1 Introduction

In online social networks (OSNs), news travel fast and reach a large number of users within a short period of time [24, 33]. Such a rapid information diffusion comes with valuable social benefits such as using OSNs to help save lives during the 2011 tsunami disaster in Japan [35] as well as the Red River flood and Oklahoma fires in 2009 [43]. However, although having a great potential to do good for the society, OSNs have also been recognized as a convenient tool to (positively or negatively) influence people. For example, a number of recent studies indicated that Twitter, Facebook, and YouTube have been used to spread terrorist propaganda [44] and negatively influence users (online radicalization) [2, 7, 26, 36].

In this context, emotions have generally been recognized as an important factor in influencing or manipulating people’s opinions and beliefs [29]. In particular, recent studies indicated that emotions can be passed through online interactions from one user to another [11, 21], resulting in the so-called emotional contagion. In addition, numerous studies reported on user reactions to emotionally charged messages. For example, Faraon et al. [10] and Stuart [38] found that users tend to pay more attention to the negative messages, while Bayer et al. [4] and Stieglitz and Linh [37] presented a contradictory finding (i.e., positive messages receive more attention). The common denominator in either case is that emotions conveyed in OSN messages have the potential to trigger a strong emotional reaction in people [5, 9, 25, 28].

This paper extends our prior analysis presented in [23]. In particular, we study the impact of emotions on the messaging behavior of OSN users on Twitter, Facebook, and YouTube. To this end, we performed an emotion analysis over 5.6 million social media messages that occurred in 24 systematically chosen real-world events. For each of these messages, we derived emotion scores concerning the eight basic emotions according to Plutchik’s wheel of emotions [32]. In general, we found that people tend to conform to the base emotion of a particular event. However, we also found empirical evidence that in all three OSNs prospectively, negative real-world events are accompanied by a substantial amount of shifted (i.e., positive) emotions in the corresponding messages. In order 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 remainder of this paper is organized as follows. In Sect. 2, we discuss related work. Next, Sect. 3 describes our data analysis procedure followed by a detailed report on our results in Sect. 4. Subsequently, Sect. 5 discusses our findings, and Sect. 6 concludes the paper.

2 Related Work

Prior studies predominantly examined the impact of sentiment polarities and emotions on information diffusion over OSNs. For example, Zhang and Zhang [46] examine the impact of emojis on message diffusion patterns over a data set containing about 12 million Weibo messages. In particular, they found that positive and negative emojis result in the same effects with respect to retweets and replies. In fact, both groups of emojis have a positive effect on the number of replies a message receives and a negative effect on the retweet count. Other studies examined textual cues to identify a set of emotions or sentiment polarities. For example, Kim et al. [20] conducted a questionnaire-based study to examine the role of emotional valence on the diffusion of anti-tobacco messages. They found that positive emotions boost the transmission of messages, while negative ones had the opposite effect.

In [12], Ferrara and Yang extracted emotion polarities for about 19 million tweets by applying the SentiStrength algorithm [40]. In particular, they studied four aspects of information diffusion over Twitter: retweet count, like count, the speed of diffusion, and the scope of the diffusion. The results of the study show a clear evidence of the Pollyanna hypothesis [8], which refers to the human preference to like positive messages more than negative and neutral ones. Moreover, in terms of the scope of the diffusion, the study showed that positive messages spread wider than negative and neutral ones. However, it also indicates that messages carrying negative and neutral sentiments spread faster than positive ones.

Another study that utilized SentiStrength to obtain emotion polarities [37] studied the effects of polarities on the retweet count and the speed of retweeting during the 2011 German state parliament elections. The findings suggest that emotionally charged tweets tend to be retweeted more often than the neutral ones, which is in line with the findings presented in [30, 41, 42]. In particular, tweets carrying a negative sentiment are strongly associated with an increase in the retweet rate.

In [16], Gruzd et al. analyzed sentiment polarities from tweets related to the 2010 Winter Olympics. They found that a user’s position in the social network can be regarded as an indicator of the user’s tendency to post positive or negative messages. Specifically, Gruzd et al. showed that users who tweeted predominantly positive messages generally have more followers on Twitter, while users who tweeted more negative messages exhibited a higher tweet-per-user rate.

In [41], Trung et al. assigned sentiment polarities to a data set of about 11,000 tweets by using a Bayesian classifier trained on the annotated tweets from three domains: news, industry, and entertainment. In particular, they studied three aspects of information diffusion: the number of retweets, speed of diffusion, and the scope of diffusion. In contrast to the findings from [12], Trung et al. found that all emotionally charged messages (i.e., positive as well as negative ones) spread wider (i.e., to more users) than the neutral ones. However, in terms of the speed of diffusion, they found no significant difference among the neutral, positive, and negative messages.

Such a dissonance in the findings might result from the fact that existing papers predominantly study the diffusion patterns only with respect to one particular domain of interest (such as health care, politics, popular culture, or sports; see, e.g., [20, 37]) which makes it difficult to generalize the respective findings across domain borders. While some papers report on diffusion patterns of messages belonging to different domains, the corresponding papers do not follow a systematic approach for studying the differences between domains with respect to information diffusion patterns (see, e.g., [12, 41]).

Even though a number of effects relating to sentiment polarities have been studied in the related work, aspects beyond the effects of sentiment polarities on the information diffusion in OSNs have rarely been investigated. For example, Berger [6] discusses the effects of emotional arousal on information sharing. In particular, the study distinguishes between dimensions of emotions other than emotion polarities only. Berger found that arousal increases the likelihood for sharing an information, regardless of whether the respective information conveys a positive or a negative sentiment. Two other studies [18, 39] consider anger, anxiety, awe, and sadness, as annotated by human encoders. The results of both studies indicate that anger and awe increase the content sharing behavior, while sadness and anxiety were negatively associated with content diffusion.

3 Data Analysis Procedure

In this section, we outline the four main phases of our study (see Fig. 1). In Sect. 3.1 we describe our data extraction procedure. Section 3.2 provides more details on cleaning the raw data set followed by Sect. 3.3 in which we outline the heuristics used to identify emotions and their corresponding emotion scores. Finally, Sect. 3.4 provides details on our data analysis procedure and the scope of this paper.

Fig. 1
figure 1

Research phases

Full size image

3.1 Data Extraction

In order to study the impact of emotional valence shifts on information diffusion, we systematically identified 24 real-world events that belong to five different domains (sports, politics, popular culture, war and terrorism, and others) and collected more than 5.6 million corresponding messages published on Twitter, Facebook, and YouTube. The 24 events have been selected such that they fall in one of the following categories:

  1. 1.

    events that potentially trigger positive emotions (e.g., birthday celebrations, festivities)

  2. 2.

    events that potentially trigger negative emotions (e.g., war, terror, death)

  3. 3.

    emotionally polarizing events (e.g., presidential elections, controversial topics)

We summarize the events extracted for each category in Table 1, where N refers to the number of messages in each category, followed by the relative size of each category in our data set (in percent).

Table 1 List of events extracted from Twitter (T), Facebook (F), and YouTube (Y)
Full size table

To extract the public messages, we used the corresponding APIs offered by Twitter, Facebook, and YouTube. In particular, we used Twitter’s Search APIFootnote 1 to extract publicly available tweets. For this extraction, we used a pre-defined list of hashtags and restricted the search to English-language tweets only. Moreover, we collected the tweets related to the 24 events by starting with the date of an event announcement and stopped 7 days after. In total, the extraction resulted in 4,418,655 tweets. Furthermore, we used YouTube Comment APIFootnote 2 to extract 152,614 publicly available YouTube comments on a set of 98 manually selected YouTube videos related to the 24 events in our study. In addition, we used Facebook’s Graph APIFootnote 3 to extract 1,128,635 publicly available Facebook comments on 69 manually selected Facebook posts related to the 24 events.

3.2 Data Preprocessing

After the data extraction, we cleaned the raw data set by removing entries that contained uninformative content with respect to emotion extraction (e.g., entries that consisted of URLs only). Thus, after preprocessing our data set included messages that contained either text, emoticons, or a combination of both. Moreover, since our Facebook and YouTube data sets included messages that were written in languages other than English, we used Python’s langdetectFootnote 4 language detection library to identify the language of a particular message and removed messages that were not written in English.

The final number of messages for each OSN after preprocessing is shown in Table 1.

3.3 Emotion Extraction

After preprocessing the overall data sets, we further processed each message by lemmatizing it and tagging words to their corresponding part-of-speech category. We then identified the presence of specific emotions conveyed in the messages and computed an emotion intensity score for each message by applying a customized emotion extraction script (see [22] for further details on the heuristics used and an evaluation). In general, the procedure is encoded in the script:

  1. 1.

    identifies the presence of Plutchik’s eight basic emotions (anger, fear, disgust, sadness, joy, trust, anticipation, surprise) [32] by relying on the NRC word-emotion lexicon [27],

  2. 2.

    assigns an intensity score for each emotion in every tweet by counting the number of words in the NRC lexicon that are associated with an emotion and multiplies them with a score provided in the AFINN lexiconFootnote 5 [31],

  3. 3.

    deals with negation (e.g., “I am not happy.”) by shifting the valence of a word (e.g., the term not happy results in a negative emotion score for joy: joy = −1),

  4. 4.

    deals with intensifiers (e.g., very happy), downtoners (e.g., hardly happy), and maximizers (e.g., absolutely happy),

  5. 5.

    identifies misspellings and repeated letters to find “hidden” boosters (e.g., “I am sooooo happy” is regarded as “I am so happy.”),

  6. 6.

    as noted in [19], emoticons are often used instead of words to express emotions. Thus, our script also identifies emoticons and categorizes them as positive (e.g., happy face :) and laughing face :D ), negative (e.g., sad face :( and crying face :’( or broken heart < /3 ), or conditional (heart < 3 ) (see also [1]). Note that we regard a heart ( < 3 ) either as a positive or a negative emotion carrier because its meaning depends on the context of its use. For example, in a sentence “You will be missed < 3,” it is used in a negative context (sadness), while in the sentence “I love him so much < 3,” the emoticon is used in a positive context (joy). Thus, to correctly interpret the emoticon < 3, we first identify the dominant emotion in the tweet and then assign the corresponding emotion score.

Moreover, for our analysis we also extended the NRC dictionary with a list of common acronyms used in social media (such as LOL, WTF, and YOLO).

3.4 Data Analysis and Research Questions

First, we analyzed how social network users express specific emotions during positive, negative, and polarizing events. Next, we separated each data set into a subset that conveys expected emotions and a subset that conveys shifted emotions in terms of their valence. In particular, we treat emotions of a shifted valence as unexpected emotions (e.g., a positive event receives messages that predominantly convey negative emotions). We then analyzed how the user behavior in each subset is influenced by expected and shifted emotions.

Our analysis was guided by the following research questions.

RQ1: Which emotions are expressed during positive, negative, and polarizing events?

For RQ1, we searched for emotions communicated during positive, negative, and polarizing events. For this research question, it was of particular interest whether emotions belonging to a specific emotional valence (i.e., positive or negative) dominate in an event category.

To answer the first research question, we obtained the average intensity of each of the eight basic emotions for each event (see Table 1). Moreover, we computed the bivariate correlation between each pair of emotions.

RQ2: Which messaging behavior do users exhibit during positive, negative, and polarizing events?

For RQ2, we studied how users react to the three types of events (positive, negative, polarizing) in terms of platform-specific user actions. Thus, for Twitter we consider the number of retweets, number of likes, tweeting rate, tweeting count per user, and one-to-one communication. For Facebook, we study the number of replies to a comment, like count, daily time rate, and number of comments per user. And for YouTube, we examine the number of replies and likes to a comment, as well as a daily time rate, and the number of comments per user.

RQ3: Are there differences in the messaging behavior when users are faced with messages that convey expected emotions and those with a shifted emotional valence?

For RQ3, we study how users respond to the emotions conveyed in messages. In particular, we contrast the behavior toward the expected emotions and the shifted emotions and provide a time series analysis of each.

4 Results

In this section, we first show the intensities of emotions expressed in each OSN during positive, negative, and polarizing events (Sect. 4.1). We then examine the user behavior as a reaction to emotionally charged messages and show a time series analysis of the shifted emotions with respect to the expected emotions (Sect. 4.2).

4.1 Emotion Intensity During Positive, Negative, and Polarizing Events

Our analysis shows that OSN users express emotions with a similar intensity over Twitter, Facebook, and YouTube upon encountering polarizing, positive, and negative events (see Figs. 2, 3, and 4 where negative emotions are colored red (anger, sadness, disgust, fear), positive are green (joy, trust), and conditional (i.e., context-dependent) are yellow (surprise, anticipation)).

Fig. 2
figure 2

Emotions expressed during polarizing events on Twitter (T), Facebook (F), and YouTube (Y)

Full size image
Fig. 3
figure 3

Emotions expressed during positive events on Twitter (T), Facebook (F), and YouTube (Y)

Full size image
Fig. 4
figure 4

Emotions expressed during negative events on Twitter (T), Facebook (F), and YouTube (Y)

Full size image

Furthermore, for each category (positive, negative, polarizing), Fig. 5 shows the respective difference between the expected and the shifted emotions. Thus, in positive events, the negative emotion score (shifted) is subtracted from the positive emotion score (expected). In negative events, the positive emotion score (shifted) is subtracted from the negative emotion score (expected). Moreover, since there is no expected emotion for polarizing events, we chose to subtract the positive emotion score from the negative emotion score.

Fig. 5
figure 5

Difference between the emotional intensity of expected and shifted emotions

Full size image

To mitigate bias in the results which may emerge due to the length of a message (i.e., tweets are restricted to 140 characters,Footnote 6 while Facebook and YouTube posts can be considerably longer), 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) with the message count in the data set (N):

$$\displaystyle \begin{aligned} \scalebox{1.5}{\ensuremath{\frac{\sum_{i=1}^{n}\frac{e_i}{S_i}}{N}}}. \end{aligned}$$

We found that messages sent during polarizing events exhibited no tendency of a particular group of emotions to greatly dominate over the other, as compared to the positive and negative events. As shown in Fig. 2, OSN users expressed positive and negative emotions with a similar intensity. For polarizing events, Fig. 5 further shows that the relative difference between the scores assigned to negative and positive emotions only exhibits a low difference (0.02 for Twitter, 0.08 for Facebook, and 0.04 and YouTube). These results were expected to a certain degree, as users tend to either approve/support or disapprove/oppose a topic of interest during polarizing events (e.g., political campaigning).

With respect to OSN-related differences in emotional intensities during polarizing events, we found that our Facebook data set contained 39% emotionally neutral messages, while YouTube and Twitter messages were more emotionally charged (24% and 21% emotionally neutral messages, respectively). These platform-related differences are depicted in Fig. 2.

In contrast, and as shown in Fig. 3, positive events exhibited a higher intensity of positive emotions (joy, trust) as compared to negative emotions (anger, fear, disgust, sadness). In fact, in positive events the differences between the intensities of positive and negative emotions are considerably higher (0.70 for Twitter, 0.40 for Facebook, and 0.26 for YouTube) as compared to the data sets for polarizing and negative events.

Interestingly, when comparing the intensities of specific emotions communicated over the three OSN platforms, we found that a single tweet carries on average a more intense positive emotion, when compared to messages sent via the other two platforms. However, when observing the shifted emotions in positive events (i.e., negative emotions related to positive events), our results reveal that YouTube users tend to express (on average) more intense negative emotions as compared to Facebook and Twitter users. This difference is particularly evident in Fig. 5, where the difference between positive and negative emotional intensities on YouTube is only 0.26, compared to 0.70 on Twitter and 0.40 on Facebook.

Figure 4 shows emotional intensities communicated during negative events. As expected, negative events showed a comparatively higher intensity of negative (expected) emotions on Twitter. However,we also found a considerable presence of positive emotions (see Fig. 4) and only a low difference between the intensities of negative and positive emotions (see Fig. 5). In contrast to Twitter, emotions in messages related to negative events communicated on YouTube and Facebook are even predominantly positive on average (see Fig. 5, the difference between negative and positive emotions is −0.05 on YouTube and −0.04 on Facebook, i.e., the shifted emotion is slightly dominant over the expected emotion).

Next, we examine whether different emotions belonging to the same emotional valence are communicated jointly in a single message. To this end we performed a bivariate correlation analysis for each pair of emotions (e.g., anger with disgust, anger with joy). Our results show a high Spearman’s ρ coefficient between disgust and anger (ρ = 0.81) as well as sadness and fear (ρ = 0.87) on YouTube, between sadness and anger (ρ = 0.92) as well as joy and trust (ρ = 0.86) on Facebook, and between sadness and anger (ρ = 0.70) on Twitter.

During positive events, anger and disgust were highly correlated (ρ = 0.83) on YouTube, while the same holds for sadness and fear (ρ = 0.89) on Facebook. Negative events exhibited a high correlation between disgust and anger (ρ = 0.83), as well as fear and sadness (ρ = 0.82) on YouTube, a high correlation between fear and anger (ρ = 0.83) as well as fear and sadness (ρ = 0.81) on Facebook, and a high correlation between sadness and fear (ρ = 0.71) on Twitter.

Based on the aforementioned results, we conclude that emotions belonging to the same emotional valence tend to be communicated together in a single OSN message. This observation is particularly evident in our Twitter data set (see Fig. 6a), where users are limited to 140 characters only, i.e., Twitter users only have limited space to express their emotions and opinions. We also observed that when users are allowed to post longer messages, there is a higher correlation between positive and negative emotions (e.g., posts that convey joy also convey anger) (see Fig. 6b, c).

Fig. 6
figure 6

Correlation between pairs of negative emotions (N,N), negative and positive emotions (N,P), and pairs of positive emotions (P,P) on (a) Twitter, (b) Facebook, and (c) YouTube during negative (red cross), positive (green dot), and polarizing (yellow star) events

Full size image

However, it is worth mentioning that emotions belonging to two different categories in terms of emotional valence (positive vs. negative) are weakly or at most moderately correlated as compared to emotions belonging to the same emotional valence (e.g., joy and trust; anger and disgust). In a similar way, we found that different negative emotions are only weakly or moderately correlated during positive and polarizing events.

4.2 User Behavior

For the purposes of this paper, OSN user behavior is defined as all user actions that result in sending or forwarding a public message/comment as well as user actions related to appreciating (“liking”) a public message/comment. Moreover, in our analysis we also consider the number of messages per user and the speed of message generation (i.e., the number of messages per time unit).

User Behavior on Twitter

As shown in Table 2, positive events trigger the highest number of retweets and likes. This shows a tendency of users to prefer engaging in positive discussions rather than negative (which confirms the Pollyanna hypothesis, see also [8, 12]). Moreover, we found that users tend to engage in a one-to-one communication (via @username) more frequently during positive events than during negative or polarizing ones. However, the results also 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. This finding corresponds to the ones discussed in [6], which suggests that emotions of a high arousal (such as anger) increase the social transmission of information. Interestingly, in our data set polarizing events exhibited the highest tweeting rate per minute (49.42 tweets/minute).

Table 2 Twitter user behavior in positive, negative, and polarizing events (confidence interval 95%; the highest value for each category is highlighted in bold font respectively)
Full size table

The results of the Welch’s t-test (with a 95% significance level) indicate that there is a significant difference in how users respond to tweets conveying expected emotions and those containing shifted emotions. Apart from one exception in the like count (p > 0.05), all other tweeting behaviors that we considered exhibit a clear pattern: expected emotions receive more retweets. Moreover, for each of the 24 events that we analyzed, OSN users tend to send more tweets per minute that convey an expected emotion. Another interesting insight can be observed in the one-to-one communication (social sharing) of expected emotions between OSN users. During positive events, users tend to engage in a one-to-one communication by sharing predominantly positive emotions. Analogously, negative events exhibit a comparable pattern (see Table 2).

Figure 7a shows the time series of tweets separated into those that carry a predominantly positive emotion and those that carry a negative emotion. In particular, there are a noticeable smaller number of tweets that convey negative (i.e., shifted) emotions during positive events. However, our data also shows that, although small in size, negative tweets occurred consistently throughout the extraction period (mean(set difference)=36668.11, sd(set difference)=45844.64).Footnote 7

Fig. 7
figure 7

Occurrences of negative messages during positive events on Twitter and vice versa. Positive messages are depicted in green and negative messages in red. (a) Positive events and (b) negative events

Full size image

In contrast, we found that during negative events events a considerable number of tweets conveying positive (i.e., shifted) emotions occur (mean(set difference)=4231.67, sd(set difference)=5603.162). This observation was consistent over the entire extraction period (see Fig. 7b). Interestingly, our data set also revealed unexpected cases where the positive tweets (i.e., the shifted emotions) even exceed the (expected) negative tweets (the largest difference between the two subsets is 6403 tweets).

User Behavior on Facebook

Table 3 shows that Facebook users also slightly prefer replying to and liking Facebook posts that have a positive emotion score, while in polarizing events we again found the highest average number of comments per unit of time.

Table 3 Facebook user behavior in positive, negative, and polarizing events (confidence interval 95%; the highest value for each category is highlighted in bold font respectively)
Full size table

In particular, the results of the Welch’s t-test indicate a significant difference in the effects of expected vs. shifted emotions in each data set (see Table 3). For negative events, we found that users tend to reply and comment predominantly on negative posts and also send more messages that convey negative emotions per day, as compared to positive posts (replies t = 2.39, p < 0.05; comments t = 1.19, p < 0.05; time rate t = 2.69, p < 0.05). For positive events, we found one statistically significant result for the comment rate per user (t = 18.32; p < 0.05), which indicates that users tend to comment more on positive posts during positive events rather than on negative posts.

By observing the time series plots in Fig. 8b, we can see that the temporal patterns of expected and shifted emotions during negative events resemble those we found on Twitter. In particular, a considerable number of messages conveying positive emotions are sent during negative events. The positive (shifted) emotions even dominate the negative emotions at certain dates (see the green dots in Fig. 8b).

Fig. 8
figure 8

Occurrences of negative messages during positive events on Facebook and vice versa. Positive messages are depicted in green and negative messages in red. (a) Positive events and (b) negative events

Full size image

With respect to the temporal patterns of negative messages sent during positive events, Fig. 8a shows that positive emotions dominate over the negative ones throughout the entire data extraction period. Again, this observation is analogous to the temporal patterns observed on Twitter (see Fig. 7a).

User Behavior on YouTube

Similar to Facebook and Twitter, YouTube users also prefer to “like” comments on YouTube videos relating to positive events (see Table 4), as compared to comments on YouTube videos relating to polarizing or negative events.

Table 4 YouTube user behavior in positive, negative, and polarizing events (confidence interval 95%; the highest value for each category is highlighted in bold font respectively)
Full size table

However, unlike Facebook and Twitter users, YouTube users exhibit higher reply counts to comments on YouTube videos that are depicting a polarizing event (e.g., political campaigning, such as TV debates). Moreover, we also observed that YouTube users exhibit the highest rate of comments per time unit for videos on positive events, while Facebook users exhibited this behavior for polarizing events and Twitter users for negative events.

Analogously to the results for Facebook and Twitter, Table 4 shows that YouTube users tend to comply with the base mood of an event by replying more to negative messages during negative events (t = 3.55, p < 0.05) and positive messages during positive events (t = 4.16, p < 0.05). In the same way, YouTube users also tend to “like” positive messages during positive events (t = 2.13, p < 0.05) and send more negative comments per time unit during negative events (t = 2.15, p < 0.05).

For our YouTube data set, Fig. 9b shows that during negative events, we again found a considerable number of messages with a positive (i.e., shifted) emotion score. Similar to our findings for Facebook and Twitter, positive messages even dominate over negative messages on certain dates.

Fig. 9
figure 9

Occurrences of negative messages during positive events on YouTube and vice versa. Positive messages are depicted in green and negative messages in red. (a) Positive events and (b) negative events

Full size image

In positive events though, we again predominantly found messages conveying positive emotions (see Fig. 9a).

5 Discussion

Our results bring forth interesting insights into how OSN users behave during positive, negative, and polarizing events when faced with shifted emotions. In our study, we considered OSN user behavior in terms of retweets and one-to-one communication (on Twitter), replies (on YouTube and Facebook), as well as likes, the number of messages per user, and the speed of message generation (time rate) (on all three OSNs). Consistent with previous work from the field of psychology, we found that to a considerable extent, positive emotions also occur during negative events. An explanation for the observed phenomenon can be attributed to social connection [3, 14] as one of the fundamental human needs. In fact, previous studies indicated that even in the times of sorrow and anxiety, people tend to eventually be supportive and positive toward one another (see, e.g., [34]).

In our data set, we found examples of people explicitly calling for social bonding during emotionally tough events (e.g., after the 2016 earthquake in central Italy, people posted: 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). Moreover, our data set indicates that people tend to show appreciation and love for a 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., earthquake in Italy) or war (e.g., Aleppo bombing). Thus, we found empirical evidence that supports the undoing hypothesis [13], which states that people tend to use positive emotions as an antidote to undo the effects of negative emotions.

For Twitter, our results further indicate that expected emotions result in more retweets. We thereby confirm findings from [17], which suggested that people prefer sharing messages that correspond to the emotional valence of the respective event (i.e., users tend to pass along negative tweets during negative events). This might be attributed to the human tendency to conform to the situation. However, we were also able to observe a similar phenomenon on Facebook and YouTube. Beyond the mere sharing of existing messages, we also observed that users in general prefer replying to and liking messages that convey emotions which correspond to the base emotion of an event. The same holds for the message generation time rate and message per user rate on all three OSN platforms we considered in our study.

Other studies bring an additional interesting insight into the interpersonal interactions over social media, which might explain our observations of users to conform to the base emotion. According to [45], emotional messages tend to influence the emotions conveyed in other users’ messages. This phenomenon, called emotion contagion in [45], emerges from the social connections of OSN users (or their position in the network). In this context, we observed that messages sent by “fans” (we follow an assumption that fans follow their idols on OSNs with a high probability) tend to be congruent with the messages sent by their idols. For example, a tweet posted by Pentatonix in which they announce the release of their new album triggered positive reactions from their fans. Note, however, that considering structural network properties alone might not be sufficient to study emotional contagion. For example, OSN users might form a connection (e.g., “follow”) with an influential user (e.g., a politician) even though they do not actually agree with the person’s ideology or point of view. Thus, the emotions passed by an influential user might also be shifted by his/her followers due to disagreement or sarcasm [15]. We leave the study of this issue for our future work.

6 Conclusion

In this paper, we presented a systematic study concerning the influence of emotional valence shifts on the messaging behavior of OSN users on Twitter, Facebook, and YouTube. Our study is based on a data set including 5.6 million messages belonging to 24 real-world events. The events have been subdivided into positive, negative, and polarizing, and for each of these event categories, we analyzed the intensity of Plutchik’s eight basic emotions (sadness, fear, anger, disgust, joy, trust, surprise, anticipation). Thereby, our paper complements existing studies by not only considering polarizing emotion scores (positive vs. negative) but also the influence of eight basic emotions according to Plutchik’s wheel of emotions. In order to study the impact of the eight emotions on user behavior in OSNs, we considered user reactions to emotionally charged messages.

Our findings indicate that people generally prefer sharing messages that correspond to the emotional valence of the respective event. Furthermore, we conducted a time series analysis and found a clear distinction between positive and negative events, with respect to shifted emotions. In particular, we found that positive events trigger a comparatively smaller number of negative messages. However, while negative events exhibit predominantly negative messages, they are accompanied by a surprisingly large number of positive messages. In fact, our analysis shows that in negative events positive messages may even exceed the negative ones on all three OSN platforms. To the best of our knowledge, this is the first study which found empirical evidence that supports the undoing hypothesis in online social networks.

In our future work, we plan to extend our analysis to studying messages written in languages other than English. In addition, we plan to investigate how sarcasm is related to shifts in the emotional valence.