효율적인 프롬프트 엔지니어링과 사용자 피드백 기반 인터랙티브 Text-to-SQL 시스템 설계
Design of an Interactive Text-to-SQL System with User Feedback Based on Lightweight Models

초록

Purpose This study proposes an interactive framework to enhance the accuracy and practicality of Text-to-SQL by combining a lightweight LLM with user feedback. Methods Built on Llama 3.2_3B, we use structured Chain-of-Thought prompting and instruction tuning from simulated feedback logs on the Spider dataset for training and evaluation. Results The framework improves Execution Accuracy on Spider by 18.06 percentage points over the baseline and shows additional zero-shot gains on BIRD. Conclusion The approach enables reliable database interaction for non-experts and demonstrates the feasibility of secure, local, lightweight LLM-based Text-to-SQL systems.

키워드

Text-to-SQLPrompting Engin.User FeedbackLarge Languge ModelIntruction Tuning
제목
효율적인 프롬프트 엔지니어링과 사용자 피드백 기반 인터랙티브 Text-to-SQL 시스템 설계
제목 (타언어)
Design of an Interactive Text-to-SQL System with User Feedback Based on Lightweight Models
저자
조예진김무철이남연
DOI
10.35373/KMES.31.1.3
발행일
2026-03
유형
Y
저널명
한국경영공학회지
31
1
페이지
37 ~ 47