Items Outperform Adjectives in a Computational Model of Binary Semantic Classification.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psychology

Background:

• Semantic memory stores world knowledge, with distributional semantic models (DSMs) representing word meanings in vector spaces.
• Classic DSMs lack mechanisms for concept property specification, limiting their application in general semantic knowledge theories.
• Understanding how humans perform binary semantic classification tasks is crucial for refining semantic memory models.

Purpose of the Study:

• To develop a computational model for binary semantic classification tasks (e.g., size, animacy).
• To evaluate the efficacy of various DSMs and classification mechanisms.
• To propose a model consistent with the instance theory of semantic memory.

Main Methods:

• Developed a family of computational models using distributional semantic models.
• Implemented mechanisms for binary semantic classification based on concept properties.
• Evaluated models on a large dataset of over 1,500 words for classification accuracy and response time prediction.

Main Results:

• The most successful model created composite representations for semantic extremes (e.g., 'big' vs. 'small').
• This model achieved human-range performance in classifying words based on size and animacy.
• The model accurately predicted human response times in the classification task.

Conclusions:

• Humans likely use task-relevant instances representing semantic extremes to make classification decisions.
• The proposed model provides a computational framework for understanding semantic classification within instance theory.
• This work advances DSMs by incorporating a mechanism for concept property representation and utilization.