This study asks whether politicians’ “public personality” traits—specifically agency and communion—can be measured at scale from their public statements using computational text analysis. The authors propose a theory-driven, domain-specific operationalization of agency and communion and compare multiple computational approaches to extract these traits from a large corpus that includes parliamentary speeches, social media posts, and interviews. They validate their approach with human-labeled data, functional tests, and targeted analyses of German politicians and the 2024 U.S. presidential debate featuring Donald Trump and Kamala Harris. The paper reports that prompting-based techniques, especially with advanced models like DeepSeek-V3, performed better than supervised and semisupervised alternatives, while noting this is a multimethod feasibility study rather than a definitive measurement standard.
Bold approach: theory-driven operationalization of agency and communion Validation: human-labeled data and functional tests Key comparison: prompting techniques vs. supervised methods
Data and methods
The authors compile a large textual corpus of politicians’ public statements drawn from parliamentary speeches, social media posts, and interviews. They operationalize two political personality traits—agency and communion—within a domain-specific framework and apply several computational text-analysis methods to infer trait prominence from text. The paper evaluates methods using human-labeled validation data and functional tests; the abstract highlights prompting-based techniques and DeepSeek-V3 as especially competitive.
Main findings
Across their multimethod comparisons, prompting-based approaches (notably with advanced models like DeepSeek-V3) outperformed the supervised and semisupervised methods the authors tested. The authors illustrate trait prevalence patterns in statements by German politicians and in the 2024 U.S. presidential debate between Donald Trump and Kamala Harris, while framing the work as a feasibility study that points to promising directions rather than definitive measurement claims.
- Methods compared: prompting (large models), supervised, and semisupervised techniques.
- Validation: comprehensive human-labeled data and functional tests.
