대조 학습을 이용한 유사 특허기술 탐색 시스템

초록

Prior to patent application, relevant existing patents must first be searched. Patent attorneys investigate previous patents and technology trends, but this can be a time consuming process; thus, automating the process has become imperative. Previous research has used natural language processing models to recommend similar patents by vectorizing these patents in latent space. In the current study, we propose using PatentBERT, a pretrained language model for patents, to vectorize words and reduce the search scope of patents using the contrastive loss function. This model can reduce the time needed for preliminary patent investigation and can be applied to other areas, such as screening research papers or searching for similar references.

키워드