Predicting Meme Success with Linguistic Features in a Multilayer

Backpropagation Network
Keith T. Shubeck (kshubeck@memphis.edu)
Stephanie Huette (shuette@memphis.edu)
Department of Psychology, 202 Psychology Building
Memphis, TN 38152 USA

Abstract individuals, the more we will understand cultural trends that

may have been previously considered bewilderingly
T he challenge of predicting meme success has gained
attention from researchers, largely due to the increased anomalous. The challenge then becomes for researchers to
availability of social media data. Many models focus on develop robust and valid methods for detecting memes,
structural features of online social networks as predictors of tracking their mutations, and predicting their success. The
meme success. T he current work takes a different approach, current model attempts to develop a method for predicting
predicting meme success from linguistic features. We propose meme success by analyzing its linguistic features. The
predictive power is gained by grounding memes in theories of linguistic features used represent dimensions known to be
working memory, emotion, memory, and psycholinguistics.
T he linguistic content of several memes were analyzed with
important in communication and encompass emotion and
linguistic analysis tools. T hese features were then trained arousal the language may elicit, as well as more basic
with a multi-layer supervised backpropagation network. A set features such as length, concreteness, and orthographic
of new memes was used to test the generalization of the features such as misspellings .
network. Results indicated the network was able to generalize The challenge of detecting and tracking memes has been
the linguistic features in order to predict success at greater approached in a variety of ways, with varying success. The
than chance levels (80% accuracy). Linguistic features appear broad and encompassing nature of the definition for meme
to be enough to predict meme transmission success without
any information about social network structure.
has resulted in the term being operationalized differently
from study to study. In addition to the changing operational
Keywords: meme prediction; distributed cognition; definitions, the domains of meme studies also vary. For
evolutionary psychology; neural networks example, some studies focus on visual or video content such
as YouTube memes (Shifman, 2012; Xie, Nastev, Kender,
Introduction Hill & Smith, 2011), and others on textual memes, like
The term “meme” was originally coined by Richard quoted text in the news cycle (Simmons, Adamic, & Adar,
Dawkins in his book, The Selfish Gene. Dawkins, an 2011; Leskovec, Backstrom, & Kleinberg, 2009). Other
evolutionary biologist, describes “meme” as a unit for research has focused on microblogging memes in social
carrying cultural ideas or behavior, similar to how genes networks such as Twitter or Yahoo! Meme (Ratkiewicz et
carry genetic information from one generation to the next. al., 2010; Adamic, Lento, Adar & Ng, 2014; Tsur &
Just as genes propagate from organism to organism, memes Rappoport, 2012; Ienco, Bonchi, & Castillo, 2010).
propagate from mind to mind by way of communication and For our purposes here, we will focus on popular text-based
social learning (Dawkins, 1989). Under this lens, memes are memes, of which some have visual components that were
also subject to mutations, where each mutation either not included in the model, and others simply contain text.
strengthens or weakens the meme’s fitness. Susan Not all memes are retransmitted: the process of deciding
Blackmore (1998) argues that memes compete with one whether to retransmit a meme is at the level of the
another for an individual’s limited cognitive resources for individual user. With this in mind, memes that are shorter
the chance to replicate again. Thus, some memes will fall and easier to remember should have inherent advantages
into obscurity where others will flourish. With this in mind, over memes that are lengthier and are therefore more
successful memes should be those that are easily memorable difficult to remember. Copy-and-paste memes often contain
(Blackmore, 1998). Analyzing the properties and features of instructions within the text that encourage others to copy
memes that may influence their fitness has proven to be a and paste the meme, or to “repost” the meme in its entirety
challenging endeavor, especially prior to the establishment (Adamic et al., 2014). The copy-and-paste method for meme
of various online social networks. spreading circumvents some of the cognitive memory
The internet, and more specifically social media, provides constraints, allowing for lengthier memes to succeed and
researchers interested in the study of information diffusion, become popular. However, there still must be a decision to
meme propagation, and cultural transmission a means to retransmit or repost the meme which involves cognition and
observe these concepts in an ecologically valid setting and action.
on a massive scale. Our understanding of meme propagation Another recent study set out with the goal of predicting
runs parallel with our understanding of human culture; the meme success by observing the meme’s early spreading
more we understand about memes and their mutations, their patterns within Twitter (Weng, Menczes, Ahn, 2014). The
origins, and how quickly thes e are accepted by other
authors chose to focus on the structure of the meme’s essential feature, as they contain a collection of words that
environment because previous research has shown that the fit into 80 validated word categories, ranging from emotion
structure of underlying networks impacts the spreading word categories to deception word categories. Using a
process of information (Daley & Kendall 1964; Barrat, regression model, with the above mentioned features, they
Barthelemy, and Vespignani, 2008). Design features of the found that the cognitive category of words from LIWC was
website itself (i.e., user voting feature on Digg) can also be positively correlated with the hashtag’s popularity, when the
used to improve meme prediction (Hogg & Lerman, 2012). hashtag’s content was also taken into account. For example,
Weng et al. (2014) operationalize meme success by the word “think”, a cognitive process, would predict
observing the meme’s overall popularity, relative to the increased popularity compared to a non-cognitive word, like
other memes in their dataset. They operationalize “meme” “ball”. They also found that lengthier hashtags were not as
as any hashtag observed in their dataset. Hashtags are popular as shorter hashtags. They attributed this finding to
strings of text following a “#” users insert into their tweets cognitive load theory and physical constraints for tweets
(i.e., short user submitted posts within Twitter) for labeling (i.e., 140 character limit per tweet). Cognitive load theory
purposes. Popular hashtags are tracked by Twitter and said posits that during an instance of complex learning, an
to be “trending”. Here, the definition of a successful meme individual may be underloaded or overloaded with
is determined by the frequency of usage and overall information, due to the working memory limitations. While
popularity of that meme. these findings are promising, Tsur and Rappoport (2012)
Weng et al. (2014) found that using topographic, or point out that future studies using the content of memes to
structural, features of the network enabled their model to predict success should delve deeper into the
accurately predict a meme’s popularity up to two months in psycholinguistic aspects of the content and the cognitive
advance. These topographical features included “community constraints of the receiver of the meme.
size”, where a community is a set of nodes (i.e., individual These models often posit the relevant connections of
users) who are followers of one another, and “network meme transmission are between people, but this neglects
surface” (i.e., neighbors of the audience of users).The model what happens within an individual’s mind when a meme is
used by Weng et al. (2014) is similar to other studies that encountered. The factors contributing to whether the meme
include user influence in understanding information is transmitted, or not transmitted, is entirely a factor of the
diffusion (see Romero, Meeder, & Kleinberg, 2011). individual, and thus relies more on cognitive factors than
Unfortunately, studies that include user influence (i.e., number of connections to others in a social network. If the
number of followers a given user has, number of those person decides to not transmit the meme further, the number
followers’ followers, etc.) as a key component of their of connections to the user no longer matter. The current
meme predicting model add little to our understanding of work is at the cognitive level of analysis, where connections
why certain memes are selected to be spread, and why other constitute an information space inside of an individual, and
memes are not chosen to be spread. We argue that an “success” is whether or not the individual is likely to engage
important question remains unanswered: are there linguistic in further transmitting the meme.
features and aspects of cognition that can predict the The advantage of neural networks over other rule-based
ultimate success of a meme, outside of the characteristics of systems is they are able to solve more complex problems
the social network? and carve up the solution’s space in unanticipated ways. For
Tsur and Rappoport (2012) attempt to answer that example, cognitive process words may somewhat predict
question by taking a closer look at the content of Twitter meme success, but a combination of cognitive process
hashtags in order to predict their popularity. They explain words, emotion words, concreteness, etc. might be
that their work provides two major contributions to the interacting in non-intuitive ways that contribute to
current meme literature. First, their study places emphasis transmission or non-transmission of the meme. To
on the content features of a meme in determining its demonstrate this, we predicted a binary logistic regression
popularity, something that prior to their 2012 study, has would not yield as much predictive power as the neural
been largely ignored. Secondly, by stepping away from the network model. Neural networks are able to come up with
costly graph based algorithms, used in the studies solutions that do not rely on linear or singular relationships
mentioned above, Tsur and Rappoport (2012) provide a or causality, allowing for complex interactions which are
simple and more global approach for modeling meme well known to be commonplace in thinking,
acceptance and popularity. The content features that were communication, and behavior. Performance of a binary
examined included: hashtag length (number of characters logistic regression will be compared to neural network
and words), hashtag orthography, emotional content and performance to test this prediction.
linguistic cognitive features taken from the Linguistic
Inquiry and Word Count Tool, or LIWC. LIWC Model
(http://www.liwc.net/) is a linguistic tool that counts the Meme Corpus
number of words in various categories that have been built Memes were collected from the meme wiki-style website,
upon relevant communicative dimensions (Tausczik & knowyourmeme.com, and were represented as 15 input
Pennebaker, 2010). The categories of the program are the nodes with binary values. Each element of the input vector
represented a linguistic or cognitive variable of the meme Psycholinguistic Features. Eight psycholinguistic features
that was theoretically and empirically motivated to have an were chosen as meme features. These features were selected
impact on the meme’s popularity. The target outputs based on current cognitive psychology and psycholinguistic
consisted of two binary winner-takes-all nodes, where one theories centered on sentence recall, working memory, and
represented “successful” and the other represented how emotion and arousal affect memory.
“unsuccessful”. Meme success was determined by using the Mean word concreteness was determined through the use
number of Google search results of a meme phrase, of Coh-Metrix, (http://cohmetrix.com/) a validated linguistic
verbatim. This was similar to the way that hashtag searches analysis tool that is able to automatically analyze text for
were used in the aforementioned Twitter meme studies. features such as text cohesion, parts of speech, word
In order to reduce noise in the number of inaccurate result frequency, lexical diversity, and syntactic complexity
hits, a time range filter was placed on each meme search, (McNamara, Kulikowich, & Graesser, 2011). Concreteness
based on the month the meme search queries first spiked. was chosen as a psycholinguistic feature for the current
This was determined by using Google Trends, which allows model because previous research has shown that concrete
users to show how often a particular search term is entered words are easier to recall than abstract words during a short-
in Google search, over time. If a meme’s search queries first term serial recall task (Walker & Hulme, 1999). Memes that
began to spike in October of 2009, then the search was are easier to recall and more concrete should have a distinct
limited to October 2009 to the present date. Many of the advantage over memes that are more difficult to recall. If a
memes in the current dataset are words or word phrases that given meme had more concrete terms than abstract terms
were initially esoteric, but were in use prior to the spike in then it was coded as concrete (1), if it contained no concrete
popularity of the new or updated meaning. After a certain terms, or more abstract terms, then it was coded as abstract
event (e.g., popular YouTube video, viral social networking (0).
post, etc.) the meaning of the word phrase shifted. For the The overall emotional arousal of a meme was determined
current model, we were only interested in the most current through the use of the LIWC (Linguistic Analysis and Word
generation of the meme’s meaning. After determining the Count; Pennebaker, Francis, & Booth, 2001). LIWC’s affect
total number of search results provided for each individual dictionaries were based on the emotion rating scales
meme, a median split was applied to the data to separate developed by Watson, Clark, and Tellegen (1988). For this
successful memes from unsuccessful memes. For this feature, if a meme included an emotional word, either
particular data set, memes that had 37,400 or more search positive or negative, it was considered an emotional meme
results were considered successful, and any memes below (1), and if the meme contained no emotion words then it
that threshold were considered unsuccessful. Of course all was considered a non-emotion meme (0). The emotional
memes were retransmitted to some degree, so this label arousal feature was included in the current model because
might be something more akin to “more popular” and “less previous research has shown emotional arousal, in general,
popular” when discussing memes as a whole. has an impact on long term declarative memory (Cahill &
McGaugh, 1998).
Training set. The dataset used to train the network Four other finer-grained emotional features were also
consisted of 268 established memes collected from recorded for each meme. These features were used to
knowyourmeme.com, a meme encyclopedia, which uses the determine 1) whether or not positive emotion was present,
wiki web application to collect and categorize various 2) whether or not negative emotion was present, 3) whether
internet memes. The memes included in our corpus contain there was more positive emotion than negative emotion and,
hashtag memes (e.g., #YOLO), copy-and-paste memes (e.g., 4) whether there was more negative emotion than positive
Repost this if you're a big black woman who don't need no emotion. Negative emotion has been found to enhance
man), as well as lesser known memes commonly used in memory accuracy for specific details during a recall task
smaller online communities (e.g., burst into treats). The (Kensinger, 2007). However, the broaden-and-build
average meme word length was roughly four words per hypothesis posits that positive moods broaden an
meme, with the longest meme having 31 words. Copy-and- individual’s scope of attention and thought-action
paste memes were divided into smaller chunks of text, each repertoires, whereas negative moods tend to narrow an
chunk having at most one complete sentence. In general, the individual’s scope of attention and associations between
memes used for the current study are phenotypic memes, thoughts and actions (Fredrickson & Branigan, 2005).
meaning their raw text contains the best estimate of the In their study, Tsur & Rappoport (2012) chose to include
“original” meme. Variants of these phenotypic memes were LIWC’s “cognitive” categories. They explained that this
not included. If it could not be clearly determined which category should contain words that prompt or encourage
meme came first, then both memes were included separately specific behaviors (e.g., cause, know, ought). However,
in the dataset. The linguistic and cognitive properties of the overall Tsur & Rappoport found that the more general
meme text were broken down into 15 binary features , cognitive category only marginally improved the MSE over
broken down into four categories: psycholinguistic features, the baseline. For the current study we chose to include the
physical features, orthographical features and meme type. more specific “CogMech” LIWC category (i.e., cognitive
mechanism) with the hope of improving the overall model.
 The last psycholinguistic feature included involves the
presence (1) or absence (0) of curse words, or taboo words,
in the meme. LIWC was used to determine the presence of
curse words in the set memes. LIWC’s swear word category
includes a set of socially proscribed derogatory or profane
words. A slew of previous research has shown that
emotionally arousing words, particularly taboo words, are
remembered better than neutral or nonarousing words (Jay,
Caldwell-Harris, & King, 2008, Kensinger, 2007; Kinsinger
& Corkin, 2003; LaBar & Phelps, 1998). Memes with curse
words should have a distinct advantage over memes without
curse words, in terms of the meme’s ability to be recalled.

Physical & Orthographical Features. Two physical

features of the meme text were also recorded. Intuitively,
memory span is inversely related to word length, and words Figure 1: An example of a template meme. The text varies
that take longer to read or speak are more difficult to recall from iteration to iteration, but the image remains static.
in simple recall tasks (Baddeley, Thomson, & Buchanan, Text here emphasizes awkward social behaviors.
1975). Memes that contained less than four words were
considered short (1) and memes that contained four or more Network Structure
words were considered long (0). Additionally, memes that The current model used a 4-layer backpropagation
contained words that all had less than three syllables were network that was designed to take linguistic features as
considered short (1), and memes that contained a word with inputs and classify them as either successful or
3 or more syllables were considered long (0). Shorter and unsuccessful. The neural network used to predict meme
less complex memes should be easier to recall, improving success consists of four layers: an input layer with 15 nodes,
their fitness and overall success. two hidden layers with 20 nodes each, and an output layer
Two orthographical features were included based on the with two winner-takes-all nodes that represents the
intuition that slang terms, purposeful word misspellings, or probability of success of the meme. The output nodes are
purposeful incorrect grammar usage should set some memes mutually exclusive. One node “fires” or activates whenever
apart from others. Words with incorrect spelling, or novel it thinks a meme is successful and the other node “fires” or
words and phrases should stand out more than correct word activates whenever it thinks a meme is unsuccessful. For
spellings and established words and phrasings. If memes are any given meme, if its output activation is higher on the
competing for attention, then memes with novel words or successful output node than the unsuccessful output node,
phrases should tend to be more popular or successful than then that meme is considered successful, and vice versa.
memes using traditional spelling and phrasing. There was a total of 268 memes used to train the network.
Network weights were trained on each meme 3000 times
Meme Type. Finally, three meme type features were coded.
and were modified using the delta rule. Network weights
The three meme types consist of template memes, copy-and- were initially random and trained with the delta rule, a
paste memes, and game memes. These were three different gradient descent learning algorithm. The target activation
features all mutually exclusive and determined during the value of 1 represented “successful” and values of 0
search process. Examples of game memes are “The object to represent “unsuccessful”. This value was determined by
your left will be your only weapon during a zombie using a median split on the popularity of the meme, where
apocalypse” or “You are now manually breathing”. An highly transmitted memes were considered successful and
example of a template meme is provided in Figure 1. frequently retransmitted at the level of the user, and more
infrequent memes while likely observed by many, were less
likely to be retransmitted. Learning rate was set to .001, and
the momentum term was set to 0.2. The network was trained
to convergence. The network was trained on all training
items 3000 times, at which the MSE reached an average of
.228. Matlab coding of the network is available from the
first author upon request.

Results
In order to test the accuracy of the network, a random
subset of 25 coded memes was left out of the training set to
test generalization to new items using a fully trained set of
connection weights. This is a test of the network’s
predictive power and generalization to new memes . The may contribute more or less to the prediction of success in
resulting output activation values were compared to the the network, and as with other neural networks it is difficult
expected target values. If the meme’s output activation on to see what is driving these results. However, comparing the
the “successful” output node was greater than the output network’s results with a binary logistic regression helped to
activation on the “unsuccessful” output node then the provide some insight. Meme length, whether or not a meme
classification was considered accurate. If the meme’s output is a template meme, and the presence or absence of swear
activation on the “unsuccessful” output node was greater words within the meme contributed significantly to
than the output activation on the “successful” output node predicting success in the logistic model. However, the
then the classification was considered inaccurate. The logistic model did not have prediction accuracy as high as
network achieved 80% prediction accuracy, or 20% higher the neural network model, pointing to the potential
than chance. Specifically, the network was able to contribution of other variables that on their own are not
accurately predict a successful meme to be successful with predictive in a regression, but in an interactive context like a
73% accuracy, and was able to accurately predict an neural network, have some predictive power. Further
unsuccessful meme to be unsuccessful with 90% accuracy. analysis of the network’s principle components is clearly
While the model appears able to learn with an overall MSE needed in future work.
of .228, it struggles to generalize accurately. The neural network model presented here has several
major limitations. The first limitation is the operationalized
Regression analysis. In addition, a binary logistic definition of success. Google search results offer a quick
regression was performed. The target values (successful or rough grained estimate for overall meme usage, but
unsuccessful) were considered the dependent variable and searching for specific phrases can still sometimes include
each input node was considered an independent variable. inaccurate search results. Without extensive and
Because all data is binary, binary logistic regression is computationally expensive web-crawlers, determining
appropriate for analyzing the factors that contribute to meme context from Google search results can be extremely
predicted success of a meme. The overall logistic regression difficult. Memes that can be used in multiple domains can
model was statistically significant, X2(14) = 48.893, p < be considered “flexible memes”, a quality that is likely
.0005. The model explained 22.3% (Nagelkerke R2) of the related to overall meme fitness. Another limitation to the
variance in meme success and correctly classified 54.1% of current study is the input set and test set are relatively small.
the successful memes as successful and 80.6% of the Many studies attempting to predict meme success have
unsuccessful memes as unsuccessful. Overall the binary access to millions of memes, albeit with a broader
logistic regression model had a prediction accuracy of operational definition. If the success of textual memes is
67.4%. Three predictor variables were statistically largely dependent on the average person’s ability to
significant. First, shorter memes were significant (p <.005), remember them, then many more cognitive variables can
and 2.802 times more likely to contribute to success. Memes and should be included.
that contained a swear word were .177 times less likely to
be successful than unsuccessful (p <.05), a small but Conclusion
significant contribution. Finally, template memes were The ability to detect and track memes and to predict their
2.223 times more likely to be successful than unsuccessful success is essential in order to improve our understanding
(p <.05). cultural evolution. Observing textual memes in particular
offers unique insights into the evolution of language. Social
Discussion media provides a petri dish environment for rapid meme
The results of the current study demonstrate the utility of generation and mutation. The current study categorized
using linguistic information as a means of predicting meme content based on 12 features grounded on cognitive
successful transmission of a meme. These preliminary theories of memory, emotion, and working memory
results warrant more in depth analyses, particularly a limitations. This experiment helped support the idea that
principle component analysis to determine whether the meme content should be considered when attempting to
network is easily carving out a solution space and clustering predict meme success. Future studies on meme prediction
success and non-success, or whether it is unclear based on should benefit from a more robust operational definition of
the input. A highly clustered solution space that exhibits success. This can likely be achieved by limiting the scope
linear seperability after learning in its first few principle from a global internet search to a specific social network. If
components would argue for the issue being with the a feed-forward backpropagation neural network can achieve
inaccuracy of the input and target information structure, relative success in predicting meme popularity, then a more
whereas a more overlapping or complex seperability would robust network that takes into account working memory
mean it is simply a very difficult problem to solve, with the limitations should provide more accurate results.
linguistic information contributing a rich source of This model demonstrates that it is not only possible to
information that could be used in models that incorporate predict overall success of a meme at greater than chance
multiple domains of information (user-level, visual feature, levels, but also argues for there being important parameters
social structure, etc.). Some of the features in the network at the level of what other models typically neglect: whether
or not the node transmits the information further. Other Lobe in Humans. Psychological Science, 9(6), 490–493.
models of meme transmission typically only take into doi:10.1111/1467-9280.00090
account the change of the meme over time (evolution), the Leskovec, J., Backstrom, L., & Kleinberg, J. (2009, June).
rates of transmission (viral) or the number of connections Meme-tracking and the dynamics of the news cycle. In
(small world networks). By incorporating cognitive Proceedings of the 15th ACM SIGKDD international
processes into models that also include information about conference on Knowledge discovery and data mining (pp.
the network at large, greater levels of prediction could be 497-506).
achieved in future instantiations of meme transmission Pennebaker, J. W., & Francis, M. E. (1999). Linguistic
models. inquiry and word count (LIWC). Mahwah, NJ: Erlbaum.
Ratkiewicz, J., Conover, M., Meiss, M., Gonçalves, B.,
References Patil, S., Flammini, A., & Menczer, F. (2011, March).
Adamic, L. A., Lento, T. M., Adar, E., & Ng, P. C. (2014). Truthy: mapping the spread of astroturf in microblog
Information Evolution in Social Networks. streams. In Proceedings of the 20th international
arXiv:1402.6792 [physics]. Retrieved from conference companion on World wide web (pp. 249-252).
http://arxiv.org/abs/1402.6792 ACM.
Baddeley, A. D., Thomson, N., & Buchanan, M. (1975). Romero, D. M., Meeder, B., & Kleinberg, J. (2011).
Word length and the structure of short-term memory. Differences in the mechanics of information diffusion
Journal of Verbal Learning and Verbal Behavior, 14(6), across topics: idioms, political hashtags, and complex
575–589. doi:10.1016/S0022-5371(75)80045-4 contagion on twitter. In Proceedings of the 20th
Barrat, A., Barthelemy, M., & Vespignani, A. (2008). international conference on World wide web (pp. 695-
Dynamical processes on complex networks. Cambridge, 704). ACM. doi:10.1145/1963405.1963503
UK; New York: Cambridge University Press. Shifman, L. (2012). An anatomy of a YouTube meme. New
Blackmore, S. (1998). Imitation and the definition of a Media & Society, 14(2), 187–203.
meme. Journal of Memetics-Evolutionary Models of Simmons, M. P., Adamic, L. A., & Adar, E. (2011). Memes
Information Transmission, 2(11), 159–170. online: Extracted, subtracted, injected, and recollected. In
Cahill, L., & McGaugh, J. L. (1998). Mechanisms of In Proceedings of the Fifth International AAAI
emotional arousal and lasting declarative memory. Trends Conference on Weblogs and Social Media.
in Neurosciences, 21(7), 294–299. doi:10.1016/S0166- Tausczik, Y. R., & Pennebaker, J. W. (2010). The
2236(97)01214-9 Psychological Meaning of Words: LIWC and
Daley, D. J., & Kendall, D. G. (1964). Epidemics and Computerized Text Analysis Methods. Journal of
Rumours. Nature, 204(4963), 1118–1118. Language and Social Psychology, 29(1), 24–54.
doi:10.1038/2041118a0 Tsur, O., & Rappoport, A. (2012, February). What's in a
Dawkins, R. (1989). The selfish gene. Oxford; New York: hashtag?: content based prediction of the spread of ideas
Oxford University Press. in microblogging communities. In Proceedings of the fifth
Fredrickson, B. L., & Branigan, C. (2005). Positive ACM international conference on Web search and data
emotions broaden the scope of attention and thought ‐ mining (pp. 643-652). ACM.
action repertoires. Cognition & Emotion, 19(3), 313–332. Walker, I., & Hulme, C. (1999). Concrete words are easier
Hogg, T., & Lerman, K. (2012). Social dynamics of Digg. to recall than abstract words: Evidence for a semantic
EPJ Data Science, 1(1). doi:10.1140/epjds5 contribution to short-term serial recall. Journal of
Ienco, D., Bonchi, F., & Castillo, C. (2010, December). The Experimental Psychology: Learning, Memory, and
meme ranking problem: Maximizing microblogging Cognition, 25(5), 1256–1271. doi:10.1037/0278-
virality. In Data Mining Workshops (ICDMW), 2010 7393.25.5.1256
IEEE International Conference (pp. 328-335). IEEE. Watson, D., Clark, L. A., & Tellegen, A. (1988).
Jay, T., Caldwell-Harris, C., & King, K. (2008). Recalling Development and validation of brief measures of positive
Taboo and Nontaboo Words. The American Journal of and negative affect: The PANAS scales. Journal of
Psychology, 121(1), 83. Personality and Social Psychology, 54(6), 1063–1070.
Kensinger, E. A. (2007). Negative Emotion Enhances doi:10.1037/0022-3514.54.6.1063
Memory Accuracy: Behavioral and Neuroimaging Weng, L., Menczzer, F., Ahn, Y.-Y. Predicting successful
Evidence. Current Directions in Psychological Science, memes using network and community structure. In
16(4), 213–218. Proceedings of the Eigth International AAAI Conference
Kensinger, E. A., & Corkin, S. (2003). Memory on Weblogs and Social Media (ICWSM’14), Ann Arbor,
enhancement for emotional words: Are emotional words MI, USA, June 2014.
more vividly remembered than neutral words? Memory & Xie, L., Natsev, A., Kender, J. R., Hill, M., & Smith, J. R.
Cognition, 31(8), 1169–1180. doi:10.3758/BF03195800 (2011). Visual memes in social media: tracking real-world
LaBar, K. S., & Phelps, E. A. (1998). Arousal-Mediated news in YouTube videos (p. 53). In Proceedings of the
Memory Consolidation: Role of the Medial Temporal 19th ACM international conference on Multimedia (pp.
53-62). ACM Press. doi:10.1145/2072298.2072307