https://orcid.org/0000-0002-9313-2007
![Sample our Politics & International Relations journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5947/TOC-Subject-Sample-2021-Politics-International-Relations.jpg"})
Abstract
Political parties address the public through multiple communication channels simultaneously, but this is not reflected in contemporary research. It is largely unclear how party competition plays out across different communication channels and whether issue salience strategies depend on the channel used. In order to answer this question, this article trains a state-of-the-art language model (BERT) on labelled manifestos and applies it for cross-domain topic classification of press releases, parliamentary speeches and tweets from parties and individual party members in Austria, Germany and Switzerland. The results show that certain channel characteristics influence parties’ issue salience. The extent to which a party addresses its issue preferences (ideal agenda) is moderated by the degree of centralised communication (party vs. individuals) and the presence or absence of a pre-given agenda, whereas a channel’s primary audience (direct vs. mediated channel) plays a much smaller role than expected. These findings illustrate the complexity of party competition in contemporary multi-channel and hybrid media environments.
Acknowledgements
I wish to thank the editor and the anonymous reviewers for their valuable comments and guidance. I am also very grateful to Heiko Giebler, Pola Lehmann, Christoffer Green-Pedersen, Christian Czymara, Lisa Zehnter and the participants at the ECPR General Conference 2022 in Innsbruck for helpful comments on earlier versions of this article, as well as Tobias Burst for his great input regarding the methodological approach. Furthermore, I am grateful to Juliane Hanel, Sarah Hegazy and Leonie Schwichtenberg for their fantastic research support.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Data availability statement
Replication material for this study is openly available in Harvard Dataverse at https://doi.org/10.7910/DVN/JNQUOD.
Notes
1 In the case of Switzerland, I only include texts in German.
2 Some press releases are drafted by party offices, but many are drafted by individual party members (e.g. MPs). However, the party office remains important. Even if press releases are drafted by individual party members, the press team of the party or parliamentary party group publishes them and is often given as contact for correspondence. Hence, press releases are some sort of hybrid communication channel with regard to the characteristic of centralisation and are therefore assigned the value of 0.5. Similarly, parliamentary speeches are assigned the value of 0.5. Although parliamentary speeches are held by individual MPs, in many countries parliamentary party group leaders are in firm control over the selection of speakers in legislative debates (e.g. Proksch and Slapin Citation2015).
3 The Twitter data set only includes original tweets. Similar to Barberá et al. (Citation2019) and Gilardi, Gessler, et al. (Citation2022), I exclude replies and retweets.
4 A more detailed description of the corpus and its properties is provided in Online Appendix A.1. Further information on the included Twitter accounts from individual party members is described in Online Appendix A.2.
5 The adapted codebook used in this article is displayed in Online Appendix A.3.
6 The pre-trained model is available via the HuggingFace python library (Wolf et al. Citation2020).
7 A detailed step-by-step explanation of the cross-domain application with the BERT model is provided in Online Appendix A.4.
8 Information on the manually coded ‘gold standard’ is available in Online Appendix A.5. The ‘gold standard’ consists of manually labelled press releases, parliamentary speeches and tweets from party accounts.
9 A more detailed breakdown of the model performance is given in Online Appendices A.6 and A.7.
10 Osnabrügge et al. (Citation2021), for example, also use annotated manifestos from the Manifesto Project as training data for their classification model. Their model achieves an accuracy of 41% (44 categories) and 51% (8 categories) in a cross-domain topic classification of parliamentary speeches from New Zealand.
11 Parliamentary speeches are for example not held every single month. Parliaments usually have a summer break and other longer breaks spanning multiple weeks, especially in the case of Switzerland.
12 As noted earlier, press releases are a special case with regard to the characteristic of centralisation. Some press releases are drafted by party offices, but many are drafted by individual party members (e.g. MPs) as well. In both cases, however, the press team of the party or parliamentary party group is important as they publish the press releases and are often given as contact for correspondence.
13 Online Appendix A.8 provides the overall percent for issue salience in manifestos, parliamentary speeches, press releases, tweets from party accounts and tweets from individual party members across all three countries and parties.
14 To account for potential biases resulting from limited data, I exclude those observations when a party has published less than 10 press releases, tweets or parliamentary speeches during a quarter.
Additional information
Funding
Notes on contributors
Christoph Ivanusch
Christoph Ivanusch is a research fellow in the Manifesto Research Project on Political Representation (MARPOR) and the H2020 project ‘OPTED Observatory for Political Texts in European Democracies’ at the WZB Berlin Social Science Centre. Furthermore, he is a doctoral student in political science at the Berlin Graduate School of Social Sciences (BGSS), Humboldt-Universität zu Berlin. His research interests include political parties, party competition, political communication and text-as-data. [[email protected]]