Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Aligning Large Language Models with Diverse Political Viewpoints

Dominik Stammbach, Philine Widmer, Eunjung Cho, Caglar Gulcehre, Elliott Ash

TL;DR

This work aligns LLMs with diverse political viewpoints from 100,000 comments written by candidates running for national parliament in Switzerland to generate more accurate political viewpoints from Swiss parties, compared to commercial models such as ChatGPT.

Abstract

Large language models such as ChatGPT exhibit striking political biases. If users query them about political information, they often take a normative stance. To overcome this, we align LLMs with diverse political viewpoints from 100,000 comments written by candidates running for national parliament in Switzerland. Models aligned with this data can generate more accurate political viewpoints from Swiss parties, compared to commercial models such as ChatGPT. We also propose a procedure to generate balanced overviews summarizing multiple viewpoints using such models. The replication package contains all code and data.

Aligning Large Language Models with Diverse Political Viewpoints

TL;DR

This work aligns LLMs with diverse political viewpoints from 100,000 comments written by candidates running for national parliament in Switzerland to generate more accurate political viewpoints from Swiss parties, compared to commercial models such as ChatGPT.

Abstract

Large language models such as ChatGPT exhibit striking political biases. If users query them about political information, they often take a normative stance. To overcome this, we align LLMs with diverse political viewpoints from 100,000 comments written by candidates running for national parliament in Switzerland. Models aligned with this data can generate more accurate political viewpoints from Swiss parties, compared to commercial models such as ChatGPT. We also propose a procedure to generate balanced overviews summarizing multiple viewpoints using such models. The replication package contains all code and data.
Paper Structure (23 sections, 2 equations, 9 figures, 5 tables)

This paper contains 23 sections, 2 equations, 9 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data
  4. Methods
  5. Results
  6. Diversity of Generations
  7. Quantitative Evaluation
  8. Human Evaluation
  9. Discussion and Conclusion
  10. Choice of models and Alignment algorithms.
  11. Choice of metadata for conditional generation.
  12. Data availability.
  13. Further data.
  14. Contested topic.
  15. Intended use case.
  16. ...and 8 more sections

Figures (9)

  • Figure 1: Prompt for conditional generation. Varying attributes are party P, language L, and political issue Q.
  • Figure 2: Average diversity of replies within a political issue, measured with Jaccard similarities (lower Jaccard similarity means higher diversity).
  • Figure 3: Win rates by different models.
  • Figure 4: Pseudocode for generating and synthesizing answers.
  • Figure 5: Overview of political preferences of ChatGPT in Switzerland for the 2023 national elections of parliament (source: smartvote.ch)
  • ...and 4 more figures