Multi-stage fine-tuning of patent domain-specific DeBERTa for advanced patent landscape on SDGs/Decarbonization

Abstract

This study presents a multi-stage fine-tuning approach using DeBERTa for advanced patent analysis and landscaping on SDGs and decarbonization technologies. The method incorporates FI subclass estimation with the significant improved accuracy on extracting relevant technologies from patent documents. The model outperformed previous BERT-based approaches in various tasks and was applied to analyze Japanese and PCT international patent applications. Key findings include the continued leading R&D by Japanese companies in SDGs/decarbonization area and the rapid emergence of Chinese firms. The study also introduced the "Japio-Decarbonization Patent Index" which can identify companies filing highly decarbonization-oriented patents. This research demonstrates the effectiveness of advanced NLP techniques in patent analysis, providing valuable insights for innovation promotion and technology trend prediction in sustainable development.