Transformer networks of human conceptual knowledge

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Linguistics

Background:

• Understanding human knowledge representation is crucial for cognitive science.
• Existing models often struggle to capture the richness of real-world conceptual knowledge.
• Bridging computational models with psychological data offers a promising avenue for progress.

Purpose of the Study:

• To develop and validate a computational model that simulates human knowledge for numerous real-world concepts.
• To assess the model's ability to predict human judgments on concept features and semantic relationships.
• To identify key model properties essential for accurate prediction of human semantic cognition.

Main Methods:

• Utilized a pretrained transformer network, fine-tuned on participant-generated feature norms.
• Applied the model to stimuli from 25 semantic cognition experiments.
• Compared model performance against several variants to determine necessary predictive properties.

Main Results:

• The model successfully extrapolated and predicted human knowledge for novel concepts and features.
• It reproduced findings across semantic verification, concept typicality, feature distribution, and semantic similarity.
• Model properties crucial for accurate prediction were identified through comparative analysis.

Conclusions:

• Combining language and psychological data enables the creation of AI models with robust world knowledge.
• The developed model offers new applications for simulating naturalistic semantic verification and knowledge retrieval.
• This approach facilitates advanced modeling of real-world categorization, decision-making, and reasoning processes.