💡 We’d love to hear what you think about the paper! And once again, a big thank-you to all reviewers and colleagues who offered invaluable input and guidance along the way.

We also test the predictive ability of our measures by confirming previous findings on the relationship between personality and ideology, and provide extensive further exploratory analysis (e.g., on gender differences) in the Appendix.

Next, we conducted "functional tests", inspired by software development practices, to verify whether models can accurately classify intentionally designed statements, showing that only DeepSeek-V3 and GPT-40 were able to pass all functional tests.

How well do different approaches perform on human-labeled data?
Our findings show that state-of-the-art instruction-tuned models perform best, reaching macro F1 scores of nearly 0.8 for both traits.

We build a full pipeline, from conceptualizing political personality traits to codebook development, human annotation, model comparison, and systematic validation. We also use multiple data sources (interviews, social media, speeches) and multiple model architectures.

@clint0475.bsky.social @sebstier.bsky.social @johannesbreuer.com @alessandronai.bsky.social @adamramey.bsky.social @kgattermann.bsky.social

Our results suggest that computational methods can capture meaningful personality cues in text. We also highlight several limitations and next steps—curious to hear your thoughts and feedback on the paper! :)

For 2) we confirm previous findings that Donald Trump signals more agency-related cues, whereas Kamala Harris signals more communion-related cues during their presidential debate

For 1), we see a clear and consistent negative relationship between the parties CHES ("left-right") score and the politicians’ share of communion. This association is particularly
pronounced for the economic dimension (left panel), but also present in the cultural dimension (right panel).

We also demonstrate that our measure is sensitive to partisan differences for (1) aggregated communion scores of politicians across political parties in the German Bundes-
tag, and (2) individual scores for the U.S. Presidential Debate between Donald Trump and Kamala Harris on September 10, 2024

We observe that Deepseek-V3 achieved the strongest performance, closely followed by GPT-4o. At the same time, traditional methods (SVM, XLM-RoBERTa) showed weaker results across validation steps (e.g., comparison with human labels (left plot) & functional tests with designed examples (right plot)).

💻 We then apply a systematic research design that includes
1) extensive operationalisation of two traits (agency and communion) using a validated framework
2) Human labelling and validation, and
3) multiple methods (SVM, XLM-RoBERTa, GPT-4o, Llama-3-8B, Deepseak-V3) and measurement strategies!

We started conceptualizing how "politicians' personality" can be measured by focusing on observable personality cues as the main empirical indicators in language. These cues reflect an amalgamation of politicians' true intrinsic traits shaped by (unobservable) strategic considerations.

🎉 🎉🎉