ÉlSIl Ell 280 Handbook of digital politics Noveck, Belli Simone (2003). 'Designing deliberative democracy in cyberspace: the role the cyber-lawycr'. Boston University Journal of Science and Technology Law 9: 1 Okada, Alexandra, Simon J. Buckingham Shum and Tony Sherborne (2008). Knowled Cartography: Software Tools and Mapping Techniques. Amsterdam: Springer. g" Parkinson, John and Jane Mansbridge (eds) (2012). Deliberative Systems: Deliberate Democracy at the Large Scale. New York: Cambridge University Press. Pickard, Victor W. (2008). 'Cooptation and cooperation: institutional exemplars of demn cratic internet technology'. New Media and Society 10(4): 625-645. Pingree, Raymond J., T. Davies and S.P. Gangadharan (2009). 'Decision structure- a new approach to three problems in deliberation'. In T. Davies and S.P. Gangadharan (eds) Online Deliberation: Design, Research and Practice. San Francisco, CA: CSLI Publications Rhee, June W. and Eun-mee Kim (2009). 'Deliberation on the net: lessons from a field experi' ment'. In T. Davies and S.P. Gangadharan (eds), Online Deliberation: Design Research and Practice. San Francisco, CA: CSLI Publications. Rittel, Horst W.J. and Melvin M. Webber (1973). 'Dilemmas in a general theory of okn ning'. Policy Sciences 4(2): 155-169. Schlosberg, David, Stephen Zavestoski and Stuart W. Shulman (2007). 'Democracy and e-rulemaking: web-based technologies, and the potential for deliberation'. eRulemakin» Research Group, Paper 1, University or Massachusetts - Amherst. Schüler, Douglas (2009). 'Online civic deliberation with E-Liberate'. In T. Davies and S.P. Gangadharan (eds), Online Deliberation: Design, Research and Practice San Francisco, CA: CSLI Publications. Sobieraj, Sarah and Jeffrey M. Berry (2011). 'From incivility to outrage: political discourse in blogs, talk radio, and cable news'. Political Communication 28(1): 19-41. Steiner, Jürg (2012). The Foundations of Deliberative Democracy: Empirical Research and Normative Implications. New York: Cambridge University Press. Sunstein, Cass R. (2002). 'The law of group polarization'. Journal of Political Philosonhv 10(2): 175-195. y y Thimm, Caja, Mark Dang-Anh and Jessica Einspänner (2014). 'Mediatized politics - structures and strategies of discursive participation and online deliberation on Twitter' In Andreas Hepp and Friedrich Krotz (eds), Mediatized Worlds: Culture and Society in a Media Age. London: Palgrave. Tuccy, Cindy Boyles (2010). 'Online vs. face-to-face deliberation on the global warming and stem cell issues'. Western Political Science Association 2010 Annual Meeting Paper Upadhyay, Meenakshi (2014). 'Political deliberation on Twitter: is Twitter emerging as an opinion leader?', International Conference on People, Politics and Media (ICPPM), 25-26 April, Jagran Lakecity University. Wilhehn, Anthony G. (2000). Democracy in the Digital Age: Challenges to Political Life in Cyberspace. New York: Routledge. Wright, Scott (2006). 'Government - run online discussion fora: moderation, censorship and the shadow of control'. British Journal of Politics and International Relations 8(4): 550-568. Wright, Scott (2009). 'The role of the moderator: problems and possibilities for government-run online discussion forums'. In T. Davies and S.P.. Gangadharan (eds), Online Deliberation: Design, Research and Practice. San Francisco, CA: CSLI Publications. Wright, Scott and John Street (2007). 'Democracy, deliberation and design: the case of online discussion forums'. New Media and Society 9(5): 849-869. 16. Computational approaches to online political expression: rediscovering a 'science of the social' Dhavan V. Shah, Kathleen Bartzen Culver, Alexander Hanna, Timothy Macafee and JungHwan Yang It is a curious fact that the empirical study of political talk, particularly online exchanges, is increasingly traced back to the social interactionism of nineteenth-century French sociologist Gabriel Tarde. As Terry Clark (1969) and Elihu Katz (2006) remind us, Tarde argued for conversation's place at the center of sociological inquiry, articulating a complex theory of 'inler-mcntal activity' concerning how people influence one another. In so doing, he developed the concepts that later became known as the two-step flow of communication and opinion leadership, among other propositions of interpersonal influence (Lazarsfeld et al., 1944; Berelson et al., 1954: Katz and Lazarsfeld, 1955). As Tarde writes about the late nineteenth century (1969 [1898]: 313), 'newspapers have transformed . . . the conversations of individuals, even those who do not read papers but who, talking to those who do, are forced to follow the groove of their borrowed thoughts. One pen suffices to set off a million tongues'. His thesis still holds true, with televised events such as the first 2012 presidential debate in the USA generating more than 10 million tweets in just a few hours, including many retweets of major accounts (Hanna et al,, 2013). Tarde was particularly concerned with the relationship between mass communication and interpersonal conversation for the formation of publics and their opinions. Katz (2006: 267) describes this mediated process as follows: To the press, he assigned the role of creating a public ... The press, then, sets an agenda for the conversation of the cafes. Opinions are clarified and crystallized in these conversations, and then translated into actions in the world of polities'. The central tenets of this ordered model - press, conversation, opinion, and action - are supported by research on multi-step flow (Rogers and Shoemaker, 1971), opinion leadership (Shah and Scheufele, 2006), and communication mediation (Lcc et al., 2013). In digital media environments, the sources of information and sites of conversation have begun to converge - amplified and 281 282 Handbook of digital politics reinforced within seemingly polarized ecologies - suggesting important directions for research on opinion and action in networked societies. The Laws of Imitation (Les his de limitation) (1903 [1890]), Tarde's most widely known work in English, speaks to communication and social influence within such settings. It also marks him as the 'founding father of innovation diffusion research' (Kinnuncn, 1996), charting processes of communicative invention, reproduction, and opposition. Methodologically, his calls for attention to observable interpersonal interactions, particularly within conversational processes, presage the approaches central to computational social science, where each interaction 'leaves digital traces that can be compiled into comprehensive pictures of both individual and group behavior, with the potential to transform our understanding of our lives, organizations, and societies' (Lazer et al., 2009: 721). The technological affordances of social media permit tracking of message creation and expression within a network, as well as reception and diffusion through systems (Namkoong et al., 2010; Han et al., 2011). Tarde's insights about 'invention' and 'imitation' provide ways to study online talk as it intersects with deliberative democracy, informed opinion, and participatory citizenship (Schudson, 1978; Barber, 1984; Habermas, 1984; Kim et al., 1999; Price and Cappella, 2002; Mutz, 2006). SITUATING POLITICAL TALK Tarde, often presented as the foil to Emile Durkheim's efforts to distinguish the study of society from that of human psychology.1 was profoundly concerned with the interplay of mental and social forces, particularly as seen in the locus of conversational exchanges. As Berelson et al. wrote in Voting (1954: 300), Tarde 'was convinced that opinions are really formed through the day-to-day exchanges of comments and observations which goes on among people ... by the very process of talking to one another, the vague dispositions which people have are crystallized, step by step, into specific attitudes, acts, and votes'. From this perspective, conversation is not simply a site of networked information exchange, but also an opportunity for the composition and clarification of one's own views (see Pingree, 2007). In the present day, this position seems all the more correct, especially in online environments, where political expression and conversation are increasingly common and visible. These environments also lend themselves to the sort of large-scale, highly detailed interactional analysis that Tarde's approach advocated. As Bruno Latour (2010) recently recognized, 'it is indeed striking that at this very moment, the fast expanding fields of "data Computational approaches to online political expression 283 visualisation," "computational social science," or "biological networks" (Lazer et al., 2009; Wimsatt, 2007) are tracing, before our eyes, just the sort of data Tarde would have acclaimed'. 'Big data' provide a way to understand everyday political talk online, its triggers, content, and structures. Such close analyses of political talk, whether face-to-face or online, have typically been restricted to ethnographic or content analytic research. Works such as William Gamson's (1992) classic study, Talking Politics, used analysis of small-group discussions around affirmative action, nuclear power, the Arab-Israeli conflict, and US industry to counter the conventional wisdom of an uninformed and inactive electorate. Taking a similarly granular approach, Papacharissi (2004) examined the quality of online political talk by studying the level of civility in 287 discussion threads drawn randomly from 147 political newsgroups, concluding that discussions were civil but heated. Somewhat similarly, Walsh (2012) used participant observation of 37 reoccurring groups from 27 distinct communities across the state of Wisconsin to show how class- and place-based identity is linked with perceptions of relative deprivation. Complementing this work on the actual content of conversations, research has also employed cross-sectional and panel survey methods to examine the causes and consequences of political talk (Huckfeldt and Sprague, 1995). Most notably, work by Jack McLeod et al. has examined the role of interpersonal conversation for community activism and political engagement (McLeod et al., 1999). Emphasizing the mediating roles of news and talk for participation in public life, communication behaviors are thought to shape and amplify the impact of background characteristics on citizens' engagement in democratic societies (McLeod et al., 1996). Various inquiries about the roles of media and conversation have coalesced into 'communication mediation models', which conclude that news consumption and political talk largely channel the effects of demographics, ideology, and social structure on outcome orientations and participatory responses (Sotirovic and McLeod, 2004; Cho et al.. 2009). This process has been further specified in the form of a 'citizen communication mediation model' (Shah et al, 2005; Shah et al., 2007), which theorizes and finds, consistent with Tarde's framework, that media influences are strong, but largely indirect, shaping opinion and action through effects on face-to-face and online discussion about news. Informational use of media, particularly newspaper and online news use, is found to stimulate expression and discussion through interpersonal and computer-mediated political talk, channeling effects on to engagement. The power of digital pathways to participation from both conventional and online news sources through digital messaging - was particularly strong among the youngest generational groups (Lee et al., 2013). 284 Handbook of digital politics POLITICAL TALK VIA DIGITAL MEDIA Face-to-face and online talk share many virtues. Both spur compositional processing, mental elaboration, attitude crystallization, cross-cutting exposure, media reflection, knowledge gain, and mutual understanding (Shah et al., 2005; Mutz, 2006; Lee et al., 2013; Valenzuela et al., 2012). But political conversation via digital platforms may have certain advantages over face-to-face talk, as well. First, digital media often provide a source of political information and a sphere for political expression, readily facilitating their interplay and participatory consequences (Dahlgren, 2000, 2005). Second, the functionality of many online media emphasize discursive elements such as walls posts, online chat, photo sharing, and social networking, highlighting opportunities for public-spirited talk (Bennett et al., 2011; Boyd and Ellison, 2007). Third, generating user-created content, whether images, videos, or posts, may demand deeper forms of reflective and compositional processing (Freelon, 2010; Ekstrom and Ostman, 2013). Regardless of the setting, 'conversation provides people with the opportunity to think through their "idea elements" and reduce cognitive inconsistency' (Kim et al., 1999: 363). Less is known about the specific consequences of online political talk, such as whether deeper dialogue contributes to the enhancement of opinion quality and the development of efficacy (Kim et al, 1999). There is also some question as to the value of measuring behaviors in the digital world using analogue tools such as survey instruments, which rely on self-reported measures of network composition, structural heterogeneity, and discussion frequency (see Sotirovic and McLeod, 2001; Eveland and Hively, 2009; Kwak et al., 2005), rather than precise measures of content-specific message expression, reception, response, and repetition (see Han et al, 2011; Namkoong et al., 2010). In addition, it may be important to look across different conversational settings, such as messaging platforms (Shah et al., 2007; Hardy and Scheufele, 2005), political blogs (Gil de Zuniga et al., 2009), and social networking sites (Bode et al., 2014), all of which have been linked to civic engagement and political action. Social networking sites bring together the most powerful features of online interaction, messaging in real time and asynchronously, posting to smaller circles and to full networks, exchanging information and providing emotional support, creating new content and sharing the ideas of others, starting a group and joining those created by others; all seemingly strengthening social ties, albeit in unique registers to differing degrees (Boyd and Ellison, 2007; Pasek et al., 2009; Gil de Zuniga et al., 2012). One recent study examines the consequences of political social networking behaviors, such as displaying a political preference on a profile, becoming 1 Computational approaches to online political expression 285 a fan or a friend of a politician, joining a cause or political group, and using a news or politics application (Bode et al., 2014). It finds that such uses shape political participation above and beyond the effects of offline talk and online expression. SOCIAL NETWORKING AROUND POLITICS There is little doubt that social media provide a forum for the discussion of political issues, a system through which to recruit individuals to participate in pertinent political issues, and a means to find people who share similar political opinions (Papacharissi, 2010). In this sense, political engagement through social media may represent a shift from top-down political communication and participation to one that is interpersonal or bottom-up (Thackery and Hunter, 2010). It is also characterized by feedback mechanisms to elites, providing opinion leaders with an opportunity to shape the thinking of political and media elites if they can 'set off a million tongues'. Social media have certainly increased candidates' digital exposure, permitting communication with volunteers, donors, and constituents (Gueorguieva, 2008). Alongside Facebook (Vitak et al., 2011) and YouTube (Dylko et al., 2012), Twitter has emerged as a major location of political interaction (Hanna et al., 2013; Tumasjan et al., 2010). Much of this work emphasizes the use of social media channels by politicians during elections (Bruns and Highfield, 2013). Studies of members of the US Congress have found that they use Twitter primarily for information dissemination, with little retweeting or hashtag use, suggesting elite efforts to lead opinion rather than echoing the ideas of others (Golbeck et al, 2010; Gainous and Wagner, 2014). Efforts to examine whether Twitter can predict election results are more mixed. Based on a sentiment analysis of more than 100,000 messages referencing a political party or politician in the 2009 German federal election, Tumasjan et al. (2010) found that Twitter 'is used extensively for political deliberation and that the mere number of party mentions accurately reflects the election result'. However, an alternative analysis of the same election found no link and argued that the predictive power of the prior study was a consequence of the 'arbitrary choices of the authors' (Jungherr et al., 2012: 229). Regardless of its predictive power, it is widely acknowledged that political talk on Twitter is triggered by media happenings and news coverage (Hanna et al., 2013; Graham and Hajru, 2011). Only a handful of studies have examined the types of political expression that occur within social networking sites, with even fewer considering how citizens self-organize under 286 Handbook of digital politics Computational approaches to online political expression 287 certain banners or hashtags (that is, user-generated keywords organized around the # symbol). Given that 'hashtags arc used to bundle together tweets on a unified, common topic, and thai the senders of these messages arc directly engaging with one another', they provide a potentially powerful way to track everyday political expression (Bruns and Burgess, 2011: 5), Hashtags can become identified with specific issue positions and protest efforts. Along these lines. Segerberg and Bennett (2011) analyzed hashtag usage around the 2009 United Nations Climate Change Conference, observing the strategic deployment of certain terms to organize action. As noted above, Twitter also allows for the tracking of 'inventive' message expression and "imitation' by others in the network as indicators of their recirculation and diffusion through systems (Tarde, 1903 [1890]; Rogers and Shoemaker, 1971). At the most basic level, this can be understood in terms of tweets and rctweets (Boyd et al., 2010), although more sophisticated analyses have examined the flow of influence within social networking systems. These studies conclude that Twitter is an excellent medium for message propagation, finding that 37.1 percent of message flows spread more than three degrees of separation away from the original sender (Ye and Wu, 2010). Of course, not all issues are created equal. Tracing the diffusion of hashtags on Twitter, Romero et al. (2011: 695) iind significant variation across topics, with hashtags on politically controversial topics 'particularly persistent, with repeated exposures continuing to have unusually large marginal effects on adoption'. Influence in online social networks is not simply a function of size of follower networks, at least as measured in retweets or user mentions (Cha et al., 2010). In fact, a recent analysis of 74 million diffusion events by 1.6 million Twitter users questioned the emphasis on online elites as influencers, concluding lhat 'word-of-mouth information spreads via many small cascades, mostly triggered by ordinary individuals' (Bakshy el al., 2011: 73). COMPUTATIONAL APPROACHES As these studies suggest, the next phase of research on political talk online may benefit from moving beyond ethnographic, content-analytic, or survey-based assessments of these phenomena, instead exploring the value of computational approaches to questions of issue attention, social influence, opinion formation, and political mobilization. The use of'big data' may provide unique insights into how political elites and their followers deploy particular language around controversial political issues, and how these forms of expression splinter into polarized factions, coalesce into action, and get recirculated through networks. Using such approaches to help illuminate how particular events or controversies spur different types of political talk and social activism, we explore two news controversies that erupted in 2012 in the United States: (1) Rush Limbaugh's statements about Sandra Fluke, a law student and women's rights activist who rose to national prominence when advocating for a contraception mandate on health care plans; and (2) the shooting of Trayvon Martin, an unarmed African-American teen killed by neighborhood watch volunteer George Zimmerman, who claimed self-defense under Florida's 'stand your ground' law. The data used to examine these cases began with a purposive sample of politically active users. We identified the most prominent political user accounts in five categories: political advocacy groups; politicians and candidates; political party operatives; journalists and pundits; and political satirists and celebrities. Recency and frequency of activity, size of follower network, ideological diversity, and political prominence were considered when creating lists within each category. From this process, we compiled a final list of 165 political elites, including more than 40 candidates for national office, 40 major advocacy groups, and more than 50 journalists and pundits. We call these the 'top-level' users (see supplemental Online Appendix, available at http://bit.ly/shah-etal-2014). We then collected follower lists for each of these users and drew a random sample of 80 of their followers, called the 'second-level' users. This resulted in 13,200 user accounts that we tracked. In practice, given the nature of Twitter user accounts, this number shrank slightly because some users' accounts were suspended or deleted. In fact, we began with the assumption that political elites gain and lose followers given their shifting prominence and visibility. Accordingly, we re-collected follower information at three distinct points during the US presidential election cycle: the initial collection in late 2011, mid-June 2012, and September 2012. Follower information was collected using the RESTful Twitter application programming interface (API) (https://dev.twitter.com/docs/api). Using an account with expanded ('whitelistcd') access to Twitter data, we gathered tweets from these users using the Streaming Twitter API (https://dcv.twittcr.com/docs/streaming-apis).2 This collection gave us the following information for each user we specified: tweets created by the user, tweets which were retweeted by the user, replies to any tweet created by the user, retweets of any tweet created by the user, and 'manual' replies to the user created without using Twitter's 'Reply' button. This resulted in a collection of more than 431 million tweets between December 30, 2011, and January 22, 2013, the day after President Obama's second inauguration. Given this sampling methodology, it is inevitable that we would pick up some 'elite' users among our follower 2X8 Handbook of digital politics Table 16.1 Follower and following counts at level 1 and 2 of collection Level 1 Level 2 Followers Mean 275,483 3.832 Median 36,324 398 Following Mean 10,551 2,243 Median 578 689 Ratio Mean 26.1 1.7 Median 62.8 0.6 sample. For example, our second-level sample included four users with more than 1 million followers: rhythm and blues singer John Legend, Whole Foods, the Hujfinglon Post, and National Public Radio (NPR)'s Scott Simon. Still, the vast majority of those included at this level were not political elites. Table 16.1 shows the means and medians of these two levels, comparing number of follower accounts, number of accounts following, and the ratio of these two numbers. Given the timing of the cases considered in this chapter, this analysis centers on the first wave of data collection. Table 16.1 compares the two levels. Focusing on median values, which avoid inflated means due to outliers, we see that level 1 users (political elites) have nearly 100 times more followers than level 2 users (follower network). Level 1 users also follow fewer users than level 2 users, resulting in a ratio of followers to following that differs dramatically between our elite and follower groups. While it may be true that follower counts do not denote influence, per se (Cha et al., 2010), this certainly differentiates our elite users from those who followed them. UNDERSTANDING THE CONTROVERSIES To analyze these data, we turn to computational methods, namely computer-aided content analysis of keywords and hashtags, along with social network mapping of these online tokens. This chapter will proceed as follows: First, we examine the frequency of the appearance of particular keywords. We then identify which level of users are using these keywords. Next, we focus on the co-occurrence of hashtags as an indicator of public sentiment and political organizing, tracking these among elites and their Computational approaches to online political expression 289 followers. Finally, we map networks of message retweeting to understand the social structure of everyday political talk. Before moving to this computational analysis, we first provide some context on the two cases, the Sandra Fluke Rush Limbaugh controversy and the Trayvon Martin George Zimmerman shooting. Sandra Fluke: A Washington, DC Conflict Georgetown law student and women's health activist Sandra Fluke landed in the center of US national controversy - and a Twitter firestorm - in February 2012 shortly after testifying in a hearing staged by Congressional Democrats. An original hearing before the House Oversight and Government Reform Committee focused on the Affordable Care Act ('Obamacare') and provisions mandating coverage of contraceptives. Committee Chairman Darrell Issa (Republican, California) refused to allow Fluke to speak in a three-hour hearing concerning the contraceptive mandate during which only men testified, all opposing this policy. Led by former House Speaker Nancy Pelosi (Democrat, California). Congressional Democrats convened an unofficial hearing February 23, 2012. to take Fluke's testimony. She spoke about the high cost of contraceptives and adverse effects on women's health, drawing praise from Democrats. Her appearance at the unofficial hearing drew derision shortly afterward, beginning with attacks from the conservative blogosphere, such as, 'Sex-Crazed Co-Eds Going Broke Buying Birth Control, Student Tells Pelosi Hearing Touting Freebie Mandate', from CNS News. The issue bubbled in those circles for five days before coming to a boil with conservative talk show radio host Rush Limbaugh. On his February 29, 2012, show, Limbaugh called Fluke a 'slut' and a 'prostitute', saying she wanted to be paid to have sex. Swift and critical response to the statement came immediately from the left, and built over days to statements from Republicans calling the remarks inappropriate, including House Speaker John Boehner (Republican, Ohio) and National Republican Senatorial Committee Vice-Chair Carly Fiorina. Limbaugh was initially unmoved, saying on his March 1 show, 'If we are going to pay for your contraceptives, thus pay for you to have sex, wc want something for it, and I'll tell you what it is: We want you to post the videos online so we can all watch'. His intransigence proved costly. Responding to online outrage directed at the program's sponsors, advertisers began pulling their spots from the show on March 2, with Sleep Train Mattress Centers announcing its decision on Twitter: 'We don't condone negative comments directed toward any group. In response, we are currently pulling our ads from Rush with 1 290 Handbook of digital politics Rush Limbaugh'. As online outrage and pressure on advertisers grew, Limbaugh apologized on his March 3 show: 'My choice of words was not the best, and in the attempt to be humorous, I created a national stir'; though some found the mea culpa lacking. The apology and a subsequent clarification did nothing to stem the sponsor losses, with some outlets reporting nearly 100 advertisers backing out of Limbaugh's show, an effect that stretched into 2014, more than two years later.' Pressure on advertisers was a clear trend on Twitter, as users called on Sleep Train, Nettlix and others to pull their spots from the show. Trayvon Martin: A Grassroots Firestorm The Florida shooting of unarmed African-American teen Trayvon Martin in February 2012 slowly built into a national conversation on race and the US justice system, discourses that resurfaced to frame the August 2014 killing of Michael Brown in Ferguson, Missouri. Martin, 17, was shot and killed by George Zimmerman, a neighborhood watch volunteer, in Sanford, Florida, the night of February 26. Zimmerman was taken into custody immediately following the shooting but was released after claiming he acted in self-defense. Florida is a so-called 'stand your ground' state, providing legal protection for individuals who use deadly force to defend themselves when they feel their lives are in danger outside of iheir homes. About ten days after Zimmerman's release, Martin's parents posted a petition to Change.org, urging Zimmerman's arrest, and the first mainstream news coverage was published the following day, when the parents tiled suit to get records in the case. National media attention followed on March 13 to 15. A week later, the Sanford police chief stepped down from the investigation and a special prosecutor was assigned to the case. Scrutiny of the case ranged from the actions of the police and prosecutors to the media's inattention to the story. The controversy was charged both racially and politically, showing considerable polarization. In-person protests and demonstrations accompanied intense social media attention to the case in the early weeks. Attention quickly narrowed to Martin's attire and questions of whether youths wearing 'hoodie' sweatshirts are inherently menacing. On March 23. speaking on Fox and Friends, Geraldo Rivera said, T think the hoodie is as much responsible for Trayvon Martin's death as George Zimmerman was'. Conservative blogs published excerpts from Martin's Twitter feed, including references to marijuana use (Graeff et al., 2014). Shortly after, a hacker broke into Martin's e-mail and social networking accounts and released personal information, which was subsequently covered by mainstream media. At the same time, Zimmerman's legal team set up a Computational approaches to online political expression 291 social media presence for their client, just as he earned scrutiny for posts about Mexicans made on his abandoned MySpace account. Nearly a month later, on April 11, 2012, the prosecutor announced second-degree murder charges against Zimmerman. EXPLORING POLITICAL TALK ONLINE Keyword Volume To analyze the relative volume of keywords on Twitter, we graphed their concentration relative to all other content. Each line graph represents the proportion of the tweets within that level mentioning a keyword or hashtag. This allowed some insight into the intensity of political talk on these topics at these different levels and its flow between elites and their followers. As can be seen in Figure 16.1, our analysis of the top-trending keywords on Twitter concerning the Sandra Fluke-Rush Limbaugh controversy %. c' Boycott — LinAnug'1 MM Figure 16.2 Proportional volume of hashtag clusters for Sandra Fluke case and Sleep Train announced its withdrawal of advertisements. This call to action trailed off more quickly among the followers than the elites, who discussed the developing story, carrying the theme in about 1 percent of all tweets well into the next week. Proportionally, the elite group tweeted more about the case, with spikes in their activity marked by spikes among followers shortly afterward in most other ways. These patterns may point to a number of possible influences. The Fluke case was a particularly Washington, DC-centric affair, beginning with Congressional hearings and quickly emerging as a flashpoint in the presidential campaign. The Washington, DC insider nature of the elite panel makes the group more likely to focus a greater proportion of Twitter activity on issues in their own arena. We also suspect Limbaugh's national profile may have led to quick and expansive focus among level 1 elites, perhaps indicated by the far greater use of Limbaugh's name than Fluke's (nearly 4.5 percent of tweets to 1.5 percent of tweets at the peak on March 3. 2012). Conversation among elites may also have been influenced by efforts from the left to highlight the case as an example of the 'war on 296 Handbook of digital politics Computational approaches to online political expression 297 Figure 16.3 Proportional volume of hashtag clusters for Trayvon Martin case women' and fundraise on the point. In contrast, it is notable that the followers were ahead of the elites in pointing toward a boycott, suggesting grassroots leadership on this matter. Turning to the Trayvon Martin case, parallel analysis Finds that among elites and their followers, #trayvon and #trayvonmartin were the most frequently used hashtags, followed by #georgezimmerman and #zim-merman (see Figure 16.3). In addition, a set of call-to-action hashtags appeared, including #iamtrayvon and #millionhoodiemarch, but was used far less frequently (less than 1 percent at its peak). It is surprising, then, that reporters reviewed the importance and effects of Internet activism through social media tools for this case (Graeff et al., 2014). As an example of everyday political talk, the timing and volume of activity in the Martin case differed from what we found with Limbaugh and Fluke. Where elites quickly and clearly led and dominated the conversation surrounding Fluke, the Martin case shows non-elite followers picking up the story earlier and discussing it with more intensity before elites reacted and came to lead the conversation. Again, without a more detailed content analysis and message tracking - issues we return to below - we can only speculate on the reasons for this. It seems likely that the difference reflects early grassroots efforts to urge media and political actors to notice the case and direct attention to it. Followers began tweeting the keyword 'Trayvon' by March 8 and began circulating the hashtags #trayvon and #trayvonmartin by March 11, in the days following Martin's parents posting their petition on Change.org. The elite group first tweeted these hashtags on March 15, after national media coverage had begun. However, once the elite group began discussing the case on Twitter, those users paid far more attention proportionally than the follower networks, coming close to double the percentage of tweets with #trayvon and #trayvonmartin on the March 23 spike. Although the elites spiked higher in tweets related to the case, their attention dropped off more quickly. The non-elite followers displayed more sustained conversation about the case, with drops that were less precipitous than those observed among the elites. Notably, the call to action implicit in the #iamtrayvon and #millionhood-iemarch hashtags were proportionally a very small part of the online conversation among elites and even less so among followers, indicating that this grassroots effort was somewhat smaller than the one behind the Rush Limbaugh boycott. Network Mapping For each of the topics, we generated a retweet network based upon tweets that included the major keywords and hashtags. Rctwccts are tweets in which users share content produced by another user, a near perfect reflection of Tarde's 'invention' and 'imitation'. Visualizations of the retweet network in both cases show clear distinctions between core groups discussing the issues. The graph depicts users and their centrality within the retweet network, with each node (bubble) in the network representing a user; the size of the node denotes how much that person retweets and is retweeted within the network; and the edges denoting a retweet. These visualizations represent the largest connected components of activity for each case. Users outside these components are not represented because of minimal interaction with this core group.4 Upon visual inspection of the Fluke network, this case shows two core clusters speaking almost entirely separately from each other (see Figure 16.4). One cluster is focused heavily on the boycott-focused hashtags of #stoprush, #boycottrush and #flushrush, represented in black. Retwccls containing these hashtags clearly dominate that core, showing a group of users heavily focused on calls to action. This cluster has fewer instances of the grouping of #standwithsandra and #sandrafluke (denoted 298 Handbook of digital politics Boycott Limbaugh Fluke Computational approaches to online political expression 299 Solidarity | Zimmerman Trayvon ■ • * • • Figure 16.4 Retweet network for Sandra Fluke case with major hashtags with light gray), as well as activity with keywords without hashtag (white). The calls to boycott are denser at the center of the network than the non-activist tweets, indicating activity focused more squarely on a set of central users. The activity among the nodes representing discussion of the actors in the event appears more disperse and less hierarchical. The contrasting core of users tweeting about the Fluke case has a retweet network far less focused on political action. The most dominant hashtag use is the cluster for #standwithrush, #istandwithrush and #slutgate (dark gray). This network cluster also has relatively heavy use of the hashtag grouping of #standwithsandra and #sandrafluke, also apparent from our factor analysis, suggesting a spirited discussion about the Figure 16.5 Retweet network for Trayvon Martin case with major hashtags various players. This cluster was considerably less dense and active than the boycott-dominated core. The network in the Trayvon Martin case also split into two distinct clusters (see Figure 16.5). Both were far less activist in their hashtag use and focused almost entirely around the use of #trayvon and #tray-vonmartin (black), with contending characterizations of the case and its players. Each node represents a user who used this hashtag or retwected a message containing the hashtag. The grouping for #zimrnerman and #georgezimmerman (dark gray) appears in one cluster but with nowhere near the dominance of the Trayvon-focused grouping. The solidarity-focused grouping of #iamtrayvon and #millionhoodiemarch (light gray) 300 Handbook of digital politics Computational approaches to online political expression 301 is virtually invisible in the network, showing almost no retweet activity among the hashtags. The two clusters, both relying on the centrality of the Trayvon hashtags, have substantially more betwecn-cluster interaction than we found in the Fluke case. Hashtag use within the network does not focus on calls to action though further analysis may reveal these in tweet content, rather than hashtag use. Overall, the cluster with the Zimmerman-focused grouping (dark gray) is less dense than the cluster containing the Trayvon-focused grouping. As observed with the use of keywords and the clustering of hashtags, these two cases functioned quite differently, despite the fact that both addressed controversies regarding inequalities. Both apparently featured social influence by elites and followers - suggesting both a two-step flow and a two-way flow of communication - though who appears to exercise influence differs by topic and stages in each controversy's evolution, especially as it related to calls to action. The visualizations of these networks confirm these differences, and suggest that measuring message propagation, information diffusion, and interpersonal influence across these fluid levels is the future of research on everyday political talk. LOOKING FORWARD Rather than reflecting on the specifics of these cases, which were mainly deployed to illustrate the most basic potential of computational approaches to understanding online political talk, we close by using them to offer some broader suggestions about the future of online politics research and the implications of big data for this coming age of computational social science. In doing so, we look to the past, returning to the work of Gabriel Tarde that began this chapter. The digital traces contained in a single tweet provide a myriad of directions for research on interpersonal interaction and social influence linked directly to Tarde's main theoretical contributions. Yet the contemporary environment also complicates this linear formulation of elite influence through the press on to the constituted public, setting the conversational agenda and thereby distilling opinion into action. Rather, as we have observed here, within the feedback loops provided by the new information environment and social networking technologies, elite agendas appear to be influenced by their followers under certain circumstances, substantially complicating the framework Tarde proposed. It may be that the most useful insights from Tarde are not derived from his early formulation of two-step flow or communication mediation, but instead from his assertion about the types of data that should be collected as part of a 'science of the social'. His approach, its full potential unrealized at the time, refuses the schism between individual and society as we have come to understand them after Dürkheim. Moreover, technology and methodology has caught up to Tarde's theory, for as Latour writes (2010: 160): What we are witnessing, thanks to the digital medium, is a fabulous extension of [Tarde's] principle of traceability. It has been put in motion not only for scientific statements, but also for opinions, rumors, political disputes, individual acts of buying and bidding, social affiliations. .. What has previously been possible for only scientific activity - that we could have our cake (the aggregates) and eat it too (the individual contributors) - is now possible for most events leaving digital traces, archived in digital databanks. As this suggests, computational approaches that stress syntactical coding such as natural language processing may prove critical to understanding the nature of social influence and interpersonal dynamics in everyday political talk (Bird el al., 2009; Jurafsky and Martin, 2009; Shah et al., 2002; Han et al., 2011).5 In addition, the growth of supervised and unsupervised machine learning tools focusing on the statistical co-occurrence of words or phrases may also prove powerful for managing large volumes of political messages (Blei and Lafferty, 2009; Hopkins and King. 2010). Such language processing allows for the coding of vast amounts of communications with considerable subtlety and qualification. Complementing these language-processing approaches are tools for tracking their movement through social networks. Efforts to examine the role of audience size, pass-along value, and conversational ability on social influence now have ways of examining these questions at a large scale (Cha et al., 2010) and can do so alongside measures of whether the message is deemed interesting or elicits positive feelings (Bakshy et al., 2011), blending the structural and the psychological. Indeed, social networking sites like Twitter and Facebook provide 'a setting where many different kinds of information spread in a shared environment' (Romero et al., 2011:695), permitting new insights regarding invention and expression of ideas and their imitation and diffusion with communication networks. These methods will provide new vistas onto political talk as it increasingly occurs through online channels, calling into question the accepted wisdom about message flows and offering new accounts of emergent forms of civic and political expression. 302 Handbook of digital politics NOTES As Terry Clark writes in the introduction of Gabriel Tarde On Communication and Social Influence: Selected Papers (1969: 25): 'Dürkheim refused to accept that sociological principles should be grounded in psychology. Sociology as a distinctive science, he held, must take as its object of study social facts; and these social facts must find their causes as well as their consequences in other distinctly social facts'. A brief tutorial of how to collect Twitter data using a simple Python script can be found at http://badhessian.org/20l2/10/collecting-real-timc-twitter-data-with-lhc-streaming-api/. Online activists have maintained detailed databases of Rush Limbaugh sponsors, tracking continued supporters (who remain targets of a boycott), and noting which brands have pulled advertisements. Al the time of writing in Fall 2014, these campaigners claim more than 2,700 local and national advertisers have withdrawn support (http://stoprush. net/rush__limbaugh_sponsor_list.php#current_a). The Fluke network overall comprises 13,365 users, with this largest component including 870. For Trayvon, the overall network comprises 13,365, with 856 users in the largest component. The vast majority of users in both are 'isolates', nodes not tied with any other users. Two main lines of modeling that have been developed do not rely so heavily on word counts and the manual curation of specific words: language modeling and statistical modeling (Monroe and Schrodt, 2008). Language modeling attempts to leverage as much information as it can out of a single document; it attempts to identify parts of speech in a given document and allows us to see the who, what, when, where, and how of a message. REFERENCES Bakshy, E.. Hofman, J.M.. Mason, W.A. and Watts, D.J. (2011). Kveryone's an influ-enccr: quantifying influence on twitter. In Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, February (pp. 65-74). ACM. Barber, B. (1984). Strong Democracy: Participatory Politics for a New Age. Berkeley, CA; University of California Press. Bennett. W.L., Wells, C. and Freelon, D. (2011). Communicating civic engagement: contrasting models of citizenship in the youth web sphere. Journal of Communication, 61, 835-856. Berelson, B., Lazarsfeld, P.F. and McPhee, W.N. (1954). Voting: A Study of Opinion Formation in a Presidential. Campaign. Chicago, IL: University of Chicago Press. Bird, S.. Klein, E. and Toper, E. (2009). Natural Language Processing with Python. SebastopoL CA: O'Reilly Media. Blei. D.M. and Laffertv, J.D. (2009). Topic models. Text Mining: Classification, Clustering, and Applications, 10, 71. Bode, L., Vraga, E K., Borah, P. and Shah, D.V. (2014). A new space for political behavior: political social networking and its democratic consequences. Journal of Computer-Mediated Communication. 19(3), 414-429. Boyd, D.M. and Ellison, N.B. (2007). Social network sites: definition, history, and scholarship. Journal of Computer-Mediated Communication, 13, 210-230. Boyd. D.. Golder, S. and Lotan, G. (2010). Tweet, tweet, relweet: conversational aspects of rctweeting on twitter. In 2010 43rd Hawaii International Conference on System Sciences (HICSS) (pp. 1 10). January, IEEE. Bruns, A. and Burgess, J.E. (2011). The use of Twitter hashtags in the formation of ad hoc publics. Paper presented at the 6th European Consortium for Political Research General Conference. August 25-27, University of Iceland, Reykjavik. Uruns, A. and Highfield, T. (2013). Political networks on twitter: tweeting the Queensland state election. Information, Communication and Society, 16(5), 667-691. Computational approaches to online political expression 303 Cha, M., Haddadi, H., Benevenuto. F. and Gummadi, P.K. (2010). Measuring user influence in Twitter: the million follower fallacy. ICWSM, 10, 10-17. Cho, J. Shah, D.V., McLeod. J.M., Mcleod. D.M., Scholl, R.M. and Gotlicb, M.R. (2009). Campaigns, reflection, and deliberation: advancing an O-S-R-O-R model of communication effects. Communication Theory, 19, 66-88. Clark, Terry (ed.) (1969). Gabriel Tarde On Communication and Social Influence: Selected Papers. Chicago, IL: University of Chicago Press. Dahlgren, P. (2000). The Internet and the democratization of civic culture. Political Communication, 17, 335-340. Dahlgren, P. (2005). The Internet, public spheres, and political communication: dispersion and deliberation. Political Communication, 22(2), 147-162. Dylko, I.B., Beam, M.A., Landrevitle, K.D. and Gcidner, N. (2012). Filtering 2008 US presidential election news on YouTube by elites and nonelites: an examination of the democratizing potential of the internet. New Media and Society, 14(5), 832-849. Ekstrom, M. and Ostman, J. (2013). Information, interaction, and creative production: the effects of three forms of Internet use on youth democratic engagement. Communication Research, online pre-publication, 0093650213476295. 1-22. Eveland, W.P. and Hively, M.H. (2009). Political discussion frequency, network size, and 'heterogeneity' of discussion as predictors of political knowledge and participation. Journal of Communication, 59(2), 205-224. Freelon. D.G. (2010). Analyzing online political discussion using three models of democratic communication. New Media and Society, 12(7), 1172-1190. Gainous, J. and Wagner, K.M. (2014). Tweeting to Power: The Social Media Revolution in American Politics. New York: Oxford University Press. Gamson, W.A. (1992). Talking Politics. New York: Cambridge University Press. Gil de Ziihiga, H.G., Jung, N. and Valenzuela, S. (2012). Social media use for news and individuals' social capital, civic engagement and political participation. Journal of Computer-Mediated Communication, 17(3), 319- 336. Gil de Zuniga, H.G., Puig-I-Abril, E. and Rojas, H. (2009). Weblogs, traditional sources online and political participation: an assessment of how the internet is changing the political environment. New Media and Society. 11(4), 553 574. Golbeck, J., Grimes, J. and Rogers, A. (2010). Twitter use by the L'C Congress. Journal of the American Society for Information Science ant! Technology, 61(8), 1612. Graeff, E., Stempeck, M. and Zuckcrman, E. (2014). The battle for 'Trayvon Martin': Mapping a media controversy online and off-line. First Monday, 19(2). http://firstmonday. org/ojs/index.php/fm/article/view/4947. Graham, T. and Hajru, A. (2011). Reality TV as a trigger of everyday political talk in the net-based public sphere. European Journal of Communication, 26(1), 18-32. Gucorguieva, V. (2008). Voters. MySpace, and YouTube the impact of alternative communication channels on the 2006 election cycle and beyond. Social Science Computer Review, 26(3), 288-300. Habermas, J. (1984). The Theory of Communicative Action: Vol. 1. Reason and the Rationalization of Society. McCarthy. T. (trans.). Boston, MA: Beacon. Han, J.Y., Shah, D.V., Kim, E., Namkoong, K., Lee, S.Y., Moon, T.J., Cleland, R., McTavish, F.M. and Gustafson. D.I I. (2011). Empathic exchanges in online cancer support groups: distinguishing message expression and reception effects. Health Communication, 26(2), 185-197. Ilanna, A., Wells, C, Maurcr, P., Friedland, L., Shah, D. and Matthes, J. (2013, October). Partisan alignments and political polarization online: a computational approach to understanding the French and US presidential elections. In Proceedings of the 2nd Workshop on Politics, Elections and Data (pp. 15-22). ACM. Hardy, B.W. and Scheufele, D.A. (2005). Examining differential gains from Internet use: comparing the moderating role of talk and online interactions. Journal of Communication, 55(1), 71-84. Hopkins, D.J. and King, G. (2010). A method of automated nonparametric content analysis for social science. American Journal of Political Science, 54(1), 229-247. 304 Handbook of digital politics I luckfeldt, R. and Sprague. J. (1995). Citizens, Politics, and Social Communications Information and Influence in an Election Campaign. New York: Cambridge University Press. Jungherr, A.. Jürgens. P. and Schoen, H. (2012). Why the Pirate Party won Ihe German election of 2009 or the trouble with predictions: a response to Tumasjan, A., Sprenger, T.O., Sander, P.G. and Welpe, I.M. 'Predicting Elections with Twitter: What 140 Characters Reveal about Political Sentiment'. Social Science Computer Review, 30(2), 229-234. Jurafsky, D. and Martin, J.H. (2009). Speech and Language Processing: An Introduction to Natural Language Processing. Speech Recognition and Computational Linguistics. 2nd edn. Upper Saddle River, N'J: Prentice Hall. Katz, E. (2006). Rediscovering Gabriel Tarde. Political Communication, 23(3), 263-270. Katz, E. and Lazarsfeld, P. (1955). Personal Influence: The Part Played by People in the Flow of Mass Communications. Glencoe, IL: Free Press. Kim, J., Wyatt, R.O. and Katz, E. (1999). News, talk, opinion, participation: the part played by conversation in deliberative democracy. Political Communication, 16, 361-385. Kinnunen, J. (1996). Gabriel Tarde as a founding father of innovation diffusion research. Acta Sociologies, 39(4), 431-442. Kwak, N., Williams, A., Wang, X. and Lee. H. (2005). Talking politics and engaging politics: an examination of the interactive relationships between structural features of political talk and discussion engagement. Communication Research, 32, 87-111. Latour, B. (2010). Tarde's idea of quantification. In Candea, M. (ed.), The Social After Gabriel Tarde: Debates and Assessments. London: Routledge. Lazarsfeld. P.F., Berelson, B. and Gaudet, H. (1944). The People's Choice: How the Voter Makes Up His Mind in a Presidential Campaign. New York: Duell, Sloan & Pearcc. Lazer, D., Pentland, A.S.. Adamic, L., Aral, S., Barabasi, A.L.. Brewer, D.....and Van Alstyne, M. (2009). Life in the network: the coming age of computational social science. Science, 323(5915), 721. Lee, N.J., Shah, D.V. and McLeod, J.M. (2013). Processes of political socialization a communication mediation approach to youth civic engagement. Communication Research, 40(5), 669-697. McLeod. J.M., Daily, K., Guo, Z., Eveland, W.P., Bayer. J., Yang, S. and Wang, H. (1996). Community integration, local media use, and democratic processes. Communication Research, 23(2), 179-209. McLeod, J.M., Schcufclc, D.A. and Moy, P. (1999). Community, communication, and participation: the role of mass media and interpersonal discussion in local political participation. Political Communication, 16, 315-336. Monroe, B.L. and Schrodt, P.A. (2008). Introduction to the special issue: the statistical analysis of political text. Political Analysis, 16(4), 351-355. Mutz, D.C. (2006). Hearing the Other Side: Deliberative Versus Participatory Democracy. New York: Cambridge University Press. Namkoong, K., Shah, D.V., Han, J.Y.. Kim. S.C.. Yoo, W., Fan, D.....and Gustafson, D.H. (2010). Expression and reception of treatment information in breast cancer support groups: how health sclf-cfficacy moderates effects on emotional well-being. Patient Education and Counseling, 81, S41-S47. Papacharissi, Z. (2004). Democracy online: civility, politeness, and the democratic potential of online political discussion groups. New Media and Society, 6(2), 259-283. Papacharissi, Z. (ed.). (2010). A Networked Self: Identity, Community, and Culture on Social Network Sites. New York: Routledge. Pasck, J., More, E. and Romer, D. (2009). Realizing the social Internet? Online social networking meets offline civic engagement. Journal of Information Technology and Politics, 6(3-4), 197-215. Pingree, R.J. (2007). How messages affect their senders: a more general model of message effects and implications for deliberation. Communication Theory, 17(4), 439 461. Price, V. and Cappella, J.N. (2002). Online deliberation and its influence: the electronic dialogue project in campaign 2000. IT and Society, 1(1), 303-329. Computational approaches to online political expression 305 Rogers, E.M. and Shoemaker, F.F. (1971). Communication of Innovations: A Cross-Cultural Approach. New York: Free Press. Romero, D.M.. Meeder, B. and Klcinberg, J. (2011, March). Differences in the mechanics of information diffusion across topics: idioms, political hashtags, and complex contagion on Twitter. In Proceedings of the 20th International Conference on World Wide Web (pp. 695-704). ACM. Schudson, M. (1978). The ideal of conversation in the study of mass media. Communication Research, 5(3). 320-329. Segcrberg, A. and Bennett, W.L. (2011). Social media and the organization of collective action: using Twitter to explore the ecologies of two climate change protests. Communication Review, 14(3), 197-215. Shah, D.V., Cho, J., Eveland, W.P. and Kwak, N. (2005). Information and expression in a digital ace: modeling Internet effects on civic participation. Communication Research, 32, 531-565. Shah, D.V.. Cho, J., Nah, S., Gotlieb, MR. Hwang, H., Lee. N., Scholl, R.M. and McLeod. D.M. (2007). Campaign ads. online messaging, and participation: extending the communication mediation model. Journal of Communication, 57, 676-703. Shah, D.V. and Scheufele, D.A. (2006). Explicating opinion leadership: nonpolitical dispositions, information consumption, and civic participation. Political Communication, 23(1), 1-22. Shah, D.V., Walls, M.D., Domke, D. and Fan, D.P. (2002). News framing and cueing of issue regimes: explaining Clinton's public approval in spite of scandal. Public Opinion Quarterly, 66(3), 339-370. Sotirovic. M. and McLeod, J.M. (2001). Values, communication behavior, and political participation. Political Communication. 18, 273-300. Sotirovic, M. and McLeod, J.M. (2004). Knowledge as understanding: the information processing approach to political learning. In Kaid, L. (ed.). Handbook of Political Communication Research. Mahwah, NJ: Lawrence Erlbaum Associates. Tarde, G. (1969 [1898]). Gabriel Tarde On Communication and Social Influence: Selected Papers, Vol. 334. Chicago, IL: University of Chicago Press. Tarde, G. (1903 [1890]). The Laws of Imitation, trans. E.C. Parsons. New York: Henry. Holt. Thackeray, R. and Hunter, M. (2010). Empowering youth: use of technology in advocacy to affect social change. Journal of Computer-Mediated Communication, 15(4), 575-591. Tumasjan. A., Sprenger. T.O., Sandner, P.G. and Welpe, I.M. (2010). Predicting elections with Twitter: what 140 characters reveal about political sentiment. ICWSM, 10, 178-185. Valenzuela, S.. Kim, Y. and de Zufiiga. H.G. (2012). Social networks that matter: exploring the role of political discussion for online political participation. International Journal of Public Opinion Research, 24(2), 163-184. Vitak, J., Zubc, P., Smock. A., Carr, C, Ellison, N. and Lampe, C. (2011). It's complicated: Facebook users' political participation in the 2008 election. Cyberpsvchology, Behavior and Social Networking, 14(3), 107-114. Walsh, K.C. (2012). Putting inequality in its place: rural consciousness and the power of perspective. American Political Science Review, 106(03), 517-532. Wimsalt, W.C. (2007). Re-engineering Philosophy for Limited Beings: Piecewise Approximations to Reality. Cambridge, MA: Harvard University Press. Ye, S. and Wu. S.F. (2010). Measuring message propagation and social influence on Twitter. com. In IEEE (ed.). Social Informatics. Berlin and Heidelberg: Springer. 1 23. Visibility and visualities: 'ways of seeing' politics in the digital media environment Katy Parry On contemplating how to approach writing a chapter on visual politics online, my initial thoughts turned to the potential slipperiness of each of these terms. How to think about politics online as distinct from its offline manifestations? How narrowly or broadly to define politics and the political? In terms of the recognized political actors and institutions of official politics and policy-making, or more broadly, as a public space in which meanings, identities and values are contested? How productive is it to separate the visual dimension from other qualities or modalities (text, sound) across a range of digital media forms? Indeed 'the visual' is about more than images alone, so how to place notions of the visual within wider concerns of visibility and visuality? Such questions around indistinct boundaries and instability form the basis for this chapter, then, and in exploring the slipperiness of these terms I address old and new concerns about visibility, vision and visuality in political communication and culture. Drawing on relevant insights from political studies, media and communications and social movement studies, the present chapter places what has come to be known as 'visual culture studies' at the heart of its approach and spirit, for reasons further explained below. It is in the interplay of online and offline political practices, serious and comedic mediations, and their authoritative or subversive purposes, that senses of convergence and collision exist, and through which new opportunities for analysis emerge. But it is also in the productive sharing of resources and tools from across the academic disciplines that we might better scrutinize, understand and appreciate the varied forms of visual politics online. The chapter is organized into five sections: first, there is a brief outline of how varied disciplinary perspectives have informed the discussion of politics and mediated imagery, tracing the tensions and anxieties associated with emergent media technologies and the concerns over the corrupting influences identified with such media. Second, elaborating on my argument for a central role of 'visual culture studies' in exploring online politics, the subsequent section outlines notions of visibility and visuality. The following two sections are broadly split into politics 'from below' and 417 418 Handbook of digital politics politics 'from above'. Arguably, the separation of protest from 'official politics' can create another division of research agendas which belies the multiple civic activities and practices each of us chooses to undertake (or not) in our everyday lives. In the hybrid media environment through which many of us experience politics, both established politicians and protesters appear to be adopting more personalized and expressive forms of engagement (Bennett, 2012; Chadwick, 2013). The sections are designed to be illustrative and exploratory, as it is beyond the scope of the current chapter to provide detailed empirical analysis. The final section sets out future research questions and my concluding comments. Before proceeding to the next scene-setting section, I outline two guiding stipulations. First, as indicated above, studying visual politics online requires openness to a variety of approaches, based on the research questions and methods that most intrigue and provoke us. In his article entitled 'There are no visual media', W.J.T. Mitchell warns against 'visual culture as the "spectacle" wing of cultural studies', noting that its very promise is in its insistence 'on problematizing, theorizing, critiquing and historicizing the visual process as such' (Mitchell, 2005: 264). A broad interdisciplinary interest in varied cultural practices and artefacts does not equate to a flattening out or homogeneity of approach, or a misrecognition of how one medium's affordances are qualitatively different to another's. As Mitchell has argued elsewhere, 'the opening out of a general field of study does not abolish difference, but makes it available for investigation, as opposed to treating it as a barrier that must be policed and never crossed' (Mitchell, 2002: 173). Second, the fast-evolving role of the Internet in our everyday lives has further disrupted many of the traditional parameters for studying images in mediated contexts. Such disruptions are not entirely new: for example, Elihu Katz (1988) recognized many years ago how the technological theory of 'disintermediation', or cutting out the middleman, could be applicable for media sociology. Social media use and peer-to-pcer sharing online merely represent the latest technological and cultural practices embodying a reinvigorated sense of connectivity and direct communication. The promiscuity of digital media images in a 'cut-and-mix' culture (van Zoonen et al., 2010) creates both possibilities and risks for producers, depicted subjects and audiences who view and share such images. The processes of upload and display add to the ephemeral quality, as images circulate and become divorced from original captions or audio, remediated in ever-mutating 'circuits of culture' (du Gay et al., 1997), seemingly unbounded by shared viewing contexts or clearly defined 'imagined communities' (Anderson, 1991). However the often celebratory claims for disintermediation and a sense of less mediation can be misleading and obfuscate the shifts that Visibility and visualities 419 digital technologies enable; new and old intermediaries may be adapting and altering their role but they remain 'vitally' important (Thumim, 2012). All digital images encountered via the Internet are mediated in some form; the contexts may be increasingly varied, with images constituted as a mix of both amateur and professional in origin, but this multifarious jumble of image-text circulates within a discursive public space of framing practices, semiotic recipes, rhetorical challenges and ironic gestures. In an inevitably brief review, the next section provides contextual background to the broad fields of study which inform the chapter, setting out the traditional anxieties, valuable perspectives and emergent tensions through which we might incorporate the study of visual culture in understanding politics and digital media. THE SPECTRE OF THE SPECTACLE: THE HAUNTING ANXIETIES AROUND THE VISUAL IMAGE IN POLITICAL COMMUNICATION The role of the image in mediated political communication has long provoked suspicion and unease. With its concerns for governance, democracy and citizenship, political studies provides theoretical tools for assessing the health of the polity and public sphere, often bringing a defensive posture to guarding the integrity of politics against debasing forces (Crick, 1968; Flinders, 2012). Although a concern of political philosophers and critical theorists in earlier centuries, fears of a distracted and passive public or citizenry have more recently been associated with the dominant role of television in political campaigning and as the main source for public knowledge. Throughout the twentieth century, as politicians increasingly addressed publics through the mass medium of television, concerns were raised over the diminishing of political life into a spectacle, distorted by a media-driven shift promoting conflict, sensationalism and inauthentic celebrity politicians. In such 'audience democracies' (Manin, 1997) citizen-viewers are characterized as passive, apathetic spectators, monitoring the actions of political leaders but merely reactive to the theatre of political life performed by a set of interchangeable elites, rather than socially engaged actors with a significant decision-making role (see also Edelman, 1988; Meyer and Hinchman, 2002; Postman, 1987; Putnam, 2000). Writers concerned with the role of media in democracy note trends towards a politics evermore shaped by 'media logic' or 'mediatization', accompanied by an overemphasis on stylization, presentation, performance and image (Blumler and Kavanagh, 1999; Mazzoleni and Schulz, 1999; Stromback, 2008). 420 Handbook of digital politics Visibility and visualities 421 Where some authors see politics tarnished by the blurring lines of information and entertainment that such 'media logic' brings about, with a public service ethos crowded out by consumer-driven "infotainment' and lifestyle programming (McChesney, 1999; Thussu, 2007), others see an enhancement for democratic life in the popular engagement with politics encouraged by a variety of formats offering a mix of serious and more playful generic recipes (Corner and Pels, 2003; van Zoonen, 2005; Richardson et al., 2012). The shifting interests and concerns reflect a cultural turn across humanities and social sciences, in which other cultural factors and contested sites of meaning and identity arc considered alongside political structures, for example, in debating the role of comedy as a catalyst for civic engagement (Baym and Jones, 2012), or the intersections of celebrity culture and political culture (Street, 1997). Alongside this, there is a turn to the emotional or personal aspects of political life, where the private lives and personal-psychological qualities of political leaders are emphasized in discussion of their political performances across the broader media environment (Corner, 2003; Langer, 2011; Stanyer, 2012). In short, many of the discussed trends in political communication and culture - whether on the appeal to the emotions and entertainment, a personalization of politics, or the role of humour and satire - often include an implicit or explicit concern with the visual or symbolic dimension. The nature of politics as spectacle is often central to these perspectives, despite proponents rarely engaging with the visual or symbolic forms or properties in any detail. The concerns for a healthy and vibrant public sphere are laudable, but such perspectives arc in danger of oversimplifying the identified problem. The 'iconophobia' or 'iconoclasm' at the heart of the Western tradition has now attracted critical attention from authors calling for a rethink towards spectatorship and democracy (Finnegan and Kang, 2004; Green, 2010). In summary, I suggest three potential pitfalls. First, there is often a sense of the emotive image in contest with the rational word; that the expressive, symbolic or affective dimensions work to degrade the crucial rationality of the political realm. This negates the interplay of word and image in mediated communications and the constitutive role of both in interpreting our social worlds. Second, an idealized notion of the citizen is contrasted against the deficiencies of the spectator in an unproductive dichotomization which fails to explore how watching and interpreting are also active and creative: 'The spectator also acts, like the pupil or scholar. She observes, selects, compares, interprets' (Ranciere, 2011: 13). Third, a traditional emphasis on official politics which narrowly defines political engagement or political actions is in danger of overlooking alternative roles for citizens outside of recognized political structures (see Coleman and Blumler, 2009: 156). This final point brings the discussion back to the distinctions between the top-down politics of institutions and policy-makers against the 'from-bclow' character of grassroots politics and dissent. Where the visual performance of political leaders has been greeted with unease and suspicion (if examined at all), those writing on protest and dissent provide a much more fruitful scholarly engagement with the sense of the visual and the artistic as powerful cultural tools in political action (for a recent collection see McLagan and McK.ee, 2012). For those struggling to gain attention on the political stage, 'image politics' (DeLuca, 1999) offer a striking way to promote a cause and spark imagination. Operating with a freedom of expression that traditional party politicians are unlikely to risk, the politics of dissent, or contentious politics, can embrace the symbolic and theatrical, and even be cheered for combating the 'pseudo-events' (Boorstin, 1962) of the political elite with heartfelt and humorous image events and political artwork. Choice of imagery and expression can, of course, have repercussions for those hoping to move from 'outsider to 'insider' status. Fighting the 'image of power with the power of imagery' (Doerr and Teune, 2012) not only identifies you as part of a collective '99%' (to use the Occupy movement's slogan), but, through an adoption of the symbols of contentious politics, can reinforce your status as a heckler rather than an orator; just as notions of 'the political class' can reinforce a sense of an unreachable elite in the distant echelons of high society. Finally, visual culture studies or 'image studies', with its ancestry in art history, museum studies and media studies, provides the younger sister to the current melange for this chapter. It is through the visually inflected framing of visual culture studies that we can better recognize the rhetorical, aesthetic, expressive and ironic qualities of politically themed imagery. One particular strength of this perspective is that it reflects on its own analytical usefulness, with key theorists sensitive to the limitations of even labelling media texts as visual media (Bal, 2003; Mitchell, 2005). It is in studying the interplay and interdiscursivity of sound, word and image in a variety of media platforms, formats and genres that we can consider visual images as resources for making meaning within a mix of modalities; that is, as resources for producers, viewers or users. Placing an emphasis on the visual here is an attempt to redress the traditional text-based research bias in political communication. Given the above summary, this chapter aims to 'scope out' the 'ways of seeing' and ways of understanding politics online. The visual metaphors in the preceding sentence are deliberate: conflations of seeing and believing, and seeing and understanding ('Ah, I see') are thought to betray a favouring of vision as the superior sense through which to access 422 Handbook of digital politics Visibility and visualities 423 the world (see Jay, 1993 and Mirzocff, 2011 for historical and critical accounts), and yet it is only more recently that social scientists have embraced an iconic turn or 'pictorial turn' within their research agendas (Mitchell, 1994). Visual experience encompasses a great deal more than pictures or imagery and, indeed, includes the written word in graphic form; yet the study of images has been separated from scientific and literary enquiry historically, marked with a rationalist suspicion dating back to Plato's allegory of the cave (see Jay, 1993: 27; Mitchell, 2005: 86). It is not only material cultural artefacts that are of interest here, it is the way such images interact with the 'pictures in our heads' (Lippmann, 1922), the mental images of our mind's eye which also guide how we see the world and how we place ourselves within social spaces and political structures. Our ways of seeing the world arc not only about vision (what the eyes observe), but 'visuality' or 'visualities'; 'what is made visible, who sees what, how seeing, knowing and power are interrelated' (Bal, 2003: 19). Notions of control, knowledge and power are crucial here: visuality is also about a political struggle, the 'right to look' and be seen as citizens (Mirzoeff, 2011). It is necessary then to further outline the intersecting notions of vision, visuality and visibility in relation to both mediated political performances and our ways of seeing the political world. RECONCILING NOTIONS OF VISIBILITY AND VISUALITY IN POLITICAL COMMUNICATIONS John B. Thompson's work on the visibility of politicians is central to thinking about how visibility and politics link to visuality and vision, and is often cited by those interested in how our politics has become more mediatized, personalized or intimate. Thompson writes of how the development of mass media technologies in the twentieth century, and especially television as the central platform, linked issues of visibility with vision (Thompson, 1995). In the age of mediated visibility, the field of vision is shaped 'by the distinctive properties of communication media, by a range of social and technical considerations (such as camera angles, editing processes and organizational interests and priorities) and by the new types of interaction that these media make possible' (Thompson, 2005: 35-36). Thompson characterizes mediated visibility as a double-edged sword for politicians; a 'source of a new and distinctive kind of fragility' (2005: 42) in a more complex and less controllable information environment, and where public-private boundaries are rewritten. More recently, Daniel Dayan has argued that historically visibility was desired as an enviable right enjoyed by the few; a form of attention-gaining which publics, as spectators, have been denied: 'Being anonymous has become a stigma, and visibility has become a right frequently and sometimes violently claimed; a right that all sorts of people feel entitled to obtain. The exclusive visibility once contcrrcd upon some is perceived by the anonymous as an injustice in need of redress' (Dayan, 2013: 139). Dayan proposes a 'paradigm of visibility', which, as distinct from the media effects paradigm, 'stresses the role of media in coordinating collective attention' (p. 139). This offers a narrative of'deprivation followed by a conquest' (p. 139) initially acquired by citizens as a form of visibility offered on a conditional basis (as reality TV contestants, for example), or by violent means, such as terrorist acts designed with media in mind. But this is often 'the wrong type of visibility' (p. 141). Further along in Dayan's narrative, new media technologies and platforms emerge, such as Facebook and YouTube, in which the quest for visibility can be taken a step further: 'Not only do such media allow publics to acquire visibility, and to acquire visibility on their own terms, but they also allow them to define the visibility of others, to become organizers of visibility' (p. 143). Such 'visibility entrepreneurs' can challenge the narrative offered by mainstream media, promoting debate and even scandal, but they can also encourage a battlefield mentality in which professionals assert their legitimacy against 'uninvited intruders' (p. 145). In either case, this is 'an attempt at interfering with a silencing process' (p. 150), in which various forms of expression can be heard or seen, and compared on an equal footing. New representational practices offer innovative ways for users and spectators to find social and political meanings while negotiating issues of truth, trust and credibility. At the same time, the traditional facilitators of the public sphere face increasing economic pressures to perform such a role. In some cases this means the loss of established newspapers, or in a recent example specific to news images, the Chicago Sun-Times sacking its entire photography department (BBC, 2013). Rather than thinking about conditions of visibility from the politicians' perspectives, Dayan's paradigm of visibility helps us to conceptualize online and physical spaces as contested sites of political meaning, values and identity, in dialogue with other more traditional mediated forms. The question then becomes one of investigating the different patterns of involvement and the kinds of visual display produced and circulated across the mediascapc. 424 Handbook of digital politics PUBLICS AND SPECTATORS: BECOMING VISIBLE CITIZENS IN VISUAL DISPLAYS Dayan's paradigm of visibility, with its emphasis on the media coordination of collective attention, offers a useful way to think about how citizenship is performed and the reactions of publics to different attempts to disrupt the status quo. In her book, Revolting Subjects, Imogen Tyler (2013) explores how certain groups in society are figured as 'revolting' and how such stigmatization can be resisted through aesthetic and political strategies. In the chapter on the August 2011 riots in the UK, feelings of abandonment and alienation expressed by the rioters are intricately linked to their sense of invisibility in the social body. Tyler cites the Guardian-London School of Economics (LSE) research project, Reading the Riots, which interviewed 270 people who had participated in the riots: 'This sense of being invisible was widespread' (Lewis et al., 2011: 25, cited in Tyler, 2013: 197). But not all attempts to combat societal invisibility - and especially those of a chaotic and violent nature - lead watching publics to reassess their prior attitudes; indeed in the case of the riots, the outraged media attention, public fear and harsh sentences handed down to participants all suggest that rather than effecting an 'alternative aesthetics', the rioters 'became the abjects they had been told they were' (Tyler, 2013: 204). One way in which the rioters and bystanders attempted to make sense of the disturbances and looting in their own locales was through the sharing of camera images via Twitter; interestingly, these shared images included captured TV screenshots in addition to on-the-strect user-generated images, indicating a sense of involvement through 'practices of remote witnessing' alongside the impulse to capture the here and now (Vis et al., 2013: 396). The sharing of such images also records the 'second screen' phenomenon, whereby users participate via websites whilst also watching television; live media events and elections tend to attract this activity but a regular UK example is the high levels of tweeting during the BBC political panel show Question Time (using the #bbcqt hashtag) (Anstead and O'Loughlin, 2011). This struggle for visibility is often most compelling in the 'image politics' (DeLuca, 1999) of political protest and social movements, with those less powerful utilizing the rhetoric of the visual in order to gain the eyes and cars of the public and, ultimately, society's decision-makers. With the diverse groups under the anti-globalization banner dominating the scene in the later twentieth century, the truly transnational character of protest coalesced in 2011 as a mixture of movements with revolutionary aims (for example, Occupy, 15-M and the 'Arab Spring' demonstrations) took to the streets in spectacular style, claiming their right to be visible and vocal Visibility and visibilities 425 in, at times, carnivalcsque displays (see Castells, 2012; Gerbaudo, 2012; Khatib, 2012). In their analysis of YouTube videos uploaded and shared over the 18 days of the Egyptian uprising in 2011, starting with the mobilization of demonstrators on 25 January, Mohamed Nanabhay and Roxane Farmanfarmaian write of an 'amplified public sphere' created through the complex interplay of the 'inter-related spaces of the physical (protests), the analogue (satellite television and other mainstream media) and the digital (internet and social media)' (Nanabhay and Farmanfarmaian, 2011: 573). Warning against simplistic separations between amateur and mainstream image-making, or singling out social media such as Facebook or Twitter, their study points to a symbiotic relationship between journalists and activists, producers and consumers. Their point on the importance of physical place is reinforced in Paulo Gerbaudo's (2012) Tweets and the Streets, with the significance of assembly, solidarity and corporeality emphasized in the reappropriation of public space, such as the 15 May 2011 demonstrations in cities across Spain (and in what became known as the 15-M movement, or los indigna-dos). Crucially, the indignados of Spain were also expressing their collective indignation, and as Gerbaudo argues, while social media was central in mobilizing the demonstrations and protest camps that followed, it was in the symbolic and material concentration in physical spaces, such as in Puerta del Sol in Madrid, that protesters rediscovered 'a sense of physical communion' (Gerbaudo, 2012: 96). In adopting the chant of'We are not on Facebook, we are on the streets', this was about appearing as a tangible public, and signals how visibility is central to this debate (p. 96). In the summer of 2013, Taksim Square in Istanbul, and the adjoining Gezi Park, became the latest symbol of an extended protest and clash with authority, with the initial protests against the square's development mutating into widespread demonstrations against the authoritarian nature of the government. Critically, it is the imposition of a commercial redevelopment in an iconic public space that first sparked the protests, while the reaction of Turkish Prime Minister Recep Tayyip Erdogan was to take a dismissive stance towards the protesters and the solidarity expressed via social media, characterizing Twitter as a 'menace' to society (Shafak, 2013). In setting up camps and demonstrations in the heart of cities around the world, such 'spectacles of dissent' (D'Arcus, 2006) are both about mobilizing local people and embodying revolutionary zeal in the immediate physical place, but also about amplifying their countercultural message through the documentation and circulation of images and videos online. As Tina Askanius argues, we are seeing an 'acsthctization of public 426 Handbook of digital politics Visibility and visualities All protest' in the theatrical displays and vast body of images and videos created: 'This emerging audio-visual repository of interconnected narratives stages popular contestation within a coherent framework and constitutes the basis from which collective identity formation is forged among activists scattered around the world' (Askanius, 2010: 341). The hope of such activists is to challenge the legitimacy of existing power relations through alternative forms of organization and political identity, adopting visually strong practices that range from the raw and antagonistic to the absurdly humorous. Mobile media devices also allow protesters to subvert the surveillance tactics of police, playfully mimicking their recording practices and even capturing violent police behaviour which has led to public inquiries and arrests (Archibald, 2011; Shaw, 2013). The 15-M and Occupy movements are emblematic of the challenge to democratic authority through networked action, but also of the desire to create an experiential and sensorially rich public space for imaginative reworking of what 'the political' might mean. POLITICIANS ONLINE: SHARING A HUG AND GETTING THE JOKE When Barack Obama realized he had secured victory in the 2012 US presidential election, his team chose to announce his second term by tweeting an image of the President hugging his wife Michelle with the simple statement: 'Four more years'. The image became the most retweeted in Twitters history, signalling the desire of more than 816000 people who shared the image in those first few days (Ries, 2012) to join in some kind of communicative political action or affinity; but also signalling a political leader comfortable with announcing his victory via social media, before his televised speech. The photograph had been taken months earlier in mid-August by Scout Tufankjian, a campaign photographer, and was selected by digital staffer Laura Olin working as part of Obama's social media team. Olin puts Obama's successful social media campaign down to choosing diligent staff who also 'knew their social-media shit', and letting those staff 'talk to people like people' (cited in Ries, 2012). In this way, Obama's campaign harnessed the visual and verbal potential of online mediation, and achieved the much-sought-after 'virality' (the copying and multiplying of a message or phenomenon through social networks) which generally eludes professional political marketing. The particular expressive qualities of the photograph are significant and strategic: its selection speaks to the personalized and emotional appeal of a president who also appears as a loving husband, embedded in a political style that 'weaves together matter and manner, principle and presentation, in an attractively coherent and credible political performance' (Pels, 2003: 57). Effective use of social media, as with other media genres, is contingent on embracing socio-technical knowhow and competences in mediated visibility. This enables a projected image of authenticity and integrity and is a recipe that few politicians accomplish, at least with any consistency. Images especially thrive in a digital world of mash-up, montage, juxtaposition, repetition and manipulation. While Obama's pictured embrace of his wife went viral and became an iconic image of the campaign, Limor Shifman would distinguish this viral image from the 'memetic video' or image which 'lures extensive creative user engagement in the form of parody, pastiche, mash-ups or other derivative work' (Shifman, 2012: 190). These images are much more indicative of the participatory culture of the Internet, according lo Shifman. The most popular Internet memes, whether user-generated or popular culture-related, are humorous and playful but not necessarily political. Those dealing directly with politics can express a range of expressive modes, from light-hearted mockery to oppositional fervour. Visual memes are particularly suited to travelling across national borders, whether mocking in tone (such as the 'Pepper Spray Cop', http://peppersprayingcop.tumblr.com/. PhotoShopped after Lieutenant John Pike casually sprayed peaceful Occupy protesters in California), or symbolic of a struggle against corruption and state violence, (for example, the 'We are all Khaled Said' Facebook page and Ihc later appropriations of Khaled's photograph as a 'visual injustice symbol' by activists during the Egyptian uprising; Olesen, 2013; see also Khatib, 2012). The success of the viral dance video by South Korean pop-star, PSY, 'Gangnam Style', subsequently inspired its own political parodies, with 'Mitt Romney Style' and 'Kim Jong Style' examples of 'prosumer-generated material with political intent (Miiller and Kappas, 2013). In some cases, politicians refer to their own memes. signalling that they are in on the joke and conversant in the socio-technological practices of online platforms: On 26 August 2012 Barack Obama signed off from his Reddit 'Ask me anything' session with a reference to the 'Not Bad' meme. The 'Not Bad' meme is based on a photograph of the US President pulling a strange expression known as a 'sturgeon face' during a visit to the UK in May 2011, and which is thought to convey general satisfaction. The combination of media platform and intertextual playfulness signal a rare mix of approachability and self-assurance. UK Deputy Prime Minister, Nick Clegg, showed he too could get the joke when a video he recorded apologising to university students for breaking his election pledge on tuition fees was subsequently set to music and released on satirical website 'The 428 Handbook of digital politics Visibility and visualities 429 Poke' (http://www.thepoke.co.uk): Clegg gave his consent for the video to be released as a charity single, seemingly content to join in the mocking of his own insincerity. Nevertheless, emergent hostility against a strong cult of visual iconography can signal more than satirical or light-hearted ridiculing of political leaders. The destruction of material posters, murals and statues, along with the subsequent circulation of the images and videos depicting such protests online, provides a symbolic rejection of the visual legacy of a regime; as happened in Syria in March 2011, when footage of the ruling Assad family posters being ripped from buildings was shared across YouTube, Twitter and Facebook (Caldwell, 2011). FUTURE RESEARCH QUESTIONS AND CONCLUDING COMMENTS This chapter has so far attempted to set the scene for researching visual politics online. In: (1) outlining the traditional tensions between politics, popular culture and images; (2) emphasizing the interrelated notions of visibility and visuality (including how we appear as political actors and the representational forms employed); and (3) providing some recent illustrations of visual politics 'from below' and 'from above', a number of questions emerge for further consideration and examination: images or other visual artifacts' (Pauwels, 2008: 84). As indicated earlier, pulling together approaches and methods from a variety of research fields can offer an enriched perspective from which to question and problema-tize the nature of visibility and visuality as encountered in the political realm. Levels of attentiveness and interest are not assured by the expanse of networked political information available, and participation remains unequal across different communities and socio-economic groups, but the affordances of the Internet-based technologies undoubtedly enable access to an abundance of visual display from around the world, and offer cheap and easy ways to generate new material. Paradoxically, the motivations behind the patterns and practices of re-presentation, linking and sharing might work to question the digital images' supposed inherent ambiguity and the 'post-photographic' disruption to truth claims. Our investment in digital images as forms of communicative expression suggests a more complex picture than a simple characterization of shallow naivety or post-photography scepticism. Whether perceived as compelling evidential material, or as profound and transformative expressions of solidarity or affinity, the role of digital images in political discourse has undoubtedly been enhanced through Internet-based presentation and the resulting collective (and at times disorderly) debates they inspire on meanings, values and identities. • How might we better understand imaging practices and online activity (posting, viewing, commenting) as meaningful political participation? • What kinds of political performance are best suited to the hybrid, potentially global, political information environment? And how do we analyse their effectiveness? • Are the most visible and visual elements of mediated politics online representative of our political cultures, or do they offer a distortive perspective? Such questions are best approached with keen attention to representational forms, alongside contexts of production, mediation and consumption. Similarly to the old debates on the spectacle of politics, there is a danger that the current fascination with how images or videos circulate online and their role in mobilizing support leads to a lack of attention given to the actual images as expressive forms: 'Rather remarkably, the systematic study of the structure and the expressive means of the image itself ("image studies") is relatively rarely practised . . . there is a persistent misunderstanding that one can go without insight into the structure of FURTHER READING Dayan, D. (2013). Conquering visibility, conferring visibility: visibility seekers and media performance. International Journal of Communication, 7, 137-153. Finnegan, C.A. and Kang, J. (2004). 'Sighting' the public: iconoclasm and public sphere theory. Quarterly Journal of Speech, 90 (4), 377^102. Gerbaudo, P. (2012). Tweets and the Streets: Social Media and Contemporary Activism. New York: Pluto Press. Khatib, L. (2012). Image Politics in the Middle East: The Role of the Visual in Political Struggle. London: I.B. Tauris. McLagan, M. and McK.ee, Y. (2012). Sensible Politics: The Visual Culture of Nongovernmental Activism. New York: Zone Books. Mitchell, W.J.T. (2005). There are no visual media. Journal of Visual Culture, 4 (2), 257-266. Pels. D. (2003). Aesthetic representation and political style: re-balancing identity and difference in media democracy. In Corner, J. and Pels, D. (eds), Media and the Restyling of Politics (pp. 41-66). London: Sage. van Zoonen, L., Vis, F. and Mihelj, S. (2010). Performing citizenship on YouTube: activism, satire and online debate around the anti-Islam video Fitna. Critical Discourse Studies, 7 (4), 249-262. 430 Handbook of digital politics Visibility and visualities 431 REFERENCES Anderson, B. (1991). Imagined Communities; Reflections on the Origin and Spread of Nationalism. London, UK and New York, USA: Verso. Anstead, N. and O'Loughlin, B. (2011). The emerging Viewertarial and BBC Question Time: television debate and real-time commenting online. International Journal of Pressl Politics, 16 (4), 440-462. Archibald, D. (2011). Photography, the police and protest: images of the G-20, London 2009. In Cottle, S and Lester, L. (eds), Transnational Protests and the Media (pp. 129 139). Oxford: Peter Lang. Askanius, T. (2010). Video Activism 2.0 - spate, place and audiovisual imagery. In Hedling, E., Hedling, O. and Jonsson, M. (eds), Regional Aesthetics: Locating Swedish Media (pp. 337-358). Stockholm: Kungliga Biblioteket. Bal, M. (2003). Visual essentialism and the object of visual culture. Journal of Visual Culture, 2(1), 5-32. Baym, G. and Jones, J. (2012). News parody in international perspective: politics, power and resistance. Popular Communication, 10 (1-2), 2-1 3. BBC (2013). Chicago Sun-Times sacks entire photo department. BBC News US and Canada, 30 May. Retrieved 7 June 2013 from http://www.bbc.co.uk/news/world-us-canada-22723725. Dennett, W.L. (2012). The personalization of politics: political identity, social media, and changing patterns of participation. Annals of the American Academy of Political and Social Science, 644(1), 20-39. Blunder, J.G. and Kavanagh, D. (1999). The third age of political communication: Influences and features. Political Communication, 16 (3), 209-230. Boorstin, D.J. (1962). The Image- New York: Atheneum. Caldwell, L. (2011). The new face of President Asad on YouTube. Arab Media and Society, Issue 13, retrieved from http://www.arabmediasociety.com/index. php?article=776&p=0. Castells, M. (2012). Networks of Outrage and Hope: Social Movements in the Internet Age. Cambridge, LIK and Maiden. MA. USA: Polity. Chadwick, A. (2013). The Hybrid Media System: Politics and Power. Oxford: Oxford University Press. Coleman, S. and Blunder, J.G. (2009). The Internet and Democratic Citizenship: Theory, Practice and Policy. Cambridge: Cambridge University Press. Corner, J. (2003). Mediated persona and political culture. In Corner, J. and Pels, D. (eds), Media and the Restyling of Politics (pp. 67-84). London: Sage. Corner, J. and Pels, D. (2003). Media and the Restyling of Politics: Consumerism. Celebrity and Cynicism. London: Sage. Crick, B. (1968). In Defence of Polities. London: Penguin. D'Arcus, ft. (2006). Boundaries of Dissent: Protest and Stale Power in the Media Age. New York: Routledge. Dayan, D. (2013). Conquering visibility, conferring visibility: visibility seekers and media performance. International Journal of Communication, 1, 137 153. DeLuca, K. (1999). Image Politics. New York: Guilford Press. Doerr, N. and Teune, S. (2012). The imagery of power facing the power of imagery: towards a visual analysis of social movements. In Fahlenbrach, K., Klimke, M., Scharloth, J. and Wong, L. (eds), The 'Establishment' Responds: Power and Protest During and After the Cold War (pp. 43 55). Cambridge, UK and New York, USA: Berghahn Books. Du Gay, P., Hall, S., Janes. L., Mackay, H. and Negus, K. (1997). Doing Cultural Studies. London: Sage/The Open University. Edelman, M. (1988). Constructing the Political Spectacle. Chicago, IL: University of Chicago. Finnegan, C.A. and Kang, J. (2004). 'Sighting' the public: iconoclasm and public sphere theory. Quarterly Journal of Speech, 90 (4), 377-402. Flinders, M. (2012). Defending Politics: Why Democracy Matters in the Twenty-First Century. Oxford: Oxford University Press. Gerbaudo, P. (2012). Tweets and the Streets: Social Media and Contemporary Activism. New York: Pluto Press. Green, J.E. (2010). 77ie Eyes of the People: Democracy in an Age of Spectatorship. New York: Oxford University Press. Jay, M. (1993). Downcast Eyes: The Denigration of Vision in Twentieth-Century French Thought. Berkeley, CA: University of California Press. Katz, E. (1988). Disintermediation: cutting out the middleman. Inter Media, 16 (2), 30-31. Khatib, L. (2012). Image Politics in the Middle East: The Role of the Visual in Political Struggle. London: I.B. Tauris. Langer, A.l. (2011). The Personalisation of Politics in the UK: Mediated Leadership from Atllee to Cameron. Manchester: Manchester University Press. Lippmann, W. (1922). Public Opinion. New York: Macmillan. Manin, B. (1997). The Principles of Representative Government. New York: Cambridge University Press. Mazzoleni, G. and Schulz, W. (1999). 'Mediatization' of politics: a challenge for democracy?. Political Communication, 16 (3), 247-261. McChesney, R. (1999). Rich Media, Poor Democracy: Communication Politics in Dubious Times. Urbana, IL: University of Illinois Press. McLagan,M. and McKee, Y. (2012). Sensible Politics: The Visual Culture of Nongovernmental Activism. New York: Zone Books. Meyer. T. and Hinchman, L. (2002). Media Democracy: How the Media Colonize Politics. Cambridge; Polity. Mirzoeff, N. (2011). The Right to Look: A Coumerhistory of Visuality. Durham, NC, USA and London, UK: Duke University Press. Mitchell. W.J.T. (1994). Picture Theory: Essays on Verbal and Visual Representation. Chicago, IL, USA and London, UK: University of Chicago Press. Mitchell, W..I.T. (2002). Showing seeing: a critique of visual culture. Journal of Visual Culture, 1 (2), 165-181. Mitchell. W.J.T. (2005). There are no visual media. Journal of Visual Culture, 4 (2). 257-266. Müller, M.G. and Kappas, A. (2013). Politics Mitt Romney style: Gangnam style as a cross-cultural visual meme - online citizen creativity and the power of digitally facilitated political prosumer participation. Presented at 63rd Annual International Communication Association Conference, 21 June, London. Nanabhay, M. and Farmanfarmaian, R. (2011). From spectacle to spectacular: how physical space, social media and mainstream broadcast amplified the public sphere in Egypt's 'revolution'. Journal of North African Studies, 16 (4), 573-603. Olesen, T. (2013). 'We are all Khalcd Said": visual injustice symbols in the Egyptian revolution, 2010-2011. In Doerr, N., Mattoni, A. and Teune, S. (eds), Advances in the Visual Analysis of Social Movements, Research in Social Movements, Conflicts and Change, Vol. 35 (pp. 3-25). Bingley: Emerald Group. Pauwels, L. (2008). Visual literacy and visual culture: reflections on developing more varied and explicit visual competencies. Open Communication Journal, 2, 79-85. Pels, D. (2003). Aesthetic representation and political style: re-balancmg identity and difference in media democracy. In Corner, J. and Pels, D. (eds), Media and the Restyling of Politics (pp. 41-66). London: Sage. Postman, N. (1987). Amusing Ourselves to Death. London: Methuen. Putnam, R.D. (2000). Bowling Alone: The Collapse and Revival of American Community. New York: Simon & Schuster. Ranciere, J. (2011). The Emancipated Spectator. London, UK and New York, USA: Verso. Richardson K., Parry, K. and Corner, J. (2012). Political Culture and Media Genre: Beyond the News. Basingstoke: Palgrave. Ries. B. (2012). The story behind the most viral photo ever. Daily Beast, 19 November. Retrieved from http://vnvw.thedailybeast.eom/articlcs/20l2/l 1/19/the-story-behind-the- most-viral-photo-ever.html. Shafak. R. (2013). The view from Taksim Square. Guardian, 4 June. G2 supplement, pp.6-8. 432 Handbook of digital politics Shaw, F. (2013). 'Walls of seeing': protest surveillance, embodied boundaries, and counter-surveillance at Occupy Sydney. Transformations, 23. Retrieved from http://www. trans forma tionsjournal.org/journal/issuc_23/article_04.shtml. Shifman, L. (2012). An anatomy of a YouTube meme. New Media and Society. 14 (2), 187 203. Stanyer, J. (2012). Intimate Politics: The Rise of the Celebrity Politician and the Decline of Privacy. Cambridge: Polity. Street, J. (1997). Politics and Popular Culture. Philadelphia. PA: Temple University Press. Stromback, J. (2008). Four phases of mediatization: an analysis of the mediatization of politics. International Journal of Pressl Politics, 13 (3), 228-246. Thompson, J.B. (1995). The Media and Modernity. Stanford, CA: Stanford University Press. Thompson, J.B. (2005). The new visibility. Theory, Culture and Society, 22 (6), 31-51. Thumim, N. (2012). Self-Representation and Digital Culture. Basingstoke: Palgrave. Thussu, D.K. (2007). News as Entertainment: The Rise of Global Infotainment. London: Sage. Tyler, I. (2013). Revolting Subjects. London, UK and New York, USA: Zed Books. van Zoonen, L. (2005). Entertaining the Citizen: When Politics and Popular Culture Converge. Lanham: Rowman & Littlefield. van Zoonen, L., Vis, F. and Mihelj, S. (2010). Performing citizenship on YouTube: activism, satire and online debate around the anti-Islam video Fitna. Critical Discourse Studies, 7 (4), 249-262. Vis, F., Faulkner. S., Parry, K., Manyukhina, Y. and Evans, L. (2013). Twitpic-ing the riots: analysing images shared on Twitter during the 2011 UK. riots. In Weller, K., Bruns, A., Puschmann, C, Burgess, J. and Mahrt, M. (eds), Twitter and Society (pp. 385-398). New York: Peter Lang. 24. Automated content analysis of online political communication Ross Petchler and Sandra Gonzalez-Bailon INTRODUCTION Content analysis has a long tradition in the social sciences. It is central to the study of policy preferences (Budge, 2001; Laver et al., 2003), propaganda and mass media (Krippendorff. 2013 [1980]; Krippendorff and Bock, 2008), and social movements (Delia Porta and Diani, 2006; Johnston and Noakes, 2005). New computational tools and the increasing availability of digitized documents promise to push forward this line of inquiry by reducing the costs of manual annotation and enabling the analysis of large-scale corpora. In particular, the automated analysis of online political communication may yield insights into political sentiment which offline opinion analysis instruments (such as polls) fail to capture. Online communication is constantly pulsating, generating data that can help us uncover the mechanisms of opinion formation - if the appropriate measurement and validity methods are developed. Several linguistic peculiarities distinguish online political communication from traditional political texts. For a start, it is often far less formal and structured. In addition, automated content analysis techniques are not always as reliable or as valid as manual annotation, which makes measurements potentially noisy or misleading. With these challenges in mind, we provide an overview of techniques suited to two common content analysis tasks: classifying documents into known categories, and discovering unknown categories from documents (Liu, 2012; Blci, 2013). This second task is more exploratory in nature: it helps to identify topic domains when there are no clear preconceptions of the topics that are discussed in a certain communication environment. The first task, on the other hand, can help to label a large volume of text in a more efficient manner than manual annotation; for instance, when the research question requires identifying the emotional tone of political communication (as positive, negative or neutral) or its ideological slant (liberal or conservative). This chapter focuses on the application of these automated techniques to online political communication, and suggests directions for future research in this domain. 433 434 Handbook of digital politics METHODS FOR AUTOMATED CONTENT ANALYSIS The application of automated text analysis techniques requires the prior acquisition and preprocessing of data. This section discusses the logic of preprocessing texts to then provide an overview of techniques to classify documents in known categories or discover topics when no categories are known. Acquiring and Preprocessing Online Political Texts Political scientists have applied automated content analysis techniques to many kinds of offline political texts, including newspaper articles (Young and Soroka, 2012), presidential and legislator statements (Grimmer and King, 2011), legislature floor speeches (Quinn et al., 2010), and treaties (Spirling, 2012). Until recently, though, political texts remained relatively understudied because they were difficult to parse and process for analysis. Acquiring online political texts is becoming simpler as more sites store and transmit them in machine-readable formats such as Extensible Markup Language (XML) and JavaScript Object Notation (JSON) or make them publicly available via application programming interfaces (APIs). When such options are unavailable, researchers familiar with statistical software or scripting languages can use new packages for automated HyperText Markup Language (HTML) scraping (Python and R, for instance, have built-in packages and libraries). Finally, when neither machine-readable nor easy-to-parse HTML data are available, researchers can crowdsource data acquisition and parsing via sites like Amazon Mechanical Turk (Berinsky et al., 2012). Overall, these new technologies enable communication scholars to access and study previously unavailable indicators of public opinion. In order to perform automated content analysis researchers must transform texts into structured data that can be quantified (Franzosi, 2004). Prior to the advent of new computational tools this was performed by human coders using a pre-determined scheme (Krippendorf, 2013 [1980]; Neuendorf, 2001). Initially, a codebook is written guided by a research question and a theoretical context. It is iteratively improved until coders no longer notice ambiguities, at which point it is applied to the data set. Automated approaches preprocess the text to reduce the complexity of language, often using a bag-of-words model to eliminate the most frequent words and to reduce words to their morphological roots (Jurafsky and Martin, 2009; Hopkins and King, 2010; Porter, 1980). After preprocessing, documents are represented as a document-term matrix in which rows Automated content analysis 435 correspond to documents, sentences, or expressions (depending on the unit of analysis), and columns correspond to words or tokens. Cells can contain continuous values (representing how frequently each term occurs in each document) or binary values (representing whether each term occurs in the document). Which of the two approaches is more appropriate (to code documents manually or to apply automated preprocessing) depends on the complexity of the research question at hand, the number of documents collected, and the tolerance for error. Although manually annotated data remain the gold standard for content analysis, the sections that follow focus mostly on cases in which data are automatically preprocessed, since this is more common when dealing with large volumes of text. Once online political texts are converted to a structured form, several methods for automated content analysis can be applied. We divide these methods into two groups to reflect the two most common content analysis tasks: classifying documents into known categories, and discovering theoretically important categories from the content. The former task encompasses techniques such as lexicon-based classification and supervised learning. The latter task encompasses unsupervised learning and the analysis of text as networks of concepts. The following sections explain the details of these techniques and highlight their relative strengths and weaknesses to help communication researchers choose the approach best suited to their specific data and research question. Classifying Documents into Known Categories The goal of supervised content analysis techniques is to classify documents into a number of known categories. For example, news articles may have left-leaning or right-leaning ideological biases (Gcntzkow and Shapiro, 2010) or have positive or negative coverage (Eshbaugh-Soha, 2010). This section offers an overview of the techniques that allow that sort of classification. There are two main methods. The first is a lexicon-based approach, which uses relative keyword frequencies to measure the prevalence of each category in a document. The second is supervised learning, which uses a training data set of manually annotated documents to classify new, unlabeled documents. Lexicon-based classification The lexicon (or dictionary)-based approach to document classification is the simplest automated content analysis technique (Liu, 2012). It is based on a list of words and phrases and their associated category labels. For example, a lexicon for classifying micro-blog posts according to sentiment i 436 Handbook of digital politics Automated content analysis 437 may map the words 'good' and 'beautiful' to the positive category and the words 'bad' and 'ugly' to the negative category. A lexicon for classifying blog posts according to ideological subject may map the words 'healthcare' and 'environment' to the left-leaning category and 'foreign policy' and 'taxes' to the right-leaning category. Off-the-shelf lexicons include the Linguistic Inquiry and Word Count, or LIWC (Pennebaker et al., 2001), and the General Inquirer (Wilson et al., 2005). Not all lexicons are based on binary categories. Some sentiment lexicons have positive, neutral, and negative terms, measured on a several points scale. The Affective Norms for English Words (ANEW) lexicon, for instance, labels words and phrases according to psychometric categories which rate words on three emotional dimensions: valence, arousal, and dominance (Bradley and Lang, 1999; Osgood et al., 1957). This lexicon helps to analyze documents by counting the relative frequency with which words appear and averaging the scores associated to each word in each dimension, from 0 to 9. This approach has been applied effectively to extract sentiment measures from a number of online data sources (Dodds and Danforth, 2009; Dodds et al.. 2011). The success of a lexicon-based content analysis relies on the quality of the lexicon; that is, how appropriate it is in the context of the specific research question and data being analyzed (González-Bailón and Paltoglou, 2015). Using 'off-the-shelf lexicons compiled with generic research goals may produce poor results when applied to specific types of political communication (Loughran and McDonald. 2011). It is always best to generate lexicons specific to a research question, and there are three main approaches for doing so. The first is to manually annotate the sentiment of all adjectives in a dictionary of all the words in a corpus, in line with the information domain under scrutiny (that is, 'warming' can be labeled differently if used in environmental policy or foreign affairs communication). This is time-consuming but tunes the lexicon to specific communication contexts. Researchers concerned with efficiency as well as accuracy have used online crowdsourcing platforms such as CrowdFlower, Amazon Mechanical Turk, and Taskcn to quickly and accurately label large sentiment lexicons. For example, Dodds et al. (2011) created a lexicon of 10222 words by merging the 5000 most frequently occurring words in a Tweet corpus, Google Books, music lyrics, and the New York Times; they then used Amazon Mechanical Turk to obtain 50 sentiment ratings of each word on a nine-point scale from negative to positive. They found that the sentiment lexicon labeled by crowdsourcing workers was highly correlated with the ANEW lexicon. The second way to generate a sentiment lexicon is dictionary-based. The general approach is to manually label the sentiment of a small set of seed words, and then search a dictionary (the most frequently used is WordNet; see Miller et al., 1990) for their synonyms and antonyms; these snowballed terms arc then labeled with the same or opposite sentiment as the corresponding seed word and then are added to the set of seed words. The process is iterated until no words remain unlabeled. For example, the seed word 'excellent' is labeled positive; synonyms such as 'beautiful', 'fabulous', and 'marvelous' arc labeled as positive as well; while antonyms such as 'awful', 'rotten', and 'terrible' are labeled as negative. An example of a lexicon generated using the dictionary-based approach is Sentiment Lexicon, constructed by Hu and Liu (2004). This dictionary-based approach quickly generates a large list of labeled sentiment words, but requires manual cleaning and ignores ambiguity due to context, which is particularly important in the analysis of political communication. The third way to generate a sentiment lexicon is corpus-based. The general approach is lo manually label the sentiment of a small set of seed words and then define linguistic rules to identify similar or dissimilar sentiment words. A seed word may be 'beautiful' and its label 'positive'; linguistic rules based on connective words (such as 'and' or 'but') help to assign labels to subsequent words. For instance, if a document in a corpus contains the phrase 'The car is beautiful and spacious' then the term 'spacious' could be assigned the label 'positive' based on the connective word 'and'. Conversely, if a document in a corpus contains the phrase 'The car is spacious but difficult to drive' then the term 'spacious' could be assigned the label 'negative' based on the connective word 'but'. This methodology requires clearly defined linguistic rules in order to achieve good results; and linguistic rules assume sentiment consistency across documents, which is not necessarily the case for most empirical domains: the same word can express opposite sentiments in different communication contexts (Liu, 2012). Overall, though, the corpus-based methodology to lexicon generation is useful in two cases: to discover other sentiment words and their orientations on the basis of a hand-made seed list; and to adapt a general-purpose lexicon to a specific communication domain. The corpus-based approach is less useful for building a general-purpose sentiment lexicon than the dictionary-based approach because dictionaries encompass more words. These three techniques are based on different assumptions that affect the results they produce. None of these sentiment lexicons is perfect because they are too general to suit the specific needs of different communication domains. In addition, certain words and phrases in online political communication are too informal, specific, or novel (and therefore infrequent) to be contained in existing lexicons. A corpus-based technique can capture and label these distinct words; for instance. Brody and Diakopoulos (2011) 438 Handbook of digital politics Automated content analysis 439 find that lengthened words in microblog posts (for example, 'looove') are strongly associated with subjectivity and sentiment; and Derks et al. (2007) find that emoticons (for example, ':)') strengthen the intensity of online communication. Researchers have already incorporated the peculiarities of online communication into their sentiment models (Paltoglou et al, 2010; Paltoglou and Thelwall, 2012), but often additional manual labeling is needed to add other novel words to the seed list. These limitations make validation a crucial component of automated content analysis (Grimmer and Stewart, 2013). Having the appropriate validation strategies in place is necessary to increase confidence in measurement. Supervised learning The second main approach to document classification using pre-existing categories is supervised learning. Supervised algorithms learn from a training data set of manually annotated documents how to classify new, unlabeled documents. The supervised learning approach has three steps. First, it constructs a training data set. Second, it applies an automated algorithm to determine the relationships between features of the training data set and the categories that are used to classify documents. And third, it predicts (or assigns) categories for unlabeled documents and validates that classification. The remainder of this section reviews these three steps in turn. The first step in supervised learning is to construct a training data set. As described above, this involves transforming unstructured textual data into structured quantitative data. In addition to preprocessing, it is common for researchers to manually code documents for features that the bag-of-words model ignores; for instance, they may add features accounting for the source or the author of a document. The larger the training data set, the more information supervised learning algorithms have with which to make predictions, but scaling up can be computationally costly. The specific research question and data source inform the balance between the need for a large training data set and the costs of compiling training data. The second step in supervised learning is to apply an algorithm that will associate text features to each category in the classification scheme. There are many different algorithms and the field of machine learning and natural language processing is quickly growing in this area; Hastie et al. (2009) offer a good overview of the techniques available. Each model has specific characteristics and parameters, which makes a general discussion difficult, but popular algorithms include (multinomial) logistic regression, the naive Bayes classifier (Maron and Kuhns, 1960), random forests (Breiman, 2001), support vector machines (Cortes and Vapnik, 1995), and neural networks (Bishop, 1995). Each of these algorithms uses the information gathered from the training data to assign new examples of text into the classification categories. This assignment takes place in the third and final step, where supervised approaches predict the categories for unlabeled documents and validate the results. A model that performs well will replicate the results of manual coding, which still offers the gold standard; a model that performs poorly will fail to replicate Ihese results. The standard method to validate models is cross-validation. This entails splitting the labeled documents into equally sized groups (usually about ten) and then predicting the categories of the observations in each group using the pooled observations in the other groups. This method avoids overfitting to data because it focuses on out-of-sample prediction. Overall, the supervised approach systematically performs better than unsupervised approaches in the analysis of online communication because it is able to capture more accurately the contextual features of the text and language used (Gonzalcz-Bailon and Paltoglou, 2015). Discovering Categories and Topics from Documents Unsupervised learning In contrast to supervised approaches, unsupervised techniques do not require manually annotated training data; consequently, they are much less costly to implement. They are good exploratory techniques but their results can be difficult to evaluate: concepts such as validity and consistency compared to human labeling do not immediately apply because these techniques are used, in part, to overcome the lack of predefined labels or categories - hence their exploratory nature. This section briefly discusses three categories of unsupervised techniques: cluster analysis, dimensionality reduction, and topic modeling. The goal of cluster analysis is to partition a corpus of documents into groups of similar documents, where 'similar' is measured in terms of word frequency distributions. The most widely used clustering algorithm is /c-means (MacQueen, 1967), which partitions documents into k disjoint groups by minimizing the sum of the squared Euclidean distances within clusters; distance is measured as the number of words that any two documents share. Other clustering algorithms use different distance metrics or objective functions (which are used to optimize or find the best clustering classification out of all possible classifications). Given that few papers provide guidance on which similarity metrics, objective functions, or optimization algorithms to choose, Grimmer and Stewart (2013) caution social scientists from importing clustering methods developed in other, more technical fields like machine learning. The computer-assisted cluster 440 Handbook of digital politics Automated content analysis 441 analysis technique suggested by Grimmer and King (2011) offers a more intuitive tool for the task of fully automated cluster analysis. The goal of dimensionality reduction is to shorten the number of terms in the term-document space while maintaining the structure of the corpus. One dimensionality reduction technique is principal component analysis, which transforms a document-term matrix into linearly uncorrelated variables that correspond to the latent semantic topics in the data set. The technique is not different from more conventional uses in multivariate modeling where a subset of variables are selected to represent a larger data set (Dunteman, 1989). A related dimensionality reduction technique is multidimensional scaling, which projects a corpus of documents into TV-dimensional space such that the distances between documents correspond to dissimilarities between them. These methods provide good intuition of the topics that characterize a corpus of text but are best used as exploratory techniques; principal component analysis, in particular, is a typical data reduction step performed prior to subsequent, more substantive analysis. Finally, the goal of topic modeling is to represent each document as a mixture of topics. Each topic is a probability mass function over words that reflect a distinct information domain. For instance, the topic 'foreign policy' may assign high probabilities to words such as 'war', 'treaty', and 'Iraq'; while the topic 'economy' may assign high probabilities to words such as 'unemployment', 'GDP', and 'labor'. The most widely used topic model is called latent Dirichlet allocation (LDA) (Blei et al., 2003). This technique has recently been applied to the analysis of newspaper content to dissect the framing of policies (DiMaggio et al., 2013). The method provides a new computational lens into the structure of texts and, as the authors state: finding the right lens is different than evaluating a statistical model based on a population sample. The point is not to correctly estimate population parameters, but to identify the lens through which one can see the data more clearly. Just as different lenses may be more appropriate for long-distance or middle-range vision, different models may be more appropriate depending on the analyst's substantive focus, (ibid.; 20) Again, the crucial step in the analysis comes with validation; that is, with the substantive interpretation of the themes identified. Network Representations of Text As the sections above have illustrated, content analysis is essentially a relational exercise: words that relate to the same topic are associated by co-appearing frequently in documents and they tend to cluster; likewise, positive words tend to be connected to other positive words, and as shown above, language connectors might change the affective tone of words by setting them in a different linguistic context. Networks offer a mathematical representation of the relational nature of language, and provide yet another tool for the analysis of its structure. Networks have been used to model narratives, and to analyze identity formation (Bearman and Stovel, 2000); to represent mental models (Carley and Palmquist, 1992); and to map semantic associations (Borge-Holthoefer and Arenas, 2010). A network approach has also been used with Twitter data to identify entities by looking at the co-occurrence of words and the clusters that emerge from those connections (Mathiesen et al., 2012). The nodes in these networks are words; what changes depending on the approach is the definition of the links that connect those words: co-occurrence is one of the options, but links can also be used to track the temporal evolution of narratives, as when political movements change their framing or candidates change their positions during an election campaign. These networks can be constructed and visualized using standard network analysis tools. One of the by-products of generating a dictionary-based lexicon (discussed above) is that the method also creates a network of words that researchers can use to label the strength as well as the sign of the sentiment expressed. For example, Kamps et al. (2004) determined the strength and sentiment of words according to their distances in WordNet from labeled seed words; in this case, two words are linked if they are synonyms, and distance is measured as the number of links that need to be crossed to go from one word to another. Blair-Goldensohn et al. (2008) also used WordNet to construct a network of positive, negative, and neutral sentiment words, and then labeled the strength of the words using a propagation algorithm: starting from a seed word, its sign (positive, negative, or neutral) is propagated to all its neighboring words in the network (its synonyms); following a majority rule, that sign is further propagated to the neighbors of the neighbors, and so on, recursively, until all words have a sign assigned - the valence of which gets weaker the further apart the word is from its seed. These network-based techniques help to extend the dictionary-based approach by suggesting measures of sentiment strength. APPLICATIONS TO THE ANALYSIS OF ONLINE POLITICAL COMMUNICATION Sentiment in Online Political Talk When applying sentiment analysis to political communication, it is important to remember that different methods inherit different assumptions 442 Handbook of digital politics from psychological theories of emotions. The ANEW lexicon, for instance, derives from now classic psychological research suggesting that three dimensions account for variance in the expression of emotion: valence (which ranges from pleasant to unpleasant), arousal (which ranges from calm to excited), and dominance (which ranges from domination to control; Osgood et al., 1957). Neurological research, on the other hand, suggests that five emotional dimensions underlie most brain activity: fear, disgust, anger, happiness, and sadness (Murphy et al., 2003). Reducing the breadth of human emotions to just a few dimensions is arguably a crude simplification, but necessary to make problems tractable; however, it also introduces measurement error that has to be taken into consideration when operationalizing research questions about the affective tone of political communication. Sentiment analysis of online political communication must take into account not only measurement error but also sampling bias. Internet users, and in particular those present in social media, are typically not representative of the population: they tend to be female, young, and urban (Duggan and Brenner, 2013); in addition, the bias might be more or less important depending on the context and subject of communication. For some dimensions of public opinion, the bias might not matter, but for others it can be crucial. Again, it is only through validity tests that the measures of public opinion extracted from online communication can be relied upon (Grimmer and Stewart, 2013). The increasing number of Internet users who join social media sites and discuss politics means that the volume of online political communication is growing, and the profile of users involved is changing. Analyses of how online sentiment changes over time must therefore account for these non-stationary characteristics, typically by comparing short, adjacent periods of online communication rather than the entire history of communication on a given site. The assumptions made by automated methods about emotional mechanisms and the nature of the samples analyzed demand a thoughtful research design when studying online communication. In many cases basic methods produce useful results that rival more sophisticated approaches; in particular, simple word frequencies and analysis of how the volume of communication fluctuates over time often yield good insights while preserving efficiency. Carvalho et al. (2011), for instance, found that in some cases these basic descriptive statistics predict sentiment as accurately as more advanced statistical techniques. This suggests that exploratory analysis can be crucial to avoid rushing into the implementation of more complex solutions when a simpler, more intuitive approach would perform as well. In addition to the lexicons introduced above, a number of alternative Automated content analysis 443 approaches have also been developed to facilitate the study of online communication. These include OpinionFinder (OF), which rates expressions as strongly or weakly subjective (Wilson et al., 2005); and the Profile of Mood States (POMS) questionnaire (Lorr et al., 2003), in which respondents rate each of 65 adjectives on a five-point scale. The questionnaire produces emotion scores in six dimensions: Tension-Anxiety, Anger-Hostility, Fatigue-Inertia, Depression-Dejection. Vigor-Activity, and Confusion-Bewilderment. Like ANEW, the POMS lexicon is suited for analyzing more complex emotions in online communication; the OF lexicon, like LIWC, is used for simpler tasks such as the identification of polarity in sentiment analysis. Other prominent lexicons optimized for the analysis of online communication include SentiWordNet (Adrea and Sebastiani, 2006) and SentiStrength (Thelwall et al., 2010). SentiStrength is particularly useful for online political communication because it includes misspellings and emoticons which abound in online talk. Recent empirical applications of these approaches include Connor et al. (2010), Bollen et al. (2011) and Castillo et al. (2013). Connor et al. (2010) derive sentiment valence from Twitter posts using a subjectivity lexicon based on a two-step polarity classification. They compare Twitter sentiment to consumer confidence and election polling data. They find high correlations (between 0.7 and 0.8) and evidence that smoothed Twitter sentiment predicts consumer confidence (but not election) poll results with relatively high accuracy. However, Bollen et al. (2011) find that the intersection of a tweet corpus and their subjectivity lexicon is not a good leading indicator of the direction of shifts in the Dow Jones Industrial Average. This highlights how sentiment analysis of online communication may not work in all contexts: some lexicons are better suited to particular problem domains, such as consumer confidence, but not financial markets. Finally, Castillo et al. (2013) apply the SentiStrength lexicon to measure sentiment in cable news coverage; although this is traditional media content, the data were accessed through a software company that develops applications for smartphones and tablets that display extra information about TV shows, including captions of content. Unsupervised Learning Applications As explained above, many unsupervised learning methods are used as exploratory tools rather than testing techniques, and thus are less common in published literature on online communication. Nevertheless, a few-prominent examples exist, although many arc still peripheral to the core research questions of political communication. Turney (2002), for instance, classifies online reviews as positive or 444 Handbook of digital politics Automated content analysis 445 negative by estimating the semantic orientation of sentences containing adjectives or adverbs. Specifically, the paper makes use of the pointwise mutual information-information retrieval (PMI-IR) algorithm to measure the number of co-occurrences between words and the seed words 'excellent' and 'poor' on AltaVista search engine results. This co-occurrence frequency determines the semantic orientation of words, and thus can be used to rate online reviews as positive or negative. Quinn et al. (2010) use a technique similar to LDA in order to analyze the daily legislative attention given to various topics in 118000 United States Senate floor speeches from 1997 to 2004. They found 42 topics, the most prominent being legislative procedures, armed forces, social welfare, environment, and commercial infrastructure. Yano et al. (2009) use LDA in order to model topics in political blog posts and their corresponding comments sections. They found five topics: religion, (election) primary, Iraq War, energy, and domestic policy. Associated with each topic are a set of words that appeared in blog posts and a set of words that appeared in comments. Additionally, the authors predict which users are likely to comment on which blogs. Finally, another recent example applies the same method to the analysis of issue salience in the Russian blogosphere (Kolstova and Koltcov, 2013). The authors use the method to identify a shift in topics during the political protests that took place during the parliamentary and presidential elections in late 2011 and early 2012. In sum, unsupervised methods are less frequently used because they are exploratory techniques employed to charter communication domains that lack predefined boundaries. They are good for estimating the structure of a corpus of text when no a priori classifications exist, but they still require a posteriori theoretical and subjective labeling of categories. This stands in contrast to supervised techniques: whereas manual annotation is the starting point for supervised techniques, it is the ending point for the unsupervised approach. FUTURE LINES OF WORK This chapter has given an overview of techniques for the automated analysis of large-scale texts, especially as they are generated in online communication. Although this is a massive area of research, and is fast evolving, a few facts have already been established. One is the consistent evidence that the effectiveness of automated classifiers is not independent from the communication domain being analyzed: the meaning of words or their emotional load varies with the context in which they are used. More work is required to build tailored dictionaries that can capture the nuances of political communication as it takes place in different information contexts; for this, supervised-learning approaches offer the most accurate (and promising) solutions. Likewise, more work is needed to consolidate validation strategies, for instance by measuring the strength to which online measures of public opinion are correlated with more traditional measures, such as polls and surveys. A more systematic account of the efficiency and robustness of different algorithms is also needed: some corpora of text are better analyzed by certain techniques than others. Supervised methods, for instance, are more appropriate for content expressed in Twitter messages, whereas for longer communication, such as blog entries, unsupervised methods might be more appropriate. More research is needed to assess the robustness of each method for different data sources, as facilitated by-online communication. In any case, the appropriateness of each technique has to be assessed in the light of each particular research question. Validation is a crucial step in the application of automated content analysis, and this implies finding ways of assessing the accuracy, precision, and reliability of automated classifiers as compared to human coding. For instance, researchers who choose a lexicon-based approach face several design considerations. The first is what type of lexicon to generate or adapt. Some sentiment lexicons have binary categories (positive and negative), some have ternary categories (positive, negative, and neutral), and some have ordinal categories (-5 to +5, for example). A second design consideration is what word features to include in a lexicon. Some lexicons simply have word valence (ranging from positive to negative), while others have additional features such as arousal (ranging from calm to excited). The type of sentiment lexicon a researcher chooses should be based on the features a lexicon offers and the specific research question they seek to answer. Researchers who choose a lexicon-based approach also face several implementation considerations. Most of these have to do with how to detect and resolve the complexities of text. The algorithm that implements a lexicon-based approach should often not just naively match words but also be sensitive to their local context. For instance, it should be aware of negating words (such as 'no', 'not', and 'none') and strengthening punctuation (such as exclamation marks, question marks, and ellipses). Some of the lexicons revised in this chapter, such as ScntiStrength, already take these language modifiers and intensifiers into account. Good lexicon-based approaches to sentiment detection do not just rely on word matching: they are also sensitive to how the local context of each word affects the overall sentiment. The advantage of automated content analysis is that it helps to scale up the amount of text analyzed by lowering the costs of coding and the 446 Handbook of digital politics efficiency of document classification; but it still needs to be reliable. Many sentiment lexicons are based on psychological theories of language use but it is still unclear whether these psychometric instruments work for written communication and large-scale text analysis. In addition, these techniques are still not very good at capturing essential features of political talk, such as sarcasm. A document may contain many strong sentiment words but the author might actually have intended the opposite sentiment to that captured by the automated approach. This means that automated methods might be more appropriate when applied to text in which sarcasm and figurative language are rarely used, for instance news reports; communication through social media, on the other hand, might be more vulnerable to measurement error. As the tools for automated content analysis become more prevalent in communication research, more unified standards for evaluation and assessment will have to be consolidated. The advantages of automated methods are. overall, too great to dismiss. ACKNOWLEDGEMENTS We would like to thank the participants of the workshop 'Extracting Public Opinion Indicators from Online Communication', sponsored by the Oxford John Fell Fund under project 113/365 while the authors were based at the University of Oxford. We arc especially grateful to Scott Blinder, Javier Borge-IIolthoefer, Andreas Kaltenbrunner, Patrick McSharry, Karo Moilanen, and Georgios Paltoglou for insightful discussions. LEARNING MORE Methods for automated content analysis are fast evolving, and any list of available resources is likely to be soon outdated. What follows are a few recommendations on where to start to learn more about the methods and applications of automated tools. Rather than an exhaustive list, these references offer entry points to what is a vast and quickly expanding area of research. FURTHER READING KrippendortY (2013 [1980]). Now in its third edition, this book is a classic in content analysis, a long-standing reference that precedes the explosion of automated methods for the Automated content analysis 447 analysis of large-scale data. Even though the book does not consider emerging methods, the discussion on validity and reliability slili applies. Liu (2012). This monograph is one of the most up-to-date reviews of opinion mining methods. It offers an accessible discussion of state-of-the art tools for automated content analysis, and it defines basic terminology as well as research standards. Dáubler et al. (2012). This research note oilers an interesting comparison of the validity of automated versus human coding in identifying basic units of text analysis. The discussion considers how automated methods offer an improvement to human coding schemes without loss of validity. Grimmer and Stewart (2013). This article offers an interesting overview of methods that analyze text at the document level. In addition to discussing in an accessible way the basic features of different approaches, the article also emphasizes the need to develop new validation methods. Dodds and Danforth (2009). One of the first examples that used unsupervised methods to extrapolate opinion measures from large-scale communication. It offers a good schematic example of how unsupervised methods work, and how il can be applied to several data sols to track aggregated sentiment dynamics. Tools for Content Analysis • R packages: - ReadMe: hllp://gking.harvard.edu/readme - TextMining: http://cran.r-project.org/web/packages/tm/vignet tes/tm.pdf LDA Topic Modeling: http://www.cs.princeton.edu/~blei/topic modeling.html - TextTools: http://www.rtexltools.com/ • Other software: - LexiCoder: http://www.lexicoder.com - SentiStrcngth: http://sentistrength.wiv.ac.uk LIWC: http://www.liwc.net. REFERENCES Adrca, E. and Sebastiani, F. (2006). SentiWordNet: a publicly available lexical resource for opinion mining. Paper presented at the 5th Conference on Language Resources and Evaluation, Genoa, Italy. Bcarman, P. and Stovek K. (2000). Becoming a Nazi: a model for narrative networks. Poetics, 27(1), 69-90. Berinsky, A.J., Iluber, G.A. and Lenz, G.S. (2012). Evaluating online labor markets for experimental research: Amazon.com's Mechanical Turk. Political Analysis, 20(3), 351-368. doi: 10.1093/pan/mpr057. Bishop, CM. 11995). Neural Networks for Pattern Recognition. New York: Oxford University Press. Rlair-Goldensohn, S., Hannan, K., McDonald, R.. Neylon, T., Reis, G.A. and Reynar, J. (2008). Building a sentiment suminarizer for local service reviews. WWW Workshop on NLP in the Information Explosion Era, Beijing. Blei, D. (2013). Topic modeling and digital humanities. Journal of Digital Humanities, 2(1). 448 Handbook of digital politics Blci, D.. Ng, A. and Jordan, M. (2003). Latent dirichlet allocation. Journal of Machine Learning and Research. 3, 993 1022. Bollen, J., Mao, H. and Zeng, X.-J. (2011). Twitter mood predicts the stock market. Journal of Computational Science, 2(1), 1-8. Borge-Holthoefer, J. and Arenas, A. (2010). Semantic networks: structure and dynamics. Entropy, 12(5). 1264-1302. Bradley. M,M. and Lang, P.J. (1999). Affective Norms for English Words (ANEW): Instruction Manual and Affective Ratings. Gainesville, FL. Brciman, L. (2001). Random forests. Machine Learning, 45(1), 5-32. doi: I0.1023/A:10109334O4324. Brody, S. and Diakopoulos, N. (2011). Cooooooooooooooollllllllllllll!!!!!!!!!!!!!!: using word lengthening to detect sentiment in microblogs. Paper presented at the Conference on Empirical Methods in Natural Language Processing, Stroudsburg, PA. Budge, 1. (cd.). (2001). Mapping Policy Preferences. Estimates for Parties, Electors and Governments 1945 1998. Oxford: Oxford LIniversity Press. Carley, K. and Palmquist, M. (1992). Extracting, representing, and analyzing mental models. Social Forces, 70(3). 601-637. doi: 10.2307/2579746. Carvalho, P., Sarmento, L., Teixeira, J. and Silva, M.J. (2011). Liars and saviors in a sentiment annotated corpus of cornmems to political debates. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: Association for Computational Linguistics. Castillo. C, dc Francisci Morales, G.. Mendoza, M. and Khan, N. (2013). Says who? Automatic text-based content analysis of television news. Paper presented at the MNLP Workshop. ACM International Conference on Information and Knowledge Management, San Francisco, CA. Connor, B.O., Balasubramanyan, R., Routledgc, B.R. and Smith, N.A, (2010). From Tweets to polls: linking text sentiment to public opinion time series. Fourth International AAA! Conference on Wehlogs and Social Media. Menlo Park, CA: AAAI. Corles. C. and Vapnik, V. (1995). Support-vector networks. Machine Learning. 20(3), 273 297. doi: 10.1023/A: 1022627411411. Daubler, T., Bcnoil, K.. Mikhaylov, S. and Laver, M. (2012). Natural sentences as valid units for coded political texts. British Journal of Political Science, 42(04), 937-951. Delia Porta, D. and Diani, M. (2006). Social Movements: An Introduction. 2nd edn. London: Wiley-Blackwell. Oerks, D., Bos, A.E.R. and von Grumbkow. J. (2007). Emoticons and online message interpretation. Social Science Computer Review, 26, 379-388. doi: 10.1177/0894439307311611. DiMaggio, P., Nag, M. and Blei, D. (2013). Exploiting affinities between topic modeling and the sociological perspective on culture: application to newspaper coverage of government arts funding in the US. Poetics, 41(6), 570-606. Dodds. P.S. and Danforth, CM. (2009). Measuring the happiness of large-scale written expression: songs, blogs, and presidents. Journal of Happiness Studies, DOI: 10.1007/ S10902-009-9150-9. Dodds, P.S., Harris, K.D., Kloumann, I.M., Bliss, CA. and Danforth, CM. (2011). Temporal patterns of happiness and information in a global social network: hedonomct- rics and Twitter. PloS ONE, 6(12), e26752. doi: 10.1371/journal.pone.0026752. Duggan, M. and Brenner, J, (2013). The demographics of social media users. http://www. pewinternet.org/Reports/2013/Social-mcdia-users.aspx. Dunteman, G.H. (1989). Principal Components Analysis. London: Sage. Eshbaugh-Soha, M. (2010). The tone of local presidential news coverage. Political Communication. 27(2), 121-140. doi: 10.1080/10584600903502623. Franzosi.R. (2004). From Words to Numbers. Narrative, Data, and Social Science. Cambridge: Cambridge University Press. Gentzkow, M. and Shapiro, J.M. (2010). What drives media slant? Evidence from US daily newspapers. Econometrica, 78(1), 35-71. Gonzalez-Bailon, S., Banchs, R.E. and Kaltcnbrunner. A. (2012). Emotions, public opinion Automated content analysis 449 and US presidential approval rates: a 5-ycar analysis of online political discussions. Human Communication Research, 38, 121-143. Gonzalez-Bailon, S., Paltoglou, G. (2015). Signals of public opinion in online communication: a comparison of methods and data sources, The Annals of the American Academy of Political and Social Science, in press. Grimmer, J. and King G., (2011). General purpose computer-assisted clustering and conceptualization. Proceedings of the National Academy of Sciences, 108(7), 2643-2650. Grimmer, J. and Stewart, B. (2013). Text as data: the promise and pitfalls of automatic content analysis methods for political texts. Political Analysis, 21(3), 267-297. Hastie, T., Tibshirani, R. and Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edn. New York: Springer. Hopkins. Daniel and King, Gary (2010). Extracting systematic social science meaning from text. American Journal of Political Science, 54(1), 229-247. Hu. M. and Liu, B. (2004). Mining and summarizing customer reviews. Proceedings of A CM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD- 2004). New York City: ACM Press. Johnston. H. and Noakes, J.A. (eds) (2005). Frames of Protest. Social Movements and the Framing Perspective. Lanhara, MD: Rowman & Littlefield. Jurafsky, Dan and Martin, James (2009). Speech and Natural Language Processing: An Introduction to Natural Language Processing. Computational Linguistics, and Speech Recognition. Upper Saddle River, NJ: Prentice Hall. Kamps, J., Marx, M., Mokken. R.J. and de Rijke, M. (2004). Using WordNet to measure semantic orientations of adjectives. Paper presented at the LREC, http://dblp.uni-trier.de/ db/conf/lrec/lrec2004.html#KampsMMR04. Kolstova, O. and Koltcov, S. (2013). Mapping the public agenda with topic modelling: the case of the Russian LivcJournal. Policy and Internet, 5(2), 207-227. Krippendorff. K. (2013 [1980]). Content Analysis. An Introduction to its Methodology. Los Angeles, CA: Sage. Krippendorff, K. and Bock, M.A. (eds) (2008). The Content Analysis Reader. Thousand Oaks, CA: Sage. Laver, M., Benoit, K. and Garry, J. (2003). Extracting policy positions from political texts using words as data. American Political Science Review, 97(2), 311 331. Liu, B. (2012). Sentiment Analysis and Opinion Mining. Chicago, IL: Morgan & Claypool. Lorr, M., McNair, D.M., Heuchei t. J.W.P, and Droppleman, L.F. (2003). POMS: Profile of Mood States. Toronto: MHS. Loughran, T. and McDonald, B. (2011). When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks. Journal of Finance, 66(1), 35 65. MacQueen, J. 1967. Some methods for classification and analysis of multivariate observations. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, 1: 281 297. London: Cambridge University Press. Maron, M.E. and Kuhns, J.L. (1960). On relevance, probabilistic indexing and information retrieval. Journal of the ACM, 7(3), 216 244. doi: 10.1145/321033.321035. Malhiesen, J., Yde, P. and Jensen, M.H. (2012). Modular networks of word correlations on Twitter. Scientific Reports, 2. Miller, G.A., Beckwith, R., Fellbaum, C, Gross, D. and Miller, K.J. (1990). Introduction to wordnet: an on-line lexical database. International Journal of Lexicography, 3(4), 235-244. Murphy, F.C., Nimmo-Smith, I. and Lawrence, AD. (2003). Functional neuroanatomy of emotions: a meta-analysis. Cognitive, Affective and Behavioral Neuroscience, 3, 207-233. Neuendorf, K. (2001). The Content Analysis Guidebook. London: Sage. Osgood, C.E., Suci, G.J. and Tannenbaum, P.H. (1957). The Measurement of Meaning, Vol. 47. Urbana, IL, University of Illinois Press. Palloglou. G., Gobron, S., Skowron. M.. Thelwall, M. and Thalmann, D. (2010). Sentiment analysis of informal textual communication in cyberspace. Proc. ENGAGE. 13-23. Paltoglou. G. and Thelwall. M. (2012). Twitter, MySpace. Digg: unsupervised sentiment 450 Handbook of digital politics analysis in social media. ACM Transaction!: on Intelligent Systems and Technology (TIST), 3(4):66: 1-66: 19. Pennebaker. J.W.. Booth. R.J. and Francis. ME. (2001). Linguistic Inquiry and Word Count: UWC. Mahwah, NJ: Erlbaum Publishers. Porter, M. (1980). An algorithm for suffix stripping. Program, 14(3), 130-137. Quinn, K.M., Monroe, B.L.. Colaresi, M.. Crespin, M.H. and Radev, DR. (2010). How to analyze political attention with minimal assumptions and costs. American Journal of Political Science, 54(1), 209 228. Spirling, A. (2012). US treaty making with American indians: institutional change and relative power. 1784-1911. American Journal of Political Science, 56(1), 84-97. doi: 10.1111 /j.l 540-5907.2011.00558.x. Thclwall, M., Buckley, K. and Paltoglou, G. (2010). Sentiment strength detection in short informal text. Journal of the American Society for Information Science and Technology, 61(12). 2544-2558. I'urney, P.D. (2002). rhumbs up or thumbs down? Semantic orientation applied to unsupervised classification of reviews. Paper presented at the Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. Philadelphia, PA. Wilson, T., Hoffmann, P., Somasundaran. S.. Kessler, J., Wiebe, J., Choi, Y.....Patwardhan, S. (2005). OpinionFinder: a system for subjectivity analysis. Proceedings of HLTIEMNLP on Interactive Demonstrations, Vancouver, British Columbia. Canada. Yano, T., Cohen, W.W. and Smith. N.A. (2009). Predicting response to political blog posts with topic models. Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL (pp. 477 485). Association for Computational Linguistics. Young L and Soroka, S. (2012). Affective news: the automated coding of sentiment in political texts. Political Communication. 29, 205 231. doi: 10.1080/10584609.2012.671234. 25. On the cutting edge of Big Data: digital politics research in the social computing literature Deen Freelon Most of this volume's chapters review studies rooted in political science, communication, and closely related disciplines. Indeed, many reference a small clique of foundational authors in agreement and/or disagreement, including Castells, Benkler, Hindman, Jenkins, Morozov, and Shirky. In the current chapter I diverge from this norm to examine a body of literature only rarely acknowledged by mainstream digital politics scholarship. This literalure contains politically relevant research by computer scientists and information scientists and is published under a variety of disciplinary labels, but will be referred to here as 'social computing research". As its name implies, social computing research includes any aspect of human behavior involving both digital technology and more than one person (Parameswaran and Whinston, 2007; Wang et al., 2007). Politics accounts for a small but thriving subset of this literature, which also encompasses health, business, economics, entertainment, artificial intelligence, and disaster response, among other topics. Social computing research on politics holds relevance for scholars of digital politics and political communication for two related reasons, one methodological and the other theoretical. First, social computing researchers have for many years led the vanguard in computational and ;Big Data' methods (sometimes in combination with other methods), in which the disciplines of political science and communication have both expressed great interest of late.1 Reviewing how social computing researchers have applied such methods to politically relevant datasets will help digital politics readers to consider how the methods could be applied to their own research. The field's methods and findings also hold a number of theoretical implications, but its researchers devote only sporadic attention to such concerns. For the benefit of those with a more theoretical scholarly orientation, and perhaps also for some social computing researchers with social science leanings, I explore major theoretical trends in the literature. 1 conclude with suggestions for future research, focusing on how digital politics researchers can best adapt the insights of social computing research to their own ends. 451 452 Handbook of digital politics On the cutting edge of Big Data 453 Before proceeding to these sections, however, it is necessary to more thoroughly describe social computing and its goals, which differ in key ways from those of the social science mainstream. The following section is devoted to this task. SOCIAL COMPUTING: A BRIEF INTRODUCTION A caveat before I begin: this section is written from the perspective of one who was trained in and still operates within a social science-based research orientation that emphasizes abstract theory as a guide and justification for empirical work (Fink and Gantz. 1996). My participation in social computing research up to this point in my career has been minimal. Accordingly, the description of social computing 1 offer here is intended as an introduction for those of a similar scholarly orientation to me, which I imagine will include many if not most of this volume's audience. In a widely cited overview article, Wang et al. (2007) define social computing as 'computational facilitation of social studies and human social dynamics as well as the design and use of 1CT [information and communication technology] technologies that consider social context' (p. 79). A similar characterization by Parameswaran and Whinston (2007) establishes social computing as a highly ubiquitous activity of study: Social computing shifts computing to the edges of the network, and empower [sic] individual users with relatively low technological sophistication in using the Web to manifest their creativity, engage in social interaction, contribute their expertise, share content, collectively build new tools, disseminate information and propaganda, and assimilate collective bargaining power, (p. 763) Both of the above quotes emphasize the two essential elements of social computing: digital tools ('computing', broadly construed) and social interaction. Of course, researchers in communication, sociology, anthropology, and other social-scientific disciplines have explored topics such as 'computer-mediated communication' and 'cyberculture' for decades. This similarity in subject matter invites the question of how social computing research differs from approaches with which we are more familiar. The main difference between social computing research and research traditions grounded in social science is as paradigmatic as that between social science and critical theory (Fink and Gantz, 1996; Potter et al., 1993). Whereas social science's goals are to explain empirical outcomes while promulgating theory, and critical theory's is to foment social change, social computing research is devoted to the development of new techniques for organizing, analyzing, and improving the user experience and output of social computing software. As such, social computing studies are usually published in highly technical articles that focus on methods, analysis, and evaluation at the expense of what we would consider 'theory' (Freelon. 2014). The call for papers for the 2014 conference on Computer-Supported Collaborative Work (CSCW). a prominent social computing publication venue, expresses this in its introduction: 'We invite submissions that detail existing practices, inform the design or deployment of systems, or introduce novel systems, interaction techniques, or algorithms' (CSCW, n.d.). Further evidence for this claim can be seen in the strong presence of employees of well-known technology companies such as Google, Microsoft, and Yahoo on major social computing conferences' program committees. Of course, theory is not always entirely absent: some articles include a few theoretical references of relevance to the project at hand, but the discussions tend to be much shorter than in most social science fields. And articles are often accepted without referencing any social science theories al all. In addition to downplaying theory, social computing research relies heavily on computational methods such as social network analysis, machine learning, computational linguistics, and algorithmic preprocessing of raw web data. These methods are common in computer science and information science, disciplines which many (though by no means all) social computing researchers call home. Programming serves at least two major purposes in social computing: (1) to develop and improve digital platforms for social interaction; and (2) to evaluate their performance efficiently and at scale. Qualitative methods such as ethnography and in-depth interviews are occasionally seen, often as part of a multi-method approach with one or more computational methods. However, such studies are relatively rare, as the next section will demonstrate. The field places the highest value on research techniques and metrics that can be implemented algorithmically. The ability to visualize quantitative results in intuitive and innovative ways is also highly prized. The final characteristic of social computing research of relevance to the digital politics researcher may seem rather obvious: the field is not principally concerned with politics per se, but rather with social computer use. In other words, social computing researchers typically analyze political cases to make broader points about social computing systems and affordances rather than about politics. Matters of system development and algorithm optimization almost always come first, and broader implications for politics are discussed secondarily if at all. Consequently, the results sections of social computing research papers often leave many implications of theoretical interest unexplored. Later in this chapter I will attempt to reclaim some of these implications in order to clarify their value for students of 454 Handbook of digital politics political science and communication. But first, I will examine in detail the most common methods social computing researchers employ. METHODS IN SOCIAL COMPUTING! RESEARCH ON POLITICS Social computing research is sometimes published in journals, but many of'the most relevant studies for our purposes are published in the archived proceedings of prominent conferences in computer science, information science, and human-computer interaction. Haphazardly selecting papers from these conferences would bias my discussion, so instead I chose them using a systematic and replicable method. First, I focused on conferences and publications sponsored by the Association of Computing Machinery (ACM) and the Institute of Electrical and Electronic Engineers (IEEE), the premier professional organizations in computer science and computer engineering. Using Google Scholar, I searched for the term "political" exclusively within such outlets. I then ranked the results in descending order by number of citations in order to capture the most widely referenced articles. Being interested only in articles that address politics as a central concern, I qualitatively assessed the most-cited items in each list, flagging articles that empirically analyze political messages, opinions, attitudes, and/or other content as their main focus. (Thus, for example, I excluded articles that analyzed political content as only one of three or more other content categories.) I continued this process until 1 had flagged 20 articles within each group, for a total of 40 articles (see Table 25.1). These form the basis of the discussions in this section and the next. After finalizing the sample, I informally classified the articles based on the methods they employed.2 Three methodological categories were used: traditional quantitative, qualitative, and computational. The traditional quantitative category includes long-established quantitative methods in social science such as surveys, experiments, content analysis, and statistical analysis of secondary data. The qualitative category includes in-depth interviews, field observations, and close readings of texts, among others. The computational category includes any method that entailed the creation of original source code whose purpose was to collect, preprocess, or analyze data. The rest of the chapter will focus mainly on this last category, as the others are much more familiar to scholars of digital politics. Unsurprisingly, the most prevalent methodological category throughout the sample was by far computational (29/40, 72.5 percent), followed by traditional quantitative (19/40, 47.5 percent) and then qualitative (5/40. 12.5 percent). All but one of the qualitative studies used a mixed U O 3 1 I a i_ O c3 £ ;j0 ? v 2 2 2 SP 60- — 2. - all* ■P ~ cd_ = 0 •a S | -a u a n ! .a 3II o "a a. fi -a M - - n arnumber=5432779. Boyd, D., Golder, S. and Lotan, G. (2010). Tweet, tweet, rctweet: conversational aspects of retweeting on twitter. In 43rd Hawaii International Conference on System Sciences (HICSS) (pp. 1-10). Conover, M.D., Goncalves, B„ Ratkiewicz, J.. Flammini, A. and Menczer, F. (2011). Predicting the political alignment of Twitter users. In Privacy, Security, Risk and Trust (PASSAT), 2011 IEEE Third International Conference on Social Computing (SocialCom) (pp. 192-199). doi:10.1109/PASSAT/SocialCom.2011.34. CSCW (n.d.l. Call for participation papers. ACM. Retrieved from http://cscw.acm.org/ participation_papers.html. DeNardis, L. and Tarn. E. (2007). Interoperability and democracy: a political basis for open document standards. In 5th International Conference on Standardization and Innovation m Information Technology, 2007. S1IT 2007 (pp. 171-180). doi:10.1109/ SIIT.2007.4629327. Diakopoulos, N.A. and Shamma, D.A. (2010). Characterizing debate performance via aggregated Twitter sentiment. In l>roceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 1195-1198). New York: ACM. doi: 10.1145/1753326.1753504. Diaz-Aviles, E., Orellana-Rodrigucz, C. and Nejdl, W. (2012). Taking the pulse of political emotions in Latin America based on social web streams. In 2012 Eighth Latin American Web Congress (LA-WEB) (pp. 40-47). doi:!0.1109/LA-WEB.2012.9. Fang, Y., Si, L., Somasundaram, N. and Yu, Z. (2012). Mining contrastive opinions on political texts using cross-perspective topic model. In Proceedings of the Fifth ACM International Conference on Web Search and Data Mining (pp. 63-72). Retrieved from http://dl.acm.org/citation.cfm7id=2124306. Fink, E.J. and Gantz, W. (1996). A content analysis of three mass communication research traditions: social science, interpretive studies, and critical analysis. Journalism and Mass Communication Quarterly. 73(1), 114-134. doi:10.1177/107769909607300l 11. Fisher, K.E., Becker. S. and Crandall, M. (2010, January). eGovemment services use and impact through public libraries: preliminary findings from a national study of public access computing in public libraries (pp. 1-10). IEEE. Freelon, D. (2014). On the interpretation of digital trace data in communication and social computing research. Journal of Broadcasting and Electronic Media, 58(1), 59 75. Fumholt, B. and Wahid, F. (2008, January') E-govemment challenges and the role of political leadership in Indonesia: the case of Sragen. In Proceedings of the 41st Annual Hawaii International Conference on System Sciences, IEEE. 470 Handbook of digital politics Garcia, A.. Standlee, A., Beckhoff, J. and Cut. Y. (2009). Ethnographic approaches to the Internet and computer-mediated communication. Journal of Contemporary Ethnography. 38(1). 52 84. Garcia, D„ Mendez. F.. Scrdiilt, U. and Schweitzer. F. (2012). Political polarization and popularity in online participatory media: an integrated approach. In Proceedings of the first Edition Workshop on Politics, Elections and Data (pp. 3-10). Retrieved from http:// dl.acm.org/cita lion.cfm?id=2389665. Golbeck, J. and Hansen, D. (2011). Computing political preference among Twitter followers. In Proceedings of the 2011 Annual Conference on Human Factors in Computing Systems (pp. 1105-1108). New York: ACM. doi:10.1145/1978942.1979106. Graesser, A.C., McNamara, D.S. and Louwerse, M.M. (2010). Methods of automated text analysis. In Kamil, M.L., Pearson, D., Moje, E.B, and Affierbach, P. (eds). Handbook of Reading Research, Vol. 4 (pp. 34 53). New York: Routledge. Gulati, G.J.. Yates, D.J. and Williams, C.B. (2012, January). Understanding (he impact of political structure, governance and public policy on e-government. In 2012 45th Hawaii International Conference on System Sciences (HICSS) (pp. 2541-2550). IEEE. Hong. S. and Nadler. D. (2011). Docs the early bird move the polls?: The use of the social media tool 'Twitter' by US politicians and its impact on public opinion. In Proceedings of the 12th Annual International Digital Government Research Conference: Digital Government Innovation in Challenging Times (pp. 182-186). New York: ACM. doi: 10.1145/2037556.2037583. Jiang, M. and Argamon, S. (2008). Exploiting subjectivity analysis in blogs to improve political leaning categorization. In Proceedings of the 31 si Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 725-726). New-York: ACM. doi: 10.1145/1390334.1 390472. Jürgens, P., Jungherr, A. and Schoen, H. (2011). Small worlds with a difference: new gatekeepers and the filtering of political information on Twitter. Proceedings of the ACM WcbSci'll (pp. 14-17). Kannabiran, G. and Petersen, M.G. (2010). Politics at the inlerface: a Foucauldian power analysis. In Proceedings of the 6th Nordic Conference on Human-Computer Interaction: Extending Boundaries (pp. 695-698). New York: ACM. doi:10.l 145/1868914.1869007. Kaschesky, M. and Riedl. R. (2011). Tracing opinion-formation on political issues on the Internet: a model and methodology for qualitative analysis and results. In 2011 44th Hawaii International Conference on System Sciences (HICSS) (pp. 1-10). doi:10.H09/ HICSS.2011.456. Kaschesky. M., Sobkowicz. P. and Bouchard, G. (2011). Opinion mining in social media: modeling, simulating, and visualizing political opinion formation in the Web. In Proceedings of the 12th Annual International Digital Government Research Conference: Digital Government Innovation in Challenging Times (pp. 317-326). New York: ACM. doi: 10.1145/2037556.2037607. Kim, B.J., Kavanaugh, A. and Perez-Quinones, M, (2007). Toward a model of political participation among young adults: the role of local groups and ICT use. In Proceedings of the 1 st International Conference on Theory' and Practice of Electronic Governance (pp. 205-212). ACM. Kim, S., Valente, F. and Vinciarclli, A. (2012). Automatic detection of conflicts in spoken conversations: ratings and analysis of broadcast political debates. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 5089 5092). doi:10.1109/ICASSP.2012.6289065. Leetaru, K. (2Ü12). Data Mining Methods for the Content Analyst: An Introduction to the Computational Analysis of Content. London: Routledge. Mascaro, C. Black, A. and Goggins. S. (2012). Tweet recall: examining real-lime civic discourse on Twiner. In Proceedings of the 17th ACM International Conference on Supporting Group Work (pp. 307-308). Retrieved from http://dl.acm.org/citation. cfm?id=2389233. On the cutting edge of Big Data 471 Munson. S.A. and Resnick, P. (2010). Presenting diverse political opinions: how and how much. In Proceedings of the 28th International Conference on Human Factors in Computing Systems (pp. 1457-1466). Nahon, K. and Hemsley, J. (2011). Democracy.com: a talc of political blogs and content. In 2011 44th Hawaii International Conference on System Sciences (HICSS) (pp. 1-11). doi:!0.1U)9/HICSS.2011.140. Parameswaran, M. and Whinston, A.B. (2007). Social computing: an overview. Communications of the Association for Information Systems, 19(37), 762-780. Park, S., Ko, M, Kim, J., Liu, Y. and Song, J. (2011). The politics of comments: predicting political orientation of news stories with commenters' sentiment patterns. In Proceedings of the ACM 2011 Conference on Computer Supported Cooperative Work (pp. 113 122). New York: ACM. doi: 10.1145/1958824.1958842. Potter, W.J., Cooper, R. and Dupagne, M. (1993). The three paradigms of mass media research in mainstream communication journals. Communication Theory, 3(4), 317-335. doi: 10.1111/j. 1468-2885.1993. tb00077.x. Ratkiewicz. J., Conover, M., Meiss, M., Goncalves, B., Patil, S., Flammini, A. and Menczcr, F. (2011). Truthy: mapping the spread of astrolurf in microblog streams. In Proceedings of the 20lh International Conference Companion on World Wide Web (pp. 249-252). New York: ACM. doi: 10.1145/1963192.1963301. Russell, M.A. (2013). Mining the Social Web. Stebastopol. CA: O'Reilly Media. Sarmento, L., Carvalho, P.. Silva, M.J. and de Oliveira, H. (2009). Automatic creation of a reference corpus for political opinion mining in user-generated content. In Proceedings of the 1st International C1KM Workshop on Topic-Sentiment Analysis for Mass Opinion (pp. 29-36). New York: ACM. doi: 10.1145/1651461.1651468. Shmueli, G. and Koppius. O. (2011). Predictive analytics in information systems research. Management Information Systems Quarterly, 35(3), 553 572. Singh. V.K., Mahata, D. and Adhikari. R. (2010). Mining the blogosphere from a sociopolitical perspective. In 2010 International Conference on Computer Information Systems and Industrial Management Applications (CISIM) (pp. 365-370). IEEE. Skoric, M., Poor, N., Achananuparp, P.. Lim. F.-P. and Jiang, J. (2012). Tweets and votes: a study of the 2011 Singapore general election. In 2012 45th Hawaii International Conference on System Science (HICSS) (pp. 2583 2591). doi:10.1109/ HICSS.2012.607. Stieglitz, S. and Dang-Xuan. L. (2012). Political communication and influence through microblogging: an empirical analysis of sentiment in Twitter messages and retweet behavior. In Hawaii International Conference on System Sciences (Vol. 0, pp. 3500-3509). Los Alamilos, CA: IEEE Computer Society. doi:http://doi.ieeecomputcrsociety.orii/l0.1109/ HICSS.2012.476. Ulicny, B., Kokar, M.M. and Matheus, C.J. (2010). Metrics for monitoring a social-political blogosphere: a Malaysian case study. IEEE Internet Computing, 14(2), 34-44. doi:10.1109/MIC.2010.22. Vallina-Rodriguez, N., Scellato, S., Haddadi, H.. Forsch, C, Crowcroft, J. and Mascolo, C. (2012). Los Twindignados: the rise of the indignados movement on Twitter. In Privacy, Security, Risk and Trust (PASSAT). 2012 Internationa] Conference on Social Computing (SocialCom) (pp. 496-501). doi: 10.1109/SocialCom-PASSAT.2012.120. Wallsten, K. (2011, January)- Beyond agenda setting: the role of political blogs as sources in newspaper coverage of government. In 2011 44th Hawaii International Conference on System Sciences (HICSS) (pp. 1-10). IEEE. Wang, F.-Y., Carley, K.M., Zeng, D. and Mao, W. (2007). Social computing: from social informatics to social intelligence. Intelligent Systems, IEEE. 22(2), 79 83. Weber, I., Garimclla, V.R.K, and Borra, E. (2012). Mining web query logs to analyze political issues. In Proceedings of the 3rd Annual ACM Web Science Conference (pp. 330-334). New York: ACM. doi: 10.1145/2380718.2380761. Wei, L. and Y'an, Y\ (2010). Knowledge production and political participation: reconsidering the knowledge gap theory in the Web 2.0 environment. In 2010 The 2nd IEEE International 1