AI-Enhanced Semantic Feature Norms for 786 Concepts

Area of Science:

• Cognitive Science
• Psycholinguistics
• Computational Linguistics

Background:

• Semantic feature norms are crucial for understanding human conceptual knowledge.
• Traditional norming methods are labor-intensive, limiting concept and feature coverage.
• Existing datasets may not fully capture the richness of human conceptual knowledge.

Purpose of the Study:

• To introduce a novel approach for augmenting human-generated semantic feature norms using large language models (LLMs).
• To create an AI-enhanced feature norm dataset (NOVA: Norms Optimized Via AI) with verified quality.
• To evaluate the performance of the AI-enhanced dataset against human-only norms and word-embedding models.

Main Methods:

• Augmenting human-generated feature norms with LLM responses.
• Verifying the quality of norms against reliable human judgments.
• Comparing the AI-enhanced dataset (NOVA) with human-only datasets and word-embedding models in predicting semantic similarity.

Main Results:

• The NOVA dataset exhibits significantly higher feature density and concept overlap compared to human-only datasets.
• NOVA outperforms both human-only norm datasets and traditional word-embedding models in predicting semantic similarity judgments.
• The study validates the quality of LLM-generated norms through human judgment verification.

Conclusions:

• Human conceptual knowledge is more extensive than previously captured in norm datasets.
• Large language models (LLMs), when properly validated, are powerful tools for cognitive science research.
• The NOVA dataset offers a richer and more comprehensive resource for studying semantic representations.