LoRA 파인튜닝과 RAG의 하이브리드 결합 모델 기반 특허요약 프레임워크

초록

This paper investigates the data memorization issue in patent document summarization using Large Language Models (LLMs) and proposes a hybrid framework of LoRA fine-tuning and Retrieval-Augmented Generation (RAG) to overcome this problem. During the experimental process, a significant amount of duplicated data was identified within existing patent datasets, demonstrating that the LoRA model, a Parameter-Efficient Fine-Tuning (PEFT) technique, performs simple memorization rather than inference, thereby distorting performance metrics. To address this, this study constructed a clean dataset of 500 samples ensuring data integrity and defined a summarization task reflecting the core logical structure of patents: 'Problem-Solution-Benefit (PSB).' Experimental results show that the proposed model achieved a BERTScore of 0.985. Notably, it demonstrated substantial inference capabilities rather than simple memorization by maintaining high semantic similarity despite low lexical similarity (0.2040) between the training and test sets.

키워드