Core Concepts
ChOiRe presents a four-step framework to predict human opinions by differentially modeling explicit and implicit personae, achieving state-of-the-art effectiveness.
Abstract
ChOiRe introduces a novel approach to align language models with human opinions. The framework consists of four steps: analyzing explicit personae, ranking implicit persona opinions, Chain-of-Opinion reasoning, and ensuring answer consistency. By filtering out irrelevant attributes and selecting the most valuable opinions, ChOiRe improves opinion prediction significantly. The method enhances fine-tuning of opinion-aligned models and demonstrates strong performance in aligning models with individual opinions.
Stats
ChOiRe achieves new state-of-the-art effectiveness by improving previous techniques significantly by 3.22%.
ChOiRe's Steps (i) and (ii) can significantly better fine-tune opinion-aligned models by up to 18.44%.
Quotes
"Aligning language models with human opinion is challenging yet vital to enhance their grasp of human values, preferences, and beliefs."
"We propose ChOiRe, a four-step solution for opinion prediction leveraging LLMs’ strong data evaluation and analytic capabilities."
"ChOiRe achieves new state-of-the-art (SOTA) in opinion alignment effectiveness and reliability."