The aim of the current paper is to formulate a conception of pragmatic patterns characterizing the construction of individual and collective identities in virtual communities (in our case: the Twitter community). We have explored several theoretical approaches and frameworks and relevant empirical data to show that the agents building virtual communities are 'extended selves' grounded in a highly dynamic and compressed, linguistically mediated virtual network structure. Our empirical evidence consists of a study of discourse related to the Latvian parliamentary elections of 2010. We used a Twitter corpus (in Latvian) harvested and statistically evaluated using the Pointwise Mutual Information (PMI) algorithm and complemented with qualitative and quantitative content analysis.
In this paper, we explore the pragmatics of political messages in Latvian Twitter communication during the 2010 general election.
The results contain a topical analysis of election discussions as well as an analysis of hashtags and retweeted messages. The fast pragmatic dynamics in Titter communication can be observed through hashtags, showing a rapid reaction of Twitter users to the elections, while top retweets support the findings of content analysis with regard to political sentiment. Content analysis reveals the possibility of significant discrepancies in terms of the cognitive and physical distances between a group and its individual members in their identity generation processes. In view of the results, we propose a hypothesis that reveals correlations between a group and its individual members, the richness of topics, channels of communication, frequency of mention, and connotations and effects of messages.
We assume that the generation of identity takes place through two simultaneous and
mutually interdependent social categorization processes – belongingness and
differentiation [3,4]. Our study undertakes to examine these two processes in action,
constrained by two selection criteria: (
Twitter is a particularly fruitful resource for this type of analysis because its brevity constraint gives rise to an abundance of shortcut techniques including expressive lexis, the use of abbreviations and hyperlinks for proper names and keywords. Rigid information hierarchies reveal what users presume to be already known and/or shared by their in-group, and are a fertile soil for the investigation of presuppositions, cultural common ground, and cultural discrepancies [6,7,15]. This is especially prominent in Twitter discourse about politics, a topic where speakers generally exhibit willingness to report their opinions despite the fact that their perspectives are often conflicting. Although political opinions are usually articulated explicitly, belongingness to an identity group1 may be partly implicit [19].
We focus on mechanisms of self-identification, formation and maintenance of ingroups and their differentiation from out-groups. The findings attempt to answer the following questions: 1) How are virtual political identities generated and maintained in a condensed public mode of communication? 2) What are the pragmatic instruments that help to achieve these processes?
Twitter can also help to understand implicit social categorization. Typically, research on social categorization is conducted using questionnaire or focus group methodologies, mainly addressing explicit political categorization. This study has incorporated some implicit factors of analysis, often crucial in political communication. Approaching human-generated digital content as empirical material for categorization analysis is not new (cp. [9]). Analysis of political messages on Twitter, although not directly focused on categorization, is also provided by several studies (cp. [22]). Several recent studies explore possible correlations between election outcomes and the level of Twitter activity of politicians (US Congress: [14], South Korea: [12]). This study, however, also analyzes political messages created by media organizations and other active users.
Co-occurrence statistics allow to quantitatively project some of a word’s semantics grounded in users’ categorization performance ([18]). Collocations show the relative most frequent (sometimes stereotypical, implicit) social categories in communication, but the research must be complimented with concordance analysis for semantic complexity. Of course, the output of such a combination of methods concerns the group (and not individual) patterns of social categorization, and pragmatic effects are related to statistical frequency of language used in communities and not to individual patterns of communication2. 1 We define identity as a continuous process where the sense of belongingness to a community interacts with the desire to be a unique individual. A community has an internal and an external structure (relationships within the group and relationships with other groups), and community identity can generate polarization effects. 2 A pragmatic pattern is a typical way of using language in a linguistic community (e.g., in social media). #MSM2011
Studies show that people frequently have difficulty explicitly articulating their ideology [19]. Thus self-report, focus groups, and questionnaires may often prove inadequate for analyzing political categorization. Ideological labels, moreover, may not correspond to subjective conceptions of beliefs, and undecided voters exhibit a much clearer opinion via implicit tasks than via explicit ones [19]. Political categories are distinguished above all by their extreme polarization (cp. [11,17]). On Twitter, initially informative messages are modified to become increasingly polarized [23].
The Saeima (the parliament of Latvia) is elected using a proportional multi-partisan representation system for 100 seats. The 2010 election saw 13 competing political parties or their alliances. Candidates from 5 parties were elected: 33 seats for “Unity” (Unity), 29 seats for “Harmony Centre” (HC), 22 seats for the “Union of Greens and Farmers” (UGF), and 8 seats each for the National Association “All For Latvia!”“TB/LNNK” (NA) and “For a Good Latvia” (FGL). The turnout for the 2010 elections was 63.12% or around 967 000 people. 3
The aims of this study are: (
We collected a dataset of tweets covering the election week, performed careful manual extraction work and numerous statistical comparisons. We also created custom tools for analyzing Latvian Twitter content including a concordance tool. We believe that this makes our results, in several respects, even more precise than, e.g., [22] who automatically translated their corpus of empirical data (German tweets) into English and only then processed it with LIWC (Linguistic Inquiry and Word Count).
The dataset consists of one week of Twitter messages (from 28-Sep-2010 to 04-Oct2010) from a subset of Latvian Twitter users, including 4 days before the election, the day of the general election (October 2) and 2 days following the election. The total size is 50'032 messages, consisting of: 50% regular tweets; 18% retweets; and 32% replies. There are no publicly available official data about the total number of Twitter participants in Latvia. According to local media experts, the estimate is approximately 40'000 users (November 2010).
In order to choose a topically relevant set of Twitter accounts, we started with a
manually selected set that included (
We did not choose a random sample to avoid large amounts of redundant data consisting of ordinary discussions unrelated to our research interests - politics, identity generation, and the media. This intentionally selected dataset allows for a more precise analysis of the above research topics.
Collected tweets are processed using the NLTK library [1]. The processing of tweets consists of: cleaning the dataset; saving the full tweet data for structure analysis; tokenizing tweets; replacing keywords, where we consolidate the various ways to write the same word or expression and replace it with a single keyword identifier.
Latvian is an inflected language in which the same word may appear in many forms. In the keyword replacement step, we collapse these forms into one keyword. We also replace different ways of writing the same expression (e.g. abbreviations and full names of party names). Since there was no stemming or lemmatizing software for Latvian that we could use, we created our own keyword replacement map for keywords related to elections.
Having processed the tweets, we performed: (
We made a list of all 1234 candidates competing for seats in the parliament, exploring their representations in selected tweets during the 4 days leading up to the election. Since only a small part of all candidates were represented in Twitter communication (in our dataset) four days before the election, we wished to compare our findings with publicity coverage of the candidates in other media in Latvia.
We identified 79 family names of the candidates occurring in collocations in the
Twitter dataset, and 170 family names of the candidates occurring in the media
monitoring dataset. We distinguish four groups of candidates: (
Further, we listed how many personal tweets, collocations and publications occur
with every of the family names in various time periods (the average number of
collocations of every family name of the candidates four days before the election is
4.68; later, we included only those (
For names of political parties (Table 1) we listed: (
Two political parties – FGL and “Unity” - have significantly higher rankings of mention than other parties. Moreover, their candidates for the post of Prime Minister (Ainārs Šlesers (FGL) and Valdis Dombrovskis (Unity)) have similar rankings of mention. In spite of these similarities, the two have strikingly different election results (“Unity” won the election and got 33 seats in the parliament, with Valdis Dombrovskis approved as the prime minister, while FGL got only 8 seats in the parliament). This led us to investigate more closely the identity generation of these individuals and organizations through political categorization in pre-election tweets. First, we identified 10 collocations of significantly high ratings for the four name keywords. Secondly, we used concordance analysis to examine the semantics in each collocation.
We have listed in Table 2 what percentage of the topics bear positive, neautral or negative connotations and how many topics are covered by each of the keywords. As Table 2 demonstrates, the individual and the organization are categorized similarly in the case of Šlesers and his political party FGL: both are more related to negative topics than positive ones. The case of Valdis Dombrovskis and his political party “Unity” is different: the individual is mostly categorized in positive or neutral topics, while the political party is categorized in negative or neutral ones. This shows that the generation of identity of an organization and that of its individual members may #MSM2011 involve significant discrepancies in terms of cognitive versus physical distances4. In this case, the cognitive distance between Dombrovskis and “Unity” is bigger than the ‘physical’ one. This may be in part due to the fact that the “Unity” election campaign focused exclusively on Dombrovskis, promoting him as the principal benefit to the voters. Thus the individual became more cognitively important than the whole (an organization).
This allows us to expand our hypothesis regarding politicians and political parties
as follows: (
The most popular hashtag was #velesanas (“election”), used in 459 tweets (17.2% of tweets containing hashtags). Other election-related hashtags that were used at least 20 times include #nobalsoju ("i voted"), #politsports (political sport), #pietiek 4 In the Spreading-Activation Theory, assuming a correlation between the collocational structure of the corpus and the mental models of its users, collocational structure reflects the cognitive distance between conceptual entities such as political parties and individuals. Indirectly connected nodes are more distant than directly connected ones. 5 Using manual concordance analysis, connotations are determined and generalized according to three categories (positive, neutral, and negative), determined individually for each tweet. Examples include: “Friends, tomorrow I shall vote for Dombrovskis, because I trust his professionalism …” (positive); “Šlesers doubts the objectivity of social media …” (neutral); “Dombrovskis: a protégé of corruption or a racketeer?” (negative). #MSM2011 ("enough!"), #vēlēšanas (#velesanas with Latvian diacritics), #cieti (“solid” – a slogan of FGL), #twibbon (twibbons were used to show party support).
For the purposes of this paper, we limited Table 3 to hashtags related to politics. Most of the top 10 hashtags on election day were related to politics (9 out of 10) and appear in the table. Other days had less election-related tags, but also a lower hashtag usage activity in general. The #velesanas (“election”) hashtag appeared the day before the election and had a remarkable spike in its usage on election day and the day following it, receding back to background level the day after that.
Hashtags that retained popularity for at least 4 days in this 5 day period were the journalism tags #pietiek and #ir. Both refer to publications seen by top Twitter users as prestigeous and integral organizations for investigative journalism. The hashtag #sleptareklama ("hidden advertising") coincided with the appearance of controversial hockey-related advertisements that were suspected of containing hidden political advertising. A creative usage of a hashtag is its syntactic integration into a sentence: a notable example is using #ir, the magazine whose name literally means “is”, as a verb: e.g., “There #is still time to form a new coalition”.
Apart from the obvious purpose of attracting attention to major topics, hashtags carry the connotation of familiarity with the object of the tag, be it a topic, an individual, or an organization – at the very least, one must know what is worth tagging. Tags help to define group identity in two recursive ways: by highlighting issues considered important by the group, and by presenting the group as the kind of community where such issues are considered important. 5.2
We considered a retweet any Twitter message that contains the string "RT @nickname" (17.68% of the selected dataset). Most retweets start with "RT @nickname", i.e. are marked as such and point to the original message. These results shows more uniformity of retweet formats than reported in [2], possibly a result of more officialized retweet functionality. For further analysis, we used retweets which #MSM2011 contained information about the original tweet (i.e. 90.46% of all retweets). An analysis of the top 20 most retweeted posts reveals that 70% of these posts are directly related to the elections; 10% are loosely related; 20% are unrelated.
There were 14 election-related messages among the 20 most retweeted messages. Of these, the majority (8 out of 14) were satirical tweets criticising a political party or a politician. Seven refer to FGL or its prominent members Ainārs Šlesers and Andris Šķēle. Other parties mentioned in these retweets were HC and FHRUL (one tweet each). The two most retweeted messages are related to the election. 5.3
The content of top retweets and hashtags reveals that the opinion leaders in the Latvian Twitter-sphere, the in-group that enjoys the highest popularity and prestige, can be vaguely defined as a group of centrists who see themselves as positioned between two perceived polarities. The cognitive space, as regarded by the in-group, can be characterized thus: to the left are krievi (“the Russians”), the parties and their supporters commonly perceived as pro-Muscovite and representing the interests of the Russian-speaking population (HC, FHRUL). To the right are nēģi (“the parasites” – an imprecise translation of the word taken from a popular tweet criticising this group), the nationalist alliance (FGL, NA) that the Twitter opinion leaders see as outdated and highly corrupt, exploiting their privilege for personal gain. The in-group supports the political alliance “Unity” and particularly its leader, Valdis Dombrovskis, who was subsequently elected Prime Minister.
The fact that the in-group appears to take a centrist position is significant: their output is less polarizing than could be expected of a highly politicized group. Still, there is a clear demarcation of the in-group from both out-groups described above. This is achieved by the opinion leaders of the in-group through several groupidentity-generating mechanisms and strengthened by the heightened emphasis on the social self [4], typical of both online communities and political discourse.
Manipulating cognitive distances is relatively easy in the dematerialized virtual space, which facilitates impressions of togetherness and mutual identification within the in-group, on the one hand, but also the distancing of the in-group from out-groups. Perhaps surprisingly, the brevity constraint of Twitter messaging, rather than complicating political categorization, can facilitate it: the format is well suited to the in-group’s simplified tripartite view of the political space. Thus, through repeated tweeting of negative content containing the letters “PLL” or “PCTVL” (acronyms of the names of political parties on the two sides of the perceived spectrum), it is soon enough to write “PLL” or “PCTVL” to evoke a cognitive frame [8] associated with negative content. Clearly, the details of this content will be unique for each user; but as long as there is a basic understanding of a commonality of reference – in this case, of the negativity of the referents – a mention of a party acronym will effectively serve as an invitation to ‘fill in the gaps’ with each reader’s own meaning [13]. #MSM2011
Political jokes6, abundant in top retweets, work in a similar manner. Provided that the humourous effect is usually achieved by inviting the audience to frame-shift through an unexpected element [7], political jokes on Twitter are doubly rewarding because they give the audience the feeling of belongingness through having understood the frame shift without surrounding linguistic context and through a very limited number of signs. Similarly to a hashtag, a retweet works recursively by simultaneously flaunting an individual’s understanding (and hence his belonging to the in-group) and helping to define his individual identity through the content of what is understood and retweeted.
Our corpus shows that power and control are very much the preoccupation of Twitter users, and the independently formed, ‘grass-roots’ community of top tweeters quickly forms their own behaviour canons. This is typical of online communities, where a myriad of rules and expectations underlie seemingly free, chaotic communication [10]. A popular political message on Twitter is at once an expression of individual and group identity, an invitation to the in-group members to share the opinion expressed, and a warning about the consequence of deviating from the group’s norms. By way of illustration, a message retweeted 15 times reads: “I heard that Šlesers won’t vote for PLL either, because they’re said to be thiefs” (our emphasis). In addition to cleverly poking fun at the politician by suggesting he will not vote for his own party, the message succeeds in conveying that the author will not vote for Šlesers, that he assumes that his in-group members will not do so, and that anyone who does vote for Šlesers will be seen as voting for a thief and undermining his or her in-group membership. In short, Twitter conformity mechanisms are just as compact as the medium itself.
Yet without a conforming audience, such successful guidance toward a rigidified, formal categorisation would not be possible (we may well judge the above message as successful, since it is on the list of top retweets). The tension between individual opinion and in-group identification (the personal vs. the interpersonal/social self) is resolved through a balance of stereotyping processes: just as the political parties and actors are stereotyped to fit into one of the few cognitive categories carved out for the occasion of the election, so the individual members engage in a certain degree of selfstereotyping [20]. Members will be more willing to overlook differences of opinion and concentrate on their commonalities (real or imagined) when membership is seen as beneficial, and particularly if the group is seen as working toward a common goal of some sort – in this case, victory in the parliamentary elections [4]. Because intragroup attraction on Twitter in the run-up to parliamentary election is ideational rather than interpersonal, the in-group achieves a high degree of political cohesion in part simply through perceiving itself as a cohesive unit. 6 An example that is comparatively demure and reproducible in an academic paper refers to the leader of the party perceived as being in the “parasites” group: “A little boy falls. Šlesers helps him up. ‘So, I guess now you will vote for me?’ ‘I only hurt my foot, not my head!’”. #MSM2011
We have formulated a correlation according to which three factors contribute to the efficiency of political messages in the electoral discourse – in particular, for a given collocation bigram: (a) the variety of thematic contexts of occurrence, (b) the frequency of mention, (c) positive connotations. (While there are other factors determining efficiency, this study has focused on popularity-oriented facets.) We have therefore extended the results stated by [12, 14] regarding the correlation between minority parties, Twitter activity, and election results. The dynamics of Twitter users’ interest in the event (the election) can be observed through hashtag usage and the most retweeted messages. Top retweets, in turn, convey user sentiment toward political parties and individuals.
We have noted instances of discrepancy between attitudes toward individual politicians as opposed to attitudes toward political groups, and observed that frequent positive mention of individuals can lead to a heightened cognitive significance of this individual, causing the perception of the significance of the relevant organization to recede into the background.
We envision possible applications of this work in analysis tools correlating Twitter dynamics with the structure generated from the parameters: (a) the variety of occurrence contexts, (b) the frequency of mention, (c) positive connotations (generated semi-automatically). The items which fit into the highest ranking of such analysis results can be further analyzed manually and a variety of pragmatic effects (stereotyping, presupposition generation a.o.) might be observed.
Finally, we can hypothesize that the user of a microblogging resource such as Twitter extends the sphere of his or her cognitive processing by involving additional interactive structures of communication. Thus, if we assume that the social categorization in a community consists of (a) self-categorization as the most crucial and basic level of identity building, (b) interpersonal communities of individuals, and (c) large-scale social communities (e.g., national identity communities) including subcommunities [4], we could argue that self-categorization involves a substantial amount of extended cognitive processing offloaded onto the digital environment (in our case, Twitter). In this sense, the results provided by our study can complement research on the extented mind [5, 21]. A more detailed analysis of the extended self and offloading effects in cognitive processing is a topic for another study.
This work has been supported by the European Social Fund project «Support for Doctoral Studies at the University of Latvia». #MSM2011