AI-Powered Patent Classification: A Hybrid Approach Merging Algorithms and LLMs

Large Language Models (LLMs) offer potential for patent analysis but are challenged by the length, complexity, and specialized language of patent documents, hindering systematic analysis. This study introduces an LLM-based approach to generate structured technical summaries of patents, aiming to improve classification efficiency and accuracy. The methodology involved comparing LLM-based patent classification performance using these summaries against full-text and claims-only representations across various test cases. Results demon-strate that summaries significantly improve classification: precision increased by approximately 11% over full-text and 20% over claims, while accuracy rose by 10% and 14%, respectively. F1-scores also showed substantial gains, with compa-rable recall, indicating effective retention of crucial information. A case study on wind turbine patents validated the method’s practical utility. The study concludes that LLM-generated structured technical summaries offer a robust and efficient input for patent classification, providing a promising pathway for scalable and reliable patent intelligence.