BERT를 활용한 기술 특허 분류 성능 평가

Despite the widespread impact on various fields such as finance, insurance, human resource management, accounting, manufacturing, supply chain management, IT management, and customer service due to the advent of Intelligent Process Automation(IPA) combining RPA and artificial intelligence, research on IPA applying deep learning-based natural language processing to technology patent analysis has not sufficiently developed. This study designs a technology patent classification system based on patent data on carbon dioxide capture and utilization, natural language preprocessing techniques, and BERT and BERT-derived models, and compares and analyzes accuracy, kappa coefficient, and F1-score. The main results are summarized as follows: First, among the five CCU technology classifications, the BERT model showed better performance than the ELECTRA model, suggesting that the BERT model's MLM method can understand the overall context information more effectively. Second, in patent summary classification, it showed higher performance than the first claim classification, which is judged because the language models are trained with various types of texts and based on general language use and context, thus more effectively processing patent summaries that summarize the entire patent content. This study presents the potential to significantly impact business strategy development and enhance technological competitiveness as meaningful IPA research applying BERT and BERT-derived models to technology patent classification.