Conceptual similarity as aggregation over feature sets in geometric spaces

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psychology

Background:

• Conceptual similarity judgments are fundamental to cognition, learning, and reasoning.
• Classical models struggle with judgment asymmetries, while modern geometric models lack explanatory power for these effects.
• Existing models fail to capture nuanced human judgments like asymmetry and triangle inequality violations.

Purpose of the Study:

• To develop a novel modeling framework integrating classical and contemporary approaches to conceptual structure.
• To create a similarity function capable of explaining classic judgment effects like asymmetry.
• To validate the model's predictions against human judgments and established datasets.

Main Methods:

• Representing concepts as sets of high-dimensional feature embeddings derived from natural language descriptions.
• Developing a novel similarity function tailored for set-based embeddings.
• Evaluating model predictions using a behavioral study with abstract concepts (countries) and the Nelson free word association dataset.
• Formalizing the connection between the proposed model and Tversky's Contrast Model.

Main Results:

• The proposed model successfully accounts for classic conceptual similarity judgment effects, including asymmetry.
• The model demonstrates superior performance compared to alternative approaches in predicting human judgments.
• A formal link was established between the new framework and Tversky's Contrast Model, bridging classic and modern theories.

Conclusions:

• The developed framework effectively integrates geometric and set-based methods for modeling conceptual structure.
• This approach offers a scalable and data-driven alternative that captures complex human judgment phenomena.
• The findings provide a generally applicable framework for understanding conceptual similarity in cognitive science.