ScholarWorks@중앙대학교

초록

In the era of technological hegemony, protecting national core technology (NCT) has become critical, particularly in the display industry, where continuous leaks prompt the need for precise and automated classification. However, existing automatic labeling approaches, such as cooperative patent classification (CPC) codes, threshold-based labeling, and generative pretrained transformer (GPT)-based annotation, lack clear criteria and overlook domain-specific terms. To overcome these limitations, we propose Core-TechAnnotator, a novel reproducible framework, to assess and label NCT related to active-matrix organic light-emitting diodes by measuring the semantic similarity between embedded patent-document vectors. Rigorous quantitative and qualitative analyses using the analysis of variance, silhouette scores with the elbow method, and the Calinski-Harabasz Index, identify optimal label counts and clustering algorithms. The selected Kmeans method with cosine similarity and four-level rating balances distribution better than CPC, threshold, and GPT-4.1-nano methods, ensuring strong intraclass cohesion and high interclass separation. Evaluated on 14,639 patents from three anonymized companies and a multilayer perceptron, the Core-TechAnnotator method achieves 95.22% accuracy and a 0.95 F1-score. Beyond classification, the proposed method visualizes a patent-level graph, extracts NCT-related keywords, and assigns security ratings that guide resource allocation. These findings suggest that Core-TechAnnotator can enhance NCT protection and reduce leak risks and operational costs by pinpointing high-risk technology.

키워드