Converging on a nativist core? Comparing issues on the Facebook pages of the Pegida movement and the Alternative for Germany

Abstract

Computational methods offer a new perspective on the evolving agendas of right-wing movements and parties online. This article showcases computational approaches to text analysis (specifically so-called topic models) to diachronically investigate nativist right-wing issues in social media by comparing comments posted on the Facebook page of the Pegida movement to those of the Alternative for Germany. After describing topic modelling as an increasingly popular method and drawing on the literature on right-wing nativism online, we investigate a set of shared issues relevant to the mobilization of commentators, including opposition to Islam, migration, the government and the media. We furthermore show contrastively how issue prevalence differs between the two groups, and how issue shares change over time, in some instances converging on a shared nativist core. We close with a series of suggestions on the utility of computation content analysis for the study of rapidly evolving political agendas.

Keywords

Content analysis Facebook Germany political talk STM topic models

Introduction

The possibilities and challenges afforded to empirical research by novel digital data sources and new computational tools are under intense discussion in media and communication studies, as elsewhere in the social sciences (Golder and Macy, 2014; Salganik, 2017; Sloan and Quan-Haase, 2016). Datafication looms large, both by opening up new analytical avenues into the study of political discourse and by providing innovative computational techniques for the study of communicative actors in a variety of social domains (Lomborg and Bechmann, 2014; Shah et al., 2015). Computational methods represent a set of tools for the analysis of digital traces which are increasingly being used in the interdisciplinary study of communication and media. Examples include network and link analysis supported by crawlers and application programming interfaces (APIs; Himelboim et al., 2013; Kaiser and Puschmann, 2017), sentiment analysis (Ceron et al., 2014; Young and Soroka, 2012), the use of supervised machine learning methods for automated content analysis (Schwartz and Ungar, 2015; Van Atteveldt et al., 2008) and topic modelling (Jacobi et al., 2016; Rauchfleisch, 2017).

In this article, we demonstrate the value of computational text analysis for the study of novel forms of political movements. Specifically, we rely on digital trace data to approach comparatively the discourse of a right-wing nativist movement (Patriotic Europeans Against the Islamization of the Occidenct/Patriotische Europäer gegen die Islamisierung des Abendlands; Pegida) and a political party (the Alternative for Germany/Alternative für Deutschland; AfD) through the lens of user comments posted to public Facebook pages of these actors. Relying on the technique known as topic modelling, we present a repertoire of issues discussed on these pages over an 18-month period. We study user comments, rather than statements made by group organizers or political candidates (cf. Stier et al., 2017, for such a study), allowing us to compare a political party with traditional leadership structures with a grass-roots social movement. Our approach is validated by political hindsight: Anyone following German politics will attest that the movement (Pegida) and the party (the AfD) have converged markedly over time, with the AfD tacking strongly to the right, a development that was less evident at the time when the comments were posted.

Why choose this particular example? The last decade has seen a global rise of both left- and right-wing populist parties and groups, from Italy’s Five Star Movement and the Front National (FN) in France to the UK Independence Party (UKIP) and Germany’s Pegida (Calhoun, 2013; Conti and Memoli, 2015; Pirro and Van Kessel, 2017). Frequently, social media is an essential component of the mobilization efforts of these movements, enabling their internal and external communication as well as connecting them with constituents (Bennett and Segerberg, 2012; Engesser et al., 2016; Margetts et al., 2016). Social media is deployed in distinct ways by different actors engaging in political discourse, such as politicians (Ausserhofer and Maireder, 2013; Gil De Zúñiga et al., 2009) and activists (Bastos et al., 2015; Freelon et al., 2016), in addition to taking on a very different significance across national contexts (Magin et al., 2016; Tufekci and Wilson, 2012). Instances of right-wing nativist protest movements point to web forums, Twitter and Facebook as their political point of origin. Through the nativist lens, traditional media is regarded as biased and politically correct at best and corrupt and beholden to an aloof liberal elite at worst (Engesser et al., 2016). By contrast, social media is regarded as a more accessible and cost-effective channel for coordination and communication (Dostal, 2015, Margetts et al., 2016). As populist movements claim to represent the will of the people, studying user comments enables us to identify and track themes that reflect a particular set of constituent issues, rather than a traditional political programme. Our approach also has the advantage that the supporters of the political movements and parties are frequently unwilling to participate in interviews or surveys, making it difficult to assess their motives. We focus on the following questions:

What issues are discussed in comments on the Facebook pages of Pegida and the AfD and how do the shares of these issues fluctuate over time?

How do topic distributions differ between the two pages, and to what extent do these shares converge or diverge over an extended period?

To what extent can the specifics of this case regarding the ability to track long-term shifts in the political agenda of informal groups and newly formed parties be generalized?

In summary, the case selected here is useful because right-wing activists show particular enthusiasm for social media, because polling them is extraordinarily difficult and because their agendas are particularly volatile and subject to dynamic and surprising shifts which may be evident only retrospectively.

Pegida, the AfD and the Internet

The movement known as Patriotic Europeans Against the Islamization of the Occidenct (Patriotische Europäer gegen die Islamisierung des Abendlands; Pegida) became a force in German politics after organizing grass-roots rallies in the city of Dresden that drew close to 20,000 protesters in 2015, and surging again at the height of the European refugee crisis in 2016 (Dostal, 2015; Vorländer et al., 2016). Attendance dropped considerably from late 2016 onwards, though the movement is still active. Programmatically, Pegida ‘combined fear of “Islamisation” with general criticism of Germany’s political class and the mainstream media’ (Dostal, 2015: 523). The in/out group distinction made by populist actors and noted by Mudde (2007) is a consistent feature of Pegida’s weekly rallies in Dresden, which often feature rejection of the established parties across the political spectrum. In other respects, however, the movements’ central issues are less easy to pinpoint. For example, Vorländer et al. (2016) found only 24% of respondents in a survey conducted at Pegida rallies listed Islam, Islamism or Islamization as their main motive to participate in the rallies, a relatively low number for a movement carrying this issue in its name (p. 77). Both the formal founding of Pegida and its first rally on 20 October 2014 were preceded by the launch of a Facebook group on 11 October 2014, with a peak of activity and the launch of demonstrations in a number of other German cities in December 2014 coinciding. The Facebook page was subsequently banned in the summer of 2016 on grounds of the use of Nazi iconography in its logo.

The Alternative for Germany (Alternative für Deutschland, AfD) was founded in 2013 with a programme that centred on abandoning the Euro and radically altering Germany foreign policy (Arzheimer, 2015; Schmitt-Beck, 2016). Initially dubbed ‘the professors’ party’ in the media for the large number of conservative economists among its ranks, it polled at 12.6% in the 2017 German national elections and has attained even better results in subsequent state elections, particularly in East Germany. The party was founded by a combination of breakaways from the conservative Christian Democratic Union (CDU) and those who had previously not been members of any party. The founders around economics professor Bernd Lucke were by and large individuals of high socioeconomic standing and the party avoided association with other European right-wing parties, such as the French FN, the Italian Lega Nord, Austria’s Freedom Party (FPÖ), Belgium’s Vlaams Belang or the Dutch Partij voor de Vrijheid (PVV).

A unifying feature of both Pegida and the AfD is their enthusiastic embrace of social media, in particular Facebook, for communication and coordination, which contrasts markedly with more established political actors in Germany. Vorländer et al. (2016: 2) highlight the foundational importance of Facebook for Pegida, both because the movement initially developed out of a private Facebook group, and because of the continued importance for mobilization of the group’s pages, while Arzheimer (2015) emphasizes the relevance of Facebook for the AfD’s communication strategy by stressing that ‘the AfD uses its Facebook page to direct attention to news articles that support the party’s positions, to “spin” stories on issues that will chime with the party’s supporters, and occasionally to poke fun at political adversaries’ (p. 550). One does not need to go as far as Gerbaudo (2014) who argues for a distinct ‘populism 2.0’ to assume that social media has enabled both Pegida and the AfD to engage more effectively with their constituents than had previously been possible.

Such affinities follow a consistent pattern. Social media is widely seen by populists as a formative new instrument for the production and dissemination of political opinions, and for mobilizing support within the scope of connective action, thereby advancing the integration of the inherent technical and market logics of online media into political discourse (Bennett and Segerberg, 2012; Klinger and Svensson, 2015). As Mazzoleni (2014) argues,

Facebook, Twitter, YouTube and the blogosphere, have acquired a very special function [ … ] namely, of picking up the populist output of the mass media and relaunching it in the wider context of the electronic public sphere, thus generating a favorable climate of opinion (p. 51).

The increasingly hybrid nature of the media system (Chadwick, 2017) and the strengthening of the feedback loop between politics and social media, thus appears to fit with a populist discursive style that impacts both the mainstream and the fringes of politics, potentially bringing the two closer together stylistically (Gerbaudo, 2014).

It would be a mistake to characterize such developments as new (Schulz, 2014). The specific trajectory of research on right-wing political extremism online parallels the development of the internet itself, spanning political science, sociology and media and communication studies (Burris et al., 2000; Engesser et al., 2016; Zhou et al., 2005). Studies of the right-wing radicalism and extremism use different approaches, but often attempt to list sites, platforms or services used by extremist right-wing groups and describe critical functions for their coordination and communication (Atton, 2006). One explanation for the relative lack of research into the content and discourse generated by and around right-wing groups is the novelty of methods for the acquisition of web and social media platform data and the difficulty of reliably categorizing different actors in the right-wing spectrum. As our approach will show, the application of computational methods to such data is often hampered by restrictions in data access, enforced by companies such as Facebook and Twitter through their API usage terms.

Methods and data

Topic modelling

Topic modelling describes a family of computational techniques for analysing textual content that is increasingly popular in both industry research and across different academic fields (Blei, 2012, Mohr and Bogdanov, 2013). The foundation of topic modelling is the bag-of-words approach (sometimes bag-of-features approach) to written data, in which words are considered to be features that possess distinct distributions in a text or collection of texts. Latent Dirichlet allocation (LDA; Blei et al., 2003) and correlated topic models (CTMs; Blei and Lafferty, 2007) are among the most prominent forms of topic modelling derived from LSA and pLSA (probabilistic LSA). Among the commercial applications of topic modelling are information retrieval tasks and recommender systems (Wang and Blei, 2011; Zhao et al., 2011). LDA is a generative model in which the occurrence of individual words in a document is explained by topics that generate the words. Each document is characterized by a multinomial combination of topics and each topic has a probability of generating certain words, with high-frequency function words such as ‘the’ or ‘and’ occurring in roughly even distribution, while words that probabilistically characterize a topic exhibit a skewed distribution among topics. Like other computational approaches, LDA is quite scalable, but in-depth qualitative knowledge of the data under analysis is still crucial, particularly to determine whether the association of terms with topics is the result of a sampling error or some other interference.

LDA also encodes a number of very specific assumptions about meaning in that the words of each document arise from a mixture of topics, where each topic is a multinomial over a fixed word vocabulary. The topics are shared by all documents in the collection, but the topic proportions vary stochastically across documents (Blei and Lafferty, 2007: 18). An analysis based on just two topics may result in a very coarse classification, in which the topic shares are equal across documents. By contrast, picking a large number of topics will result in a strongly skewed distribution with many topics only occurring in a small number of. Terms are likewise associated with topics by way of the log-likelihood of occurrence in them.

In contrast to most forms of content analysis common in social science, topic modelling induces functional categories purely from structural features. This can create an awkward situation in which the expert knowledge of the researcher seemingly has no place, because a topic model generates its analysis from word distributions alone. The careful selection of features is crucial and the approach assumes both that all topics are similar in the sense of possessing a unique feature distribution imprint, and that world knowledge, while important, is non-essential to distinguish topics from each other. The number of topics in LDA must be predetermined. Depending on the document type and the planned analysis, the number of topics can be very large (several hundreds). Human judgement is needed to distinguish between them and make them interpretable. In particular, LDA does not output interpretable labels of the learned topics. Usually, topics are represented by the top n (e.g. top 10) most probable terms generated by each topic. However, these terms are heavily dependent on preprocessing such as stopword removal and can repeat across topics. Only a human analyst can make sense of the topics that have been learned. Finally, the discovered latent topics may or may not correspond to what the human analyst would call ‘topics’. Depending on preprocessing and aggregation, ‘topics’ may also reflect authorship (through style or specialized vocabulary), genre or other orthogonal categories, and must be manually evaluated.

Facebook comment corpus

For our study, we extracted comments published between 1 January 2015 and 24 May 2016 in two German-language public Facebook pages using the Rfacebook library for R (Barbera et al., 2015) that draws on Facebook’s API. We excluded comments with a length of under 100 characters, as these are difficult to classify reliably when conducting content analysis. The pages thus examined were those of the Pegida movement prior to the page’s deletion by Facebook for violation of community standards (facebook.com/pegidaevdresden) and the main page of the AfD (facebook.com/alternativefuerde).¹ Table 1 provides an overview of the full data set.

Table 1.

Description of the Facebook data set.

Page name/id	Description	Type	# Comments	# Users
Pegidaevdresden	Pegida	Movement	194,488	39,517
Alternativefuerde	AfD	Party	121,034	42,801

The data we describe are no longer available to researchers at the time of writing, because Facebook has chosen to severely restrict access to its Graph API following the Cambridge Analytica scandal. While researchers have mobilized against such restrictions, it is unclear whether similar analysis using Facebook data will be possible in the future.

Applying topic modelling to identify themes in Facebook comments

We used structural topic models (STMs), a variant of LDA (Blei, 2012; Blei et al., 2003), to extract topics within the populist subcorpus. In practice, this was achieved by relying on the quanteda R package for corpus construction and the STM packages for model building (Benoit et al., 2017; Roberts et al., 2016). Because traditional LDA is a mixed-membership model, documents simultaneously belong to several topics at once, a conceptualization that does not always square well with content analysis in the social sciences. This aspect is particularly salient with comments, which are generally much shorter than scientific papers or newspaper articles, and do not cover different distinct issues in quite the same way. STM introduces a number of advantages over ‘plain vanilla’ LDA, namely the ability to initialize the Dirichlet process in a fashion that makes results reproducible and the capacity to model prevalence and content covariates, both of which are helpful for our comparative purposes.

As part of preprocessing, we removed a set of 736 German-language stopwords based on a manually compiled list. This list consists of high-frequency function words (conjunctions, articles, pronouns and interjections) along with very general content words (adjectives, nouns and verbs), as well as the terms ‘Pegida’ and ‘AfD’. Words such as ‘Mann’ (man), ‘Frau’ (woman) and ‘Zeit’ (time) were also removed because their broad distribution across contexts limits their predictive value for categorization. In a next step, we transformed the content to lower case, removed numbers and punctuation, and tokenized the text. Finally, the resulting document term matrix (DTM) was trimmed, removing those terms occurring in less than .2% of all documents (comments). This assured a manageable size of the DTM for further processing. We then calculated an STM with 10 topics.² Our choice was motivated by a previously performed manual content analysis within the same project in which 2066 comments were assigned one of 26 categories by two coders. Stier et al. (2017), using a similar approach, modelled a much larger number of topics to achieve a high degree of granularity in relation to specific events. While the results of the manual content analysis informed our choice of methods, we ultimately opted for topic modelling because of its greater scalability along with the reliability of the outcomes, which largely mirrored our earlier results.

Results

A key aim of this study is to demonstrate the value of topic modelling for better understanding the prevalence of populist themes in social media, in the chosen case in the comments on the Pegida and AfD Facebook pages. We do this in two ways: by tracing the prevalence of the topics over time and by examining the topical contrast between both groups to show their relative difference and similarity (the binary source variable, as outlined by Roberts et al., 2016).

We first describe the topic model through manually assigned English-language labels derived from keywords, that is, terms with a strong association with the topic (Bonilla and Grimmer, 2013; see Appendix 1 for the raw data and a list of different keyword metrics). Because of the brevity of the comments, we deliberately opt for a small number of relatively coarse topics, rather than a larger and more granular set of topics. A result of this strategy is that sometimes nuances (rather than entirely distinct subjects) are folded into individual topics. For example, the topic Immigration also addresses social inequality, crime and comments on immigration policy in Austria and Hungary to an extent. The topic Politicians also subsumes references to ‘do-gooders’ (Gutmenschen) and asylum seekers (Asylanten), and generally uses a number of strongly negatively connoted words, as well as expletives. The topic Refugees is also tangentially concerned with the war in Syria and Iraq, and Germany’s policy towards Russia. The topic Commenting stands out somewhat from the other issues. It captures meta-discourse about commenting etiquette and the opinions voiced by other users, a theme that is surprisingly salient.

Figure 1 shows the topic prevalence contrast between the two pages. Topics with a strong prevalence among the Pegida commentators include the Media, Islam, Politicians, Refugees and Immigration. By contrast, topics with a strong prevalence for the AfD include Poll and Election Results, Angela Merkel and her government, AfD Candidates and Money, specifically the Greek debt crisis. Much of these differences can be explained by the differing strategic aims of the movement (Pegida) and the party (AfD). Whereas Pegida lacks the necessity of developing a coherent political programme and is/was able to subsist largely on resistance to a number of internal and external enemies, the AfD competes for votes with others parties and is accordingly focussed on observing their behaviour and its own predicted electoral chances. Whereas Pegida distances itself from the entire political class, the AfD is concerned mainly with the government of Angela Merkel, which it aims to replace. Opposition towards the German government and its policies, for the AfD in relation to the bailout of Greece and the Euro, and for Pegida with respect to the refugee crisis, is the common denominator between the two movements. It is important to note that these discrepancies can also have additional reasons, such as differing moderation behaviour and usage policies on the two pages, rather than representing preferential differences.

Figure 1.

Topic prevalence contrast between the two pages (x-axis; left = Pegida, right = AfD) for 10 STM topics (y-axis, random order).

Figure 2 shows the topic prevalence over time for the 10 STM topics between January 2015 and May 2016 that we modelled for both pages. The topic Refugees peaks in October 2015, along with asylum applications in Europe in the course of the European refugee crisis of 2015 to 2016, and the topic Money is correlated with the unfolding of the Greek debt crisis, peaking in June 2015. Of interest is the relative decline in the topics Islam and the Media which seem to tail Pegida’s foundational issues, in relation to these events and their media coverage. The share of the topics Immigration, Refugees and Politicians rises to a stable share of 10 to 20% each in 2015, whereas Money and Angela Merkel both drop off in the course of 2016.

Figure 2.

Topic prevalence over time for 10 STM topics (January 2015–May 2016).

Finally, Figure 3 shows the topic prevalence contrast over time for the two pages (Pegida = red, AfD = blue; January 2015–May 2016), allowing an assessment of the relative importance of each topic for commentators on the two pages. Pegida comments are more prone to topical fluctuations than those of the AfD, that is, that the reaction to the news cycle is stronger. The topics Refugees and Immigration both gain in share over time for Pegida, while there is only a very slight increase for the AfD. The importance of the Media declines steeply for Pegida, whereas it is flat for the AfD. The topic Money declines in importance for the AfD in line with the resolution of the Greek debt crisis, but larger economic issues increase in importance for Pegida. The inverse is true with Islam, which declines for Pegida, but becomes an increasingly relevant issue for the AfD. The interest in AfD Candidates and Poll and Election Results is plausibly greater among the AfD than Pegida.

Figure 3.

Topic prevalence contrast over time for the two pages (Pegida = red, AfD = blue; January 2015–May 2016).

Our analysis reveals both differences and similarities regarding the issues raised by commentators on the two pages, with some aligning more closely to a core nativist agenda than others. Comments on the Pegida page are thematically more heterogeneous, while those on the page of the AfD are focussed more on strategic political goals (winning elections) and the government of Angela Merkel, in addition to the AfD’s foundational issue of European fiscal policy. By contrast, the Pegida commentators emphasize internal (the Media, Politicians) and external enemies (Islam, Immigration, Refugees) that Mudde (2004, 2007) regards as constitutive for populist ideology. Our study also sheds light onto changes in the agenda of commentators over an extended period of time. The salience of Islam and especially the Media declines for Pegida sympathizers, while the salience of Immigration and Refugees increases significantly. A similar development among AfD supporters takes place for the Money topic, signalling a gradual turn towards other issues, for example, a slow but steady rise of comments related to Islam.

Conclusion

Methodologically, this article presents a new approach for the comparative study of nativist agendas on Facebook, agendas which are otherwise difficult to study because their supporters do not always articulate themselves publicly or speak to the press, and because small gradual changes in them are difficult to observe. The approach’s strength furthermore lies in the precise quantification of topic proportions. Measurement error notwithstanding, this allows us to gauge both change over time and relative difference between actors (in this case, a movement and a party) that would otherwise be very hard to observe. Topic models therefore represent an additional attractive solution for comparative analyses if appropriate steps are taken for validation (Grimmer and Stewart, 2013; Stier et al., 2017). Once a satisfactory fit for the number of topics and the validity of composition have been established, their prevalence over time or among actors can be used to systematically study issue careers and users’ agendas. In the case under study, the comparison reveals that some issues have grown more salient to the supporters of both Pegida and the AfD, while others have declined in relevance. Another key finding of our study is that an eventual convergence on similar issues between two political actors that would otherwise be difficult to trace becomes feasible through an inductive computational approach, allowing us to determine both when and on what issues the two entities converge. Computational techniques such as topic modelling thus offer an attractive alternative to established procedures in communication research when studying dynamic political actors that significantly rely on social media platforms to communicate and mobilize.

Nativist right-wing movements such as Pegida and populist parties such as the AfD have embraced social media and Facebook in particular to communicate and coordinate with their supporters and sympathizers. For Pegida, the very existence of the group appears to have been contingent on Facebook, as the platform provided the organizational infrastructure to coordinate its initial physical rallies. While the AfD existed as an organization before turning to social media, it has embraced Facebook in particular more energetically than any other party.

Our study has specific limitations. Comments on a Facebook page should not be regarded as a faithful reflection of a movement’s or party’s definitive stance towards political issues. As formal membership is not required to comment on a Facebook page, such remarks reflect the beliefs and motifs of sympathizers more closely than a party programme or official speeches by politicians. They may be more extreme in their views, or reflect issues of low relevance in day-to-day politics.

We have also sought to describe how the prejudices, grievances and common objects of aggression of nativist right-wing movements are expressed in social media, and how inductive computational methods offer an attractive means of analysis for this type of data because of their ability to capture small, gradual change in heterogeneous material. An analysis of the temporal dimension of populist discourse shows its dependency on crises that call the legitimacy of established political parties into question. This is true both for the AfD in relation to the Greek sovereign debt crisis and for Pegida with regard to the European refugee crisis of 2015 to 2016, inflection point that the STM is able to faithfully capture. At the same time, opposition to Islam and criticism of the German government appear as two potential candidates for issue convergence between Pegida and the AfD, that is, common ideological denominators for both groups.

Footnotes

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the European Regional Development Fund-Project “Creativity and Adaptability as Conditions of the Success of Europe in an Interrelated World” (No. CZ.02.1.01/0.0/0.0/16_019/0000734).

Notes

Appendix 1.

Highest scoring keyword associations with topics (K = 10). The four different measures used here are described in the work of Roberts et al. (2016).

Topic # and English-lang. label	Highest probability	FREX	Lift	Score
Topic 1 ‘islam’	islam, problem, muslime, kultur, religion, europa, gesellschaft, moslems, gott, freiheit	islam, religion, muslime, koran, christen, moslems, muslim, muslimen, kirche, europe	europe, muslim, koran, religionen, islam, muslimen, religion, christen, moslem, muslime	europe, islam, muslime, religion, problem, koran, muslim, moslems, christen, kultur
Topic 2 ‘immigration’	deutschland, deutschen, länder, raus, europa, migranten, ausländer, millionen, probleme, zukunft	deutschland, österreich, migranten, fakt, ungarn, ausländer, polen, schande, fachkräfte, raus	armes, österreich, polen, bundesrepublik, fakt, schweden, kriminalität, ungarn, soziale, zuwanderer	armes, deutschland, deutschen, länder, migranten, raus, ausländer, europa, österreich, ungarn
Topic 3 ‘afd candidates’	recht, liebe, petry, herrn, verstehen, super, schlecht, falsch, versuchen, respekt	petry, herrn, wünsche, aussage, erfolg, lucke, super, respekt, liebe, aussagen	frauke, gauland, petry, höcke, sendung, wünsche, antworten, lucke, argumente, interview	frauke, petry, recht, liebe, herrn, lucke, super, wünsche, dr, höcke
Topic 4 ‘merkel’	merkel, volk, regierung, politik, grenzen, gabriel, volkes, kanzlerin, ddr, deutschlands	merkel, kanzlerin, regierung, volk, merkels, erdogan, angela, gabriel, seehofer, mutti	angela, gauck, merkels, bundeskanzlerin, mutti, merkel, rücktritt, erdogan, kanzlerin, seehofer	angela, merkel, volk, regierung, politik, grenzen, kanzlerin, gabriel, erdogan, merkels
Topic 5 ‘politicians’	politiker, deutsche, kinder, staat, gutmenschen, asylanten, arbeiten, pack, arbeit, kopf	arbeiten, rente, kotzen, kind, eltern, scheiß, kinder, lernen, dumm, arbeit	damen, arbeiten, rente, gesindel, eltern, krank, scheiß, kind, kotzen, dummen	damen, kinder, politiker, deutsche, arbeiten, pack, rente, staat, asylanten, bezahlen
Topic 6 ‘the media’	medien, polizei, dresden, schön, straße, angst, nazis, gewalt, reden, wahrheit	demo, presse, lügenpresse, montag, dresden, nazis, antifa, polizisten, straße, polizei	bachmann, gegendemonstranten, lutz, demo, demonstranten, montag, dresdner, patrioten, leipzig, demos	bachmann, dresden, polizei, medien, lügenpresse, straße, presse, nazis, demo, montag
Topic 7 ‘money’	geld, eu, euro, berlin, zahlen, griechenland, kosten, milliarden, stadt, millionen	euro, milliarden, griechenland, video, griechen, sorry, hamburg, stadt, germany, berlin	germany, griechen, euro, banken, kostet, milliarden, monat, schulden, griechenland, video	germany, geld, euro, eu, griechenland, milliarden, berlin, zahlen, kosten, stadt
Topic 8 ‘afd polls’	partei, spd, cdu, wählen, demokratie, parteien, grünen, linken, linke, wahl	spd, wählen, cdu, parteien, grünen, wahl, grüne, wähler, wahlen, partei	etablierten, fdp, grüne, bundestagswahl, rot, spd, wählen, parteien, grünen, wahl	etablierten, spd, partei, cdu, wählen, parteien, grünen, wahlen, wahl, linke
Topic 9 ‘refugees’	flüchtlinge, frauen, krieg, flüchtlingen, männer, usa, türkei, heimat, grenze, europa	krieg, männer, russland, syrien, waffen, usa, flüchtlinge, grenze, heimat, hause	irak, russen, russland, soldaten, waffen, nato, syrien, syrer, bundeswehr, krieg	irak, flüchtlinge, krieg, frauen, männer, syrien, usa, russland, türkei, grenze
Topic 10 ‘commenting’	seite, bevölkerung, kommentar, prozent, ard, grund, hand, landes, gegeben, schauen	ard, kommentar, seite, zdf, zahl, prozent, gegeben, propaganda, mdr, bevölkerung	ard, zdf, sender, kommentar, gez, propaganda, abschaffung, mdr, seite, zahl	ard, seite, kommentar, prozent, zdf, bevölkerung, mdr, zahl, sender, gez

Abbreviation: FREX, FRequency and EXclusivity.

Appendix 2.

Model and topic scores based on the criteria of exclusivity and semantic coherence (Roberts et al., 2016). In this approach, models with low likelihood values are identified during the cast net stage of the STM and semantic coherence and exclusivity, as well as sparsity, are calculated to determine model fit. Numerals represent the average for each model and dots represent topic specific scores.

References

Arzheimer

(2015) The AfD: Finally a successful right-wing populist Eurosceptic party for Germany? West European Politics 38(3): 535–556.

Atton

(2006) Far-right media on the internet: Culture, discourse and power. New Media & Society 8(4): 573–587.

Ausserhofer

Maireder

(2013) National politics on Twitter: Structures and topics of a networked public sphere. Information, Communication & Society 16(3): 291–314.

Barbera

Piccirilli

Geisler

(2015) Rfacebook. Available at: https://cran.r-project.org/web/packages/Rfacebook/.

Bastos

Mercea

Charpentier

(2015) Tents, Tweets, and events: The interplay between ongoing protests and social media: Tents, Tweets, and events. Journal of Communication 65(2): 320–350.

Bennett

Segerberg

(2012) The logic of connective action: Digital media and the personalization of contentious politics. Information, Communication & Society 15(5): 739–768.

Benoit

Watanabe

Nulty

, et al. (2017) quanteda: Quantitative analysis of textual data. Available at: http://quanteda.io.

Blei

(2012) Topic modeling and digital humanities. Journal of Digital Humanities 2(1). Available at: http://journalofdigitalhumanities.org/2-1/topic-modeling-and-digital-humanities-by-david-m-blei/.

Blei

Lafferty

(2007) A correlated topic model of Science. The Annals of Applied Statistics 1(1): 17–35.

10.

Blei

Jordan

(2003) Latent Dirichlet allocation. Journal of Machine Learning Research 3: 993–1022.

11.

Bonilla

Grimmer

(2013) Elevated threat levels and decreased expectations: How democracy handles terrorist threats. Poetics 41(6): 650–669.

12.

Burris V, Smith E and Strahm A (2000) White Supremacist Networks on the Internet. Sociological Focus 33(2): 215–235. https://doi.org/10.1080/00380237.2000.10571166

13.

Calhoun

(2013) Occupy Wall Street in perspective. British Journal of Sociology 64(1): 26–38.

14.

Ceron

Curini

Iacus

, et al. (2014) Every Tweet counts? How sentiment analysis of social media can improve our knowledge of citizens’ political preferences with an application to Italy and France. New Media & Society 16(2): 340–358.

15.

Chadwick

(2017) The Hybrid Media System: Politics and Power. Oxford: Oxford University Press.

16.

Conti

Memoli

(2015) The emergence of a new party in the Italian party system: Rise and fortunes of the five star movement. West European Politics. Epub ahead of print 13 January. DOI: 10.1080/01402382.2014.996377.

17.

Dostal

(2015) The Pegida movement and German political culture: Is right-wing populism here to stay? The Political Quarterly 86(4): 523–531.

18.

Engesser

Ernst

Esser

, et al. (2016) Populism and social media: How politicians spread a fragmented ideology. Information, Communication & Society. Epub ahead of print 8 July. DOI: 10.1080/1369118X.2016.1207697.

19.

Freelon

McIlwain

Clark

(2016) Quantifying the power and consequences of social media protest. New Media & Society 20: 990–1011.

20.

Gerbaudo

(2014) Populism 2.0: Social media activism, the generic internet user, and interactive direct democracy. In: Trottier

Fuchs

(eds) Social Media, Politics and the State: Protests, Revolutions, Riots, Crime and Policing in the Age of Facebook, Twitter and YouTube. New York: Routledge, pp. 67–88.

21.

Gil De Zúñiga

Puig-I-Abril

Rojas

(2009) Weblogs, traditional sources online and political participation: An assessment of how the internet is changing the political environment. New Media & Society 11(4): 553–574.

22.

Golder

Macy

(2014) Digital footprints: Opportunities and challenges for online social research. Annual Review of Sociology 40: 129–152.

23.

Grimmer

Stewart

(2013) Text as data: The promise and pitfalls of automatic content analysis methods for political texts. Political Analysis 21(3): 267–297.

24.

Himelboim

McCreery

Smith

(2013) Birds of a feather tweet together: Integrating network and content analyses to examine cross-ideology exposure on Twitter. Journal of Computer-Mediated Communication 18(2): 154–174.

25.

Jacobi

van Atteveldt

Welbers

(2016) Quantitative analysis of large amounts of journalistic texts using topic modelling. Digital Journalism 4(1): 89–106.

26.

Kaiser

Puschmann

(2017) Alliance of antagonism: Counterpublics and polarization in online climate change communication. Communication and the Public 2(4): 371–387.

27.

Klinger

Svensson

(2015) The emergence of network media logic in political communication: A theoretical approach. New Media & Society 17(8): 1241–1257.

28.

Lomborg

Bechmann

(2014) Using APIs for data collection on social media. The Information Society 30(4): 256–265.

29.

Magin

Podschuweit

Haßler

, et al. (2016) Campaigning in the fourth age of political communication. A multi-method study on the use of Facebook by German and Austrian parties in the 2013 national election campaigns. Information, Communication & Society 20(11): 1698–1719.

30.

Margetts

John

Hale

, et al. (2016) Political Turbulence: How Social Media Shape Collective Action. Princeton, NJ: Princeton University Press.

31.

Mazzoleni

(2014) Mediatization and political populism. In: Esser

Strömbäck

(eds) Mediatization of Politics: Understanding the Transformation of Western Democracies. Basingstoke: Palgrave Macmillan, pp. 42–56.

32.

Mohr

Bogdanov

(2013) Introduction – Topic models: What they are and why they matter. Poetics 41(6): 545–569.

33.

Mudde

(2004) The populist zeitgeist. Government and Opposition 39(4): 542–563.

34.

Mudde

(2007) Populist Radical Right Parties in Europe. Cambridge: Cambridge University Press.

35.

Pirro

ALP

van Kessel

(2017) United in opposition? The populist radical right’s EU-pessimism in times of crisis. Journal of European Integration 38: 378–390.

36.

Rauchfleisch

(2017) The public sphere as an essentially contested concept: A co-citation analysis of the last 20 years of public sphere research. Communication and the Public 2(1): 3–18.

37.

Roberts

Stewart

Tingley

(2016) stm: R Package for structural topic models. Journal of Statistical Software VV(Ii). pp. 1–40.

38.

Salganik

(2017) Bit by Bit: Social Research in the Digital Age. Princeton: Princeton University Press.

39.

Schmitt-Beck

(2016) The ‘Alternative für Deutschland in the electorate’: Between single-issue and right-wing populist party. German Politics. Epub ahead of print 15 June. DOI: 10.1080/09644008.2016.1184650.

40.

Schulz

(2014) Mediatization and new media. In: Esser

Strömbäck

(eds) Mediatization of Politics. London: Palgrave Macmillan, pp. 57–73.

41.

Schwartz

Ungar

(2015) Data-driven content analysis of social media: A systematic overview of automated methods. The ANNALS of the American Academy of Political and Social Science 659(1): 78–94.

42.

Vorländer H, Herold M and Schäller S (2016) PEGIDA: Entwicklung, Zusammensetzung und Deutung einer Empörungsbewegung. Wiesbaden: Springer Fachmedien Wiesbaden. https://doi.org/10.1007/978-3-658-10982-0

43.

Shah

Cappella

Neuman

(2015) Big data, digital media, and computational social science: Possibilities and perils. The ANNALS of the American Academy of Political and Social Science 659(1): 6–13.

44.

Sloan

Quan-Haase

(2016) The Sage Handbook of Social Media Research Methods. London: SAGE.

45.

Stier

Posch

Bleier

, et al. (2017) When populists become popular: Comparing Facebook use by the right-wing movement Pegida and German political parties. Information, Communication & Society 20: 1365–1388.

46.

Tufekci

Wilson

(2012) Social media and the decision to participate in political protest: Observations from Tahrir Square. Journal of Communication 62(2): 363–379.

47.

Van Atteveldt

Kleinnijenhuis

Ruigrok

, et al. (2008) Good news or bad news? Conducting sentiment analysis on Dutch text to distinguish between positive and negative relations. Journal of Information Technology & Politics 5(1): 73–94.

48.

Wang

Blei

(2011) Collaborative topic modeling for recommending scientific articles. In: Proceedings of the 17th ACM SIGKDD international conference on knowledge discovery and data mining – KDD’11, pp. 448–456. Available at: https://doi.org/10.1145/2020408.2020480.

49.

Young

Soroka

(2012) Affective news: The automated coding of sentiment in political texts. Political Communication 29(2): 205–231.

50.

Zhao

Jiang

Weng

, et al. (2011) Comparing Twitter and traditional media using topic models. In: Serdyukov

Braslavski

Kuznetsov

, et al. (eds) Advances in Information Retrieval. Heidelberg: Springer, pp. 338–349.

51.

Zhou

Reid

Qin

, et al. (2005) US domestic extremist groups on the web: Link and content analysis. IEEE Intelligent Systems 20(5): 44–51.