Towards efficient patent analysis: A large language model and BERT-refined methodology for keyphrase extraction

Abstract

Patents play a pivotal role in engineering design by safeguarding innovation, forecasting technical trends, and promoting knowledge sharing. However, the vast volume of patents and their complex technical descriptions pose significant challenges for effective analysis and information retrieval. To address these issues, we propose an integrated framework that combines large language models (LLM) and a BERT-refined approach for patent analysis. Specifically, patent titles and abstracts are first collected, and term frequency-inverse document frequency (TF-IDF) is introduced to extract candidate keyphrases. An LLM is then employed to refine these keyphrases by filtering irrelevant terms and identifying significant keywords. Subsequently, a fine-tuned BERT model is developed for named entity recognition (NER) to extract domain-specific keywords, which are further refined into keyphrases through our BERT-refined keyphrase extraction (BRKE) method. Experimental results on a large dataset of USPTO patents demonstrate the effectiveness of the proposed BRKE. It achieves the highest F1-score of 52.97% when the top-10 keyphrases are retained, outperforming keyBERT, YAKE, and RAKE by 9.52%, 6.1%, and 2.35%, respectively. By enhancing the accuracy of patent keyphrase extraction, our contributions make patent analysis more efficient and accessible to both analysts and design engineers.