Llama meets EU: Investigating the European Political Spectrum through the Lens of LLMs

Abstract

Instruction-finetuned Large Language Models inherit clear political leanings that have been shown to influence downstream task performance. We expand this line of research beyond the two-party system in the US and audit Llama Chat in the context of EU politics in various settings to analyze the model's political knowledge and its ability to reason in context. We adapt, i.e., further fine-tune, Llama Chat on speeches of individual euro-parties from debates in the European Parliament to reevaluate its political leaning based on the EUandI questionnaire. Llama Chat shows considerable knowledge of national parties' positions and is capable of reasoning in context. The adapted, party-specific, models are substantially re-aligned towards respective positions which we see as a starting point for using chat-based LLMs as data-driven conversational engines to assist research in political science.

Figures (15)

• Figure 1: Examples of responses to EUandI question from LLMs adapted in different euro-party speeches, i.e., left-wing GUE/NGL and far-right ID parties.
• Figure 2: The different templates we use to audit the models. Setting A and B have the same options as the Main Question Template in 3rd person. $S$ denotes a statement from the EUANDI questionnaire, $T$ is the title of a debate, $U$ an utterance (speech), $O$ a member state, $P$ a national party name and $J$ a justification on a specific topic.
• Figure 3: Results for contextualized auditing in setting C for German parties, i.e., predicted party based on justifications. Individual rows represent the target party and the bars refer to the predicted party by Llama Chat.
• Figure 4: Radar plots for the adapted models (Section \ref{['sec:adaptation']}) on EUandI. The radars depict the polarity of each model across the 7 thematic categories (Section \ref{['sec:datasets']}). The yellow areas represent the polarity of the baseline model, Llama Chat, out-of-the-box. In contrast, the gray areas represent the polarity based on the model's options (automatic evaluation). The dark-shaded areas, e.g., green for the Greens/EFA party, represent the polarity based on the party's options. In contrast, the light-shaded areas represent the polarity based on the model's justifications (manual evaluation). We present an enlarged version of the radars plots in Figure \ref{['fig:radars_big']}.
• Figure 5: Distribution of speeches in the newly released EU Debates dataset per EUandI thematic topic.