?
Interpretable approach to detecting semantic changes based on generated definitions
This paper investigates definition modeling as an approach to semantic change detection, which offers the advantage of providing human-readable explanations, unlike traditional embedding-based approaches that lack interpretability. Definition modeling leverages large language models to generate dictionary-like definitions based on target words and their contextual usages. Despite its potential, practical evaluations of this method remain scarce. In this study, FRED-T5 was fine-tuned using the Small Academic Dictionary for the task of definition modeling. Both quantitative and qualitative assessments of definition modeling’s effectiveness in detecting semantic shifts within the Russian language were conducted. The approach achieved a Spearman’s rank correlation coefficient of 0.815 on the Rushifteval task, demonstrating strong alignment with expert annotations and ranking among the leading solutions. For interpretability, a visualization algorithm was proposed that displays semantic changes over time. In the qualitative evaluation, our system successfully replicated manual linguistic analysis of 20 Russian words that had undergone semantic shifts. Analysis of the generated meanings and their temporal frequencies showed that this approach could be valuable for historical linguists and lexicographers.