Centroid analysis: Inferring concept representations from open-ended word responses.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psychology

Background:

• Understanding conceptual representations is crucial in cognitive science and psychology.
• Existing methods may lack the granularity to assess individual and group-level conceptual structures effectively.
• Mapping open-ended responses onto semantic vector spaces offers a novel approach.

Purpose of the Study:

• To propose and evaluate centroid analysis, a new method for measuring concepts.
• To assess the efficacy of centroid analysis at both individual and group levels.
• To determine optimal parameters for centroid analysis using various semantic models and datasets.

Main Methods:

• Developed centroid analysis to represent concepts as the geometric center of response semantic vectors.
• Utilized two distributional semantic models (e.g., fastText) for meaning-to-vector mapping.
• Evaluated methods across diverse datasets, response types (free association, feature generation), and lexicon sizes.

Main Results:

• At the group level, optimal performance was achieved using multiple free associations (70 unique, 245 total responses), fastText, and response frequency weighting.
• At the individual level, best results were obtained with fastText and a minimum of eight responses per item.
• Centroid analysis demonstrated robust measurement of conceptual structures across different conditions.

Conclusions:

• Centroid analysis provides a robust method for quantifying conceptual representations at individual and group levels.
• The method allows for detailed investigation into how factors like culture, language, and personal experience shape concepts.
• Optimal parameters were identified for maximizing the accuracy of centroid analysis in different research contexts.