Available via license: CC BY 4.0
Content may be subject to copyright.
When a Computer Cracks a Joke: Automated Generation of Humorous Headlines
Khalid Alnajjar
Faculty of Arts
University of Helsinki
khalid.alnajjar@helsinki.fi
Mika H¨
am¨
al¨
ainen
Faculty of Arts
University of Helsinki
mika.hamalainen@helsinki.fi
Abstract
Automated news generation has become a major in-
terest for new agencies in the past. Oftentimes head-
lines for such automatically generated news articles are
unimaginative as they have been generated with ready-
made templates. We present a computationally creative
approach for headline generation that can generate hu-
morous versions of existing headlines. We evaluate our
system with human judges and compare the results to
human authored humorous titles. The headlines pro-
duced by the system are considered funny 36% of the
time by human evaluators.
Introduction
Humor, while showcased by a wide spectrum of species
in the animal kingdom, has a subcategory that is exclusive
to the human being. Verbal humor can only exist in the
presence of language, and its generation by computational
means is far from trivial.
Humor is effectively a perceiver dependent phenomenon.
Nothing can be inherently funny, but humor is perceived and
appraised by the mind of a human perceiving it. And ul-
timately, the perceived humor, if accepted as such by the
listener, elicits an emotional response accompanied with a
vocal response of laughter that originates from our ances-
tors, the species before homo sapiens (cf. Ross, Owren, and
Zimmermann 2010).
Our paper focuses on generating humor in news head-
lines. This NLG task is not of the traditional sort, where
conveying factual information is the uttermost goal of the
system, but rather the affective content of the message is
taken into the primary focus of the study.
Automated news generation is a flourishing field with new
research being published in a timely manner. This is re-
flected by the number of recent publications on the topic
(Nesterenko 2016; Yao et al. 2017). Quite often, however,
as the generated news has to cater for the purpose of com-
municating facts the question of creativity is set aside. In
such a context, there is a trade-off between creativity and
communicativity (see H¨
am¨
al¨
ainen and Honkela 2019).
Whereas creative headline generation is not a new do-
main to computational creativity, with quite some existing
publications on the topic (Lynch 2015; Gatti et al. 2015;
Alnajjar, Lepp¨
anen, and Toivonen 2019), we aim to inter-
twine headlines, creativity and humor by proposing a novel
method for humor generation that is reasoned by the existing
theories on humor.
Our approach alters a word in an existing headline for a
humorous effect. We evaluate the method proposed in this
paper quantitatively with human judges. We take the differ-
ent constituents of humor in consideration in the evaluation
to uncover the relation of each feature to the humor produced
by our system.
Related Work
Humor has received some interest in the past for more than
a decade (Ritchie 2005; Hong and Ong 2009; Valitutti et al.
2013; Costa, Oliveira, and Pinto 2015). We dedicate the re-
maining of this section to describing some of the most recent
work conducted on the topic.
Pun generation with a neural model language model is
one of the most recent efforts on humor generation (Yu,
Tan, and Wan 2018). Their approach consists of training
a conditional language model an using a beam search to find
sentences that can support two polysemous meanings for a
given word. In addition they train a model to highlight the
different meanings of the word in the sentence. Unfortu-
nately, they evaluate their system on human evaluators based
on three quantitative metrics: fluency, accuracy and read-
ability, none of which tells anything about how funny or apt
the puns were.
Alnajjar and H¨
am¨
al¨
ainen (2018) present a genetic algo-
rithm approach for generating humorous and satirical movie
titles out of existing ones. Their method works on a word
level replacement and aims for low semantic similarity of
the replacement word with the original word to maximize
surprise and high similarity with Saudi Arabia to maximize
coherence. They consider pun as one of the fitness func-
tions of the genetic algorithm, but the output is not strictly
limited to puns. On top the genetic algorithm, they train an
RNN model that learns from the genetic algorithm and real
people.
Surprise is also one of the key aspects of a recent pun
generator (He, Peng, and Liang 2019). They model surprise
as conditional probabilities. They introduce a local surprise
model to assess the surprise in the immediate context of the
pun word and a global surprise to assess the surprise in the
arXiv:2109.08702v1 [cs.CL] 17 Sep 2021
context of the whole text. Their approach retrieves text from
a corpus based on an original word - pun word pair. They
do a word replacement for local surprise and insert a topic
word for global surprise.
An approach building on humor theories is that of Win-
ters, Nys, and De Schreye (2019). The theories are used
in feature engineering. They learn templates and metrical
schemata from jokes rated by people with a star rating. They
embrace more traditional machine learning techniques over
neural networks, which has the advantage of a greater inter-
pretability of the models.
Humor has also been tried to recognize automatically in
the past. One of such attempts is focuses on extracting
humor anchors, i.e. words that can make text humorous,
automatically (Yang et al. 2015). A similar humor an-
chor based approach is also embraced by Cattle and Ma
(2018). Both of the approaches rely on feature engineer-
ing basing on humor theories. Recently LSTM models have
been used for the task of humor detection with a different
rates of success (Cai, Li, and Wan 2018; Sane et al. 2019;
Zou and Lu 2019).
Humor
Humor is an inherent part of being a human and as such it
has provoked the interest of many researchers in the past to
formulate a definition for it (see (Krikmann 2006)). Koestler
(1964) sees humor as a part of creativity together with dis-
covery and art. In his view, what is characteristic to humor
in comparison to the other two constituents of creativity, is
that its emotional mood is aggressive in its nature. He calls
bisociation in humor the collision of two frames of reference
in a comic way.
Raskin (1985) presents a theory that is not too far away
from the previously described one in the sense that in order
for text to be humorous, it has to be compatible with two
different scripts. The different scripts have to be somehow
in opposition, for example in the sense that one script is a
real situation and the other is not real.
In Attardo and Raskin (1991) humor is seen to consist
of six hierarchical knowledge resources: language, narrative
strategy, target, situation, logical mechanism and script op-
position. As in the previous theories, the incongruity of two
possible interpretations is seen as an important aspect for
humor. An interesting notion that we will take into a closer
examination is that of target. According to the authors it is
not uncommon for a joke to have a target, such as an im-
portant political person or an ethnic group, to be made fun
of.
Two requirements have been suggested in the past as com-
ponents of humor in jokes: surprise and coherence (see
(Brownell et al. 1983)). A joke will then consist of a sur-
prising element that will need to be coherent in the context
of the joke. This is similar to having two incongruous scripts
being simultaneously possible.
Veale (2004) points out that the theories of Raskin (1985)
and Attardo and Raskin (1991) entail people to be forced
into resolution of humor. He argues that humor should not
be seen as resolution of incompatible scripts, but rather as
a collaboration, where the listener willingly accepts the hu-
morous interpretation of the joke. Moreover, he argues that
while incongruity contributes to humor, it does not alone
constitute it.
Generating Humorous Headlines
In their work, Hossain, Krumm, and Gamon (2019) identi-
fied several ways people altered news headlines to be humor-
ous. In our method, we aim to model the following ones of
their findings: the replacement forms a meaningful n-gram,
the replacements are semantically distant, the replacement
makes a strong connection with the entity of the headline
and belittles an entity or a noun and the replacement creates
incongruity. We see the n-gram finding in a broader way of
the replacement being compatible with the the existing script
(context). The semantic distance is seen as an index of sur-
prise, and the connection between the entity is assimilated
with the target of the joke.
The findings we are not focusing on in this paper are that
the replacements are sarcastic, suppress tension or have a
setup and punchline. The first two are left out as assessing
them computationally is a task worth of a paper on their own
right, and the third one is left out as it focuses on a particu-
lar kind of humor. However, the punchline structure might
emerge from the other features being modelled although not
explicitly taken into consideration.
In addition to the findings described above, we take the
concreteness of the replacement word into account. The rea-
son for this that concrete words are more likely to provoke
mental images (see (Burroway 2007)). In fact, we could see
this in the humorous training dataset by Hossain, Krumm,
and Gamon (2019), where 90% of the most humorous re-
placement words were concrete as opposed to only 75% of
the least humorous replacement words being concrete.
For the above experiment and the rest of the paper, we use
the lexicon of 40k common English words that has a con-
creteness score from 1 to 5 assigned (Brysbaert, Warriner,
and Kuperman 2014). If the score assigned with the word is
greater or equal to 3, we consider it concrete. The concrete-
ness is evaluated by lemmatizing the word with spaCy (Hon-
nibal and Montani 2017) if it does not exist in the lexicon.
Modelling Humor
Our system operates by taking an existing headline from the
corpus of altered headlines (Hossain, Krumm, and Gamon
2019). This corpus has been syntactically parsed by us by
using spaCy (Honnibal and Montani 2017), and it has been
tagged for the words that should be replaced by its original
authors. For a selected headline, our system tries to find
suitable humoristic replacement words.
We assess the different potential humorous replacements
in terms of multiple parameters, which are prosody, con-
creteness, semantic similarity of the replacement to the orig-
inal word and the semantic relatedness of the replacement to
negative words describing the target. In this section, we ex-
plain how the individual parameters are modelled. An over-
all view of our method is depicted in Figure 1.
For prosody, we look at the sound similarity between
the original word and the replacement. We assess this in
Figure 1: A diagram visualizing the process of humor generation.
terms of full rhymes, assonance, consonance and allitera-
tion. These are implemented with rules. As the written form
of English is notoriously deviant from the phonation, we use
eSpeak-ng1to produce IPA transcription for the words the
prosody of which is being assessed.
For concreteness we use the values provided in Brysbaert,
Warriner, and Kuperman (2014) to score the concreteness
of the replacement word. And for semantic similarity we
use the pretrained word embeddings from Bojanowski et al.
(2017). We use the semantic similarity to assess surprise,
in other words, we want to minimize the similarity of the
replacement word to the original.
To measure how a new replacement connects to the word
selected to be the target of the joke in the headline, a target
must first be found. We consider recognized entities in the
headline as the potential targets. In case no entities were
recognized, we use the subjects in the headline. If neither of
them existed, nouns in the headline are treated as target. Out
of the list of targets, a random target tis picked to focus on.
For this target t, we retrieve words that are related to it to act
as descriptive words revealing potential attributes to make
fun of. We employ two resources to obtain such knowledge
regarding the selected target:
1. The Non-Official Characterization (NOC) list (Veale
2016) which contains information about more than 1000
well-known characters (e.g. Donald Trump and Kim
Jung-un) and their expanded stereotypical properties sup-
plied by (Alnajjar et al. 2017) (e.g. Donald Trump:
[wealthy, successful, greedy, aggressive, . . .etc]).
2. A semantic relatedness model built from word associa-
tions collected from a web text corpus ukWac2, follow-
ing the approach described in Meta4meaning (Xiao et al.
1UK English voice, https://github.com/espeak-ng/espeak-ng
2https://wacky.sslmit.unibo.it/doku.php?
2016). We chose to base our relatedness model on a web-
based corpus instead of a news-based one to favor discov-
ering related words from various domains, which would
be perceived as more humorous.
If the target tis an entity, we search the first resources (i.e.
the NOC list and the expanded properties) to collect its top
kstereotypical properties. In case no available knowledge
regarding the entity existed, we attempt to acquire the top
krelated words to the rest of the potential targets (subjects
and nouns, respectively) using the second resource (i.e. the
semantic relatedness model). In our case, we empirically set
kto 100 to allow diversity and reduce noisy relations, while
ensuring the descriptiveness of the words to the target.
To be able to place the target in a humorous light, we
only regard the descriptive words that describe it negatively,
which is determined by employing a polarity classifier pro-
vided by Akbik, Blythe, and Vollgraf (2018). Lastly, the
connection of the replacement word to the target is assessed
based on the semantic relatedness between the replacement
word and the target’s negative descriptions. We desire to
maximize such connections to encourage replacements that
are associated with the target from a negative angle.
Generation and picking out the best candidate
We use the Humicroedit dataset of headlines published by
Hossain, Krumm, and Gamon (2019) as the source of orig-
inal headlines. Furthermore, the dataset contains edits per-
formed by humans to make the headlines humorous along
with a score indicating how humorous they were when per-
ceived by other people on a scale from 0 to 3. The motiva-
tion for using this dataset is that the editors were required to
make a single change to either a verb or a noun in the head-
id=corpora
reads book 187
VBZ NN Frequency
dobj
Figure 2: An example of a syntactical relation in the reposi-
tory of grammatical relations (Alnajjar 2018) along with its
frequency.
line to make it humorous, which focuses the scope when
modeling such a process computationally.
In our generation method, we only consider headlines
where the original word that is selected to be replaced is
parsed as either a noun or a verb using spaCy and is a sin-
gle token (i.e. ignoring cases such as “Illegal Immigrants”).
The rational behind is to reduce misparsing errors and con-
centrate on a single-word changes.
For an original headline howith its selected word to be
replace wo, our method converts it into a humorous one hh
by replacing wowith another word whas follows. It begins
by acquiring replacement candidates Cthat fit the syntacti-
cal position of the selected word woby querying a massive
syntactical repository of grammatical relations that have a
frequency greater than 50 in a web-based corpus (Alnajjar
2018) (see Figure 2 for an example of a grammatical rela-
tion in the repository). By considering candidates that are
apt to the existing syntactical relations in the headline, we
ensure that the new replacement has syntactic cohesion and
suits the grammatical context.
To illustrate how the method works, let’s consider the
headline ho=“City halls and landmarks turn green in sup-
port of Paris climate deal” as an example, where the word to
replace wo=climate. After parsing this headline, we find
that the to-be-replaced word wois a noun (NN) and has a
dependency (compound) on the word deal (NN). We query
the syntactical repertory to find potential replacements that
suit this relation, which yields 58 candidate replacements
(C={‘loan’, ‘business’, ‘cash’, ‘oil’, ‘holiday’, ‘peace’,
‘content’, ‘drug’ . . . etc}).
In the next phase, the method removes the original word
wofrom the candidates if it existed and prunes out any candi-
date word in Cthat is not identified as concrete (i.e. having
a concreteness score greater or equal to 3 based on (Brys-
baert, Warriner, and Kuperman 2014)). As a result, candi-
date words such as ‘peace’ and ‘content’ in the earlier ex-
ample are removed resulting in a total of 34 candidates. If
there is more than 500 replacement candidates (e.g. in sit-
uations where the token to replace is a verb and is the root
of the phrase), we randomly select 500 candidates in Cto
be examined. This is performed to reduce the search space
that the method will traverse and to efficiently discover lo-
cal optimal solutions as there is no particular global optimal
solution for the task we are addressing.
Replacement candidates are then evaluated on the four
humour aspects we are modeling, which are 1) prosody, 2)
concreteness score, 3) inverted (i.e. minimized) semantic
similarity between the original word woand the candidate
c, and 4) the semantic relation between the candidate cand
the negative words of the selected target t. As we are deal-
ing with multiple criteria for modeling humour, we adopt
a non-dominant multi-objective sorting approach (Deb et al.
2000) to find and select candidates in the Pareto front. Addi-
tionally, applying a non-dominant sorting for creative tasks
(e.g. generating humour) increases the chances of finding
balanced and diverse solutions that are more likely to be
deemed good (Alnajjar, Hadaytullah, and Toivonen 2018).
Applying the evaluation and the non-dominant sorting
on the example headline, the method highlights candidates
such as ‘cash’, ‘meal’, ‘drug’ to be chosen as replacements.
For the same example, the original word climate was re-
placed with marijuana by a human editor in the Humicroedit
dataset. Interestingly, marijuana is a drug and our method
was able to suggest it.
Results and Evaluation
To evaluate our method, we randomly select 83 headlines
from the Humicroedit dataset that meet our criteria specified
earlier. For each headline, we request our method to produce
humorous alternatives, ranked by the non-dominant sorting,
out of which we randomly select 3 to be evaluated from the
top humorous headlines.
Table 1 shows some of the headlines generated by our
approach. The humorous replacement word is marked in
bold. The original word and the replacement word suggested
by a human from the corpus are shown in their respective
columns.
We conduct our evaluation on Figure-Eight3, which is
a crowd-sourcing platform that assigns paid reviewers for
tasks such as questionnaires. We evaluate all the 3 varia-
tions produced by our system for the 83 headlines, showing
the original headline as well. In addition, we evaluate the
human edits for the same headlines from the dataset. The
reviewers were not told they were evaluating computer gen-
erated humor, as the mere knowledge of a computer being
an author of a creative artefact is known to provoke a bias
towards seeing the generated output in a more negative light
(see (Colton, Wiggins, and others 2012)).
We asked five people to rate the headlines based on the
following questions:
1. The altered headline is humorous.
2. The altered word is surprising.
3. The altered word fits into the headline.
4. The altered word is concrete.
5. The joke of the headline makes fun of a person or a group
of people (also known as the target of the joke).
6. The altered word shows the target in a negative light.
7. The altered word is a pun of the original word.
We evaluate the first two questions on the scale form 0 to
3 (Not funny,Slightly funny,Moderately funny and Funny.
Or surprising in the case of the Q2) similarly to the ques-
tions for humor in Hossain et al. (2017). The rest of the
3https://www.figure-eight.com/
Humorous headline by our system Original word Human replacement
Thieves carry out elaborate van heist to steal millions
in cereal, Swiss police say cash blouses
Trump eats the wrong Lee Greenwood on Twitter tags woos
’I was very angry’ at Trump, says Myeshia Johnson,
widow of fallen sock soldier cake
Trump Tried To Climb Heather Heyer’s Mother
During Funeral: ‘I Have Not And Now I Will Not’
Talk To Him
call proposition
U.S. says Turkey is helping ISIS by Combing
Kurds in Syria bombing feeding
Table 1: Examples of generated headlines.
questions are presented as yes/no questions. The sixth ques-
tion is only visible if the fifth question has been answered to
affirmatively.
Figure 3: Percentages for each evaluation question. H mark-
ing human authored headlines, and S computer authored
ones.
Figure 3 shows the percentages of the result for each ques-
tion from the human evaluation. The results for the human
edited titles (H) and the ones produced by our method (S)
are shown side by side. From the question 3 onward, 0
marks negative and 1 affirmative answer. All in all, our sys-
tem scored slightly lower on the questions than real people,
which is to be expected due to the difficulty of the problem.
However, our system got slightly better results in the ques-
tion number 6, which means, that when the system had rec-
ognizably picked a target, it managed to convey negativity
towards the target on a level comparable to a real human.
In terms of humor, our system managed to produce at least
slightly humorous headlines 36% of the time, whereas peo-
ple produced at least slightly humorous headlines 56% of
the time. In comparison, for a recent pun generator, (He,
Peng, and Liang 2019) report a success rate of 31% for their
system according to a human evaluation, to put our results
in a computational perspective.
Table 2 shows the results form another perspective. The
score row shows the results for human authored titles in the
original publication (Hossain et al. 2017), whereas the hu-
man row shows the results for the very same titles in our
evaluation. The max shows the average of the best scor-
ing generated headline out of the 3 ones produced for each
original headline, and min shows the average of the worst
headline in the triplets. Avg is the average of the scores for
all the generated headlines.
By looking at the results this way, we can see that at best,
our method can produce humor comparable to real humans
in the scale of funniness, with a higher amount of surprise,
better aptness of the replacement word to the context, higher
level of concreteness, higher negativity towards the target
and higher level of puniness, falling shorter only in the case
of having a perceivable target for the joke in the headline.
Focusing on the best scoring individuals might sound like
giving too good a picture of the performance of the system,
however, they set the upper boundary for the performance
of the system. This being said, with the exact same method,
better results could be obtained in the future by developing
a better way for ranking the humorous headline candidates
output by the system.
By considering the headlines produced by our method that
have the maximum score for an original headline, we see
that 47 of them were credited as humorous (i.e. having a
score ≥1) out for the 83 original title. On the other hand,
43 of the human generated were considered humorous.
In the following analysis, we aim to evaluate the different
criteria considered in our method for modeling humour. In
terms of prosody, we look at the number of times a headline
was considered to be punny by people with respect to our
method’s score on the prosody dimension. Overall, 22% of
the generated headlines were considered to have a pun in re-
lation to the original word. Out of these headlines, 88% of
them were evaluated positively on the prosody dimension by
our system. This indicates that the method exhibited capa-
bility of assessing the sound similarity and punniness to the
original word.
For the concreteness, we are considering concrete words
defined in (Brysbaert, Warriner, and Kuperman 2014) as
candidates. As a result, we expected to have headlines pro-
duced by the method score high on the fourth question. Con-
trary, only 56% of them were deemed concrete. This indi-
cates that a more robust model is required to model the con-
creteness of terms.
By observing Figure 3, we notice that 46% of head-
lines suggested by our method are considered surprising (i.e.
scoring at 1, on average). As we are using a word em-
beddings model, it is difficult to come up with a semantic
Q1 Q2 Q3 Q4 Q5 Q6 Q7
µxSD µxS D µxSD µxSD µxSD µxSD µxSD
score 0,99 0,62 - - - - - - - - - - - -
human 0,97 0,49 0,89 0,41 0,58 0,23 0,6 0,2 0,36 0,23 0,61 0,41 0,28 0,18
max 0,97 0,45 0,98 0,36 0,69 0,19 0,7 0,15 0,33 0,17 0,86 0,34 0,35 0,17
avg 0,6 0,28 0,67 0,23 0,53 0,15 0,56 0,11 0,19 0,12 0,6 0,31 0,22 0,11
min 0,28 0,27 0,37 0,24 0,36 0,18 0,4 0,16 0,06 0,12 0,27 0,39 0,1 0,11
Table 2: Mean and standard deviation of altered headlines by humans and our method.
similarity threshold that separates similar words from non-
similar ones, especially for modeling surprisingness. There-
fore, we test the scores assigned by the models on three
thresholds of similarity (0.3, 0.2 and 0.1) with respect to the
headlines viewed as surprising by online people. Out of the
46% surprising headlines, 98%, 84% and 40% headlines are
considered to be dissimilar by the semantic model by using
the three above mentioned thresholds. This indicates that
minimizing the semantic similarity increases surprise to a
degree, after which lowering the similarity results in a lower
surprise.
Lastly, we perform the same analysis regarding the con-
nection between the replacement word and the selected tar-
get with respect to question five and six. 75% of the time,
our function scored positively on headlines evaluated as
making fun of a target. Out of which, 77% were correctly
seen as negative by the method with respect to Q6.
Discussion
As the best headlines produced by our system for each orig-
inal headline can, on the average, reach to a human level in
terms of most of the factors measured by our evaluation, an
immediate future direction for our research is to develop a
better ranking mechanism to reach to the maximum capac-
ity of our system. Perhaps such ranking could be learned
by training an LSTM classifier on humor annotated corpora
such as the one used in this paper or the one proposed by
(West and Horvitz 2019).
For surprise, we opted for a rather modest approach by
assimilating it to an inverse semantic similarity to the orig-
inal word. However, different metrics have been proposed
to model this phenomenon, such as a neural network based
composer-audience model (Bunescu and Uduehi 2019) or
probabilistically modelling the likelihood of a certain word
occurring in a given context (see (Degaetano-Ortlieb and
Piper 2019)).
The particularly low score on the concreteness highlights
the inadequacy of using an annotated lexicon for its assess-
ment. Perhaps, in the future, concreteness could be mod-
elled in a more robust context dependent way. Previous work
(Naumann, Frassinelli, and im Walde 2018) exists show-
ing differences in the distributional representations of con-
crete and abstract words. As word embedding models are
based on the distributional hypothesis, this discovery could
be exploited for a context dependent classification by using
context-aware word embeddings.
If the method was to be used as a tool for assisting journal-
ists when composing news articles, the fact that employing
computational methods for headline generation might result
in offensive headlines (see (Alnajjar, Lepp¨
anen, and Toivo-
nen 2019)) has to be taken into account. Our humor model
maximizes the negative relation to its target, which might
be considered as an insult, if understood in a wrong, non
humorous fashion.
Our current approach focuses on English, in the future,
we are interested in using our method for other languages
as well such as Finnish. This would require a more robust
surface realization method to deal with morphology more
complex than that of English (H¨
am¨
al¨
ainen and Rueter 2018).
There is already a similar semantic database available for
Finnish (H¨
am¨
al¨
ainen 2018) as the one we used for English,
which greatly facilitates a multilingual port of our method.
Conclusions
We have presented a method for generating humorous head-
lines that in its current state, falls behind the human level
humor. Nevertheless the results reach to a comparable level
with an existing neural based method. The method proposed
by us has the potential of reaching to a human level humor
generation in the limited domain task of altering a word in
an existing headline if a better ranking mechanism for its
output was introduced.
The evaluation and analysis we conducted on the results
has revealed several features which can be modelled better
in the future to improve our method. As we have gathered
human judgements for headlines generated by our system
for original headlines that are based on an existing humor
annotated corpus, we are releasing our evaluation results and
the generated titles4in the same format as the corpus we
used so that our data can be easily used in research dealing
with the existing dataset.
Acknowledgments
This work has been partially supported by the European
Union’s Horizon 2020 programme under grant 825153 (Em-
beddia).
References
Akbik, A.; Blythe, D.; and Vollgraf, R. 2018. Contextual
string embeddings for sequence labeling. In COLING 2018,
4https://zenodo.org/record/4976481
27th International Conference on Computational Linguis-
tics, 1638–1649.
Alnajjar, K., and H¨
am¨
al¨
ainen, M. 2018. A master-apprentice
approach to automatic creation of culturally satirical movie
titles. In Proceedings of the 11th International Conference
on Natural Language Generation, 274–283.
Alnajjar, K.; H¨
am¨
al¨
ainen, M.; Chen, H.; and Toivonen, H.
2017. Expanding and weighting stereotypical properties of
human characters for linguistic creativity. In Proceedings
of the 8th International Conference on Computational Cre-
ativity, 25–32. Atlanta, United States: Georgia Institute of
Technology.
Alnajjar, K.; Hadaytullah, H.; and Toivonen, H. 2018. “Tal-
ent, Skill and Support.” A method for automatic creation of
slogans. In Proceedings of the 9th International Conference
on Computational Creativity (ICCC 2018), 88–95. Sala-
manca, Spain: Association for Computational Creativity.
Alnajjar, K.; Lepp¨
anen, L.; and Toivonen, H. 2019. No time
like the present: Methods for generating colourful and fac-
tual multilingual news headlines. In The 10th International
Conference on Computational Creativity, 258–265. Associ-
ation for Computational Creativity.
Alnajjar, K. 2018. The 12 million most fre-
quent English grammatical relations and their frequencies.
https://doi.org/10.5281/zenodo.1255800.
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):293–348.
Bojanowski, P.; Grave, E.; Joulin, A.; and Mikolov, T. 2017.
Enriching word vectors with subword information. Trans-
actions of the Association for Computational Linguistics
5:135–146.
Brownell, H. H.; Michel, D.; Powelson, J.; and Gardner,
H. 1983. Surprise but not coherence: Sensitivity to ver-
bal humor in right-hemisphere patients. Brain and language
18(1):20–27.
Brysbaert, M.; Warriner, A. B.; and Kuperman, V. 2014.
Concreteness ratings for 40 thousand generally known en-
glish word lemmas. Behavior research methods 46(3):904–
911.
Bunescu, R. C., and Uduehi, O. O. 2019. Learning to Sur-
prise: A Composer-Audience Architecture. In Proceedings
of the Tenth International Conference on Computational
Creativity, 41–48.
Burroway, J. 2007. Imaginative Writing: The Elements of
Craft. Pearson.
Cai, Y.; Li, Y.; and Wan, X. 2018. Sense-aware neural
models for pun location in texts. In Proceedings of the 56th
Annual Meeting of the Association for Computational Lin-
guistics (Volume 2: Short Papers), 546–551. Melbourne,
Australia: Association for Computational Linguistics.
Cattle, A., and Ma, X. 2018. Recognizing humour using
word associations and humour anchor extraction. In Pro-
ceedings of the 27th International Conference on Compu-
tational Linguistics, 1849–1858. Santa Fe, New Mexico,
USA: Association for Computational Linguistics.
Colton, S.; Wiggins, G. A.; et al. 2012. Computational
creativity: The final frontier? In Ecai, volume 12, 21–26.
Montpelier.
Costa, D.; Oliveira, H. G.; and Pinto, A. M. 2015. In reality
there are as many religions as there are papers-first steps to-
wards the generation of internet memes. In ICCC, 300–307.
Deb, K.; Agrawal, S.; Pratap, A.; and Meyarivan, T. 2000.
A fast elitist non-dominated sorting genetic algorithm for
multi-objective optimization: Nsga-ii. In Schoenauer, M.;
Deb, K.; Rudolph, G.; Yao, X.; Lutton, E.; Merelo, J. J.;
and Schwefel, H.-P., eds., Parallel Problem Solving from
Nature PPSN VI, 849–858. Berlin, Heidelberg: Springer
Berlin Heidelberg.
Degaetano-Ortlieb, S., and Piper, A. 2019. The scientization
of literary study. In Proceedings of the 3rd Joint SIGHUM
Workshop on Computational Linguistics for Cultural Her-
itage, Social Sciences, Humanities and Literature, 18–28.
Minneapolis, USA: Association for Computational Linguis-
tics.
Gatti, L.; ¨
Ozbal, G.; Guerini, M.; Stock, O.; and Strappa-
rava, C. 2015. Slogans are not forever: adapting linguis-
tic expressions to the news. In Twenty-Fourth International
Joint Conference on Artificial Intelligence.
H¨
am¨
al¨
ainen, M., and Rueter, J. 2018. Development of an
open source natural language generation tool for finnish. In
Proceedings of the Fourth International Workshop on Com-
putational Linguistics of Uralic Languages, 51–58.
H¨
am¨
al¨
ainen, M. 2018. Extracting a semantic database with
syntactic relations for finnish to boost resources for endan-
gered uralic languages. The Proceedings of Logic and Engi-
neering of Natural Language Semantics 15 (LENLS15).
He, H.; Peng, N.; and Liang, P. 2019. Pun generation
with surprise. In Proceedings of the 2019 Conference of the
North American Chapter of the Association for Computa-
tional Linguistics: Human Language Technologies, Volume
1 (Long and Short Papers), 1734–1744. Minneapolis, Min-
nesota: Association for Computational Linguistics.
Hong, B. A., and Ong, E. 2009. Automatically extracting
word relationships as templates for pun generation. In Pro-
ceedings of the Workshop on Computational Approaches to
Linguistic Creativity, 24–31. Association for Computational
Linguistics.
Honnibal, M., and Montani, I. 2017. spaCy 2: Natural lan-
guage understanding with bloom embeddings, convolutional
neural networks and incremental parsing. To appear.
Hossain, N.; Krumm, J.; Vanderwende, L.; Horvitz, E.; and
Kautz, H. 2017. Filling the blanks (hint: plural noun) for
mad Libs humor. In Proceedings of the 2017 Conference on
Empirical Methods in Natural Language Processing, 638–
647. Copenhagen, Denmark: Association for Computational
Linguistics.
Hossain, N.; Krumm, J.; and Gamon, M. 2019. “presi-
dent vows to cut <taxes>hair”: Dataset and analysis of
creative text editing for humorous headlines. In Proceed-
ings of the 2019 Conference of the North American Chapter
of the Association for Computational Linguistics: Human
Language Technologies, Volume 1 (Long and Short Papers),
133–142. Minneapolis, Minnesota: Association for Compu-
tational Linguistics.
H¨
am¨
al¨
ainen, M., and Honkela, T. 2019. Co-operation as an
asymmetric form of human-computer creativity. case: Peace
machine. In Proceedings of the 1st Workshop on NLP for
ConvAI.
Koestler, A. 1964. The act of creation. London Hutchinson.
Krikmann, A. 2006. Contemporary linguistic theories of
humour. Folklore: Electronic journal of folklore (33):27–
58.
Lynch, G. 2015. Every word you set: Simulating the cog-
nitive process of linguistic creativity with the pundit system.
International Journal of Mind Brain and Cognition 6(1-1).
Naumann, D.; Frassinelli, D.; and im Walde, S. S. 2018.
Quantitative semantic variation in the contexts of concrete
and abstract words. In Proceedings of the Seventh Joint Con-
ference on Lexical and Computational Semantics, 76–85.
Nesterenko, L. 2016. Building a system for stock news gen-
eration in Russian. In Proceedings of the 2nd International
Workshop on Natural Language Generation and the Seman-
tic Web (WebNLG 2016), 37–40. Edinburgh, Scotland: As-
sociation for Computational Linguistics.
Raskin, V. 1985. Semantic Mechanisms of Humor. Springer
Science & Business Media.
Ritchie, G. 2005. Computational mechanisms for pun gen-
eration. In Proceedings of the Tenth European Workshop on
Natural Language Generation (ENLG-05).
Ross, M. D.; Owren, M. J.; and Zimmermann, E. 2010. The
evolution of laughter in great apes and humans. Commu-
nicative & integrative biology 3(2):191–194.
Sane, S. R.; Tripathi, S.; Sane, K. R.; and Mamidi, R.
2019. Deep learning techniques for humor detection in
Hindi-English code-mixed tweets. In Proceedings of the
Tenth Workshop on Computational Approaches to Subjec-
tivity, Sentiment and Social Media Analysis, 57–61. Min-
neapolis, USA: Association for Computational Linguistics.
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 Proceedings of the
51st Annual Meeting of the Association for Computational
Linguistics (Volume 2: Short Papers), volume 2, 243–248.
Veale, T. 2004. Incongruity in humor: Root cause or epiphe-
nomenon? Humor: International Journal of Humor Re-
search 17(4):419–428.
Veale, T. 2016. Round up the usual suspects: Knowledge-
based metaphor generation. In Proceedings of the Fourth
Workshop on Metaphor in NLP, 34–41. San Diego, Califor-
nia: Association for Computational Linguistics.
West, R., and Horvitz, E. 2019. Reverse-engineering satire,
or “paper on computational humor accepted despite making
serious advances”. In Proceedings of the 33rd AAAI Confer-
ence on Artificial Intelligence.
Winters, T.; Nys, V.; and De Schreye, D. 2019. Towards a
General Framework for Humor Generation from Rated Ex-
amples. In Proceedings of the Tenth International Confer-
ence on Computational Creativity, 274–281.
Xiao, P.; Alnajjar, K.; Granroth-Wilding, M.; Agres, K.; and
Toivonen, H. 2016. Meta4meaning: Automatic metaphor
interpretation using corpus-derived word associations. In
Proceedings of the 7th International Conference on Compu-
tational Creativity (ICCC 2016). Paris, France: Sony CSL.
Yang, D.; Lavie, A.; Dyer, C.; and Hovy, E. 2015. Humor
recognition and humor anchor extraction. In Proceedings
of the 2015 Conference on Empirical Methods in Natural
Language Processing, 2367–2376.
Yao, J.-g.; Zhang, J.; Wan, X.; and Xiao, J. 2017. Content
selection for real-time sports news construction from com-
mentary texts. In Proceedings of the 10th International Con-
ference on Natural Language Generation, 31–40. Santiago
de Compostela, Spain: Association for Computational Lin-
guistics.
Yu, Z.; Tan, J.; and Wan, X. 2018. A neural approach to
pun generation. In Proceedings of the 56th Annual Meeting
of the Association for Computational Linguistics (Volume 1:
Long Papers), 1650–1660.
Zou, Y., and Lu, W. 2019. Joint detection and location of
English puns. In Proceedings of the 2019 Conference of the
North American Chapter of the Association for Computa-
tional Linguistics: Human Language Technologies, Volume
1 (Long and Short Papers), 2117–2123. Minneapolis, Min-
nesota: Association for Computational Linguistics.
