Distributional Measures of Semantic Abstraction

Area of Science:

• Computational Linguistics
• Cognitive Science
• Natural Language Processing

Background:

• Semantic abstraction is a core cognitive concept, crucial for efficient knowledge representation.
• Distributional hypothesis and computational measures (e.g., co-occurrence, neighborhood density) have been used to study semantic categorization.
• Existing research has not extensively compared computational abstraction measures across different conceptual tasks.

Purpose of the Study:

• To investigate semantic abstraction in English using variants of distributional measures.
• To compare the effectiveness of measures for abstract-concrete and generality-specificity distinctions.
• To identify reliable computational measures for different types of semantic abstraction.

Main Methods:

• Exploited variants of distributional measures, including frequency, word entropy, and neighborhood density (target-context diversity).
• Applied measures to distinguish between abstract-concrete (e.g., glory-banana) and generality-specificity (e.g., animal-fish) word pairs.
• Conducted experiments to evaluate measure performance and identify conceptual differences.

Main Results:

• Achieved precision higher than 0.7 in identifying lexical-semantic abstraction.
• Identified distinct reliable measures for abstract-concrete (neighborhood density variants) and generality-specific (frequency, word entropy) distinctions.
• Observed significant differences in results based on word class (nouns vs. verbs) and the impact of ambiguity.

Conclusions:

• More general words are more frequent and less surprising; abstract words appear in more diverse semantic contexts than concrete words.
• Distributional models require consideration of word classes and ambiguity for accurate conceptual categorization.
• Different distributional measures are effective for different facets of semantic abstraction.