LLM을 이용한 최적 특허빅데이터분석 모형 분석

Patent Evaluation Rating is an online patent rating service that objectively evaluatesthe excellence of patents by country by utilizing patent information extracted frompatent specifications, bibliographic information, administrative information, etc. Byusing Patent Evaluation Rating, a small number of core patents can be selected froma large number of patents and provides meaningful insights to patent applicants. Patent Evaluation Rating utilizes patent qualitative and quantitative factors, andquantitative factors are numerical data such as the number of citations and thenumber of family patents, while qualitative factors are established by reflectingexpert opinions based on bibliographic information. In this paper, we apply largescale language models (LLMs) to the bibliographic information of patents to solvethe above problems because expert opinions used as qualitative indicators can bebiased due to their subjective nature. Therefore, we analyze how and on what basisexperts assign patent classes based on patent bibliographies. To do this, thecollected patent bibliographies are first preprocessed and analyzed using pre-trainedlanguage models. Then, we explore the most appropriate language model for thepatent domain.