트랜스포머 기반 BERT를 활용한 비특허 문헌 자동 분류의 성능 향상 방안 연구

Purpose Non-Patent Literature (NPL) plays a crucial role in patent examination but is difficult to classify due to its vast volume and diverse formats. This study proposes an approach utilizing BERT-based Natural Language Processing (NLP) techniques to automatically classify NPL and assign Cooperative Patent Classification (CPC) codes. Design/methodology/approach NPL abstracts cited in U.S. patents were collected from KIPRIS Plus. The study applied vectorization techniques such as TF-IDF, SBERT, and anferico/bert-for-patents, and compared classification performance using Logistic Regression, XGBoost, LightGBM, BERT, RoBERTa, and anferico/bert-for-patents models. Findings The anferico/bert-for-patents model, specialized for patent documents, achieved the highest classification accuracy (56.3%) and effectively captured the semantic representation of NPL. This study contributes to improving NPL search and classification efficiency, enhancing the prior art search process in patent examination.