Bilingual Textual Similarity in Scientific Documents

Abstract

Maps of science visualizing the structure of science help us analyze the current spread of science, technology, and innovation (ST&I). ST&I enterprises can use the maps of science as competitive technical intelligence to anticipate changes, especially those initiated in their immediate vicinity. Research laboratories and universities can understand their environmental changes and use the map for their research management. However, traditional maps based on bibliometrics, such as citation and cocitation, have difficulty in representing recently published papers and ongoing projects that have few or no references; thus, maps based on contents, i.e., text-mining, have been developed in recent years for locating research papers/projects, for example, using word and paragraph vectors. The content-based maps, however, still pose difficulty in comparing documents in different languages. Therefore, aiming to construct a bilingual (English and Japanese) content-based map of science for the analyses of ST&I information resources in different languages, this article proposes a method for creating word and paragraph vectors corresponding to bilingual textual information in the same multidimensional space. In a comparison of 11 methods for generating document vectors, we confirmed that the best method achieved 87% accuracy of the bilingual content matching based on 10$\,$000 IEEE papers. Finally, we published a map of approximately 150$\,$000 funding projects of the National Science Foundation, Japan Society for the Promotion of Science, and Japan Science and Technology agency from 2013 to 2017.