Affective Polarization in Online Climate Change Discourse on Twitter
AAffective Polarization in Online Climate ChangeDiscourse on Twitter
Aman Tyagi
CASOS, Engineering and Public PolicyCarnegie Mellon University
PIttsburgh PA, [email protected]
Joshua Uyheng
CASOS, Institute for Software ResearchCarnegie Mellon University
Pittsburgh PA, [email protected]
Kathleen M. Carley
CASOS, Institute for Software ResearchCarnegie Mellon University
Pittsburgh PA, [email protected]
Abstract —Online social media has become an important plat-form to organize around different socio-cultural and politicaltopics. An extensive scholarship has discussed how people aredivided into echo-chamber-like groups. However, there is a lackof work related to quantifying hostile communication or affectivepolarization between two competing groups. This paper proposesa systematic, network-based methodology for examining affectivepolarization in online conversations. Further, we apply our frame-work to 100 weeks of Twitter discourse about climate change. Wefind that deniers of climate change (Disbelievers) are more hostiletowards people who believe (Believers) in the anthropogenic causeof climate change than vice versa. Moreover, Disbelievers usemore words and hashtags related to natural disasters duringmore hostile weeks as compared to Believers. These findingsbear implications for studying affective polarization in onlinediscourse, especially concerning the subject of climate change.Lastly, we discuss our findings in the context of increasinglyimportant climate change communication research.
Index Terms —climate change, affective polarization, stancedetection, online social networks
I. I
NTRODUCTION
Online social networks represent a powerful space for publicdiscourse. Through large-scale, interconnected platforms likesocial media, diverse communities may potentially participatein open exchanges of views and information about a vast rangeof issues. However, research has increasingly demonstratedthe dangers of polarization in online communication [1]–[3].Attributed to various psychological, social, and technologicalfactors, intergroup communication on cyberspace has dis-played tendencies to feature pathological dynamics especiallyconcerning contentious issues [4], [5].Polarization on social media could be broadly divided intodifferent categories. Opposed groups may communicate ina highly balkanized fashion, such that members of an in-group are only minimally exposed to out-group membersand their beliefs [6], [7]. This phenomenon has been termed
This work was supported in part by the Knight Foundation and the Officeof Naval Research grants N000141812106 and N000141812108. Additionalsupport was provided by the Center for Computational Analysis of Social andOrganizational Systems (CASOS), the Center for Informed Democracy andSocial Cybersecurity (IDeaS), and the Department of Engineering and PublicPolicy of Carnegie Mellon University. The views and conclusions containedin this document are those of the authors and should not be interpreted asrepresenting the official policies, either expressed or implied, of the KnightFoundation, Office of Naval Research or the U.S. government. interactional polarization . Polarization can also pertain tohighly negative sentiments toward out-groups in the formof affective polarization [8], [9]. Social scientific researchexamines how these phenomena are interconnected across avariety of contexts, such that online groups that disagree ona given topic are also more likely to be hostile toward eachother [10]. In this paper, we focus on quantifying affectivepolarization between two groups with opposing beliefs usingTwitter discourse on a significant social issue.One significant issue which has received heated attentionin online public discourse is climate change [11]–[14]. Wefocus on those who cognitively accept anthropogenic causesof climate change (
Believers ) and those who reject the same(
Disbelievers ). Previous work demonstrates not only sharpdivergences in climate change beliefs but also the emergenceof communities insulated from the opposed group [14]–[16].In other words, online discussions about climate change are interactionally polarized , implying the persistence of echochambers between Believers and Disbelievers [17]–[19].Much less work, however, engages the question of affectivepolarization in online climate change discourse. A cruciallimitation in prior work lies in the methodological optionsavailable to past researchers. Relying consistently on manuallyannotated corpora and datasets of limited size, existing schol-arship has faced barriers to measuring the emotional compo-nent of climate change discussions in a generalizable fashion[8], [17], [18]. Drawing on recent advances in computationalstance detection, targeted sentiment analysis, and networkscience measures, we present an integrated methodologicalpipeline for addressing this gap in the literature.More specifically, we show how computational methodsmay be leveraged to generate (a) automated stance labelsfor climate change Believers and Disbelievers, (b) individualmeasurements of the interaction valence between in-group andout-group members, and (c) broader assessments of group-level affective polarization. We demonstrate the utility ofour framework by applying our methodology to a large-scaledataset of 100 weeks of online climate change discussion onTwitter. Furthermore, we link our findings to natural disasterswords to explain important climate change belief constructs.In sum, this work proposes to answer the following researchquestions: a r X i v : . [ c s . S I] A ug ) How can affective polarization be computationally mea-sured on a large-scale, long-term corpus of online cli-mate change discussions?2) Do climate change Believers or Disbelievers featuregreater levels of affective polarization in online publicdiscourse?3) What is the relationship of affective polarization betweenthe two groups with use of natural disaster relatedwords ?The subsequent sections of this paper are organized asfollows. First, we provide an overview of related work inthis area, illustrating computational analysis of polarizationin general terms and then in the case of climate changespecifically. We zero in on the dearth of principled empiricalwork on affective polarization specifically in relation to onlineclimate change discourse. Second, we present our proposedmethodological pipeline which integrates machine learningmodels and network science techniques to facilitate a noveland effective framework for assessing affective polarization.Third, we share our findings on our large-scale, long-termTwitter dataset. Last, we discuss implications for understand-ing the state of climate change discourse on digital platformsas well as related empirical investigation of affective polariza-tion on online social networks.II. R ELATED W ORK
A. Computational analysis of polarization
Recognizing the ubiquity of online conflicts, rigorous schol-arship in the computational and social sciences has tackledthe problem of polarization. More traditional approaches inoffline settings have relied on survey measures to empiricallyassess divergence in beliefs between groups [9], [20]. Butwith burgeoning developments in computational methods -especially with respect to natural language processing andmachine learning - automated methods have also arisen toleverage the vast digital traces linked to online activity [21],[22].General approaches infer individual attitudes from userinformation, such as the texts associated with an accounton social media (e.g., Facebook comments, tweets). Groupmembership as well as group communication are similarlyincorporated into analyses of polarization, by examining thebeliefs of individuals in conjunction with their traceablepatterns of digital interaction with other individuals. Givenvarious conceptualizations of polarization, different frame-works have been developed to quantify pathological patternsof communication across groups holding similar or opposedstances on a given issue [23]–[26].Social network analysis has gained much methodologicaltraction in this regard. Representing online conversations asgraphical structures, numerous approaches measure polariza-tion as a function of homophily in local community structures[7], [27]. In other words, the extent to which those holding We provide the list of natural disaster related words used in our analysisin §VI similar views are more likely to interact with each other - incontrast to those with whom they disagree - allows an intuitiveand principled measure of polarization. For example, randomwalk scores quantify the probability of a random walk startingfrom a node belonging to a given stance group ending up ina node belonging to the same or a different stance group [2],[28], [29].More recent scholarship, however, emphasizes the impor-tance of examining not just pathologically isolated commu-nication, or interactional polarization; but also pathologicallyhostile communication, or affective polarization. Burgeoningevidence suggests that the problem of echo chambers repre-sents a significant, yet incomplete, picture of polarization inonline social networks [1]. People holding opposed views, infact, do interact with each other - but this does not necessarilymitigate polarization [6]. Instead, research finds that theseintergroup exposures trigger further incivility and toxicity [20].Hence, reliable measures for affective polarization are needed,although the computational literature in this area remains inits nascent stages [10].
B. Climate change and polarization
In the specific case of climate change discourse, analysisof polarization has also represented a major research topic.Numerous studies link polarized beliefs about climate changeto partisan divides, with more conservative individuals lesslikely to cognitively accept anthropogenic climate change thanliberals [11], [15]. Past work specifically demonstrates thatalthough higher levels of education and information accessmay increase the likelihood of climate change belief, theseeffects remain much lower among conservatives [13], [15].Such effects have been explained from the lens of elitesignalling - whereby followers emulate the beliefs of theirpreferred political leaders - uneven exposure to informationbased on partisan media, as well as a generalized dislike forthe members of the opposed ideological group [30]–[32].However, with time, scholars have also noted general trendstoward increasing climate change beliefs overall [16]. Even ifthese do not necessarily translate into concrete support forpolicy instruments to address climate issues [12], the long-term instability of climate change skepticism points to valuableways forward for science communication [33]. Collectively,these finding suggest the importance of accounting for thepsychological processes surrounding climate change belief anddisbelief, going beyond the transmission of information [34].These issues take on specific forms in cyberspace, whereinformation flows are inextricably entangled with communitydynamics. Studies employing social network analysis have un-covered robust evidence that online climate change discussionstend to exhibit echo chamber-like homophilic interactions[14], [19], [35]. Qualitative analysis further showed that inrare instances of intergroup communication, more negativeframes tended to prevail, featuring dismissal of climate changeinformation as hoaxes, derailment of conversations to heatedissues of identity, as well as overall higher levels of sarcasmand incivility [8], [17], [18]. Notwithstanding the valuablediographic insights derived from these studies, their samplingstrategies have tended to rely on a minuscule fraction ofthe larger conversation to facilitate in-depth content analysis.Hence, larger-scale and more generalizable findings on theaffective dynamics of online climate change discourse arenotably lacking in the literature.
C. Contributions of this work
Motivated by the foregoing insights, this work seeks tocontribute to the literature by offering a methodologicalpipeline for examining affective polarization. As the succeed-ing sections demonstrate, our framework combines machinelearning and network science methods in a novel, scalable, andgeneralizable fashion for ready application in a variety of con-tentious issues. This overcomes many of the methodologicalbarriers present in prior work, including their common relianceon expensive survey or experimental measures, or manuallyannotated datasets in the context of social media research onclimate change discourse [13], [15], [16], [19], [31].From a theoretical standpoint, we additionally contributea nuanced operationalization of affective polarization as lo-cated on a group level. We unpack how group-level metricsvaluably produce asymmetrical views of hostile behavior,thereby facilitating more fine-grained analysis of how differentstance groups engage in varied levels of affectively polarizedinteractions. This conceptually aligns with the asymmetry ofpsychological factors characterizing climate change Believersand Disbelievers, especially over time [11], [32], [33].Finally, on an empirical level, our work also extends pre-vailing scholarship on polarized climate change discourse.While established findings paint a picture of consistent echochambers between climate change Believers and Disbeliev-ers, we provide evidence for the flipside of these dynamics.We specifically quantify, over a larger-scale and longer-termdataset than previously examined in prior work, the extent towhich intergroup interactions systematically feature hostility.This may inform possible data-driven interventions for poli-cymaking beyond more prevalent frames of intergroup contactand science communication [28], [34].III. D
ATA AND M ETHODS
A. Data collection
We collected realtime tweets using Twitter’s standard API with keywords “Climate Change”, “ https://developer.Twitter.com/en/docs/tweets/search/overview/standard B. Stance labels
We use a state-of-the-art stance mining method [22] tolabel each user as a climate change Disbeliever or Believer.We use a weak supervision based machine learning model tolabel the users in our dataset. The model uses a co-trainingapproach with label propagation and text-classification. Themodel requires a set of seed hashtags essentially being used byBelievers and Disbelievers. The model then labels seed usersbased on the hashtags used at the end of the tweet. Usingthe seed users, the model trains a text classifier and uses acombined user-retweet and user-hashtag network to propagatelabels. In an iterative process, the model then labels users whoare assigned a label by both methods with high confidence .We set ClimateChangeIsReal and
SavetheEarth as Believersseed hashtags and
ClimateHoax and
Qanon as Disbelieversseed hashtags. These hashtags have been shown to be usedmostly by the respective groups [14]. Out of the total 7Musers, the algorithm labels 3.9M as Believers and 3.1M asDisbelievers. We randomly sampled 500 users from each groupto manually validate the results. We label a user as Disbelieverif we find any Tweet akin to someone who does not believein climate change or anthropogenic cause of climate change.Otherwise, we label the user as Believer. We observe thatthe average precision from manual validation of 1000 usersis 81.8%.
C. Affective polarization metrics
We measure affective polarization in this work by combin-ing outputs from an aspect-level sentiment model, a classicnetwork science measure known as the E/I index [36] andEarth Mover’s Distance (EMD) [37]. Outputs are combinedin the five steps which follow to produce dynamic group-levelmeasurements of affective polarization.
1) Aspect-level sentiment:
Aspect-level sentiment refers tothe emotional valence of a given utterance toward one of theconcepts it mentions [21]. Sentiments toward specific entitiesare vital to consider in polarized discussions such as those weconsider here. For instance, climate change Disbelievers mightexpress negative feelings toward notions of greenhouse gases,while in agreement with a fellow Disbelievers with whom theyare interacting.We utilize Netmapper to extract entities from each tweet,and predict the aspect-level sentiment of each tweet towardeach entity [38], [39]. Netmapper relies on a multilinguallexicon of positively and negatively valenced words to cal-culate sentiment values. Aspect-level sentiment relies on aheuristic of a sliding window over words in the sentence. Morespecifically, word-level sentiment is computed based on theaverage of known valences for surrounding words.For the purposes of this work, each tweet by a certain agent i which mentions or replies to agent j is assigned an aspect-levelsentiment score from − (very negative) to +1 (very positive)directed toward the concept “@[agent j ’s Twitter handle]” We use the parameter values as defined in [22] as { k = 5000 , p = 5000 , θ I = 0 . , θ U = 0 . , θ T = 0 . } .
2) Affective networks:
Given the aspect-level sentimentscores, we construct two affective networks representing theclimate change conversations on a per-week basis. Let G + =( V, E + ) denote a positive interaction network where the setof vertices V contains all Twitter accounts in our dataset andthe set of directed edges E + contains all positive-valencedmentions and replies between agents in V . Similarly, let G − = ( V, E − ) denote a negative interaction network overthe same set of agents V and the set of directed edges E − representing their negative-valenced mentions and replies.In both cases, E + and E − denote weighted edges. Weobtain their weights as follows. Let S ij denote the set of allaspect-level sentiments in tweets by agent i toward agent j ,where i, j ∈ V . Then the weight w + ij of edge e + ij ∈ E + from i to j is given by (cid:80) x ∈ S ij min (0 , x ) . Conversely, the weight w − ij of edge e − ij ∈ E − from i to j is given by (cid:80) x ∈ S ij min (0 , − x ) .
3) E/I indices:
We assess group-level differences in pos-itive and negative interactions using Krackhardt’s E/I index[36]. For a given affective network, the E/I index intuitivelycaptures the extent to which each stance group k engages incorrespondingly valenced interactions with members of theout-group relative to their in-group. Hence, for instance, highvalues of the E/I index for the negative interaction networkwould indicate that the given stance group interacts in a morenegative way to their opponents relative to those who sharetheir beliefs.To compute the E/I indices, let V k ⊆ V denote the set ofagents belonging to stance k and V k (cid:48) those who do not holdstance k . The E/I index of stance group k on the positiveinteraction network is therefore computed as follows: P + k = E + k − I + k E + k + I + k (1)where E + k = (cid:80) i ∈ V k ,j ∈ V k (cid:48) w + ij and I + k = (cid:80) i,j ∈ V k w + ij . On theother hand, the E/I index of stance group k on the negativeinteraction network is similarly computed thus: P − k = E − k − I − k E − k + I − k (2)where E − k = (cid:80) i ∈ V k ,j ∈ V k (cid:48) w − ij and I − k = (cid:80) i,j ∈ V k w − ij . Giventhe construction of P + k and P − k , we note that both values arebounded between − and +1 .
4) Polarization valence:
We find whether the interactionshave negative valence or positive valence by defining polar-ization P k by taking a difference of the two E/I indices asexpressed below: P k = P − k − P + k . (3)In this work, we operationalize our view of affective polar-ization in terms of high E/I indices on the negative interactionnetwork, and low values on the positive interaction network. P k thus captures this intuition by assigning positive valuesfor groups that display disproportionately hostile or negativeinteractions toward the out-group relative to their in-group.Values close to , on the other hand, indicate relatively evenlevels of positive and negative interactions for in-group andout-group members. Finally, negative values indicate that thoseholding stance k are more negative to their in-group butpositive to their out-group.
5) Polarization magnitude:
To find the magnitude of affec-tive polarization we use Earth Mover’s Distance (EMD) onthe distribution of weighted edges for outgroup and ingroupinteractions. This is similar to computing first Wasserstein dis-tance between two 1D distributions [40]. Similar to affectivenetworks, we define G = ( V, E ) as interaction network wherethe set of vertices V contains all Twitter accounts in our datasetand the set of directed edges E contains all valenced (positiveor negative) mentions and replies between agents in V . In thiscase, we do not separate negative and positive valence graphsand treat weight w ij of edge e ij ∈ E from i to j as given by (cid:80) x ∈ S ij x . Let u k be distribution of w ij , where i ∈ V k , j ∈ V k (cid:48) and let v k be distribution of w ij , where i ∈ V k , j ∈ V k .For a group holding stance k , we define our novel affectivepolarization metric as: l k = (cid:40) − (cid:82) + ∞−∞ | U k − V k | : P k < (cid:82) + ∞−∞ | U k − V k | : P k ≥ (4)where U k and V k are the respective CDFs of u k and v k . EMD is proportional to the minimum amount of workrequired to covert one distribution to another . We use P k to assign positive or negative valence to the EMD. Althoughthere are other techniques to find the difference in distributionsuch as KS-Test [41]. However, during our experiments, wefound that EMD is able to capture more nuanced differencesin distributions. More likely because the EMD can capturedifferences in heavy-tailed distributions better and it does notmake any parametric assumptions [40].Our novel affective polarization metric l k is positive when P k > . As noted in §III-C4, a positive value would meanmore hostility or negative sentiment in intergroup communi-cation compared to intragroup communication. On the otherhand, a negative value of l k is when P k < , meaning morepositive sentiment in intergroup communication compared tointragroup communication.IV. R ESULTS
Using the metric defined in Equation 4, in this section,we first explore how affective polarization between Believersand Disbelievers is changing over the 100 weeks. Then weexplore how hostile periods are related to natural disaster-related words.We first look at how the affective polarization metric ischanging over time in figure 1. Overall, our analysis foundthat climate change Disbelievers tended to exhibit high levels http://infolab.stanford.edu/pub/cstr/reports/cs/tr/99/1620/CS-TR-99-1620.ch4.pdfig. 1. Affective polarization metric ( l k ) for Believers and Disbelievers of climate change. Higher positive values denote more hostility towards the othergroup. The dotted lines represent mean 1 standard deviation, which for Believers is -0.091 and 0.080 and disbelievers is -0.117 and 0.106. The analysis wasdone on data collected from 26th August 2017 to 14th September 2019 as described in §III-A. of hostility toward climate change Believers. This finding wasrelatively consistent throughout the 100-week period underobservation, as the time series for climate change Disbelieversonly very rarely goes below the threshold of 0, which indicatessimilarly valenced interactions toward in-group and out-groupmembers. Some weeks displayed exceptionally high levels ofhostility toward climate change Believers, greater than onestandard deviation from the mean. The standard deviation of l k is lower for Disbelievers than for Believers. Indicating thatDisbelievers act in much more organized manner over the 100weeks than Beleivers.Climate change Believers, on the other hand, were notgenerally hostile toward Disbelievers, as the time series forclimate change Believers tends to fluctuate over and under thethreshold of 0. This indicates that climate change Believerscommunicate with in-group and out-group members with rel-atively similar emotional valence. However, on certain weeks,climate change Believers did also feature exceptionally highhostility scores. This suggests that climate change Believersmay also behave in a hostile manner toward climate changeDisbelievers, even if not over the long term.To investigate instances where hostility between Believersand Disbelievers is high we compare those weeks with weekswhere hostility is low. We define hostile weeks as those datapoints where l k is more than mean plus 1 standard deviation,i.e. from figure 1, all the weeks where for Believers l k > . and for Disbelievers l k > . . The number of such weeksfor Disbelievers where l k > . is 20 and for Believerswhere l k > . is 12. We look further into these weeks asexamples of exceptional hostilie weeks.Next, we use natural disaster-related words as a proxyto determine how natural disasters play a role in hostility Fig. 2. Percentage of the top 100 most frequent hashtags containing naturaldisaster-related words. The figure shows the percentage when the affectivepolarization metric is greater than 1 standard deviation or otherwise. Theerror bars represent 1 standard errors. between the two groups. In figure 2 we look at the top100 most frequent hashtags used within those groups to findthe percentage of hashtags related to natural disasters. Asexpected, Believers use more natural disaster-related hashtagsthan Disbelievers. However, during the exceptional hostileweeks Believers use less of these hashtags. Interestingly,Disbelievers show the exact opposite behavior. Disbelieversuse more natural disaster-related hashtags when they are morehostile towards Believers. We provide further evidence of thisfinding in figure 3. In figure 3, we look at the percentageof Tweets with at least one natural disaster-related word. Wefind similar patterns as mentioned above. Moreover, we findthat a greater percentage of Tweets from Disbelievers mentionatural disaster-related words compared to Believers. Thisindicates that Disbelievers are calling out natural disastersmore when they are exceptionally hostile towards Believerscompared to other weeks.
Fig. 3. Percentage of tweets with at least one natural disaster-related word.The figure shows the percentage when the affective polarization metric isgreater than 1 standard deviation or otherwise. The error bars represent 1standard errors.
V. D
ISCUSSION AND F UTURE W ORK
Taken together, our findings suggest the importance ofconsidering affective polarization in online discourse, particu-larly concerning the subject of climate change. Whereas paststudies had shed light on the echo chamber dynamics whichcharacterized intergroup communication surrounding climatechange [19], we show how this polarization extends also tothe realm of emotion in the form of affective polarization.We extend existing studies which highlight the role of inci-vility and personalized framing in encounters between climatechange Believers and Disbelievers [8], [18] by introducing ascalable technique for analyzing relative intra- and intergroupinteraction valence. This allowed us to quantify the extent ofhostile communications between the two groups over a large-scale, long-term dataset - thereby validating existing findingsin a generalizable manner as well as showing their relativestability over time.Furthermore, we highlight the value of viewing polarizationfrom an asymmetrical perspective. Related scholarship inpolitical psychology underscores how ideological asymmetriesunderpin conflict dynamics across a variety of social issues[42]. In other words, the participation of two groups withinpolarized discourse does not necessarily mean that both groupsengage in conflict in the same way. Prior work illustrates thatthese findings translate robustly to the digital sphere - politicalelites or opinion leaders who share moralized content behavein distinct ways depending on their ideological orientations[4]. The present work contributes to the literature by showinghow these dynamics unfold the standpoint of the public atlarge concerning online climate change discourse.Indeed, higher levels of hostility from Disbelievers present aspecifically notable finding for social scientific scholarship onclimate change discourse. Longitudinal analysis in prior work suggests that generalized climate change beliefs over timeare increasing [15], [16], and climate change Disbelievers inparticular are more susceptible to potential belief change [33].But significant cognitive barriers remain for fuller acceptanceof anthropogenic causes for climate change and the corre-sponding urgency for responsive policy changes [31], [43].Higher levels of hostility among climate change Disbelieverstoward climate change Believers constitutes one such obstaclefor further dialogue between the two groups. As past stud-ies suggest, one psychological factor which impedes climatechange Beliefs is not related to the climate at all, but anchorsprimarily on the feelings of dislike felt by one group towardsthe other [32]. Such challenges may thus persist in the formof further entrenchment of Disbelievers within interactionalsiloes and disengagement from intergroup communicationaltogether [19]. Or as emergent studies show, they can alsotrigger what have been called ‘trench warfare dynamics’ [6] -whereby Disbelievers persistently communicate with Believersbut solidify their own cognitive immovability in the process.These insights are especially important to consider givenour secondary set of findings. Our analysis suggests thatfurther asymmetries arise between Believers and Disbelieversengagement with disaster words in relation to their levels ofaffective polarization. Although comparable levels are seenwhen both groups are within average levels of our metric,moments of increased affective polarization correlate withopposite behaviors for Believers and Disbelievers. Believersappear to shift to other areas of contention, such that theiraggression is characterized by non-disaster topics. In contrast,Disbelievers’ increased invocation of disaster terms points tomore aggressive discussion of these catastrophes, albeit posi-tioned in resistance to explanations related to anthropogenicclimate change. This introduces another layer of intractableconflict in beliefs, as major climate events do not appear toinvite susceptibility of belief change for Disbelievers. Instead,they potentially incite more vigorous psychological resistance.Collectively, these findings point to significant benefitsto studying affective polarization in online climate changediscourse. Although social media discourse does not neces-sarily constitute a representative sample of a particular globalpopulation [44], digital platforms like Twitter nonetheless con-stitute a vital space for public conversations about importantissues like climate change. Hence, these findings paint auseful picture of public discourse as situated specifically incyberspace, which may also bear implications for how digitallymediated science communication and public policy may alsobe designed and implemented [30], [34].Besides the issue of demographic representativeness foronline data, other limitations attend the present analysis. First,although we have a large number of tweets to characterizegeneral affective behavior, however, it does not encompassthose interactions which do not include our collection key-words. Second, the task of getting an aspect-level sentimentof each tweet towards other entities is a non-trivial task. Weuse Netmapper which has been used with reasonable accuracyfor multiple sentiment level tasks [45], [46]. The focus of thisaper is on designing a framework to get affective polarizationscore between two competing groups and we do not makean effort to improve aspect-level sentiment scores. Last, inour analysis we use a list of natural disaster related words.Communication about the natural disasters could also happenusing specific names related to these disasters, for exampleusing “Dorian” instead of “Hurricane Dorian”. Such analysiswould require a more comprehensive list of natural disastersoccurring around the world during the 100 weeks. This is outof scope for the current work.Recognizing the foregoing limitations, we also consideravenues for future work in this area. On a conceptual level,researchers may wish to expand the binary system of climatechange beliefs assumed here. Affectively polarized dynamicsbetween multiple groups may be a more challenging yetalso potentially informative line of inquiry to explore giventhe diversity of positions held with respect to this complexissue. Methodologically, computational analysis may extendour findings by performing more fine-grained characterizationof the types of hostility expressed by both groups. Naturallanguage processing (e.g., topic models) may offer one wayforward in this regard. Acknowledging the non-neutrality ofcyberspace, it would also be important to consider whetherdisinformation maneuvers may also be involved in shapingthe wider climate change discussion. Inauthentic bot-like ac-counts and trolls may unduly influence different groups bymanipulating the flow of information or amplifying intergroupaggressions; such factors have been seen in relation to othercontentious issues and may potentially be present here as well[47]–[49].Finally, taking flight from the digital scope of our research,further studies may fruitfully examine several hypothesesopened up by our results. For instance, social scientists mayinvestigate actual levels of experienced hostility by climatechange Believers and Disbelievers toward opposed groups.These evidence bases would be valuable to accumulate incross-cultural settings, as well as over time - especially inconnection with concurrent political shifts and natural climate-related developments like anomalous weather patterns andwider-ranging disasters [15], [16], [31].R
EFERENCES[1] P. Barber´a, J. T. Jost, J. Nagler, J. A. Tucker, and R. Bonneau, “Tweetingfrom left to right: Is online political communication more than an echochamber?”
Psychological science , vol. 26, no. 10, pp. 1531–1542, 2015.[2] K. Garimella, G. De Francisci Morales, A. Gionis, and M. Mathioudakis,“Political discourse on social media: Echo chambers, gatekeepers, andthe price of bipartisanship,” in
Proceedings of the 2018 World Wide WebConference , 2018, pp. 913–922.[3] A. Tyagi, A. Field, P. Lathwal, Y. Tsvetkov, and K. M. Carley, “Acomputational analysis of polarization on indian and pakistani socialmedia,” in
Social Informatics . Springer, 2020.[4] W. J. Brady, J. A. Wills, J. T. Jost, J. A. Tucker, and J. J. Van Bavel,“Emotion shapes the diffusion of moralized content in social networks,”
Proceedings of the National Academy of Sciences , vol. 114, no. 28, pp.7313–7318, 2017.[5] D. Geschke, J. Lorenz, and P. Holtz, “The triple-filter bubble: Usingagent-based modelling to test a meta-theoretical framework for theemergence of filter bubbles and echo chambers,”
British Journal ofSocial Psychology , vol. 58, no. 1, pp. 129–149, 2019. [6] R. Karlsen, K. Steen-Johnsen, D. Wollebæk, and B. Enjolras, “Echochamber and trench warfare dynamics in online debates,”
EuropeanJournal of Communication , vol. 32, no. 3, pp. 257–273, 2017.[7] A. Matakos, E. Terzi, and P. Tsaparas, “Measuring and moderatingopinion polarization in social networks,”
Data Mining and KnowledgeDiscovery , vol. 31, no. 5, pp. 1480–1505, 2017.[8] A. A. Anderson and H. E. Huntington, “Social media, science, and attackdiscourse: How twitter discussions of climate change use sarcasm andincivility,”
Science Communication , vol. 39, no. 5, pp. 598–620, 2017.[9] J. N. Druckman and M. S. Levendusky, “What do we measure whenwe measure affective polarization?”
Public Opinion Quarterly , vol. 83,no. 1, pp. 114–122, 2019.[10] M. Yarchi, C. Baden, and N. Kligler-Vilenchik, “Political polarizationon the digital sphere: A cross-platform, over-time analysis of interac-tional, positional, and affective polarization on social media,”
PoliticalCommunication , pp. 1–42, 2020.[11] R. E. Dunlap, A. M. McCright, and J. H. Yarosh, “The political divideon climate change: Partisan polarization widens in the us,”
Environment:Science and Policy for Sustainable Development , vol. 58, no. 5, pp. 4–23, 2016.[12] D. R. Fisher, J. Waggle, and P. Leifeld, “Where does political polar-ization come from? locating polarization within the us climate changedebate,”
American Behavioral Scientist , vol. 57, no. 1, pp. 70–92, 2013.[13] A. M. McCright and R. E. Dunlap, “The politicization of climate changeand polarization in the american public’s views of global warming,2001–2010,”
The Sociological Quarterly , vol. 52, no. 2, pp. 155–194,2011.[14] A. Tyagi, M. Babcock, K. M. Carley, and D. C. Sicker, “Polarizingtweets on climate change,” in
To appear International Conference SBP-BRiMS , A. H. Halil Bisgin, C. Dancy, , and R. Thomson, Eds. Springer,2020.[15] L. C. Hamilton, J. Hartter, M. Lemcke-Stampone, D. W. Moore, andT. G. Safford, “Tracking public beliefs about anthropogenic climatechange,”
PloS one , vol. 10, no. 9, p. e0138208, 2015.[16] T. L. Milfont, M. S. Wilson, and C. G. Sibley, “The publics belief inclimate change and its human cause are increasing over time,”
PloS one ,vol. 12, no. 3, p. e0174246, 2017.[17] S. M. Jang and P. S. Hart, “Polarized frames on climate change andglobal warming across countries and states: Evidence from twitter bigdata,”
Global Environmental Change , vol. 32, pp. 11–17, 2015.[18] C. W. van Eck, B. C. Mulder, and A. Dewulf, “Online climate changepolarization: Interactional framing analysis of climate change blogcomments,”
Science Communication , p. 1075547020942228, 2020.[19] H. T. Williams, J. R. McMurray, T. Kurz, and F. H. Lambert, “Networkanalysis reveals open forums and echo chambers in social mediadiscussions of climate change,”
Global environmental change , vol. 32,pp. 126–138, 2015.[20] A. Banks, E. Calvo, D. Karol, and S. Telhami, “
TheInternational Journal of Press/Politics , p. 1940161220940964, 2020.[21] B. Huang, Y. Ou, and K. M. Carley, “Aspect level sentiment classifi-cation with attention-over-attention neural networks,” in
InternationalConference on Social Computing, Behavioral-Cultural Modeling andPrediction and Behavior Representation in Modeling and Simulation .Springer, 2018, pp. 197–206.[22] S. Kumar, “Social media analytics for stance mining a multi-modalapproach with weak supervision,” Ph.D. dissertation, Carnegie MellonUniversity, 2020.[23] K. Darwish, “Quantifying polarization on twitter: The kavanaugh nom-ination,” in
International Conference on Social Informatics . Springer,2019, pp. 188–201.[24] D. Demszky, N. Garg, R. Voigt, J. Zou, J. Shapiro, M. Gentzkow,and D. Jurafsky, “Analyzing polarization in social media: Methodand application to tweets on 21 mass shootings,” in
Proc. ofNAACL
Chaos:An Interdisciplinary Journal of Nonlinear Science , vol. 25, no. 3, p.033114, 2015.[26] I. Weber, V. R. K. Garimella, and A. Batayneh, “Secular vs. islamistpolarization in egypt on twitter,” in
Proceedings of the 2013 IEEE/ACMInternational Conference on Advances in Social Networks Analysisnd Mining , ser. ASONAM 13. New York, NY, USA: Associationfor Computing Machinery, 2013, p. 290297. [Online]. Available:https://doi.org/10.1145/2492517.2492557[27] L. G. Stewart, A. Arif, and K. Starbird, “Examining trolls and polar-ization with a retweet network,” in
Proc. ACM WSDM, workshop onmisinformation and misbehavior mining on the web , 2018.[28] K. Garimella, G. De Francisci Morales, A. Gionis, and M. Mathioudakis,“Reducing controversy by connecting opposing views,” in
Proceedingsof the Tenth ACM International Conference on Web Search and DataMining , 2017, pp. 81–90.[29] K. Garimella, G. D. F. Morales, A. Gionis, and M. Mathioudakis,“Quantifying controversy on social media,”
ACM Transactions on SocialComputing , vol. 1, no. 1, pp. 1–27, 2018.[30] T. Bolsen and M. A. Shapiro, “The us news media, polarization on cli-mate change, and pathways to effective communication,”
EnvironmentalCommunication , vol. 12, no. 2, pp. 149–163, 2018.[31] J. T. Carmichael and R. J. Brulle, “Elite cues, media coverage, and publicconcern: an integrated path analysis of public opinion on climate change,2001–2013,”
Environmental Politics , vol. 26, no. 2, pp. 232–252, 2017.[32] L. Van Boven, P. J. Ehret, and D. K. Sherman, “Psychological barriersto bipartisan public support for climate policy,”
Perspectives on Psycho-logical Science , vol. 13, no. 4, pp. 492–507, 2018.[33] H. C. Jenkins-Smith, J. T. Ripberger, C. L. Silva, D. E. Carlson,K. Gupta, N. Carlson, A. Ter-Mkrtchyan, and R. E. Dunlap, “Partisanasymmetry in temporal stability of climate change beliefs,”
NatureClimate Change , vol. 10, no. 4, pp. 322–328, 2020.[34] D. M. Kahan, H. Jenkins-Smith, T. Tarantola, C. L. Silva, and D. Bra-man, “Geoengineering and climate change polarization: testing a two-channel model of science communication,”
The ANNALS of the Ameri-can Academy of Political and Social Science , vol. 658, no. 1, pp. 192–222, 2015.[35] E. F. Bloomfield and D. Tillery, “The circulation of climate changedenial online: Rhetorical and networking strategies on facebook,”
Envi-ronmental Communication , vol. 13, no. 1, pp. 23–34, 2019.[36] D. Krackhardt and R. N. Stern, “Informal networks and organizationalcrises: An experimental simulation,”
Social psychology quarterly , pp.123–140, 1988.[37] F. L. Hitchcock, “The distribution of a product from several sources tonumerous localities,”
Journal of mathematics and physics , vol. 20, no.1-4, pp. 224–230, 1941.[38] L. R. Carley, J. Reminga, and K. M. Carley, “Ora & netmapper,”in
International Conference on Social Computing, Behavioral-CulturalModeling and Prediction and Behavior Representation in Modeling andSimulation . Springer, 2018.[39] J. Uyheng, T. Magelinski, R. Villa-Cox, C. Sowa, and K. M. Carley,“Interoperable pipelines for social cyber-security: assessing twitter in-formation operations during nato trident juncture 2018,”
Computationaland Mathematical Organization Theory , pp. 1–19, 2019.[40] A. Ramdas, N. G. Trillos, and M. Cuturi, “On wasserstein two-sampletesting and related families of nonparametric tests,”
Entropy , vol. 19,no. 2, p. 47, 2017.[41] F. J. Massey Jr, “The kolmogorov-smirnov test for goodness of fit,”
Journal of the American statistical Association , vol. 46, no. 253, pp.68–78, 1951.[42] J. T. Jost, “Ideological asymmetries and the essence of political psy-chology,”
Political psychology , vol. 38, no. 2, pp. 167–208, 2017.[43] M. T. Ballew, A. R. Pearson, M. H. Goldberg, S. A. Rosenthal, andA. Leiserowitz, “Does socioeconomic status moderate the politicaldivide on climate change? the roles of education, income, and indi-vidualism,”
Global Environmental Change , vol. 60, p. 102024, 2020.[44] F. Morstatter, J. Pfeffer, H. Liu, and K. M. Carley, “Is the samplegood enough? comparing data from twitter’s streaming api with twitter’sfirehose,” in . AAAI press, 2013, pp. 400–408.[45] K. M. Carley and D. M. Beskow, “Trident joust 2017, after actionreport,”
Center for the Computational Analysis of Social and Organiza-tional Systems , 2017.[46] D. Cardoso Llach and J. Argota S´anchez-Vaquerizo, “An ecology ofconflicts,” in
Computer-Aided Architectural Design. “Hello, Culture” ,J.-H. Lee, Ed. Singapore: Springer Singapore, 2019, pp. 198–212.[47] K. M. Carley, G. Cervone, N. Agarwal, and H. Liu, “Social cyber-security,” in
International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation inModeling and Simulation . Springer, 2018, pp. 389–394. [48] D. A. Broniatowski, A. M. Jamison, S. Qi, L. AlKulaib, T. Chen, A. Ben-ton, S. C. Quinn, and M. Dredze, “Weaponized health communication:Twitter bots and russian trolls amplify the vaccine debate,”
Americanjournal of public health , vol. 108, no. 10, pp. 1378–1384, 2018.[49] A. Bessi and E. Ferrara, “Social bots distort the 2016 us presidentialelection online discussion,”
First Monday , vol. 21, no. 11-7, 2016.
VI. A