시각언어모델 기반 맥락 추출과 한국어 대규모 언어모델 파인튜닝을 통한 여행 블로그 콘텐츠 자동 생성

초록

In this paper, we propose an automated system for generating Korean travel blog content by integrating Vision-Language Model (VLM)–based context extraction, fine-tuned Korean language model, and a text-to-image (T2I) generation model via prompt engineering. Using the Qwen2-VL model, we analyze travel photos to extract visual context and produce emotionally nuanced captions. We then leverage a large-scale crawled corpus of real-world travel blogs to fine-tune HyperCLOVA X, enabling it to create natural, storytelling-oriented blog text. In addition, we employ a travel-specific prompt engineering approach with DALL-E to generate custom postcards and stamp images, providing users with intuitive and creative value for their travel records. A robust system prompt minimizes hallucinations while preserving expressive writing. User surveys indicate that the fine-tuned model is, on average, 60.9% more specialized in travel-related content. These findings demonstrate that the proposed approach can significantly reduce manual effort while producing high-quality travel blog content, suggesting new possibilities for AI-based content creation in the tourism domain.

키워드