Embeddings for Preferences, Not Semantics

Short summary

Researchers propose embeddings that measure preference similarity rather than semantic similarity for collective decision-making. They formalize decoupling preference signals (stance, values) from semantic nuisance (style, wording) using synthetic training data. Validation across 11 deliberation datasets shows preferential embeddings significantly outperform standard embeddings for opinion clustering.

• •New embedding approach prioritizes preference similarity over semantic similarity
• •Formalized as invariance problem: decoupling preference signals from stylistic nuisance
• •Validated across 11 online deliberation datasets with improved accuracy

Generated with AI, which can make mistakes.