생성형 AI를 활용한 지역어 데이터 구축 방안 연구 - 전남 지역어를 중심으로 -

Abstract

This study explores methods for constructing regional dialect data using generative artificial intelligence (AI), focusing on the Jeonnam dialect. Recognizing regional dialects as vital cultural and linguistic heritage, the study aims to enhance awareness of their value and propose AI-based strategies for their systematic preservation and revitalization. Experiments compared outputs from ChatGPT (o3-mini) and Naver HyperCLOVA X using three Jeonnam dialect dictionaries, generating and analyzing 1,000 sample sentences. The results reveal that existing large language models (LLMs) trained primarily on standard Korean exhibit significant limitations in handling phonological and syntactic variations within dialectal sentences. Particularly, word-order patterns, sentence endings, and contextual nuances were inadequately represented, demonstrating that current vector embedding and similarity-based approaches are not optimized for dialectal data. These findings highlight the necessity of developing dialect-adaptive LLMs that can contextually learn and reproduce non-standard linguistic structures. Nonetheless, leveraging preexisting dialect resources and continuously refining high-quality data through partial retraining can serve as a practical methodology for expanding and improving reginal dialect corpora. The study concludes that generative AI offers a promising avenue for cost-effective, scalable regional language data construction. However, human evaluations and expert review remain essential to ensure naturalness and authenticity. By integrating linguistic knowledge, socio-cultural context, and computational adaptability, this research contributes to reducing linguistic imbalance within the AI ecosystem and supporting inclusive Korean language technologies.