On the Inﬂuence of Emotional Valence Shifts on
the Spread of Information in Social Networks
sen1and Mark Strembeck1,2,3
1Vienna University of Economics and Business
2Secure Business Austria Research Center (SBA)
3Complexity Science Hub Vienna (CSH)
Department of Computer Science
Sapienza - Universit`
a di Roma, Italy
Department of Mathematics
a di Padova, Italy
Abstract—In this paper, we present a study on 4.4 million
Twitter messages related to 24 systematically chosen real-world
events. For each of the 4.4 million tweets, we ﬁrst extracted
sentiment scores based on the eight basic emotions according
to Plutchik’s wheel of emotions. Subsequently, we investigated
the effects of shifts in the emotional valence on the spread of
information. We found that in general OSN users tend to conform
to the emotional valence of the respective real-world event. How-
ever, we also found empirical evidence that prospectively negative
real-world events exhibit a signiﬁcant amount of shifted emotions
in the corresponding tweets (i.e. positive messages). To explain
this ﬁnding, we use the theory of social connection and emotional
contagion. To the best of our knowledge, this is the ﬁrst study that
provides empirical evidence for the undoing hypothesis in online
social networks (OSNs). The undoing hypothesis postulates that
positive emotions serve as an antidote during negative events.
Keywords - sentiment analysis, diffusion, Twitter
In online social networks (OSNs), rapid information diffu-
sion comes with valuable social beneﬁts. For example, Twitter
was an important means in helping to save lives during the
2011 Tsunami disaster in Japan  as well as Red River
ﬂoods and Oklahoma ﬁres in 2009 . However, although
having a great potential to do good for society, OSNs have also
been recognized as a tool to inﬂuence people. For example,
a number of studies have shown that Twitter, Facebook, and
YouTube have been used to spread terrorist propaganda and
negatively inﬂuence users (e.g., online radicalization).
Recent studies have recognized the important role of emo-
tions in such information diffusion processes, indicating that
emotions in OSNs are contagious  and may lead to a high
diffusion rate [4, 5].
In this paper, we investigate the effects of speciﬁc positive
(e.g., joy) and negative (e.g., anger, sadness) emotions on user
behavior and information diffusion patterns. To systematically
address this issue, we extracted 4.4 million tweets related to 24
real-word events that induce positive, negative, or polarizing
emotional reactions. In our analysis, we distinguish between
expected emotions (a positive event is expected to trigger pre-
dominantly positive emotions) and those of a shifted emotional
valence (negative emotions that are related to positive events).
The paper is organized as follows. In Section II, we discuss
related work followed by our data analysis procedure in
Section III. We report on our results in Section IV and discuss
them in Section V. Section VI concludes the paper.
II. RE LATE D WO RK
Some prior studies did not ﬁnd signiﬁcant differences in the
effects of positive and negative messages. For example, 
indicated that both spread wider than messages with a neutral
sentiment score. Similar ﬁndings have been reported in ,
which suggested that emotionally-charged tweets in general
tend to be re-tweeted more often than neutral ones.
In contrast, other studies found empirical evidence that
supports the Pollyanna hypothesis which refers to the human
preference to like positive messages more than negative and
neutral ones . For example  found that positive emotions
boost the transmission of messages, while negative emotions
induce the opposite effect. Similar ﬁndings have been reported
in , where the authors showed that positive messages spread
wider (i.e. to more users) but also slower than negative and
neutral ones. Tsugawa and Ohsaki , on the other hand,
suggested that negative messages spread faster and wider than
positive or neutral messages.
Even though a number of effects related to sentiment polar-
ities (i.e., positive, negative, and neutral) have been studied in
the related work, aspects beyond the effects of polarities have
rarely been investigated. For example, Berger  discusses the
effects of psychological arousal on information sharing. In par-
ticular, he found that affective arousal increases the likelihood
for sharing information, regardless of whether the respective
information conveys a positive or a negative sentiment. In
addition, a limited number of studies examines the effects
of speciﬁc emotions. For example, [11, 12] consider anger,
anxiety, awe, and sadness, as annotated by human encoders.
The results of both studies showed that anger and awe increase
the content sharing behavior, while sadness and anxiety were
negatively associated with the content diffusion.
III. DATA ANA LYSI S PROCEDURE
1) Data extraction. In order to study the impact of emo-
tions on information diffusion, we systematically identiﬁed 24
events belonging to 5 different domains (as shown in Table I).
We collected 4,418,655 million publicly available tweets by
using the Twitter search API and a list of predeﬁned hashtags.
The events have been selected such that they fall in one of the
following categories: 1) events that trigger positive emotions
(e.g. festivities), 2) events that trigger negative emotions (e.g.
death), 3) emotionally polarizing events (e.g. presidential
elections). In Table I, Nindicates the actual number of tweets
in each category, followed by the overall size of each category
in our data-set (in percent), and RT is the percentage of re-
Domain Event Nr. tweets
Negative (N=1,490,495; 34%, RT=76.38%)
Politics 1) Erdogan’s threats to EU 804
2) US anti-Trump protests 381,982
Pop culture 3) Death of Leonard Cohen 89,619
4) Death of Colonel Abrams 1,253
War & 5) Aleppo bombings 995,561
terrorism 6) Seattle shooting 73
Other 7) Lufthansa strike 3,387
8) Ransomware in Seattle 2,564
9) Yellowstone incident 15
10) Earthquake in central Italy 15,237
Positive (N=1,115,587; 25%, RT=68.88%)
Sports 11) Rosberg winning Formula 1 215,703
12) Murray winning ATP 62,184
13) Rosberg retirement message 34,201
Pop culture 14) “Beauty and the Beast” trailer release 138,979
15) “Fantastic beasts” trailer release 64,264
16) ComiCon Vienna 704
17) Miley Cyrus birthday 76,270
18) New Pentatonix album released 9,341
19) Ellen Degeneres medal of freedom 73,854
Other 20) Thanksgiving 440,087
Polarizing (N=1,812,573; 41%, RT=73.90%)
Politics 21) Death of Fidel Castro 720,548
22) 2016 Austrian presidential elections 2,558
23) 2016 US presidential elections 891,425
Pop culture 24) The Walking Dead season 7 premiere 198,042
LIS T OF EV EN TS HA PPE NE D BET WE EN OCT OB ER AN D DECEMBER 2016.
In particular, we extracted the data for each of the 24 events
on a daily basis, starting the extraction when the event was
ﬁrst announced and stopped 7 days after. We recorded the text
of each tweet, its re-tweet count, like count, account name that
posted the tweet, and the date when the tweet was published.
The data extraction was restricted to English language tweets.
2) Data pre-processing. After the data was collected, we
cleaned the raw data-set by removing entries that contained
uninformative content with respect to emotion extraction (such
3) Emotion extraction. After pre-processing, we lemma-
tized the data-set, tagged parts-of-speech, and then extracted
the corresponding emotions via an R script. Our script iden-
tiﬁes the presence of Plutchik’s eight basic emotions  by
relying on the NRC word-emotion lexicon . In particular,
we considered negation (e.g., “I am not happy.”) and adverbs
of degree (e.g., very,hardly,absolutely), as well as emoticons,
which we categorized as positive (e.g., happy face “:)”) or
negative (e.g., sad face “:(”). In addition, we extended the
NRC dictionary with a list of common acronyms used in social
media (such as LOL or YOLO).
For processing our data-set, we used ﬁve machines: three
running on Windows 7 with 16 GB RAM and Intel Core i5-
3470 CPU @3.20 GHz, and two running on Linux - one with
32 GB RAM and Intel Xeon E3-1240 v5 CPU @3.5GHz and
the other with 16 GB RAM and Intel Xeon CPU E5-2620
v3 @2.40GHz. On these 5 machines, the emotion extraction
procedure took approximately a week to complete.
4) Data analysis and research questions. In our analysis,
we ﬁrst observed how Twitter users express speciﬁc emotions
during positive, negative, and polarizing events. We separate
each of the 24 event-related data-sets into a subset that conveys
expected emotions and a subset with a shifted emotion in terms
of its valence. We then observe how user behavior in each
subset is inﬂuenced by expected and shifted emotions.
Our analysis is guided by the following research questions.
RQ1: Which emotions are expressed during positive, nega-
tive, and polarizing events?
RQ2: Which tweeting behavior1do users exhibit during
positive, negative, and polarizing events?
RQ3: Are there differences in the tweeting behavior when
users are faced with tweets that convey expected emotions and
those with a shifted emotional valence?
IV. RES ULT S
A) Emotion intensity in different events. Figure 1 shows
the intensity of the eight emotions in each of the three
categories of events (i.e. polarizing, positive, and negative,
respectively). In the ﬁgure, negative emotions are colored
red (anger, sadness, disgust, fear), positive green (joy, trust),
and conditional (i.e. context-dependent) emotions in beige
(surprise, anticipation). To mitigate bias in the results, we
present the scores of each emotion averaged over the sentence
count. Finally, to show the relative presence of each emotion
in the data-set, we divide the averaged emotion scores e(based
on the sentence count S) by the tweet count (N)
In polarizing events there was no tendency of a particular
group of emotions to dominate over the other (the difference
between the negative and positive group is 0.015, see Figure
1a). In contrast, positive events showed a higher count of
positive emotions (joy, trust) (difference between positive and
negative is 0.15, see Figure 1b). As expected, negative events
showed a comparatively higher count of negative emotions.
However, and interestingly, with a noticeable presence of pos-
itive emotions (difference between the negative and positive
group is only 0.044, see Figure 1c).
B) User behavior during the analyzed events. Table II
summarizes the user behavior during positive, negative, and
polarizing events. In particular, positive events trigger the
highest number of re-tweets and likes. This shows a tendency
of users to prefer engaging in positive discussions rather than
negative (which conﬁrms the Pollyanna hypothesis, see also
1We study how users react to the three types of events in terms of re-
tweets, number of likes, tweeting rate, tweeting count per user, and one-to-one
a) Polarizing events. b) Positive events. c) Negative events.
Fig. 1. Emotions in prospectively polarizing, positive, and negative events.
[7, 9]). Interestingly, users also tend to engage in a one-to-
one communication (via @username) more frequently during
positive events than during negative or polarizing ones. How-
ever, the results indicate that users tend to send comparatively
more tweets during negative events (2.86 tweets per user)
than during polarizing (1.92 tweets per user) or positive (1.62
tweets per user) events. Interestingly, in our analysis polarizing
events exhibited the highest tweeting rate per minute (49.42
To examine the differences in how users respond to the
tweets conveying expected emotions and those containing
shifted emotions, we rely on the t-test. The results indicate
a signiﬁcant difference in user responses. Apart from one ex-
ception in the like count (p>0.05), all other tweeting behaviors
exhibit a clear pattern. In particular, expected emotions receive
more re-tweets, are distributed faster, and are dominant in a
one-to-one communication (@-count). Another interesting in-
sight can be observed while examining the differences between
negative and positive events. While negative events tend to
exhibit a higher number of negative tweets per user, positive
events dominate in the retweet count of positive tweets and
one-to-one social sharing of predominantly positive emotions.
Negative Positive Polarizing
Retweet count 1600.10 5677.63 4821.90
t=1.98, p<0.05 t=243.511, p<0.05
Like count 0.98 1.49 1.22
t=-0.97, p>0.05 t=0.98, p>0.05
Time rate 37.58 42.48 49.42
t=25.94, p<0.05 t=57.47, p<0.05
Tweet per user 2.86 1.62 1.92
t=13.33, p<0.05 t=26.04, p<0.05
@-count 1.02 1.19 1.02
t=25.77, p<0.05 t=69.93, p<0.05
USER BEHAVIOR IN POSITIVE,NEGATIVE,AND POLARIZING EVENTS.
Figure 2a) shows the time series of tweets divided into
the ones that carry a predominantly positive emotion and
those that carry a negative emotion. During positive events,
there is a noticeable smaller number of tweets that convey
negative emotions. However, the data shows that, although
small in size, negative tweets occur consistently throughout
the data-set (mean(set difference)=36668.11, sd(set differ-
a) positive events b) negative events
Fig. 2. Negative tweets during positive events and vice versa. Positive
emotions are colored green and negative emotions red.
In contrast, our analysis indicates that this is not the
case during negative events. As shown in Figure 2b), dur-
ing negative events positive tweets consistently follow the
tweets conveying expected negative emotions (mean(set dif-
ference)=4231.67, sd(set difference)=5603.162). Interestingly,
our data-set also revealed unexpected cases where (shifted)
positive tweets even exceed the (expected) negative tweets (the
largest difference between the two subsets is 6403 tweets).
Our results bring forth interesting insights into how OSN
users behave during positive, negative, and polarizing events
when faced with shifted emotions. Delving into the ﬁeld of
psychology, we found that positive emotions occur also during
negative events. An explanation for the observed phenomenon
2Set difference refers to the difference between the count of the expected
emotions and shifted emotions.
can be attributed to social connection  as one of the
fundamental human needs. In our data-set we found examples
of people explicitly calling for social bonding during emotion-
ally tough events (e.g., after the earthquake in central Italy,
people tweeted: “please join us as we #PrayforItaly”) and
a public and explicit expression of vulnerability that triggers
compassion (e.g., “Oh dear world, I am crying tonight”,
during Aleppo bombings). In particular, our data-set indicates
that people tend to show appreciation and love for the person
they care about or admire (e.g., a deceased singer, such as
Leonard Cohen) or even comfort each other and send messages
of hope during natural disasters (e.g., earthquakes in Italy)
or war (e.g., Aleppo bombing). Thus, we found empirical
evidence that supports the undoing hypothesis , which
states that people tend to use positive emotions as an antidote
to undo the effects of negative emotions.
Our results further show that expected emotions result in
more re-tweets. We thereby conﬁrm ﬁndings from , which
indicated that people prefer to share messages that correspond
to the emotional valence of the domain. This might be at-
tributed to the human tendency to conform to the situation.
Other studies bring an additional interesting insight into inter-
personal interactions over social media, which might explain
our observations. According to , emotional messages tend to
inﬂuence the emotions conveyed in other users’ messages. This
phenomenon, called emotional contagion, relies on studying
the social connections of the users (or their position in the
network). Even though studying the network structure of the
users participating in the Twitter discourse during positive,
negative, and polarizing events is out of the scope of this
study, we were able to observe that messages sent by “fans”
(we follow an assumption that fans follow their idols on
Twitter with a high probability) tend to be congruent with
the messages sent by their idols. For example, a tweet posted
by Pentatonix (announcement of their new album) triggered
positive reactions from their fans (in terms of retweeting and
original positive tweets).
VI. CONCLUSION AND FUTURE WORK
In this paper, we studied the inﬂuence of emotional valence
shifts on information diffusion in Twitter. We began our study
by extracting over 4.4 million tweets belonging to positive,
negative, or polarizing events. For each of the three categories,
we analyzed the intensity of Plutchik’s eight basic emotions
and identiﬁed shifts in emotional valence. We found that
users generally prefer sharing messages that correspond to the
emotional valence of the respective event. While conducting
a time-series analysis, we found a clear distinction between
positive and negative events, with respect to shifted emotions.
We found that positive events trigger a comparatively small
number of negative messages. However, while negative events
exhibit predominantly negative messages, they are accompa-
nied by a surprisingly high number of positive messages (in
our data-set, this occurred in 17.5% of the cases). To the best
of our knowledge, this is the ﬁrst study which found empirical
evidence that supports the theory of the undoing hypothesis in
In our future work, we plan to extend our analysis to
studying messages written in languages other than English.
Moreover, we are currently studying the effects of shifts in
emotional valence over a Facebook and YouTube data-set.
Another interesting direction would be to examine the impact
of bots on the valence shifts.
Mauro Conti is supported by a Marie Curie Fellowship
funded by the European Commission (PCIG11-GA-2012-
321980), partially supported by the TENACE PRIN Project
20103P34XC funded by the Italian MIUR and by the Project
“Tackling Mobile Malware with Innovative Machine Learning
Techniques” funded by the University of Padua, and by the
EU TagItSmart! Project (H2020-ICT30-2015-688061).
 C. St Louis and G. Zorlu, “Can Twitter predict disease outbreaks?” BMJ,
vol. 344, 2012.
 S. Vieweg, A. L. Hughes, K. Starbird, and L. Palen, “Microblogging
during two natural hazards events: What Twitter may contribute to
situational awareness,” in Proceedings of the SIGCHI Conference on
Human Factors in Computing Systems, ser. CHI ’10, 2010.
 A. D. I. Kramer, J. E. Guillory, and J. T. Hancock, “Experimental
evidence of massive-scale emotional contagion through social networks,”
Proc. of the National Academy of Sciences, vol. 111, no. 24, 2014.
 J. Berger, “Arousal increases social transmission of information,” Psy-
chological Science, vol. 22, no. 7, 2011.
 S. Stieglitz and D.-X. Linh, “Emotions and information diffusion in
social media- sentiment of microblogs and sharing behavior,” Journal
of management information systems, vol. 29, no. 4, 2013.
 D. N. Trung, T. T. Nguyen, J. J. Jung, and D. Choi, Understanding
Effect of Sentiment Content Toward Information Diffusion Pattern in
Online Social Networks: A Case Study on TweetScope, 2014.
 J. Boucher and C. E. Osgood, “The pollyanna hypothesis,” Journal of
Verbal Learning and Verbal Behavior, vol. 8, no. 1, 1969.
 H. S. Kim, S. Lee, J. N. Cappella, L. Vera, and S. Emery, “Content
characteristics driving the diffusion of antismoking messages: Impli-
cations for cancer prevention in the emerging public communication
environment,” Journal of National cancer institute. Monographs, vol. 47,
 E. Ferrara and Z. Yang, “Quantifying the effect of sentiment on
information diffusion in social media.” PeerJ Computer Science, vol. 1,
 S. Tsugawa and H. Ohsaki, “Negative messages spread rapidly and
widely on social media,” in Proceedings of the 2015 ACM on Conference
on Online Social Networks, 2015.
 B. Suh, L. Hong, P. Pirolli, and E. H. Chi, “Want to be retweeted?
Large scale analytics on factors impacting retweet in Twitter network,”
in Proc. of the 2010 IEEE Second International Conference on Social
 I. Heimbach and O. Hinz, “The impact of content sentiment and
emotionality on content virality,” International Journal of Research in
Marketing, vol. 33, no. 3, 2016.
 R. Plutchik, “The nature of emotions,” American Scientist, vol. 89, no. 4,
 S. M. Mohammad and P. D. Turney, “Crowdsourcing a word-emotion
association lexicon,” Computational Intelligence, vol. 29, no. 3, 2013.
 R. F. Baumeister and M. R. Leary, “The need to belong: Desire
for interpersonal attachments as a fundamental human motivation.”
Psychological Bulletin, vol. 117, 1995.
 B. L. Fredrickson, “The role of positive emotions in positive psychology:
The broaden-and-build theory of positive emotions.” The American
Psychologist, vol. 56, 2001.
 “Do people prefer to pass along good or bad news? valence and rele-
vance of news as predictors of transmission propensity,” Organizational
behavior and human decision processes, vol. 68, no. 2.