Polarized Ukraine 2014: Opinion and Territorial Split Demonstrated with the Bounded Confidence XY Model, Parameterized by Twitter Data
Maksym Romenskyy, Viktoria Spaiser, Thomas Ihle, Vladimir Lobaskin
aa r X i v : . [ phy s i c s . s o c - ph ] J u l Polarized Ukraine 2014: Opinion and Territorial Split Demonstrated with theBounded Confidence XY Model, Parameterized by Twitter Data
Maksym Romenskyy
Department of Life Sciences, Imperial College London, London SW7 2AZ, UK andDepartment of Mathematics, Uppsala University, Box 480, Uppsala 75106, Sweden
Viktoria Spaiser
School of Politics and International Studies, University of Leeds, Leeds LS2 9JT, UK
Thomas Ihle
Institute of Physics, University of Greifswald, Felix-Hausdorff-Str. 6, Greifswald 17489, Germany
Vladimir Lobaskin
School of Physics, University College Dublin, Belfield, Dublin 4, Ireland (Dated: July 26, 2018)Multiple countries have recently experienced extreme political polarization, which in some casesled to escalation of hate crime, violence and political instability. Beside the much discussed presi-dential elections in the United States and France, Britain’s Brexit vote and Turkish constitutionalreferendum, showed signs of extreme polarization. Among the countries affected, Ukraine facedsome of the gravest consequences. In an attempt to understand the mechanisms of these phe-nomena, we here combine social media analysis with agent-based modeling of opinion dynamics,targeting Ukraine’s crisis of 2014. We use Twitter data to quantify changes in the opinion divideand parameterize an extended Bounded-Confidence XY Model, which provides a spatiotemporaldescription of the polarization dynamics. We demonstrate that the level of emotional intensity is amajor driving force for polarization that can lead to a spontaneous onset of collective behavior ata certain degree of homophily and conformity. We find that the critical level of emotional intensitycorresponds to a polarization transition, marked by a sudden increase in the degree of involvementand in the opinion bimodality.
I. INTRODUCTION
Ukraine represents a bright example of a nearlyevenly split society with two opposing camps, wherethe East/South gravitates towards Russia while theWest/North towards European neighbors [1]. The over-all political vector in the country sways between politi-cal parties and leaders that on the one side seek closerties to the West and in particular Europe and on theother hand to the East and in particular Russia. The Or-ange Revolution in 2004 brought pro-western politiciansto power, however, in the 2010 elections a pro-easternpolitician, Viktor Yanukovych, was elected for president,not least because of major support in the eastern re-gions of Ukraine (see Fig. 1). In November 2013, af-ter Yanukovych failed to sign a political association andfree trade agreement with the European Union, protestsin Ukraine erupted. The initially peaceful rallies becameviolent in January 2014 after the government passed lawsto suppress the protests. In February 2014, the violenceescalated, which led to the removal of Yanukovych fromoffice by the parliament. Meanwhile a separatist andanti-interim-government movement rose with the sup-port of Russia in eastern and southern parts of Ukraine,and ignited a military conflict. Crimea was annexed bythe Russian Federation after a referendum that was de-nounced internationally as illegitimate and illegal. Laterin 2014, the crisis resulted in further territorial separa- tion with over 2.6 million internally displaced personsand refugees and a formation of self-proclaimed states inDonetsk and Luhansk [2]. These events escalated the po-larization in the country that has grown over the years.As the two political sides became more extreme in theirviews in the course of the events a dialogue and there-fore a peaceful solution has become increasingly difficult.The extreme opinion divide affected not only society asa whole but also destabilized multiple families and localcommunities.In this paper, we present a novel approach, in whichwe combine rich social media data with the power ofmethods of statistical physics, to study political opin-ion polarization mechanisms, seeking to understand whatmechanisms turned Ukraine into a irreconcilably polar-ized state. Though the integration of the two approacheshas been increasingly discussed [5, 6], there are only fewstudies so far that have actually attempted to combinethese [7–11]. None of them, however, studies polariza-tion and integrates the unstructured data analysis com-prehensively with the computational model. In the paststudies there has been a methodological gap between thesocial media analysis and agent-based modeling, whichhas limited the relation between the two to rather qual-itative statements. We are presenting an approach herethat elaborately links the two approaches. Combiningthem has the advantage to parameterize the computa-tional model and validate it by empirical evidence and
Figure 1: (A) Ukraine’s political divide in 2010 elections [3]. The majority of the voters in the eastern and southern regions ofUkraine supported a pro-eastern candidate, Viktor Yanukovych. (B) Ukraine’s linguistic divide according to national census2001 [4]. In the three regions most affected by Ukraine crisis in 2014, Luhansk oblast, Donetsk oblast and Crimea, Russian isa native language for more than 50% of the population. on the other hand to make use of the rich social mediadata in a theory-guided way beyond mere descriptives.This holds the potential to gain new insights into theunderlying social mechanisms of polarization. In the fol-lowing, we will describe a new computational model ofpolarization, a 2D lattice agent-based model, that is in-formed by the theoretical work on polarization discussedbelow and that brings in the spatial dimension and em-phasizes the role of regional differences. We then lookat the empirical polarization dynamics in the UkrainianTwittersphere, before parameterizing the model with theanalyzed Twitter data, hereby validating the model. Thecombination of the two approaches reveals the importantrole that emotional intensity levels play in polarizationthus far not sufficiently accounted for by classical the-oretical or empirical studies of polarization. The fewstudies that investigate the role of emotions in polar-ization have usually focussed on very specific emotions,e.g. “self-conscious” emotions like pride and embarrass-ment/shame and showed that these emotions can rein-force conformity and polarization [12]. In this paper, wedo not focus on specific emotions, but, rather examinethe role of emotional intensity levels and show how theseemotional intensity levels can be a decisive driver in po-larization.
II. OPINION POLARIZATION
Opinion polarization has been intensely studied overthe last few decades, initiated by the observation thatgroups tend to adopt positions that are more extremethan the initial individual positions of its members[13, 14]. One explanation of this phenomenon is based onthe Social Comparison Theory, which suggests that peo- ple want to be perceived in a more favorable way thanwhat we perceive to be the average tendency. Throughobservations they determine what the average tendencyis and then they express a slightly more extreme opinionthan the perceived average opinion [13, 15, 16]. Thereis clear evidence for this assumption from numerous ex-perimental studies [13, 17, 18]. Moreover, due to the ho-mophily phenomenon [19], which states that people aremore likely to interact with those who are similar to themwith respect to socio-economic background [20] as well asattitudes [21], people are more likely to socially comparethemselves to similar peers. Group polarization phenom-ena are also explained drawing on the Persuasive Argu-ment Theory, which states that people are more likely tochange their opinion when presented with persuasive ar-guments. Group polarization can occur when the groupdiscourse is manipulated through biased information ormisinformation, exposing a group to false persuasive ar-guments [13, 16, 22]. The theory is strongly supportedby empirical evidence from numerous experimental stud-ies [13, 23, 24]. The two mechanisms, social comparisonand persuasive argument usually co-occur. For instance,persuasive arguments in an environment of biased infor-mation can further push an individual to adopt a moreextreme attitude than expressed in the group they com-pare themselves to. Another mechanism with respect topolarization dynamics is the Biased Assimilation [25] inopinion formation processes, which maintains that peo-ple are likely to keep their original position and drawsupport for it if confronted with mixed or inconclusivearguments. This mechanism, which again is empiricallywell established [26–28], shows that people are only to alimited extent open to change their opinions and this cancontribute to polarization dynamics. In fact, Dandekar etal. [21] show that the two earlier described mechanismsand in particular the social comparison mechanism, is notsufficient to produce polarization, the biased assimilationmechanism has to be added.Polarization processes have been empirically investi-gated through experimental studies as mentioned previ-ously and to a lesser extent through survey based stud-ies [29, 30]. Furthermore, applied statistical physicsand agent-based modeling approaches [21, 31–33] havebeen used extensively to study polarization mechanismsthrough computer simulations. Prominent are for in-stance bounded confidence models of opinion dynamics,stochastic models for the evolution of continuous-valuedopinions within a finite group of individuals that exploreconditions for consensus and opinion fragmentation, in-troduced by Deffuant et al. [34] and further elaboratedby numerous studies [35–37]. We will draw inspirationfrom these models as well. The challenges of using sta-tistical physics models for modeling social phenomena areknown. The tractable models are usually oversimplifiedand too general and thus lack flexibility required to reflectthe features of a particular social phenomena. Moreover,while computational studies are rigorous in investigat-ing the specific mechanisms and dynamics of polarizationthey often lack empirical foundation and thus it remainsoften unclear to what extent these often abstract mod-els accurately represent phenomena we see in real world.More recently, social media (e.g. Twitter and Facebook)data have been increasingly used to study opinion polar-ization [38–41]. Gruzd and Tsyganova [42], for instance,use Ukrainian Vkontakte data to demonstrate the splitbetween the two political camps in Ukraine (pro-East vs.pro-West) during the Maidan protests. Twitter has alsoplayed a quite important role as a contested public de-bate arena throughout the crisis in Ukraine [4, 5]. Stud-ies with social media data have revealed the strong effectof homophily in online social networks, which may leadto phenomena like the echo chamber [41, 45, 46], whereopinions are amplified through communication and repe-tition inside an “enclosed” social system. Echo chamberscan prevent people from noticing contrary persuasive ar-guments and they skew the perceived average that peopletake into consideration in social comparison processes.Though social media data is potentially rich, i.e. fine-grained, time-resolved, relational, geo-coded, etc., with-out an explicit theoretical underpinning, the studies ofpolarization on social media remain usually rather de-scriptive. By combining rich social media data with thepower of methods of statistical physics a better under-standing of specific political opinion polarization mecha-nisms is sought in this paper.
III. METHODSA. Political opinion mining of Twitter data
We used the archived Twitter API Streaming datafrom October 2013 to September 2014 provided by the Archive Team [47]. The archived data is the freely avail-able Twitter Streaming API Spritzer Sample, which col-lects 1% of all public tweets in real-time. The TwitterStreaming API Spritzer Sample allows for unfiltered dataanalysis and hence for capturing the full discourse picturecomparing to a more narrow, focussed (e.g. based on cer-tain hashtags, user networks etc.) approach facilitated bydata using the Twitter REST API [48] that would disre-gard discourse contributions beyond the specified searchqueries. The data for January 2014 was missing and wedid not find any other archive providing this data. Thetweet data is stored in JSON format. The data was pro-cessed and analyzed in Python. Specifically, the data wasfiltered for Russian/Ukrainian language (excluding Twit-ter users who specified being from Russia), cleared ofSPAM, and then filtered for political content with an ex-tensive set of keywords (see the electronic supplementarymaterial for details). We used the tweets to determinethe political affiliation of each Twitter user using a set ofkeywords indicating neutral, pro- or anti-West attitudesand a set of keywords indicating neutral, pro- or anti-Russian attitudes (see the electronic supplementary ma-terial for details). Moreover, we conducted a sentimentanalysis of the tweets, in order to determine the politicalaffiliation for previously seemingly neutral Twitter users.For this purpose we utilized the SentiStrength system forautomatic sentiment analysis, built an Ukrainian sen-timent words dictionary (see the electronic supplemen-tary material) and extended the SentiStrength Russiansentiment words dictionary to make it equivalent to theUkrainian sentiment words dictionary (see the electronicsupplementary material for details). Our classificationalgorithm scanned all tweets for each user and checkedwhether the tweet contained any of the keywords and anyof the sentiment words specified and calculated overallscores for political “West” and “East” affiliation as wellas sentiment scores (see the electronic supplementary ma-terial for details). For the geo-plots in Fig. 3C,D, we fo-cused on the data of the last week in September 2014. Wescanned the Twitter data of each user for two possible ge-ographical information, the value of their “location” tagand/or geo/place “coordinates” tag which is a latitudeand a longitude coordinate value. The vast majority ofusers have the default option “geo enabled:false”, thus donot provide precise geographical information attached tothe tweet, a few more provide profile “location” informa-tion, but overall, geographical data is often not specifiedand thus missing. We therefore ended up with only 493Twitter users for whom we had a geographical informa-tion. The “West” and “East” political affiliation scoresof these 493 Twitter users were plotted in Fig. 3C,D (seethe electronic supplementary material for details). N u m be r o f u s e r s E m o t i ona l / neu t r a l op i n i on s Figure 2: (A) Time series plot of number of tweets, in total, neutral and emotional. (B) Time series plot of emotional/neutraltweets ratio with key political events tags. (C) Time series plot of number of average (weekly) users’ opinions, in total, pro-West and pro-East. (D) Average weekly number of tweets per user. No data was available for January 2014 (see Materials andMethods and the electronic supplementary material) therefore the data gaps in plots A-D between the weeks 13 and 17.
B. Bounded Confidence XY Model of OpinionPolarization
We model opinion dynamics in an assembly of inter-active agents placed on a two-dimensional lattice of fi-nite size, which emulates a regional opinion distributionand localization. Each opinion is represented by a vectorthat can freely rotate in plane, similar to that in the XY model of a magnet, and is characterized by the length p (emotional intensity, i.e. fervent strength of an opinion)and polar angle θ (orientation). This allows us to modela continuous opinion spectrum with respect to the speci-fied direction by a cosine of the angle between the vectors,which can vary in the range between -1 (agent opposesthe opinion) and 1 (agent fully supports the opinion).Each agent can interact with a fixed number of nearestneighbors. The social interactions are introduced via lo-cal mean field, similar to the Vicsek model [49]: the neworientation of each vector is calculated as a mean of theaverage direction of the neighbors and agent’s own ori-entation: θ i ( t + ∆ t ) = tan − " N P j, | j − i |≤ r p j sin( θ j ( t )) P j, | j − i |≤ r p j cos( θ j ( t )) + ξ ( η, t ) , (1)where N is the number of interacting agents, ξ is theangular noise variable uniformly distributed in the inter-val [ − η/ , η/
2] and η is the noise strength. The noise isadded to model the level of conformity of the individual: zero noise, η = 0 corresponds to full conformity, η = 2 π allows an individual to deviate from the group opinionby an unlimited value. A contribution of each interact-ing agent is weighted by the emotional intensity of itsopinion p . The interaction has a limited range r , whichsets the number of nearest neighbors considered. In 2D,interaction range r = 2 corresponds to 24 interactionpeers, r = 3 gives 48 peers, etc (see the electronic sup-plementary material for details). In addition to this, theinteraction is selective so that the vectors align only withthose neighbors, whose orientation (opinion) deviates byan angle less than some fixed value α from their ownopinion vector. This rule is inspired by the bounded con-fidence model, or Deffuant model [34], and allows one toimitate systems with different levels of opinion tolerance[50]. Scaled value α/π denotes a fraction of the opinionspectrum that is taken into account by each individual,e.g. α/π = 0 . i in the previous timestep) because using a relatively small values of η and α already allows to achieve a socially realistic behavior.We model a finite system with the boundary condi-tions set by two fixed rows of agents on each side. Theboundary vectors are fixed according to the followingscheme: the left and upper rows are oriented to the left( θ = π ), imitating a bias towards “West” and the rightand bottom rows are fixed to point in the opposite di-rection ( θ = 0), thus mimicking a bias towards “East”.When other agents interact with the boundary vectorsthis introduces a spatially dependent local bias that canaccount for geographical inhomogeneity in opinions, thusimitating cultural or ethnic differences, information bias,etc. (see the electronic supplementary material for de-tails; see also the interactive model [51]).These model settings reflect the theoretical assump-tions about polarization described above. The orien-tation represents political attitudes and the local meanfield calculation of new orientations for each agent rep-resents the social comparison theory assumptions. Thehomophily effect is included through the bounded confi-dence feature, where agents align only with those neigh-bors, whose political orientation is similar to their own.Furthermore, the biased assimilation mechanism is re-flected in the noise variable that determines the agents’willingness to change opinion (conformity level). Finally,the persuasive argument theory assumption, in particularwith respect to the contribution of biased information topolarization, is simulated via the boundary conditions.We have added a further parameter to our model, theemotional intensity, i.e. vector length, that representsthe level of emotional strength and vehemence of a po-litical opinion. This parameter reflects results we haveobtained from Twitter data sentiment analysis discussedin the next sections. It also builds on recent computa-tional models and big data analysis of opinion dynamics[52, 53], which showcase the importance of emotions, andin particular of negative emotions, for people’s engage-ment in political debates and for opinion formations andchanges.We modeled evolution of the opinion spectrum in thedescribed system starting from randomized initial distri-butions of agents’ emotional intensity and orientationsbased on general uniform distributions. The emotionalintensity levels for each agent were kept fixed in each sim-ulation while the orientations evolved due to noise andinteractions. For each opinion spectrum extracted fromTwitter data, we performed a simulation until a steadystate was reached. After that, the following statisticswere collected: steady state distribution of the opinionalong the “East-West” scale, as defined by the boundaryconditions, mean order parameter, and bimodality index.We should note here that in this setup the week-by-week series of calculated properties do not reflect the realtime, nor the actual system’s dynamics as the history ofthe individuals as well as previous steady states are ig-nored. To follow the variation of collective properties,ideally one should look at the evolution of each user’sopinion and derive the group behavior from the corre-sponding statistics. For this purpose, one would needto either parameterize the individual opinions from theempirical data directly or solve the inverse problem andintroduce the variation of opinions based on the instan-taneous statistical averages. As we could reach only arandom sample of the tweets, it was not possible to followthe former route and track individual users. The empir- ical data we have are discontinuous. Moreover, we haveno appropriate model for individual psychology. There-fore, we decided not to introduce an artificial evolutionof the opinions. As the individual history is lost, we canonly follow the variation of the averages correspondingto the snapshot of Twitter data. C. Model parameterization with Twitter data
To parameterize the Bounded confidence XY model weused weekly distributions of the overall users’ emotionalintensity scores. For each week, the overall emotionalintensity of each user was defined from the mean of theaverage sentiment scores on a continuous scale from 0to 5, with 0 corresponding to a neutral average user’sopinion and 5 reflecting an extremely emotional averageopinion. This measure did not contain any informationabout political affiliation, so that each non-zero valuecould correspond to either pro-West or pro-East user’spolitical attitude. The data was sampled with a bin sizeof 1 for all non-zero opinions and a resulting discrete dis-tribution was normalized by the total number of usersper week giving probabilities for each discrete value ofoverall emotional intensity (0 to 5). This distributionwas then applied to the simulated system to define thelength p of each opinion vector. In each simulation, val-ues of p were assigned to agents at random, accordingto obtained discrete weekly distributions of overall emo-tional intensity, and kept constant throughout the simu-lation. The simulated system consisted of 14641 agentsplaced in nodes of 120 ×
120 lattice. We performed atleast 10 update cycles to determine the structure of thesteady state. Each statistics was averaged over at least5 independent runs. We computed the bimodality coef-ficient as β = ( γ + 1) /k , where γ is the skewness and k is the kurtosis of weekly distributions of average opinionscores in Twitter data or weekly distributions of cosinesof orientations θ of opinion vectors in simulations. IV. RESULTSA. Opinion Polarization in the UkrainianTwittersphere
To validate the model assumptions, we used TwitterStreaming data from October 2013 to September 2014provided by the Archive Team. We use Twitter data be-cause it provides fine-grained, time-series and rich dataon recent political opinion dynamics in Ukraine, other-wise not available. To determine the political affiliationof the Twitter users in our data, we used a political affilia-tion classification procedure based on keyword and senti-ment analysis suggested and tested by Spaiser et al. [54](see the electronic supplementary material for details).As a result, every Twitter user was assigned two scores,a “West” score and an “East” score, representing their
Figure 3: (A) Pro-against-European (West)/pro-against-Russian (East) opinion space plot for the first week of October 2013.(B) Pro-against-European (West)/pro-against-Russian (East) opinion space plot for the second week of August 2014. Thepoints in plots C and D are Twitter users based on their two scores, the orange dot is the average. (C) Twitter users (dots)colored according to their “East” affiliation score and plotted over the Ukraine map using Twitter geo, place or locationinformation. (D) Twitter users (dots) colored according to their “West” affiliation score and plotted over the Ukraine mapusing Twitter geo, place or location information. political position in a pro-against-European (West)/pro-against-Russian (East) opinion space (see Fig. 3A,B) andan emotional intensity score, combining their sentimentanalysis scores.Our Twitter analyzes show that political polarizationdid indeed take place in the Ukrainian Twittersphere be-tween October 2013 and September 2014. Fig. 2A showsa discontinuous increase in emotional users in February2014, while the total number of users and the number ofneutral ones increased steadily. Fig. 2B depicts a ratioof emotional/neutral opinions, where a jump from valuesof about 1.0 to about 1.4 is visible around February aswell. We added to this figure the key political incidentsin Ukraine during this year, so the discontinuous changesin the data can be related to actual political events. Thisshows that the biggest discontinuous change in Februarytook place around the time when the Maidan protestsescalated and 100 people died on a single day. These re-sults inspired the inclusion of emotional intensity levelsas a parameter in the computational model. In additionto an abrupt increase in the total number of users (Figs.2A,C), users’ involvement in the topic also changed dis-continuously (Fig. 2D). The dynamics of the averagenumber of tweets per user completely resembles that ofthe emotional level of tweets (Fig. 2B), with a charac-teristic jump around week 19 (February 2014). Fig. 3A,B moreover shows the polarization in a con-tinuous two-dimensional space defined by pro-against-European (West)/pro-against-Russian (East) scores.This figure shows that the polarization resulted in ap-pearance of two main opinion clusters, those who arepro-East and against-the-West on the one side and thosewho are against East but who are also quite critical of theWest on the other side (see the electronic supplementarymaterial for details). This in fact reflects most currentsurveys that show the disappointment of many Westernoriented Ukrainians with the pro-West Poroshenko gov-ernment [55].Our Twitter data analysis confirms that political affil-iation follows the expected geographical pattern. Peoplewho are pro-East and against-West are more likely to belocated in the Eastern and Southern parts of Ukraine,while Ukrainians with rather an against-East and (crit-ical) pro-West attitude are to be found in the Westernand Northern territories (see Fig. 3C,D). This analysisis a validation of our classification procedure, since thegeographical distribution of the Twitter users with theirrespective political affiliation scores matches the actualgeographical political camp distributions in Ukraine (seeFig. 1). A Figure 4: (A)Time series plot of coefficient of bimodality of Twitter users’ opinion distribution. (B) Time series plot ofcoefficient of bimodality of vector orientations in simulations ( η = 0 . B. Validating the Bounded Confidence XY Modelof Opinion Polarization with Twitter data
To quantify the opinion divide, we present here thebimodality index for the opinion spectrum (see Materi-als and Methods for the definition), as we found it to bemost sensitive to the changes of the spectrum, therefore,the best integral characteristic of the observed behavior.The changes in the bimodality coefficient over 12 monthsfrom October 2013 are shown in Fig. 4A,B. In the Twit-ter data the bimodality keeps as low as 0.2 to 0.26 fromthe start of observation until week 18 (February 2014) butthen demonstrates sudden increase to 0.35 to 0.40 withinthe next two or three weeks, after which it stays high,and the original value is never restored. This suddenincrease of the bimodality reflects a formation of two dis-tinct political camps and clearly shows that the opiniondivide has suffered a significant deepening in this shortperiod. Moreover, we see that the deepening was non-recoverable in the short term. The computational model,parameterized with the Twitter emotional intensity spec-trum for each respective week, captures this behavior welland shows a similar qualitative trend. In a simulationwith α = 0 . π , the bimodality coefficient jumps fromca. 0.60 – 0.65 to 0.8 – 0.85 during the same period. Thequantitative difference in the values is mostly due to theuse of cosine function. We should note that, althoughthe model is parameterized by the Twitter data, it isnot bound to reproduce the distribution, as the vectorsare allowed to change the orientation, and this behav-ior follows only from the specific anisotropic interactionsbetween the agents. We repeated the simulations witha different interaction parameter α = 1 .
0, correspondingto agents without any resistance to opinion change andfound no jump in the bimodality. Therefore, the opiniondivide is conditioned by both the restriction of confidenceand by conformity (noise) levels, thus confirming the ho-mophily and biased assimilation assumptions. Moreover,the drastic change in bimodality corresponds to the sud- den increase of the emotional intensity, which we notedin the data in Fig. 2B.
C. Onset of Opinion Clustering and Formation ofTerritorial Domains
The increase in emotional intensity leads to importantconsequences in the spatial dimension. In the weeks be-fore and early at the outbreak of the crisis, the simula-tions display large diversity of the opinions (Fig. 5C)characterized by the large number (Fig. 5A) of small(Fig. 5B) opinion domains (clusters). This picture,however, changes further into the crisis, around week19, when a smaller number of larger opinion domainsis formed. While clusters rarely exceed 100 agents beforethe critical weeks 17-18, starting from week 19 we observeclusters of up to 5000 agents and the numbers rarely dropbelow 1000. That the behavior becomes expressly collec-tive is consistent with the rise of emotional intensity andof the fraction of involved agents (those with non-zero va-lence), which increases the number of interactions withineach individual’s circle and thus the local aligning field.This is further confirmed by Fig. 5F,G showing the po-lar opinion histogram for the two critical time points,week 18 and 19 obtained from simulation analyses. Thedistributions are visibly gravitating towards 0 or 180 de-grees (“East” and “West”) in both sets. The fraction ofneutral opinions drops from week 18 to week 19 and thedistribution of opinions in each subgroup becomes verynarrow.Simulation analysis moreover shows that the opiniondivide induces also territorial splitting. Subfigures Fig.5C,D show the in-plane opinion distribution. We plot-ted the opinions as predicted by the model just beforethe jump in the bimodality (week 18 – C) and immedi-ately after that (week 19 – D). The change in the dis-tribution between these points is dramatic: while thepredicted data for week 18 show a merely uniform dis-
Figure 5: (A)Time series plot of average number of clusters in simulations. (B) Time series plot of maximal cluster size insimulations. (C) Simulation snapshot for week 18, February 2014. (D) Simulation snapshot for week 19, February 2014. Eachsquare in plots C and D represents an individual agent; size of a square is proportional to emotional intensity; color of eachsquare denotes cosine of an orientation angle for each spin, with -1 and 1 corresponding to orientation towards “West” and“East”, respectively. (E) Log-log plot of distribution of cluster sizes in simulations for week 18, February 2014 (blue) andweek 19, February 2014 (red). (F) Simulation cluster orientation diagram for week 18, February 2014. (G) Simulation clusterorientation diagram for week 19, February 2014. Orientation of clusters in plots F and G is shown in degrees; 180 and 0degrees denote orientation towards “West” and “East”, respectively; length of each bin reflects cluster probability. Simulationparameters in plots A–G are α = 0 .
15 and η = 0 . tribution of both “East” and “West” orientations, thepicture for week 19 features two distinct clusters withpredominant “West” orientation in the lower left cor-ner and domination of the “East” orientation in the topright corner. These orientations correspond to vector di-rections in the preset boundary conditions. Each do-main contains practically no opposing opinion, as theyare squeezed out to the periphery and then to the op-posing domain as the steady state develops. We can seesmall islands of mixed/neutral opinion in the middle ofthe simulation domain. The prediction of the territorialdivide matches also well the geo-location data shown inFig. 3C,D. We should stress however that the biasedboundary conditions alone are not sufficient to produceany large domains even in the system with limited con-fidence (Fig. 5C, see the electronic supplementary ma-terial for details), although they definitely facilitate thiscollective behaviour.The simulation allows us to analyze the steady statesof the system and provides insights into the mechanismsof sudden onset of polarization, clustering, and territo-rial splitting (see the electronic supplementary materialfor details). We in particular examined varying levels ofnoise, restriction angles and vector lengths. High noise(low conformity) corresponds to a globally disordered be-haviour (without any prominent consensus or polariza-tion) and the range of higher α allows only states withpolar order (global consensus). Smaller restriction an- gles, α < . π , on the other hand produce regions ofprevalence of the bipolar states, thus structure the sys-tem in a polar or bipolar way. A combination of smallrestriction angle (strongly bounded confidence) and highnoise (low conformity) does not produce any global or-der and is rather unrealistic since at these conditions thesystem represents a set of selective but randomly vacil-lating agents. Polarized states are generally only possibleat low noise (high conformity) and small restriction an-gles (strongly bounded confidence). The changing levelof emotional intensity pushes the boundary between thepolarized and non-polarized societies moreover outwards,thus extending the range of the polarized states, andbrings the originally weakly polarized society to a highlypolarized one. This splitting resembles in appearance aphase separation in dissimilar liquids (e.g. oil in water).The important difference of our system from the liquidstate systems is that the agents are not dissimilar fromthe beginning but the dissimilarity and effective repul-sion between the opposite arises from strong social inter-actions that dictate cohesion between the like opinions.Another important observation here is that the transi-tion is driven primarily by the increase of emotional levelwhile the other parameters (conformity, confidence) stayconstant. We should stress that the source of and theoriginal direction of the emotional agents were not cru-cial. The key properties of the model that determine thenature of the steady states (non-polar, polar or bipolar)are the high conformity and strongly bounded confidence. V. CONCLUSION
We have analyzed the opinion dynamics over a recentperiod of political unrest in Ukraine. Based on the Twit-ter data, we registered an onset of emotional intensityof tweets that corresponded to rising levels of involve-ment of the population in the political feud, fueled bythe action of the government, collisions of the opposinggroups, and foreign military activities. The escalatingopinion divide around the time became apparent amongothers in the jump of the opinion bimodality index. Weproposed an agent-based lattice model to study politicalpolarization as a collective behavior including the spatialdimension of polarization. By parameterizing the modelwith Twitter data at distinct time points, we predictedthe onset of collective behavior and territorial splittingof the opinion. We demonstrated that the tendency ofterritorial splitting is conditioned by the high conformityand homophily in the society and is driven by the growthin emotional intensity. Our analyzes demonstrate clearlythe importance of emotional intensity for polarization, afactor that has been largely ignored thus far in classictheoretical and empirical literature on polarization witha few noteworthy exceptions as discussed earlier. Specifi-cally, while our analyzes confirm the importance of socialcomparison, homophily, persuasive argument and biasedassimilation mechanisms and their specific interactionsfor polarization, they also show that these mechanismsare not sufficient to ignite the extreme societal polariza-tion we can observe for instance in Ukraine. The emo-tional intensity is a key reinforcement mechanism thathas to be added.Our analysis seem also to suggest a link between polar-ization and separatist trends. The polarization dynamicsthat we establish for February 2014 in the data and sim-ulation increases in fact further around April and May2014, when the self-declared Donetsk and Luhansk Peo-ple’s Republics were formed in the South-East of Ukraine.This seems to suggest that the opinion split may facili-tate the separatist trends on its own. The observed phe- nomenon is not unique to Ukraine, and similar processesof deepening polarization leading to separatism are wellknown elsewhere (e.g. Northern Ireland). In most cases,however, the separatism is related to more obvious eth-nic, racial, or religious differences between the commu-nities, while in Ukraine the division is more subtle androots in small cultural differences, which were artificiallyenhanced by external factors. And our results show howdangerous targeted agitation can be when it is backedby modern information warfare techniques [56, 57] as itboils emotions making polarized world views increasinglyirreconcilable.
Ethics
Tweets were collected and analysed in accordance withthe Twitter Privacy Policy and the Twitter Terms of Ser-vice.
Data accessibility
The datasets supporting the conclusions of this articleare provided by the Archive Team [47].
Authors’ contributions
MR: Development and Testing of the Bounded Con-fidence XY Model, Parametrization and Testing of theModel with Twitter Data, Development of Ukrainian dic-tionary for sentiment analysis, Preparing graphics, Writ-ing the paper, VS: Processing and Analyzing Twitterdata, Preparing graphics, Writing the paper, TI: Devel-opment of the Bounded Confidence XY Model, Writingthe paper, VL: Development of the Bounded ConfidenceXY Model, Writing the paper.
Acknowledgements
The authors thank Ian Harper for help with program-ming and testing of the web applet. [1] James Bell, Katie Simmons, and Russ Oates. De-spite concerns about governance, Ukrainians wantto remain one country. Pew Research Center. , 2014. Ac-cessed on: 30 May 2017.[2] European Commission. Humanitarian Air andCivil Protection. Ukraine. ECHO Factsheet. , 2016. Ac-cessed on: 30 May 2017.[3] BBC. Ukraine’s sharp divisions. ,2014. Accessed on: 30 May 2017. [4] Ukraine’s 2001 census. http://2001.ukrcensus.gov.ua .Accessed on: 14 Aug 2017.[5] Dirk Helbing and Stefano Balietti. Howto do agent-based simulations in the future:From modeling social mechanisms to emer-gent phenomena and interactive systems design. ,2011. Accessed on: 30 May 2017.[6] Mark Birkin and Nick Malleson. Investigatingthe behaviour of Twitter users to construct anindividual-level model of metropolitan dynamics. , 2012. Ac- cessed on: 26 June 2018.[7] Emilio Serrano, Carlos A. Iglesias, and Mercedes Garijo.A novel agent-based rumor spreading model in Twitter. In Proceedings of the International World Wide Web Confer-ence (IW3C2) , 2015.[8] William Rand, Jeffrey Herrmann, Brandon Schein, andNeˇza Vodopivec. An agent-based model of urgent diffusionin social media.
Journal of Artificial Societies and SocialSimulation , 18(2):DOI: 10.18564/jasss.2616, 2015.[9] Bruno Gon¸alves, Nicola Perra, and Alessandro Vespig-nani. Modeling users’ activity on Twitter networks: Val-idation of dunbar’s number.
PLOS ONE , 6(8):1–5, 082011.[10] L. Weng, A. Flammini, A. Vespignani, and F. Menczer.Competition among memes in a world with limited atten-tion.
Scientific Reports , 2:00335, 03 2012.[11] Misako Takayasu, Kazuya Sato, Yukie Sano, Kenta Ya-mada, Wataru Miura, and Hideki Takayasu. Rumor diffu-sion and convergence during the 3.11 earthquake: A Twit-ter case study.
PLOS ONE , 10(4):1–18, 04 2015.[12] Elizabeth Suhay. Explaining group influence: The roleof identity and emotion in political conformity and polar-ization.
Political Behavior , 37(1):221–251, 2015.[13] Daniel J. Isenberg. Group polarization: A critical reviewand meta-analysis.
Journal of Personality and Social Psy-chology , 50(6):1141–1151, 1986.[14] Serge Moscovici and Marisa Zavalloni. The group as apolarizer of attitudes.
Journal of Personality and SocialPsychology , 12(2):125–135, 1969.[15] Morris H. DeGroot. Reaching a consensus.
Journal of theAmerican Statistical Association , 69(345):118–121, 1974.[16] Cass R. Sunstein. The law of group polarization.
TheJournal of Political Philosophy , 10(2):175–195, 2002.[17] Robert S. Baron and Gard Roper. Reaffirmation of socialcomparison views of choice shifts: Averaging and extrem-ity effects in an autokinetic situation.
Journal of Person-ality and Social Psychology , 33(5):521–530, 1976.[18] David G. Myers, Sandra Brown Wojcicki, and Bobette S.Aardema. Attitude comparison: Is there ever a band-wagon effect?
Journal of Applied Social Psychology ,7(4):341–347, 1977.[19] Miller McPherson, Lynn Smith-Lovin, and James M.Cook. Birds of a feather: Homophily in social networks.
Annual Review of Sociology , 27:415–444, 2001.[20] Sergio Currarini, Matthew O. Jackson, and Paolo Pin.An economic model of friendship: Homophily, minorities,and segregation.
Econometrica , 77(4):1003–1045, 2009.[21] Pranav Dandekar, Ashish Goel, and David T. Lee. Biasedassimilation, homophily, and the dynamics of polarization.
PNAS , 110(15):5791–5796, 2013.[22] Eugene Bernstein. Persuasion as argument processing.In Hermann Brandstatter, James H. Davis, and GiselaStocker-Kreichgauer, editors,
Contemporary problems ingroup decision-making , pages 103–124. NewYork: Aca-demic Press, 1982.[23] Ebbe B. Ebbsen and Richard J. Bowers. Proportion ofrisky to conservative arguments in a group discussion andchoice shifts.
Journal of Personality and Social Psychol-ogy , 29(3):316–327, 1974.[24] Amiram Vinokur and Eugene Bernstein. Novel argumen-tation and attitude change: The case of polarization fol-lowing group discussion.
European Journal of Social Psy-chology , 8(3):335–348, 1978.[25] Charles G. Lord, Lee Ross, and Mark R. Lepper. Biased assimilation and attitude polarization: The effects of priortheories on subsequently considered evidence.
Journalof Personality and Social Psychology , 37(11):2098–2109,1979.[26] Arthur G. Miller, John W. McHoskey, Cynthia M. Bane,and Timothy G. Dowd. The attitude polarization phe-nomenon: Role of response measure, attitude extremity,and behavioral consequences of reported attitude change.
Journal of Personality and Social Psychology , 64(4):561–574, 1993.[27] Geoffrey D. Munro, Peter H. Ditto, Lisa K. Lockhart,Angela Fagerlin, Mitchell Gready, and Elizabeth Peterson.Biased assimilation of sociopolitical arguments: Evaluat-ing the 1996 U.S. presidential debate.
Basic and AppliedSocial Psychology , 24(1):15–26, 2002.[28] Charles S. Taber and Milton Lodge. Motivated skepti-cism in the evaluation of political beliefs.
American Jour-nal of Political Science , 50(3):755–769, 2006.[29] Wayne E. Baker.
America’s Crisis of Values: Reality andPerception . Princeton, NJ: Princeton University Press,2005.[30] Michael Dimock, Carroll Doherty, Jocelyn Ki-ley, and Russ Oates. Political polarizationin the American public. Pew Research Center. , 2014. Ac-cessed on: 26 June 2018.[31] Delia Baldassarri and Peter Bearman. Dynamics ofpolitical polarization.
American Sociological Review ,72(5):784–811, 2007.[32] Claudio Castellano, Santo Fortunato, and VittorioLoreto. Statistical physics of social dynamics.
Reviewsof Modern Physics , 81(2):591–646, 2009.[33] Michael M¨as, Andreas Flache, K´aroly Tak´acs, andKaren A. Jehn. In the short term we divide, in the longterm we unite: Demographic crisscrossing and the effectsof faultilines on subgroup polarization.
Organization Sci-ence , 24(3):716–736, 2013.[34] Guillaume Deffuant, David Neau, Frederic Amblard,and G´erard Weisbuch. Mixing beliefs among interactingagents.
Advances in Complex Systems , 3:87–98, 2000.[35] Rainer Hegselmann and Ulrich Krause. Opinion dynam-ics and bounded confidence models, analysis, and simula-tion.
Journal of Artificial Societies and Social Simulation ,5(3):1, 2002.[36] Santo Fortunato, Vito Latora, Alessa Ndro Pluchino,and Andrea Rapisarda. Vector opinion dynamics in abounded confidence consensus model.
International Jour-nal of Modern Physics C , 16(10):1535, 2005.[37] Jan Lorenz. Continuous opinion dynamics underbounded confidence: A survey.
International Journal ofModern Physics C , 18(12):1819, 2007.[38] Eytan Bakshy, Solomon Messing, and Lada Adamic. Ex-posure to ideologically diverse news and opinion on Face-book.
Science , 348(6239):1130–1132, 2015.[39] Michael D. Conover, Jacob Ratkiewicz, Matthew Fran-cisco, Bruno Goncalves, Filippo Menczer, and AlessandroFlammini. Political polarization on twitter. In
Fifth Inter-national AAAI Conference on Weblogs and Social Media ,pages 89–96, 2011.[40] Sarita Yardi and Danah Boyd. Dynamic debates: Ananalysis of group polarization over time on twitter.
Bul-letin of Science, Technology & Society , 30(5):316–327,2010.[41] Molly Asher, Cristina Leston Bandeira, and Vikto- ria Spaiser. Assessing the effectiveness of e-petitioningthrough Twitter conversations. In Political Studies Asso-ciation Annual Meeting 2017 , 2017.[42] Anatoliy Gruzd and Ksenia Tsyganova. Politically po-larized online groups and their social structures formedaround the 2013–2014 crisis in Ukraine.
Internet, Poli-tics, Policy 2014: Crowdsourcing for Politics and Policy. ,2014.[43] David Stern. The Twitter war: So-cial media’s role in Ukraine unrest. , 2014. Ac-cessed on: 26 June 2018.[44] Alexander Ronzhyn. The use of facebook and Twitterduring the 2013-2014 protests in ukraine. In
Proceedingsof the European Conference on Social Media: ECSM 2014 ,2014.[45] Elanor Colleoni, Alessandro Rozza, and Adam Arvids-son. Echo chamber or public sphere? Predicting politicalorientation and measuring political homophily in Twitterusing Big Data.
Journal of Communication , 64(2):317–332, 2014.[46] Pablo Barber´a. Birds of the same feather tweet together:Bayesian ideal point estimation using Twitter data.
Po-litical Analysis , 23(1):76–91, 2015.[47] Archive Team: The Twitter Stream Grab. https://archive.org/details/twitterstream . Ac-cessed on: 30 May 2017.[48] Mi Kyung Lee, Ho Young Yoon, Marc Smith, Hye JinPark, and Han Woo Park. Mapping a twitter scholarlycommunication network: a case of the association of inter-net researchers’ conference.
Scientometrics , 112(2):767–797, 2017.[49] Tamas Vicsek, Andras Czir´ok, Eshel Ben-Jacob, Inon Cohen, and Ofer Shochet. Novel type of phase transi-tion in a system of self-driven particles.
Physical ReviewLetters , 75:1226–1229, 1995.[50] Maksym Romensky, Vladimir Lobaskin, and ThomasIhle. Tricritical points in a vicsek model of self-propelledparticles with bounded confidence.
Physical Review E ,90:063315, 2014.[51] Bounded Confidence XY Model: Online Simulator. http://maksymromensky.com/bounded_confidence_xy_model .Accessed on: 24 June 2018.[52] Pawel Sobkowicz. Discrete model of opinion changes us-ing knowledge and emotions as control variables.
PLoSONE , 7(9):1–16, 09 2012.[53] Pawel Sobkowicz. Minority persistence in agent basedmodel using information and emotional arousal as controlvariables.
The European Physical Journal B , 86(7):335,2013.[54] Viktoria Spaiser, Thomas Chadefaux, Karsten Donnay,Fabian Russmann, and Dirk Helbing. Communicationpower struggles on social media: A case study of the 2011-12 Russian protests.
Journal of Information Technology &Politics , 14(2):132–153, 2017.[55] Julie Ray. Ukrainians disillusioned with leadership. , 2015. Ac-cessed on: 30 May 2017.[56] Christina Cottiero, Katherine Kucharski, EvgeniaOlimpieva, and Robert W. Orttung. War of words: theimpact of russian state television on the russian internet.
Nationalities Papers , 43(4):533–555, 2015.[57] Jessikka Aro. The cyberspace war: propaganda andtrolling as warfare tools.
European View , 15(1):121–132,2016.
Supplementary Material for: Polarized Ukraine 2014: Opinion and Territorial SplitDemonstrated with the Bounded Confidence XY Model, Parameterized by TwitterData
Maksym Romenskyy,
Viktoria Spaiser, Thomas Ihle, Vladimir Lobaskin Department of Life Sciences, Imperial College London, London SW7 2AZ, UK Department of Mathematics,Uppsala University, Box 480, Uppsala 75106, Sweden School of Politics and International Studies, University ofLeeds, Leeds LS2 9JT, UK Institute of Physics, University of Greifswald, Felix-Hausdorff-Str. 6, Greifswald 17489,Germany School of Physics, University College Dublin, Belfield, Dublin 4, Ireland (Dated: July 26, 2018)
Supplementary MethodsFurther Model Specifications
Supplementary Figure S1 illustrates interactions in the Bounded Confidence XY model. The red opinion vectorinteracts with its 24 neighbors located within two nearest lattice nodes in any direction r = 2. In this illustration,each of the 24 neighbors has orientation relative to the focal (red) vector less than α and therefore all 24 agentscontribute to future orientation of the focal vector. For simplicity, all vectors in Supplementary Fig. S1 are shown tohave same length. In all simulations, the initial orientations of all agents except the ones at the system’s boundarywere drawn randomly from the uniform distribution in the interval [ − π, π ]. α Figure S1: The interaction parameters in the Bounded Confidence XY model with r = 2. The focal opinion vector (red arrow)interacts only with those neighbors whose relative orientation is less or equal to α and who are located within the two nearestrows/columns (blue arrows). To quantify formation of territorially isolated domains in our model, we performed a simple cluster analysis. Wedefine a cluster based on two criteria: distance and relative orientation between the agents. Therefore, a cluster isa set of connected agents, each of which is within the cutoff distance (defined by r ) from one or more other opinionvectors from the same cluster and a relative orientation between any two neighbors in the cluster is less or equal torestriction angle α . Conversely, two agents will not belong to the same cluster, if there is no continuous path on theneighbor network leading from the first agent to the second or if this path is broken because the angle between twoneighbors is larger than α and hence the opinion vectors do not interact. For each update step we calculated size ofeach cluster, maximal cluster size and total number of clusters.We characterised orientational ordering in our model using two order parameters. Polar order parameter was usedto quantify the average degree of opinion agreement between the agents ϕ = 1 N (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N X j =1 exp( ıθ j ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) , (S1) Figure S2: Phase behavior of the Bounded Confidence XY model. (A) Phase diagrams based on polar ϕ and bipolar Q order parameters for the model system without and with parameterization by Twitter data. (B) Combined phase diagramsfor the model system without and with parameterization by Twitter data showing regions of dominance of polar and bipolarorder. (C) Phase diagrams based on polar ϕ and bipolar Q order parameters for non-parameterized model system at differentsimulation parameters (see legend). (D) Combined phase diagrams showing regions of dominance of polar and bipolar orderfor unparameterized model system at different simulation parameters (see legend). The red plus marker denotes simulationparameters α and η for the parameterized version of the model used throughout the paper. In plots C and D, s denotes size ofa side of a square lattice; E stands for East and means that all boundary spins are fixed to the right ( θ = 0), W − E stands forWest – East and means that the boundary spins on the left and on top are fixed to the left ( θ = π ) and the boundary spins onthe right and on the bottom are fixed to the right ( θ = 0); r denotes the interaction range. where ı is the imaginary unit and θ j is the direction of each vector j . This order parameter turns zero in the isotropicphase, when opinions of all agents are very different from each other, and assumes finite positive values in the orderedphase, reaching unity when global consensus settles in.To characterise opinion polarization in our model we use the following bipolar order parameter Q = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N N X j =1 exp( ı θ j ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . (S2)When the two vectors are oriented perfectly collinearly, Q = 1. Note that a perfectly polarly ordered phase ischaracterized by ϕ = Q = 1, as the polar ordering implies the bipolar ordering. A bipolarly ordered phase requiresonly Q = 1 while the polar order parameter can take any value ϕ <
1. Therefore, requirements for the polar orderare more restrictive.For each weekly distribution of opinions, both for the Twitter data and in simulations, we computed the bimodality − − − Contra/Pro West C on t r a / P r o E a s t − − − − Contra/Pro West C on t r a / P r o E a s t A B Figure S3: (A) Panel Twitter users (black dots) in the pro-against-European (West)/pro-against-Russian (East) opinion space,data from first week in October 2013. (B) Panel Twitter users (black dots) in the pro-against-European (West)/pro-against-Russian (East) opinion space, data from first week in September 2014. coefficient (see main text, Methods) by first calculating the skewness and kurtosis of the distribution. Skewness isdefined as the third standardised moment around the mean γ = µ µ / , (S3)where µ and µ are the second and the third cumulants, respectively. Kurtosis is computed as the fourth centralmoment k = µ µ , (S4)where µ and µ are the second and the fourth cumulants, respectively.The model phase diagrams, shown in Supplementary Figs. S2 A-D, demonstrate three different steady states ofthe system. The high noise (low conformity) generally corresponds to a globally disordered behavior (without anyprominent consensus or polarization). The region of disordered behavior shrinks as the restriction angle α is increased(Supplementary Fig. S2 A). The lesser restriction leads to more interactions between the nearest neighbours and to theonset of order (consensus). At lower noise, below the transition line we observe either polar or bipolar structuring of thesystem (Supplementary Fig. S2 B). The range of higher α allows only states with global consensus, while at α < . π we see the region of prevalence of the bipolar polarized states. In other words, the polarized states are only possibleat low noise (high conformity) and small restriction angles (strongly bounded confidence). The change in level ofemotional intensity pushes the boundary between the polarized and non-polarized societies outwards (SupplementaryFig. S2 B), thus extending the range of the polarized states, and brings the originally weakly polarized society to ahighly polarized one. The most important observation here is that the transition happens just due to the increase ofemotional intensity while the other parameters (conformity, confidence) stay constant. Supplementary Figures S2 C,Dshow phase diagrams for unparameterized systems at different simulation parameters. For a larger system, s = 240constituting of 58081 agents, the area of bipolar ordering is slightly reduced as compared to the standard system size( s = 120) used in this study. The shrinking happens because the influence of biased boundary vectors is decreaseddue to larger system size. If the agents have larger interaction range, r = 3 (i.e. each vectors interacts with up to 48nearest neighbors), the region of bipolar ordering expands to larger restriction angles and noise values because longercorrelation range becomes possible and effect of the boundary bias is more pronounced at these conditions. Finally,if all boundary vectors are fixed in one direction (in this case to the East, θ = 0), the area of bipolar polarizationshrinks significantly but does not disappear. At these conditions bipolar ordering also displays more sensitivity withrespect to noise than to restriction angle. Twitter Data
The archived Twitter Streaming API Spritzer Sample tweet data was stored in JSON ( J ava S cript O object N otation) file format, which is the most common open standard data format to transmit data objects in asyn-chronous browser/server communication. The data was then processed and analyzed in Python using the PythonNatural Language Text Processing Toolkit (NLTK) [S1, S2]. NTLK is a collection of various classes, interfaces andfunctions for natural language processing, including important text mining methods. It was developed at the Univer-sity of Pennsylvania and is widely used in computational linguistics, machine learning, and cognitive science (see inparticular https://nltk.googlecode.com/svn/trunk/doc/api/index.html ).The collected tweets were filtered, first for Russian and Ukrainian language, identifying respective alphabet char-acters in the tweet text. Tweets, where Twitter users specified being from Russia or any region/town in Russia wereremoved from the data. We also filtered for SPAM tweets (around 17% of all the tweets in our data), using the followingkeywords, based on word count analyses: porn, phone, games, androidgames, minecraft, ipadgames, loosing weight, gold,mtvstars, crossword, barbie, sony, holidays, shop, TV series, volkswagen, sales, starlet, apartments, estate, bmw, mercedes,sex, happysex, viagra, stock, prostitutes, teen, price, diet, buy, credit in English, Ukrainian and Russian language. Tweetscontaining these words were removed from the data. Furthermore, we filtered the data for political context with anextensive set of keywords to remove irrelevant tweets and therefore unnecessary noise from the data: protest, mobiliza-tion, Ukraine, kiborgi, glory to the heroes (one word hashtag in Ukrainian and Russian - geroiamslava), junta, government,duma, kremlin, parliament, rada, premier ministry, president, minister, ministry, political action, demonstration, opposition,power, authorities, democracy, nationalists, communists, liberals, Putin, Poroshenko, political party, politics, politicians,political, policy, revolution, citizen, criticism, critics, agitation, sanctions, censoring, illegal, legal, solidarity, assembly, rally,law, regulation, resistance, civil disobedience, resist, reforms, communism, capitalism, administration, news, society, viola-tion, RNBO, Iaijtheniukh, ukr, ukry, khokhly, okraina, ruina, raguli, poproshenko, papashenko, papasha, bacon to the heroes(a wordplay in Russian with glory to the heroes - geroiamsala), civil war, nazis, Bandera, banderlogi, bandery, banderovtsy,benderovets, visitka Iarosha, titushki, EuroMaidan, U revolution, peaceful march, radio freedom, inforesist, Aronets, av-tomaijdan, Ukraine truth news (one word hashtag), SOS Maidan (one word hashtag), digital Maidan (one word hashtag),Maidan history (one word hashtag), sidemaidan, Maidan, freedom, Ukrainian, NATO for Ukraine (one word hashtag), will,choice, against, anti-Maidan, ukrop, Europe, USA, Kiev, Timoshenko, Cameron, dead, killed, negotiations, military service,Zakharchenko, police, Obama, war, activists, right sector, Azarov, upper, leader, Jatseniuk, Turchinov, Klitchko, MID,Nayyem, briginets, constitution, Avakov, MVD, Yushenko, Kuchma, Kravchuk, Kharkov, Grushevskogo, warriors, soldiers,prisoners, Euro, EU, military, Sloviansk, army, West, East, Mariupol, reconciliations, NATO, ATO, crimealook, conflict,Medvedchuk, Medvedchukov, geo-politics, confrontation, Strasbourg, crisis, Lutsenko, Tusk, peace, world, Vladimirova,Vladimir, relations, unity, national, help, oligarchs, glory to Ukraine (one word hashtag), hundred, over, own will, self-determination, power, annexion, annexing, separatism, separatist, rebels, freedom fighters, freedom fight, moskali (meaningMoscow sympathizers), [Russian] official, Putin khujlo (offensive word, one word hashtag), luganda, lungandon, daunbas,zrada, occupants, MGB, Zakhar, Russian peace (one word hashtag), Russian world (one word hashtag), Putin is murder(one word hashtag), Putler, Putin help (one word hashtag), putinism, insurgents, insurgency, small/so what Russia, Crimeais ours (one word hashtag), At least Crimea is our (one word hashtag), punitive, ptnpnkh, ptn, pnkh, vata, vatnik, glory toRussia (one word hashtag), Russia, provocation, Novorossiia, DNR, LNR, berkut, kiberberkut, glory, stopcrimeantatarsgeno-cide, russiainvadeukraine, russiaviolatedceasefire, stoprussianaggression, weapon, Russian, Medvedev, Moscow, legislation,Yanukovich, truth, Donetsk, Luhansk, Russian march (one word hashtag), Russians, Ukrainians, Non-russians, Navalny,monument, Lenin, western, eastern, geek, Lavrov, grad, humanitarian, anti-Russian, SSSR, Donbass, Bafana, elections,legislator, fire, Mirakova, ukraintsami, radio, Putin supporters (one word hashtag), Ukraine France, Zakharov, Mironov inUkrainian and Russian language. Tweets that contained at least one of these keywords were kept in the filtered data,otherwise removed.Any analysis of Twitter data faces a number of well-known difficulties [S3]. Some of them, e.g. the SPAM tweetproblems, we have addressed already above. One potential problem is that the sample only includes public tweetsfrom public Twitter accounts. This does not pose a problem in the context of our study though, since we areinterested in the use of Twitter as an instrument of communication in the public sphere. Moreover, Twitter data isnot representative, which again is rather unproblematic for our study because all political groups and their supportersare represented on Twittersphere, so the public debate on Twitter does overall mirror the general public debate andpublic opinions [S4, S5]. Another potential issue is that the sample is based solely on the 1% of all public tweets,which for instance makes it difficult to use the data as panel data (though we’ve done this to probe opinion changeswithin individuals, see Supplementary Fig. S3, the number of observations is however drastically reduced). However,other Twitter data samples offered by Twitter (e.g. Gardenhose with 10% of all pubic tweets) have to be purchased,which makes them often unaffordable for research purposes.One aspect of Twitter data ”richness” is that it is supposedly geo-referenced. However, as already mentioned inthe main manuscript quite a large proportion of Twitter users do in fact not provide any geographical information.Frequently, even if ”geo enabled” was activated by the Twitter user (thus the value was set to ”true”), the actual”geo” or ”place” ”coordinates” value would be ”null” because the device that was used to post the tweet had nogeo-referencing (e.g. GPS) activated. The data that the user provides through the profile ”location” tag is more oftenavailable, however, here we have to rely on the accuracy and honesty of the Twitter users. Moreover, if the informationprovided through ”location” was imprecise, e.g. just ”Ukraine”, we could not use this information for plotting. Didwe however have serious and specific information from the user, e.g. ”Kiev”, then we used that information togenerate a longitude and a latitude coordinate value that we could then plot on a generated Ukraine map in Fig.3C,D in the main manuscript. We deliberately ignored another potential geographical information, the ”time zone”,because it offers only imprecise geographical information. Overall, we could extract form the Twitter data of the lastSeptember week 2014 493 Twitter users for whom we had sufficiently precise geographical data. This data along withthe calculated West and East scores were used to produce the Figures 3C and D in the main manuscript. Twitter Data Analysis
In order to understand the empirical polarization dynamic in Ukraine and furthermore in order to make the datausable for model calibration, it was necessary to identify the political opinion of the Twitter users in our data. First,we identified the Twitter users in our Twitter data based on the value of their ”screen name”. We then compiled twolists of keywords, extracted from the word count analysis of the most common words that contain words associatedwith either the pro-East political camp or the pro-West political camp: annexion: -2 , annexing: -2, separatism: -2,moskali (meaning Moscow sympathizers): -2, [Russian] official: -2, Putin khuijlo (one word hashtag): -3, Luganda: -3,Lugandon: -3, Daunbas: -3, zrada: -2, insurgents: -2, occupants: -2, MGB: 1, Zakhar: 1, Russian peace (one wordhashtag): -2, Putin is murder (one word hashtag): -3, Putler: -3, Putin help (one word hashtag): 2, putinism: -1, freedomfighters (one word): 2, freedom fight (one word): 2, small/so what Russia: 2, Crimea is ours (one word hashtag): 3, Atleast Crimea is ours (one word hashtag): 3, punitive: 3, ptnpnkh: -3, ptn: -1, pnkh: -3, vata: -3, vatnik: -3, glory toRussia: 3, Russia: 0, Russian: 0, Russians: 0, provocation: 0, separatist: -3, separatism: -3, Novorossiia: 2, DNR: 1, LNR:1, sanctions: 1, berkut: 1, kiberberkut: 1, stopcrimeantatarsgenocide: -3, russiainvadedukraine: -3, russiaviolatedceasefire:-3, anti-Russian sanctions (one word hashtag): 2, stoprussianaggression: -3, peace: 0, Putin: 0, weapon: 0, Medvedev:0, Crimea: 0, Moscow: 0, Yanukovich: 0, truth: 0, Donetsk: 0, from Lugansk (one word): 0, Russian march (one wordhashtag): 0, Non-Russians: 0, dead/killed: 0, negotiations: 0, politce: 0, rebels: 0, Lugansk: 0, war: 0, duma: 0, Mockva:-1, government: 0, leader: 0, monument: 1, Lenin: 1, warriors: 0, prisonerns: 0, military: 0, East: 0, army: 0, relations:0, soldier: 0, reconciliations: 0, geek: -2, Lavrov: 0, grad: -1, humanitarian: 1, anti-Russian: 2, SSSR: 1, Donbass:0, Bafana: 0, elections: 1, referendum: 0, confrontation: 0, fire: 0, crisis: 0, conflict: 0, Putin supporters (one wordhashtag): -1, Mironov: 0 were the keywords for the East political camp and mobilization: 1, kiborgi: 2, unity: 1, gloryto the heroes (one word hashtag): 3, RNBO: 1, Iaijtheniukh: -3, junta: -2, ukr: -3, ukry: -3, khokhly: -1, okraina: -1,Ruina: -3, raguli: -3, poproshenko: -2, Papashenko: -2, papasha: -3, bacon to the heroes (a wordplay in Russian withglory to the heroes - geroiamsala): -3, civil war (one word hashtag):- 2, visit Kaiarosha (one word hashtag): 2, nazis: -3,Bandera: -3, banderlogi: -3, bandery: -3, banderovtsy: -3, titushki: 2, Ukraine: 0, Euro Maidan (one word hashtag): 1,U REVOLUTION: 2, peaceful march (one word hashtag): 1, radio freedom (one word hashtag): 1, inforesist: 1, Aronets:1, avtomaijdan: 1, Ukrainian truth news (one word hashtag): 1, SOS Maidan (one word hashtag): 1, digital Maidan (oneword hashtag): 1, Maidan history (one word hashtag): 1, sitemaidan:1, Maidan: 0, freedom:1, government: 1, Ukrainian:0, Urkrainians: 0, NATO for Ukraine (one word hashtag): 2, will: 2, will: self-determination, against: 2, revolution: 2,anti-Maidan:- 2, ukrop: -3, Poroshenko: 0, Europe: 0, USA: 0, Kiev: 0, weapon: 0, Timoshenko: 0, truth: 0, Cameron:0, democracy: 0, dead/killed: 0, negotiations: 0, military service: 0, Zakharchenko: 0, opposition: 0, Obama: 0, war:0, parliament: 0, activists: 0, political action (one word): 0, protest: 0, sector: 0, Azarov: 0, upper: 0, leader: 0, rada:0, Yatseniuk: 0, Turchinov: 0, Klichko: 0, MID: 0, Nayyem: 0, Briginets: 0, constitution: 0, president: 0, Avakov: 0,MVD: 0, Yushchenko: 0, Kuchma: 0, Kravchuk: 0, Kharkov: 0, Grushevskogo: 0, warriors: 0, Euro: 0, EU: 0, military:0, Slaviansk: 0, West: 0, western: 0, army: 0, soldier: 0, Mariupol: 0, reconciliations: 0, NATO: 0, ATO: 1, conflict: 0,Medvedchuk: 0, Medvedchukvv: 0, Strasbourg: 0, crisis: 0, Lutsenko: 0 ,Tusk: 0, peace: 0, relations: 0, national: 1, help:1, glory to Ukraine (one word hashtag): 2, hundred: 2, over: 2, power: 2, BENDERovets: 3, civil war: -2, krimealook:1, Ukraine France: 1 were the keywords for the West political camp. These keywords were scored between -3 and3, with negative scores indicating negative attitudes towards the respective political camp, positive values positiveattitudes and a zero a neutral opinion. Some keywords were unambiguously associated with a political affiliation. Forinstance, the word ”Putler” (scored -3), a composition of Putin and Hitler, is clearly a negatively annotated wordreferencing the East political camp. Similarly, the word ”banderovtsy” (scored -3) linking the West political camp toStepan Bandera, leader of the Ukrainian nationalist and independence movement during the Second World War, whocooperated with Nazi Germany, shows clearly a strong disapproval of the West political camp. On the other handthe hashtag http://sentistrength.wlv.ac.uk ) approach to determine a sen-timent score for each tweet. SentiStrength is a free Java-based automatic sentiment analysis tool, widely used inresearch, which is also available for the Russian language. We created a comprehensive dictionary for Ukrainiansentiment words (to be shared upon request) based on the SentiStrength sentiment scoring system and we reviewedthe sentiment word dictionary that SentiStrength is using for the Russian language and complemented it with other S k e w ne ss ( ab s ) S k e w ne ss ( ab s ) K u r t o s i s K u r t o s i s Figure S4: (A) Time series plot of absolute skewness for opinion distributions in Twitter data. (B) Time series plot of absoluteskewness for opinion distributions in simulations. (C) Kurtosis of the distribution of opinions in Twitter data. (D) Kurtosis ofthe distribution of opinions in simulations. Simulation parameters in plots B and D are α = 0 .
15 and η = 0 . sentiment words (among others offensive words that were missing in SentiStrength) to make it equivalent to theUkrainian dictionary. The sentiment scores in SentiStrength and thus in our two sentiment words dictionaries rangefrom -5 to +5 with 0 signifying a neutral word, negative values a negative sentiment and positive values a positivesentiment. The higher the absolute value the stronger the sentiment. For each tweet we calculated a sentiment scoreaverage based on the identified sentiment words in the tweet. We calculated a separate sentiment score for the Eastpolitical camp related keywords and one for the West political camp related keywords. The emotional intensity scoreis averaging these two sentiment scores.In the case of unequivocal keyword tweets, the tweet score would derive from these keywords. For instance, if akeyword contained the words ”Putler” (scored -3) and the hashtag ”stoprussianaggression” (scored -3), the tweet scorewould be the average of these two scores, thus -3. Whenever a tweet was scored zero because of neutral keywords, weadded to the score the respective sentiment score. Thus if an East political camp related tweet had a zero score (e.g.”Putin”), and the East political camp related sentiment score was -2 (for instance resulting from the sentiment word”zachvatil” (translated: grabbed)), then the tweet would get an East political affiliation score of -2. This is basedon the assumption that users would express positive sentiments about terms associated with their own camp and/ornegative sentiments towards terms associated with the other camp. Since most users would have posted several tweets,users were assigned a set of tweet scores, depending on the number of tweets and from these scores overall averagescores, one East score, one West score, were calculated for each user, representing their political affiliationAutomatic classification and sentiment analysis have certainly their limitations, e.g. automatic sentiment analysisoften fail to spot irony and classifications do not account for instance for complex inner-fraction dynamics, that ispolitical fractions within political fractions. However, manual classification is becoming increasingly impossible withthe growing amount of data and/or limited capacities and resources and thus automatic classification is increasinglyapplied. It is however important to work on further developing and elaborating these tools and to supervise, criticallyreflect and where required correct the process of automatic analysis and its outcomes. Additional Results
We present here some additional results that support our main conclusions in the manuscript. SupplementaryFigure S4 shows the skewness and kurtosis of opinions in the Twitter data and computer simulation. These plotsshow again the discontinuity of political opinions in the Ukrainian Twittersphere, with jumps in February 2014. [S1] S. Bird, E. Klein, and E. Loper,
Natural Language Processing with Python (Sebastopol: O’Reilly, 2009).[S2] J. Perkins,
Python Text Processing with NLTK 2.0 Cookbook (Birmingham: PACKT, 2010).[S3] D. Ruths and J. Pfeffer, Science , 1063 (2014).[S4] D. Stern,