Context Matters: Recovering Human Semantic Structure from Machine Learning Analysis of Large-Scale Text Corpora

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• Machine learning models infer concept relationships from text but often misalign with human judgments.
• Understanding how semantic knowledge is organized and used in human judgments is crucial.

Purpose of the Study:

• To improve the accuracy of machine learning models in predicting human semantic judgments.
• To investigate the role of semantic context in human judgment and AI model performance.

Main Methods:

• Trained machine learning algorithms using contextually-constrained text corpora (domain-specific Wikipedia subsets).
• Developed a novel method for dimensionality reduction of embedding models to enhance contextual relevance.
• Compared algorithm predictions with empirical human similarity and feature judgments.

Main Results:

• Contextually-constrained embeddings significantly improved predictions of human judgments.
• The dimensionality reduction method enhanced predictions from contextually-unconstrained models.
• The approach bridges the gap between large-scale data analysis and direct human perception.

Conclusions:

• Semantic context is critical for accurate AI-driven understanding of human semantic knowledge.
• This novel approach enhances AI's ability to model human judgment and semantic representations.
• Leveraging online corpora with contextual constraints can advance cognitive science research.