Owing to the difficulty of constructing large-scale datasets and the scarcity of Korean-language resources, recent deep learning-based patent retrieval research gaces limitaions in improving model performance. To address these challenges, this study proposes a methodology for automatically building a large-scale patent retrieval dataset from Korean patent documents. The method automatically extracts semantically related pairs of technical components between patent applications and cited prior art using claim comparison tables in office action notices. In addition, the sentences that are most similar to each technical component are extracted from both the patent application and the cited prior art documents. Korean patent XML parsing techniques are combined with a KorPatBERT-based CPC classification model, and a hybrid similarity measure integrating sentence embedding–based semantic similarity with lexical similarity is employed. Subsequently, a large-scale, high-quality dataset approximately 19 times larger than a manually constructed expert dataset was built and validated through large-scale experiments simulating real-world retrieval environments. Experimental results indicate that models trained on the automatically constructed dataset achieved Top-70 accuracy comparable to or better than those trained on expert-built datasets. Accordingly, this study presents a practical and cost-effective approach for constructing high-quality Korean patent retrieval datasets and demonstrates improved performance and real-world applicability.