생성형 인공지능 DeepSeek 기반 TOPIK 논증 글쓰기의 구조 분석 -ChatGPT 텍스트 분석과 비교를 통해서-

초록

This study aims to compare and analyze the structural characteristics of TOPIK argumentative writing generated by DeepSeek and ChatGPT from a text structure perspective, while identifying limitations and suggesting improvements for AI models. Unlike previous studies that focused solely on a single AI model, usually ChatGPT, this research incorporates a comparative approach that includes a newer model, DeepSeek, to explore potential for mutual advancement. The comparison revealed that both models displayed similar internal and external structural tendencies. However, DeepSeek showed relatively lower structural accuracy than ChatGPT. Both models shared three major structural issues: excessive paragraph segmentation, flawed body organization, and redundant conclusions. These limitations stem from a mismatch between the general nature of training data used in generative AI and the specific goals of targeted writing tasks, such as the TOPIK argumentative writing. To address these issues, this study proposes an AI prompt optimization method using a “condition + task prompt” format. Examples of responses generated using this method demonstrated improved structural accuracy. In conclusion, this research suggests structural improvement strategies for text generation based on a “problem prevention and effective generation” approach, offering practical insights at a time when general purpose generative AIs are widely used. Furthermore, if the National Institute for International Education, the organizer of TOPIK, develops a specialized AI model tailored for tasks like Question 54 argumentative writing, the methodology and findings of this study could serve as a valuable reference.

키워드