A Bimodal Domain-Specific Deep Learning Approach for Automated Patent Classification

• Main Authors: LIAO, WEN-HUI, 廖文慧
• Other Authors: HSIAO, RONG-SHUE
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/wtu56h

碩士 === 國立臺北科技大學 === 電子工程系 === 107 === Patent classification is fundamental to search patents in huge patent database. An automatic patent classification system is beneficial for subject technology to be effectively compared with previous cases of relevant technology, and further improve the efficiency of patent analysis. In this research, we present a bimodal domain-specific deep learning approach for automated patent classification. Classification is based on International Patent Classification (IPC), which is published and managed by the World Intellectual Property Organization (WIPO). The system comprises word embedding pre-trained by US patent documents to convert word into word vector, and adopts CNN, CNN-BiLSTM and Hierarchical Attention Model in deep learning for text classification. The research further uses a bimodal data processing method to fuse the image-processed features of patent with the text-embedded vector-converted features to try to improve the classification accuracy. Training and testing data came from patent documents granted by the United States Patent and Trademark Office, which was classified into 8 categories to IPC at section level, 131 categories at class level, and 646 categories at subclass level. The experimental results showed that the system, comprising word embedding pre-trained by US patent documents and Hierarchical Attention Model for feature extraction and classification, obtained a better performance compared with previous researches. Among them, the system reaches a micro-average precision of 61.66% at IPC class level and 43.25% at IPC subclass level using claim as input.