Campaigns and Social Media Communications: A Look at Digital Campaigning in the 2010 U.K. General Election

Abstract

Social media are said to have the potential to transform relationships between political parties, candidates, and citizens. This chapter is a study of social media use at different levels in the 2010 United Kingdom general election to see to what extent that potential is realized. The research compares the use of Twitter by the national level of the campaign, composed of the three major parties, and their leaders, as well as the campaigns of the three major parties across the nine electoral districts in England's second city, Birmingham. It examines the candidates and parties’ that various informational and engagement strategies at the national and Birmingham levels of the campaign with respect to their campaign functions. The analysis is carried out using natural language processing to computerize the content analysis. The findings reveal that social media are used at both levels, primarily for the undirectional transfer of information rather than for engagement. However, at the Birmingham level of the campaign there appears to be significantly greater emphasis on the creation of personal connections between candidates and the public than at the national level of the campaign. This suggests that lower profile candidates use social media in a compensatory manner, offsetting their limited media coverage which voters typically rely on in getting to know the candidates.

Appendix: Data Collection

Twitter posts were collected using the twitteR module (Gentry 2012) in the R language. This captures all of the public posts from each candidate’s account. Twitter posts were segmented on a daily basis and coded using natural language processing (NLP) (Bird et al. 2009). There are limitations to machine coding. The machine coding has limited ability to analyze words in relation to their contexts and to evaluate the implication of specific utterances for constituting the overall meaning of a text. We mitigate these risks as our analyses do not attempt to identify valances, selecting terms that are signified to a greater degree by the presence or absence of a term rather than the manner in which terms combine to form meanings.¹³ Nevertheless, NLP significantly advances the ability to code large volumes of text that would otherwise be inaccessible if left to human coding. NLP can provide a high degree of reliability regarding the structure of a textual artifact. Furthermore, apart from its advantages in managing large volumes of data, NLP also enables the identification of patterns across large volumes of text that go beyond or are independent of the specific meanings communicated by authors (Ramsay 2005). In our case, given the highly structured nature of Twitter messages and their limited size, the occurrence of certain words can more reliably be linked to the incidence of campaign themes and communicative operations. NLP’s algorithmic approach entails certain limitations: using too narrow a list of terms runs the risk of Type I, false negative errors, while using a more elaborate list of terms runs the risk of Type II, false positive errors.

Before applying NLP, all metadata was removed from the database leaving only the text of the messages. All terms were entered as lower case to match the normalization of the text. As the hypotheses under investigation relate to temporal and institutional structures or to differences between social media platforms, we take aggregates of posts rather than individual posts as our unit of analysis. After a series of tests to train the machine coder, the machine results were then verified against a second subsample of human-coded Twitter posts. A daily segment of tweets were selected based on a relatively even dispersion of machine scores across each of the categories so as to not bias the testing based on a high incidence of one dimension.¹⁴ Although these categories do not encompass the full range of communications in the samples, each tweet in the human-coded sample matched at least one of the dimensions of horizontal communications or information provision. The errors were summed with respect to the eight coding categories and calculated as a percentage of the total elements identified by human coding (Nolan 2002).