Extracting metadata of scientific references in patents based on combination of representation learning and machine learning

Abstract

ABSTRACTThe scientific references in patent (SRPs), one type of non‐patent references (NPRs), is the best representative of scientific knowledge cited in patents (SKCP). However, SKCP is often indicated by NPRs rather than SRPs currently, which may reduce the accuracy of the analysis of science & technology transfer and transformation. In the other hand, the content of SRPs is important for subsequent in‐depth patent analysis, which requires more accurate SRPs' metadata extraction. In order to enhance the accuracy of SRPs identification and SRPs' metadata extraction, this paper uses the representation learning method to learn vectors of the NPRs and the parts separated by delimiters with each SRP (metadata), respectively. These vectors conclude the context information and the same type of metadata may appear in the same position in vector space. Then the problem of SRPs identification and SRPs' metadata extraction are transferred as the automatic categorization with the machine learning method. The experiment is conducted in the field of gene where patents cite lots of SRPs. The result shows that in 1000 patent citations, the accuracy of SRP identification reaches 99.27% and the accuracy of SRP title extraction reaches 90.57% which is 4.51% higher than the pure machine learning method.