There are growing needs for patent analysis using Natural Language Processing (NLP)-based approaches. Although NLP-based approaches can extract various information from patents, there are very few approaches proposed to extract those parts what inventors regard as novel or having an inventive step compared to all existing works ever. To extract such parts is difficult even for human annotators except for well-trained experts. This causes many difficulties in analyzing patents. We propose a novel approach to automatically extract such keywords that relate to novelties or inventive steps from patent claims using the structure of the claims. In addition, we also propose a new framework of evaluating our approach. The experiments show that our approach outperforms the existing keyword extraction methods significantly in many technical fields.