Flexibility in conceptual combinations: A neural network model of gradable adjective modification

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Linguistics

Background:

• Conceptual combination is key to human cognition.
• Gradable adjectives (e.g., 'light', 'dark') demonstrate flexible meaning based on context.
• Understanding how semantic composition works is an ongoing challenge.

Purpose of the Study:

• Investigate neural network encoding of gradable adjectives in adjective-noun pairs.
• Use brightness as a test case for perceptual feature modulation.
• Explore how neural networks generalize to novel combinations.

Main Methods:

• Trained a neural network to predict human brightness ratings for nouns and adjective-noun pairs.
• Assessed generalization to untrained adjective-noun combinations.
• Analyzed neural network internal representations to understand information encoding.

Main Results:

• Neural networks demonstrated flexible learning of gradable adjectives.
• Models initially relied on adjective information, then modulated by noun context.
• Model outputs replicated non-additive feature modulation observed in human data.

Conclusions:

• Neural networks can flexibly encode gradable adjectives, mirroring human semantic composition.
• Findings provide insights into the computational mechanisms of conceptual combination.
• Results generate testable predictions for future research on language and cognition.