LSC-Pattern: Interpretable Lexical Semantic Change Analysis via Diachronic Sense Distributions
Abstract
AbstractLexical semantic change detection identifies words whose meanings change over time, but it often cannot explain which senses and semantic properties are responsible for the change. We propose LSC-Pattern, a framework for interpretable lexical semantic change analysis that combines dictionary-grounded diachronic sense distributions with interpretable semantic dimensions. The framework first assigns each target-word usage to an explicit dictionary sense and then analyzes the resulting sense distributions along dimensions such as valence, semantic scope, and metaphoricity. We instantiate it for Japanese with LLM-based word sense disambiguation, which reaches 0.815 accuracy against majority-vote human annotation and outperforms five baselines; its sense-distribution change scores correlate with JaSemChange human judgments (Spearman’s ρ of 0.821 on seven overlapping entries and 0.580 on all 20). Case studies illustrate how sense redistribution and operationalized semantic signals can jointly support the discovery and interpretation of candidate types of change.