deep feelings

Trento-‐RISE

Sorbonne Univ.

@m_guerini

@stjaco

GOAL

•  Inves?gate the rela?ons between virality of news ar?cles and the emo?ons they are found to evoke. –  Compare different viral phenomena to understand if they are affected by emo?ons in a similar way

–  Compare emo?ons and virality in a cross-‐lingual experimental seKng;

–  Inves?gate the effect of deep cons?tuents of emo?ons in viral phenomena.

What is Virality?

•  Virality: tendency of a content either to spread quickly within a community or to receive a great deal of aRen?on by it.

•  Growing experimental evidence that Virality is a phenomenon with many facets: several different effects of persuasive communica?on are comprised.

•  Our virality indicators: !  Plusones !  Tweets !  Comments

Virality and Persuasion Dimensions

•  NARROWCASTING: ar?cle comments. Readers who comment on the ar?cle are contribu?ng to a discussion among a small subset of the readers of that ar?cle.

•  BROADCASTING: the act of sharing an ar?cle to social networking sites as a form of broadcas(ng.

Standing on the shoulders of giants

•  Small sample of ar?cles manually annotated (∼2,500 documents)

•  Only one virality index (email-‐forward count, aka narrowcas(ng)

•  Just hin?ng to the role of one deep cons?tuents of emo?ons: arousal.

Early works on emo?ons and virality (Berger and Milkman), paving the way for this novel line of research. A few limita?ons, we try to overcome:

Berger, Jonah, and Katherine L. Milkman. "What makes online content viral?.” Journal of Marke(ng Research 49.2 (2012): 192-‐205.

How

•  The mass-‐adop?on of sharing widgets , and recent l y o f emo0onal tagging widgets, on popular websites, provides an impressive amount of data.

Dataset 65,000 news ar?cles and 1.5 million emo?on votes from rappler.com (ENG) and corriere.it (ITA)

53,226 news ar?cles -‐ rappler.com 12,437 news ar?cles -‐ corriere.it

1,145,543 emo?onal votes -‐ rappler.com 320,697 emo?onal votes -‐ corriere.it

A Linear Explana?on

We use simple linear models to analyze the rela?onship between an ar?cle emo?onal characteris?cs and the corresponding impact on virality indices

ID AFRAID AMUSED ANGRY ANNOYED DONT_CARE HAPPY INSPIRED SAD COMMENTS PLUSONES TWEETS

art_10002 75 0 0 0 0 0 25 0 231 3 8 art_10003 0 50 0 16 17 17 0 0 0 54 128 art_10004 52 2 3 2 2 6 2 31 7 12 734 art_10009 0 0 0 0 0 50 50 0 154 6 9

(ENG) Results confirmed …

Rappler CommentsEstimate Std. Err Sigf.

ANNOYED .61 .03 ***ANGRY .54 .02 ***

INSPIRED .23 .02 ***DONT_CARE .18 .04 ***

AMUSED .13 .02 ***HAPPY .10 .02 ***

SAD .07 .02 ***AFRAID -.03 .03 †

Rappler TweetsEstimate Std. Err Sigf.

INSPIRED .52 .02 ***ANGRY .32 .02 ***

ANNOYED .32 .03 ***HAPPY .31 .02 ***

DONT_CARE .30 .04 ***SAD .24 .02 ***

AMUSED .21 .02 ***AFRAID .19 .03 ***

Rappler G+Estimate Std. Err Sigf.

INSPIRED .55 .02 ***ANGRY .25 .02 ***HAPPY .24 .02 ***

AMUSED .21 .02 ***ANNOYED .20 .03 ***

SAD .16 .02 ***AFRAID .14 .03 ***

DONT_CARE .13 .04 ***

Table 3: Emotions impact on rappler.com articles viral indices. ***: p <.001; **: p <.01; *: p <.05; †: not significant – thislegend applies to all tables in this paper.

Corriere CommentsEstimate Std. Err Sigf.

ANNOYED 1.11 .04 ***HAPPY .40 .04 ***

AFRAID .19 .06 **AMUSED .13 .06 *

SAD .12 .05 *

Corriere TweetsEstimate Std. Err Sigf.

ANNOYED .33 .03 ***SAD .20 .05 ***

HAPPY .14 .03 ***AFRAID .06 .05 †

AMUSED .02 .05 †

Corriere G+Estimate Std. Err Sigf.

SAD .57 .05 ***ANNOYED .39 .03 ***

AMUSED .34 .05 ***AFRAID .33 .05 ***HAPPY .27 .03 ***

Table 4: Emotions impact on corriere.it articles viral indices.

versely, no previous studies have dealt at this scale with Italianlanguage, and it will be worth investigating in future works whatfactors may influence the trend shown for emotional dimensions inthe corriere.it data.

Turning to the statistics of the various viral indices available forthe two datasets, we provide a summary in Table 5.

Viral Index Rapplerµ CorriereµCOMMENTS 4.11 83.22THREADS 2.81 -G+ SHARES .91 3.79TWITTER SHARES 32.33 40.65FACEBOOK SHARES - 502.93

Table 5: Mean figures for virality indices.

From now on, we will only consider the intersection of the twosets of indices.

3.1 Narrow- and Broad- castingIn the literature, broadcasting refers to the act of communicating

or transmitting a content to numerous recipients simultaneously,over a communication network; on the contrary, narrowcasting hastraditionally been understood as the dissemination of informationto a narrow audience, rather than to the broader public at large.With the advent of new media, this definition as been updated (see,for instance, [15]). In fact, from the perspective of readers of onlinecontent, they can decide whether and how to share such content byeither broadcasting it to their audience or to address only a smallportion of it (narrowcasting). In [4], for example, the act of for-warding a news article by email is treated as a form of narrowcast-ing, since in such case subjects are addressing a selected section oftheir audience/contacts.

Following the above definition, we will consider article com-ments as a form of narrowcasting, as the readers who upload acomment on the article page are contributing to a discussion whichhappens at most between the readers of that article (and most prob-ably, in fact, to a small subset of it). On the other hand, we considerthe act of sharing an article to social networking sites as a form ofbroadcasting.

The rationale behind this distinction of viral facets in narrow-casting and broadcasting lies in previous research efforts showing

that virality can be assessed through different metrics which repre-sent distinct phenomena: it is not only the magnitude of spreading(virality) that depends on content but also the users reactions (com-ments, shares, tweets, etc.) [10, 11, 28].

In the following sections, all virality indices have been standard-ized in order to make them comparable (µ = 0,� = 1) across thetwo datasets. Emotion scores range between 0 and 1.

4. EMOTION ANALYSISTo analyze the relationship between an article emotional char-

acteristics and the corresponding impact on virality indices we usesimple linear models. In this section we will discuss the importanceof each emotion for the various virality indices, while the overallexplanatory power of the models will be discussed in Section 6.

Results on the rappler.com dataset, shown in Table 3, are inaccordance with the findings of Berger et al. [4]: high influence ofINSPIRING (awe) and of negative-valence and high-arousal emo-tions such as ANGER, jointly with a low influence of SADNESS.This similarity stands out for the specific case of narrowcasting,and it is maintained for broadcasting, albeit to a lower extent.

As mentioned in Section 2, other previous research works [9,16], focusing on diverse scenarios and using heterogeneous datasets,have reported results consistent with these findings.

Nonetheless, turning to the Italian resource used in our analy-ses, we notice that results obtained from the corriere.it data,summarized in Table 4, are partially in line with [4] for what con-cerns narrowcasting, while drastically diverge on broadcasting:in the latter case, SADNESS (a low-arousal emotional dimension)is found to be most relevant for virality, disproving the hypothe-sis that arousal alone can be used to explain virality phenomena.It should be noted that the lack of an INSPIRED dimension forcorriere.it cannot account for such discrepancy on broad-casting, since it seems very unlikely that its potential votes wouldhave been collected by SADNESS.

It can be hypothesized, as a consequence, that strong culturaldifferences emerge when it comes to emotions (more precisely, toexplicitly tagging one’s own feeling on a website). What factorsmight underlie this phenomenon (e.g. historical period, editorialchoices, deeper cultural sensibilities, etc.) represents a very fasci-nating research question, which is out of the scope of this work.

Rather than focusing on specific emotions, in the next section weattempt to provide a more general explanation.

Results on rappler.com in line with Berger and Milkman: •  high influence of INSPIRING (awe) •  high influence of nega?ve-‐valence and high-‐arousal emo?ons such as ANGER •  low influence of SADNESS.

(ITA) … or not? Rappler Comments

Estimate Std. Err Sigf.ANNOYED .61 .03 ***

ANGRY .54 .02 ***INSPIRED .23 .02 ***

DONT_CARE .18 .04 ***AMUSED .13 .02 ***

HAPPY .10 .02 ***SAD .07 .02 ***

AFRAID -.03 .03 †

Rappler TweetsEstimate Std. Err Sigf.

INSPIRED .52 .02 ***ANGRY .32 .02 ***

ANNOYED .32 .03 ***HAPPY .31 .02 ***

DONT_CARE .30 .04 ***SAD .24 .02 ***

AMUSED .21 .02 ***AFRAID .19 .03 ***

Rappler G+Estimate Std. Err Sigf.

INSPIRED .55 .02 ***ANGRY .25 .02 ***HAPPY .24 .02 ***

AMUSED .21 .02 ***ANNOYED .20 .03 ***

SAD .16 .02 ***AFRAID .14 .03 ***

DONT_CARE .13 .04 ***

Table 3: Emotions impact on rappler.com articles viral indices. ***: p <.001; **: p <.01; *: p <.05; †: not significant – thislegend applies to all tables in this paper.

Corriere CommentsEstimate Std. Err Sigf.

ANNOYED 1.11 .04 ***HAPPY .40 .04 ***

AFRAID .19 .06 **AMUSED .13 .06 *

SAD .12 .05 *

Corriere TweetsEstimate Std. Err Sigf.

ANNOYED .33 .03 ***SAD .20 .05 ***

HAPPY .14 .03 ***AFRAID .06 .05 †

AMUSED .02 .05 †

Corriere G+Estimate Std. Err Sigf.

SAD .57 .05 ***ANNOYED .39 .03 ***

AMUSED .34 .05 ***AFRAID .33 .05 ***HAPPY .27 .03 ***

Table 4: Emotions impact on corriere.it articles viral indices.

versely, no previous studies have dealt at this scale with Italianlanguage, and it will be worth investigating in future works whatfactors may influence the trend shown for emotional dimensions inthe corriere.it data.

Turning to the statistics of the various viral indices available forthe two datasets, we provide a summary in Table 5.

Viral Index Rapplerµ CorriereµCOMMENTS 4.11 83.22THREADS 2.81 -G+ SHARES .91 3.79TWITTER SHARES 32.33 40.65FACEBOOK SHARES - 502.93

Table 5: Mean figures for virality indices.

From now on, we will only consider the intersection of the twosets of indices.

3.1 Narrow- and Broad- castingIn the literature, broadcasting refers to the act of communicating

or transmitting a content to numerous recipients simultaneously,over a communication network; on the contrary, narrowcasting hastraditionally been understood as the dissemination of informationto a narrow audience, rather than to the broader public at large.With the advent of new media, this definition as been updated (see,for instance, [15]). In fact, from the perspective of readers of onlinecontent, they can decide whether and how to share such content byeither broadcasting it to their audience or to address only a smallportion of it (narrowcasting). In [4], for example, the act of for-warding a news article by email is treated as a form of narrowcast-ing, since in such case subjects are addressing a selected section oftheir audience/contacts.

Following the above definition, we will consider article com-ments as a form of narrowcasting, as the readers who upload acomment on the article page are contributing to a discussion whichhappens at most between the readers of that article (and most prob-ably, in fact, to a small subset of it). On the other hand, we considerthe act of sharing an article to social networking sites as a form ofbroadcasting.

The rationale behind this distinction of viral facets in narrow-casting and broadcasting lies in previous research efforts showing

that virality can be assessed through different metrics which repre-sent distinct phenomena: it is not only the magnitude of spreading(virality) that depends on content but also the users reactions (com-ments, shares, tweets, etc.) [10, 11, 28].

In the following sections, all virality indices have been standard-ized in order to make them comparable (µ = 0,� = 1) across thetwo datasets. Emotion scores range between 0 and 1.

4. EMOTION ANALYSISTo analyze the relationship between an article emotional char-

acteristics and the corresponding impact on virality indices we usesimple linear models. In this section we will discuss the importanceof each emotion for the various virality indices, while the overallexplanatory power of the models will be discussed in Section 6.

Results on the rappler.com dataset, shown in Table 3, are inaccordance with the findings of Berger et al. [4]: high influence ofINSPIRING (awe) and of negative-valence and high-arousal emo-tions such as ANGER, jointly with a low influence of SADNESS.This similarity stands out for the specific case of narrowcasting,and it is maintained for broadcasting, albeit to a lower extent.

As mentioned in Section 2, other previous research works [9,16], focusing on diverse scenarios and using heterogeneous datasets,have reported results consistent with these findings.

Nonetheless, turning to the Italian resource used in our analy-ses, we notice that results obtained from the corriere.it data,summarized in Table 4, are partially in line with [4] for what con-cerns narrowcasting, while drastically diverge on broadcasting:in the latter case, SADNESS (a low-arousal emotional dimension)is found to be most relevant for virality, disproving the hypothe-sis that arousal alone can be used to explain virality phenomena.It should be noted that the lack of an INSPIRED dimension forcorriere.it cannot account for such discrepancy on broad-casting, since it seems very unlikely that its potential votes wouldhave been collected by SADNESS.

It can be hypothesized, as a consequence, that strong culturaldifferences emerge when it comes to emotions (more precisely, toexplicitly tagging one’s own feeling on a website). What factorsmight underlie this phenomenon (e.g. historical period, editorialchoices, deeper cultural sensibilities, etc.) represents a very fasci-nating research question, which is out of the scope of this work.

Rather than focusing on specific emotions, in the next section weattempt to provide a more general explanation.

Turning to corriere.it: •  par?ally in line with (Berger and Milkman) for what concerns narrowcas(ng •  dras?cally diverge on broadcas(ng: SADNESS is found to be most relevant for virality,

disproving that arousal alone can be used to explain virality phenomena.

Idea

Rather than focusing on specific differences in the linear models, we look for configura?ons in the deeper structure of emo?ons that can account for such discrepancies.

The VAD Model

•  the Valence, Arousal, and Dominance (VAD) circumplex model of affect: –  Valence, deno?ng the degree of posi?ve/nega?ve affec?vity (e.g. FEAR has high nega?ve valence, while JOY has high posi?ve valence);

– Arousal, ranging from calming to exci?ng (e.g. ANGER is denoted by high arousal while SADNESS by low arousal);

– Dominance, going from “controlled” to “in control” (e.g. INSPIRED, highly in control, vs FEAR, overwhelming).

Mapping to the VAD model •  Exploit the work of Warriner et al.: a resource mapping 14

thousands words to the VAD circumplex model on a Likert scale, including words represen?ng our emo?onal labels.

•  To compute VAD scores for a given ar?cle doc we mul?ply the percentage of votes each emo?on e is found to evoke in doc by the corresponding VAD score, and then take the sum over the n emo?onal dimensions considered.

corriere.it commentsEstimate Std. Err Sigf.

AROUSAL .3741 .0372 ***VALENCE -.3155 .0486 ***DOMINANCE .1000 .0761 †

rappler.com commentsEstimate Std. Err Sigf.

AROUSAL .1205 .0101 ***VALENCE -.1636 .0165 ***DOMINANCE .0728 .0221 **

corriere.it tweetsEstimate Std. Err Sigf.

DOMINANCE .2158 .0709 **VALENCE -.2176 .0453 ***AROUSAL .0400 .0348 †

rappler.com tweetsEstimate Std. Err Sigf.

DOMINANCE .2244 .0221 ***VALENCE -.1495 .0165 ***AROUSAL .0065 .0101 †

corriere.it g+ sharesEstimate Std. Err Sigf.

DOMINANCE .4817 .0704 ***VALENCE -.3781 .0450 ***AROUSAL -.0242 .0345 †

rappler.com g+ sharesEstimate Std. Err Sigf.

DOMINANCE .2619 .0221 ***VALENCE -.1530 .0165 ***AROUSAL -.0371 .0101 ***

Table 7: Valence, Arousal and Dominance significant effects from simple Linear Models.

EMOTION Valence Arousal DominanceAFRAID 2.25 5.12 2.71AMUSED 7.05 4.27 5.93ANGRY 2.53 6.02 4.11ANNOYED 2.80 5.29 4.08DONT_CARE 3.53 4.27 3.62HAPPY 8.47 6.05 7.21INSPIRED 6.89 5.56 7.30SAD 2.10 3.49 3.84

Table 6: Valence, Arousal and Dominance scores for emotionlabels, as provided by [34].

5. VAD ANALYSISWe now turn to investigate how basic constituents of emotions,

such as Valence, Arousal, and Dominance (VAD), connect to vi-rality. The widely adopted VAD circumplex model of affect [6,27] maps emotions on a three-dimensional space, namely: Valence,denoting the degree of positive/negative affectivity (e.g. FEAR hashigh negative valence, while JOY has high positive valence); Arousal,ranging from calming to exciting (e.g. ANGER is denoted by higharousal while SADNESS by low arousal); and Dominance, goingfrom “controlled” to “in control” (e.g. INSPIRED, highly in con-trol, vs FEAR, overwhelming).

We thus proceed to map the emotions an article is found to evoketo the VAD circumplex model. To do so, we exploit the work ofWarriner et al. [34], who provided a resource mapping roughly 14thousands words to the VAD circumplex model, including wordsrepresenting the emotional dimensions we consider in this work –see Table 6, which for us serve as a gold standard. Such scores arein a Likert scale, ranging from 1 (low/negative) to 9 (high/positive).

Then, in order to compute VAD scores for a given article doc wesimply multiply the percentage of votes each emotion e is found toevoke in doc by the corresponding VAD score provided in Table 6,and then take the sum over the n emotional dimensions considered.The formula for Valence is provided below; the equations used forArousal and Dominance are akin.

docv =nX

i=1

votes%(ei)⇥ valence(ei) (1)

As with virality indices, each VAD dimension has been stan-dardized. Subsequently, we build linear models in order to as-sess whether and how VAD dimensions are connected to viral-ity. The results reported in Table 7 show very interesting trends:VAD dimension estimates are found to be consistent among thetwo datasets (see estimate order and sign in Table 7), in both broad-casting (tweets/g+ shares) and narrowcasting (comments) scenar-ios. Comparing these results with those provided in the previoussections, we see that the cross-cultural divergences in the relationsbetween emotional dimensions and virality indices disappear whenaccounting for their more profound VAD components.

This finding hints at a generalized, culturally indipendent phe-nomenon: readers of the articles in our datasets tend to choosecommunication forms of narrowcasting when the content is arous-ing but on which they feel less in control2.

Conversely, they turn to broadcasting (e.g. share on social net-works) when they feel more in control. This result is further sup-ported by the fact that the less important dimensions (i.e. Domi-nance for narrowcasting, and Arousal for broadcasting) are mostof the time not significant, or with a slight negative impact on themodel, indicating that these two dimensions switch roles when tran-sitioning from narrowcasting to broadcasting (and viceversa). Thisfinding is important as it appears to be valid, in our analyses, amongthe two different languages (and cultural factors) our two datasetsprovide.

Finally, the role of Valence appears to be consistent among allindices of virality in both datasets, with negative valence contribut-ing to higher virality. This result is in line with [14], who foundthat news-related content spreads more when imbued with negativeValence.

6. R2 ANALYSISIn this section we examine the explanatory power of our mod-

els and compare them with the results obtained by the models pre-sented in Berger et al. [4] – in particular, with model 2 (Positivityand Emotionality) and model 3 (Positivity and Emotionality plusthe emotions Awe, Anger, Anxiety, Sadness).

2Again, this result is in line with the intuition in [4], that arousalhas a strong effect on narrowcasting. Nonetheless, it appears ofparamount importance to take the role of Dominance (not consid-ered in [4]) into account in order to understand broadcasting phe-nomena.




















docv =nX

i=1









…

VAD Discussion

•  Cross-‐cultural divergences disappear when using VAD components.

•  A generalized, culturally independent phenomenon: –  narrowcas0ng when the content is arousing. –  broadcas0ng when there is control (dominance).

•  Further evidence: less important dimensions (i.e. Dominance for narrowcas(ng, and Arousal for broadcas(ng) are most of the ?me not significant. These two dimensions switch roles when transi?oning from narrowcas?ng to broadcas?ng (and viceversa).

VAD linear Models




















docv =nX

i=1









R2 ANALYSIS

We compare our VAD models with model 2, which was automat-ically computed starting from a lexicon of relevant affective words(the LIWC lexicon described in [25]), and our emotion-based mod-els with model 3 in [4]. It should be noted that the extensive workpresented in [4] includes other models accounting for variables(such as publication time, homepage position, author, among oth-ers) not available in our datasets. Nonetheless, it has been shownin [4] that, while augmenting the explanatory power, these variablesdo not influence weight and role of the emotions in the models.

In [4], the dependent variable is represented as binary (i.e., 1 forthose articles that make it on the most emailed list, 0 for all the oth-ers). On the contrary, our viral indices are originally represented ascontinuous variables. In order to provide a fair comparison we thusproceeded to binarize the virality indices V of article n, accordingto the following scheme:

Vn =

(1, if Vn > mean(V ) + sd(V )

0, otherwise(2)

In this way, we obtain a small sample of the most viral articlesin our dataset, specular to the most emailed list in [4]. We alsoused logistic regression in accordance with [4]. For comparison,we report in table 8 the McFadden’s R2 used by Berger et al. [4],along with the standard R2.

Viral Index R

2 McFadden’s R2

corriere.it emotion modelsComments .0813 .0922G+ .0207 .0527Tweets .0084 .0604

corriere.it VAD modelsComments .0530 .0840G+ .0195 .0446Tweets .0072 .0591rappler.com emotion models

Comments .0191 .0801G+ .0115 .0314Tweets .0112 .0300

rappler.com VAD modelsComments .0101 .0569G+ .0088 .0288Tweets .0100 .0266

Berger et al. modelsModel 2 - .0400Model 3 - .0700

Table 8: R2 scores for the various linear models described andcomparison with models presented in Berger et al. [4]

Hence, when projecting emotions to their basic VAD dimen-sions, the models experience only a slight drop in terms of explana-tory power, while gaining the advantage of language-invariance.

Moreover, our regression models for narrowcasting have greaterexplanatory power than the best performing model in [4]. This factcan be partially explained by the quality of our data: our datasetsare much bigger and the affective annotations of the news articleswere massively crowdsourced to the readers.

Finally, emotions have significant effects on both narrowcastingand broadcasting, with a stronger impact on the former. Again, thisfinding is cross-cultural consistent.

7. CONCLUSIONSIn this article we have provided a comprehensive investigation

on the relations between virality of news articles and the emotionsthey are found to evoke. By exploiting a high-coverage and bilin-gual corpus of 65k documents containing metrics of their spread onsocial networks as well as a massive affective annotation providedby readers (more than 1.5 millions votes), we presented a thoroughanalysis of the interplay between evoked emotions and viral facets.

We highlighted and discussed our findings in light of a cross-lingual approach.

Our results show differences in evoked emotions and correspond-ing viral effects across the bilingual data we crawled; amongst otherfindings, we note the remarkably discordant influence of the SAD-NESS affective dimension between the english- and italian- lan-guage datasets, in contrast with the hypothesis that arousal alonecan be used to explain virality phenomena. While these findingsseem to indicate effects of cultural differences on the relation ex-isting between emotions and virality, such hypothesis could onlybe properly tested were more information on the annotators demo-graphics available (e.g. geographic provenance).

Still, when accounting for the deeper constituents of emotionsour analyses provided compelling evidence of a generalized ex-planatory model rooted in the Valence-Arousal-Dominance (VAD)circumplex. Viral facets seem to be coherently affected by partic-ular VAD configurations (namely, the alternation between Domi-nance and Arousal), and these configurations indicate a clear con-nection with distinct phenomena underlying persuasive communi-cation. In particular, high arousal is more connected to narrowcast-ing phenomena, while Dominance to broadcasting phenomena.

Extensions of this work will include deeper investigations of therelations between VAD dimensions and virality. For instance, whilein the analyses reported herein we have used a resource [34] to mapaffective tags to their VAD constituents, we plan to increase theresolution by associating each article with a VAD representationderived from its textual content.

The results presented in this article can be very valuable in con-texts such as content marketing and native advertising, and thus berelevant not only for social science researchers interested in under-standing the factors behind virality phenomena, but also for mar-keting and industry people.

AcknowledgementsThe work of M. Guerini has been supported by the Trento RISEPerTe project. The work of J. Staiano has been partially funded bythe French government through the National Agency for Research(ANR) under the Sorbonne Universités Excellence Initiative pro-gram (Idex, reference ANR-11-IDEX-0004-02) in the frameworkof the Investments for the future program, and by the German Fed-eral Ministry of Research and Education, within the EBITA project.

8. REFERENCES[1] J. Z. Aaditeshwar Seth and R. Cohen. A multi-disciplinary

approach for recommending weblog messages. In The AAAI2008 Workshop on Enhanced Messaging, 2008.

[2] T. Althoff, C. Danescu-Niculescu-Mizil, and D. Jurafsky.How to ask for a favor: A case study on the success ofaltruistic requests. Proocedings of ICWSM, 2014.

[3] S. Aral and D. Walker. Creating social contagion throughviral product design: A randomized trial of peer influence innetworks. Management Science, 57(9):1623–1639, 2011.

[4] J. Berger and K. L. Milkman. What makes online contentviral? Journal of Marketing Research, 49(2):192–205, 2012.

•  When using VAD dimensions, only a slight drop

in terms of explanatory power, while gaining the advantage of language-‐invariance.

•  Our regression models for narrowcas?ng have greater explanatory power than in Berger and Milkman. Par?ally explained by the quality of our data: much bigger dataset and affec?ve annota?ons massively crowdsourced.

•  Emo?ons have significant effects on both narrowcas?ng and broadcas?ng: with a stronger impact on the former. Again, this finding is cross-‐cultural consistent.

Conclusions

•  A study on rela?ons between emo?ons and virality. •  While there seem to be cultural differences in these rela?ons, projec?ng documents onto deep structure of emo?ons these differences disappear. – narrowcas0ng when the content is arousing. – broadcas0ng when there is control (dominance).

Further details: M. Guerini and J. Staiano. 2015. Deep Feelings: A Massive Cross-‐Lingual Study on the Rela?on between Emo?ons and Virality. In Proceedings of WWW '15 Companion, 299-‐305.

deep feelings

Social Media