KorPatSTS 기반 특허 유사 기술 대조 학습을 활용한 특허 문헌 검색 모델에 관한 연구

This study proposes an advanced deep-learning–based patent retrieval model, NCE-KorPat, and a high-quality training dataset, KorPatSTS (Korean Patent Semantic Textual Similarity), to more precisely assess grounds for rejection arising from technological redundancy during the patent application process. The model recommends cited patent documents based on semantic and technical similarity between an application and prior art patents. KorPatSTS is a sentence-level dataset of similar patent-technology sentence pairs, drawing on the expertise of the Korea Ministry of Intellectual Property (MOIP) AI Examiner Advisory Group. The dataset aligns the technical constituent elements of claims in an application patent and the corresponding sentences in a detailed description, which explains those elements by considering the matching portions of prior art patents cited as grounds for rejection, thereby forming highly precise sentence-level correspondence pairs. In this study, the NCE-KorPat model was first developed by fine-tuning KorPatBERT, a patent-domain–specific language model, for CPC subgroup-level classification and then subsequently applying contrastive learning and optimization using the KorPatSTS dataset. When applied to Korean patent retrieval experiments, the proposed model demonstrated superior performance, outperforming both the previously best-performing Korean embedding models and state-of-the-art global embedding models. To the best of our knowledge, this study represents the first attempt to construct similar-technology sentence pairs systematically by integrating the domain expertise of Korean patent examiners and directly applying them to a practical patent retrieval model. The proposed approach is expected to substantially contribute to improving the accuracy and efficiency of patent examinations in the future.