DPO fine-tuning enhances the quality of sustainable finance term definition generation, supporting semantic change auditing of EU development financing documents.

A newly published academic study proposes that fine-tuning dictionary-definition generation models using Direct Preference Optimization (DPO) can significantly improve their performance on the Lexical Semantic Change Detection (LSCD) task. This approach aims to address the issue where existing models often generate excessive redundant definitions for the same word meaning—a phenomenon known as “over-splitting of senses.”

The research team conducted quantitative evaluations on standard academic LSCD benchmarks and further carried out qualitative validation in real-world scenarios—focusing on documents related to EU development financing projects to audit the diachronic evolution of sustainability-related terms, thereby supporting semantic consistency reviews for sustainable finance information disclosure by EU institutions.

All models, code, and prediction results have been made open source and are hosted in a GitHub repository (https://github.com/ltgoslo/dpo-for-definition-generation）。该论文以arXiv预印本形式发布，标识号为arXiv:2610.33357v1，发布时间为2026年10月27日04:00（UTC），符合时效性要求。