Synonym Identification via Category Specificity and KL Divergence

The patent segments the synonym identification process into two distinct phases: (1) learning phase using only frequent query contexts to establish reliable synonym relationships, and (2) application phase that extends these learned synonyms to infrequent contexts through category-based generalization. This segmentation allows the system to maintain high accuracy for common contexts while achieving broader coverage for rare contexts.

The patent introduces category information as an intermediary element that bridges frequent and infrequent query contexts. By identifying that entities belong to the same category (e.g., both San Francisco and Grey Bull are cities), the system can transfer synonym relationships from frequent contexts to infrequent contexts, thereby extending coverage without compromising accuracy.

The patent applies local quality by differentiating between coherent and non-coherent categories. Instead of uniformly applying category-based generalization to all categories, the system selectively applies it only to coherent categories that have been validated through statistical analysis (specificity score thresholding). This ensures that synonym substitution is extended to infrequent contexts only when the category provides meaningful contextual information.

The patent changes the parameter of category coherence by introducing a specificity score metric and threshold. Categories are evaluated based on their specificity scores, and only those exceeding the threshold are considered coherent and eligible for synonym generalization. This parameter-based filtering mechanism ensures that only categories with sufficient contextual relevance are used for extending synonym coverage.