Comparative Study of Different Approaches in Hierarchical Multi-label CPC Classification
Emre Doruk Taşkın, Baran Kılıç, Ferayenur Bozkurt, İrfan Ulaş Geçin, Özcan Somuncu, Pınar Duygulu Şahin · 2025
Accurate and consistent classification of patents is crucial for intellectual property management, analyzing technology trends and improving patent search processes. The Cooperative Patent Classification (CPC) system provides a multi-label and hierarchical structure, allowing patents to be organized according to specific technical fields. However, due to this complex structure, manual classification processes become both time-consuming and error-prone. In this paper, four different deep learning-based approaches for CPC classification are compared: (i) flat (non-hierarchical) classification, (ii) hierarchical multitask learning, (iii) hierarchical loss function, and (iv) classification using semantic similarity via label embeddings.