned in fresh news in online
media. All of these memes used the highest-ranked sentence
from the 20 generated. They are presented in the end
of the paper, together with information on their generation,
as well as their individual scores in the survey questions.
The survey opened just a few hours after generating the
last meme, and was opened for about 24 hours. This means
that some memes would only be interpreted appropriately by
someone following the daily news. Figure 3 presents overall
results, which combine the answers to the five memes.
The survey confirmed that it is often safe to replace one
word in a sentence by another of the same part-of-speech.
If inflection is handled properly, syntax remains coherent,
which makes it easier for semantics, especially when using
generic quotes. Answers on the coherence between text, image
and the present time are also positive. The meme in
figure 6 was the one with more negative answers in the first
two questions. First, possibly because it is not very easy to
find semantic connections between tumble and awake. Second,
because this meme was related to a very recent event
and, although we suggested the subjects to search for the
character in the news, most of them probably did not do it,
and were not aware of Madonna’s fall.
As for the humor aspect, while we cannot say that the generated
memes are very funny and have the ability to make
everybody laugh, the overall results are encouraging, as the
majority of the answers are positive. It is always subjective
to assess the presence of humor, especially in this case,
where world knowledge and following recent news was a requirement.
A curious fact is that the memes with clearly positive
answers in this aspect are those with Portuguese politicians.
Given that all our subjects were Portuguese, they are
probably better informed about Portuguese characters, who
probably play a more relevant role on the subjects lives, and
make them more responsive to laugh at. This is related to
another issue: the image itself or, sometimes, just the character,
might play an important role in the humor value, since
there are people for which we are more prone to laugh at
than others.
Concluding remarks
We have presented the first steps towards the development
of MemeGera, a system that generates combinations of text
and image that may be seen as Internet memes. Famous
quotes are altered according to a public figure and complemented
with their image, automatically retrieved from the
Web. Fresh information on the public figure, in the form of
frequent words, is currently obtained from Twitter. Several
altered quotes are generated and the best is selected after a
ranking that considers the humor value and the novelty of
retrieved words, in an attempt to positively discriminate the
most promising sentences. The humor value is given by an
automatic classifier, trained with positive and negative examples
of humor expressed in Portuguese. However, this
tool is still far from what we expect from it and gave very
low scores to the generated sentences (rarely more than 1%).
On the other hand, we should stress that MemeGera has
the ability of generating a different and novel sentence each
time, based on fresh news. In fact, the results of an online
survey showed that it is not only capable of generating coherent
sentences, with some relation to the character, but that
the generated combinations have some humor potential.
The work described in this paper lead to the development
of the @memegera Twitterbot that, from time to time:
(i) reads the list of current trends in the Portuguese Twitter;
(ii) checks if any of them is the name of a known person; if
so, (iii) generates a meme on that person and posts it. The
bot is still in a test phase, but we may soon start relying on
users feedback (e.g. retweets, favored) for evaluation and
adaptation of the weights in the ranking function.
Additional plans include both improvements to the system
and to its evaluation. In the scope of this and other projects,
the humor classifier shall be improved by: (i) enriching the
datasets with humorous text from the Twitter accounts of famous
Portuguese humorists; (ii) considering additional features
(e.g. ambiguity of words and adult slang, for which
there are available Portuguese resources we could use). To
increase variation, we will devise adding more quotes to our
pool, as long as they are not too specific. Regarding evaluation,
in a further survey, we aim at recording the reaction of
Proceedings of the Sixth International Conference on Computational Creativity June 2015 304
the subjects in the moment when the meme is first presented
to them, and draw conclusions from their facial expressions.
Acknowledgements
The project ConCreTe acknowledges the financial support
of the Future and Emerging Technologies (FET) programme
within the 7th Framework Programme for Research of the
European Commission, under FET grant number 611733.
<references_biblio/>
References
Barbieri, F., and Saggion, H. 2014. Automatic detection of
irony and humour in twitter. In Procs. of 5th Intl. Conf.
on Computational Creativity (ICCC).
Binsted, K., and Ritchie, G. 1994. An implemented model of
punning riddles. In Procs. of 12th National Conf. on Artificial
Intelligence (Vol. 1), AAAI ’94, 633–638. Menlo
Park, CA, USA: AAAI Press.
Blackmore, S. 2000. The Meme Machine. Oxford University
Press, USA.
Carvalho, P.; Sarmento, L.; Silva, M. J.; and de Oliveira,
E. 2009. Clues for detecting irony in user-generated contents:
Oh...!! it’s ”so easy” ;-). In Procs. of 1st Intl. International
CIKM Workshop on Topic-sentiment Analysis
for Mass Opinion, TSA ’09, 53–56. ACM.
Cook, M.; Colton, S.; and Gow, J. 2014. Automating game
design in three dimensions. In Procs. of AISB Symposium
on AI and Games.
Dawkins, R. 1976. The Selfish Gene. Oxford University
Press, Oxford, UK.
Gonc¸alo Oliveira, H., and Cardoso, A. 2015. Poetry generation
with PoeTryMe. In Besold, T. R.; Schorlemmer, M.;
and Smaill, A., eds., Computational Creativity Research:
Towards Creative Machines, Atlantis Thinking Machines.
Atlantis-Springer. chapter 12, 243–266.
Kiddon, C., and Brun, Y. 2011. That’s what she said: Double
entendre identification. In Procs. of 49th Annual Meeting
of the Association for Computational Linguistics, 89–94.
Portland, OR, USA: ACL Press.
Kuipers, G. 2010. Humor styles and symbolic boundaries.
Journal of Literary Theory 3:219–239.
Manurung, R.; Ritchie, G.; Pain, H.; Waller, A.; O’Mara,
D.; and Black, R. 2008. The construction of a pun generator
for language skills development. Applied Artificial
Intelligence 22(9):841–869.
Mihalcea, R., and Strapparava, C. 2006. Learning to laugh
(automatically): Computational models for humor recognition.
Computational Intelligence 22(2):126–142.
Morreall, J. 2013. Philosophy of humor. In Zalta, E. N., ed.,
The Stanford Encyclopedia of Philosophy. Spring 2013
edition.
Rassi, A. P.; Baptista, J.; and Vale, O. 2014. Automatic
detection of proverbs and their variants. In Procs. of 3rd
Symposium on Languages, Applications and Technologies
(SLATE 2014), Braganc¸a, Portugal, OASICS, 235–249.
Schloss Dagstuhl.
Ritchie, G. 2009. Can computers create humor? AI Magazine
30(3):71–81.
Ritchie, G. 2014. Logic and reasoning in jokes. European
Journal of Humour Research 2(1):50–60.
Rodrigues, R.; Gonc¸alo Oliveira, H.; and Gomes, P.
2014. LemPORT: a high-accuracy cross-platform lemmatizer
for portuguese. In Procs. of 3rd Symposium
on Languages, Applications and Technologies (SLATE
2014), Braganc¸a, Portugal, OASICS, 267–274. Schloss
Dagstuhl.
Santos, D., and Bick, E. 2000. Providing Internet access
to Portuguese corpora: the AC/DC project. In Procs. of
2nd Intl. Conf. on Language Resources and Evaluation,
LREC 2000, 205–210.
Sjobergh, J., and Araki, K. 2007a. Automatically creating ¨
word-play jokes in japanese. In Procs. of NL-178, 91–95.
Sjobergh, J., and Araki, K. 2007b. Recognizing humor with- ¨
out recognizing meaning. In Procs. of Applications of
Fuzzy Sets Theory: 7th Intl. Workshop on Fuzzy Logic and
Applications (WILF 2007) Camogli, Italy, volume 4578
of LNCS, 469–476. Springer.
Smith, M. R.; Hintze, R. S.; and Ventura, D. 2014. Nehovah:
A neologism creator nomen ipsum. In Procs. of 5th Intl.
Conf. on Computational Creativity (ICCC).
Stock, O., and Strapparava, C. 2005. The act of creating
humorous acronyms. Applied Artificial Intelligence
19(2):137–151.
Suslov, I. M. 1992. Computer model of a ”sense of humour”.
i. general algorithm. Biophysics 37(2):242–248.
Tagnin, S. E. O. 2005. O humor como quebra da convencionalidade.
Revista Brasileira de Lingu¨´ıstica Aplicada
5(1):247–257.
Taylor, J. M., and Mazlack, L. J. 2004. Computationally
recognizing wordplay in jokes. In Procs. of Cognitive Science
Conference (CogSci), 2166–2171.
Toivanen, J.; Gross, O.; and Toivonen, H. 2014. The officer
is taller than you, who race yourself! In Procs. of 5th Intl.
Conf. on Computational Creativity (ICCC).
Valitutti, A.; Toivonen, H.; Doucet, A.; and Toivanen, J. M.
2013. ”Let everything turn well in your wife”: Generation
of adult humor using lexical constraints. In Procs.
of 51st Annual Meeting of the Association for Computational
Linguistics, volume 2, 243–248. Sofia, Bulgaria:
ACL Press.
Veale, T., and Hao, Y. 2008. A fluid knowledge representation
for understanding and generating creative metaphors.
In Procs. of 22nd Intl. Conf. on Computational Linguistics,
volume 1 of COLING ’08, 945–952. ACL Press.
Veale, T. 2014. Coming good and breaking bad: Generating
transformative character arcs for use in compelling
stories. In Procs. of 5th Intl. Conf. on Computational Creativity
(ICCC).
Proceedings of the Sixth International Conference on Computational Creativity June 2015 305
Character Paulo Portas (Portuguese Deputy Prime-Minister)
Wikipedia words jurista, jornalista, conservador, vice-primeiroministro,
arquitecto, economista, irmao, dirigente, ˜
empresario, ... ´
(lawyer, journalist, conservative, deputy prime minister, architect,
economist, brother, leader, businessman, ...)
Top-5 nouns marc¸o (29), trajetoria ´ (23), alterac¸ao˜ (22), d´ıvida
(16), do (14)
(march, trajectory, change, debt, of)
Context He had been talking to the media about the downward
trend of the Portuguese public debt, which should start
on March.
Original quote Se nos n ´ ao entendemos a vida, como poderemos en- ˜
tender a morte?
(If we cannot understand life, how can we understand death?)
Generated quote Se nos n ´ ao entendemos a vida, como poderemos en- ˜
tender a d´ıvida?
(If we cannot understand life, how can we understand the debt?)
Humor probability 0.00253
Frequency 0.08
In Wikipedia No
Score 0.07177
Figure 4: Meme of the Portuguese Deputy Prime-Minister, Paulo Portas.
Character Yanis Varoufakis (Greek Finance Minister)
Wikipedia words Not in the used version of Wikipedia
Top-5 nouns salario ´ (37), governo (29), funcionario ´ (22),
manutenc¸ao˜ (18), ministro (17)
(salary, government, federal employee, maintenance, minister)
Context Since he became a member of the Greek government,
he has been all over the news regarding the negotiations
between Greece and the Eurogroup. This time,
he was negotiating the maintenance of the current
number of federal employees and their salaries.
Original quote Na busca da virtude, nao temas superar teu professor. ˜
(In the pursuit of virtue, do not fear overcoming your teacher.)
Generated quote Na busca da virtude, nao temas superar teu governo. ˜
(In the pursuit of virtue, do not fear overcoming your government.)
Humor probability 0.00102
Frequency 0.145
In Wikipedia No
Score 0.08696
Figure 5: Meme of the Greek Finance Minister, Yanis Varoufakis.
Character Madonna (pop singer)
Wikipedia words cantor, compositor, atriz, danc¸arino, empresario, pro- ´
dutor, album ... ´
(singer, songwriter, actress, dancer, businessman, producer, album ...)
Top-5 nouns tombo (46), queda (21), madonno (29), palco (21),
v´ıdeo (13)
(tumble, downfall, madonno, stage, video)
Context A few hours before this generation, she had fell on
stage, during a live performance.
Original quote A esperanc¸a e um sonho acordado. ´
(Hope is a waking dream.)
Generated quote A esperanc¸a e um tombo acordado. ´
(Hope is a waking tumble.)
Humor probability 0.00231
Frequency 0.23
In Wikipedia No
Score 0.10911
Figure 6: Meme of the pop singer Madonna.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 306
Character Vladimir Putin (Russian President)
Wikipedia words presidente, ex-agente, chefe, servic¸o, primeiroministro,
governo, pa´ıs, ...
(president, former agent, chief, service, prime-minister, government,
country, ...)
Top-5 nouns passo (88), risco (36), pa´ıs (36), intervenc¸ao˜ (34),
vitoria ´ (33)
(step, risk, country, intervention, victory)
Context He had been on the news due to his role in the
Ukrainian crisis. While there is not an international
intervention, people try to figure out his next step.
Original quote Em todas as coisas, o sucesso depende de preparac¸ao˜
previa. ´
(In all things, success depends on a previous preparation.)
Generated quote Em todas as coisas, o sucesso depende de intervenc¸ao˜
previa. ´
(In all things, success depends on a previous intervention.)
Humor probability 7.8 ⇥ 10!4
Frequency 0.17
In Wikipedia No
Score 0.09305
Figure 7: Meme of the Russian President, Vladimir Putin.
Character Jose S´ ocrates (former Portuguese Prime Minister) ´
Wikipedia words pol´ıtico, secretario-geral, minist ´ erio, ordena- ´
mento, mestrado ..
(politician, secretary general, ministry, planning, masters, ...)
Top-5 nouns prisao˜ (93), mesˆ (54), ex-premieˆ (41), retour (34),
novembro (33)
(jail, month, ex-Prime Minister, comeback, november)
Context He has been detained into custody for being a suspect
in a corruption case. This week he had got out
for a few hours to give his testimony in a court.
Original quote Minha vida e minha mensagem. ´
(My life is my message.)
Generated quote Minha vida e minha pris ´ ao. ˜
(My life is my jail.)
Humor probability 0.00975
Frequency 0.465
In Wikipedia No
Score 0.17308
Figure 8: Meme of the former Portuguese Prime-Minister, Jose S´ ocrates. ´
Proceedings of the Sixth International Conference on Computational Creativity June 2015 307