Professor Sadao Kurohashi of the Graduate School of Informatics, Kyoto University and Toshiaki Nakazawa, a researcher of the Information Planning Department of the Science and Technology Agency, have developed a machine translation system for Japanese-Chinese and Chinese-Japanese science and technology papers using a neural network *. From May 2018, 5, the system was provided on the website of the Japan Science and Technology Agency.
In recent years, the number of science and technology papers published by China has become the second largest in the world after the United States, and there is an increasing demand for searching and browsing science and technology literature written in Chinese in Japanese.While machine translation, which can be processed quickly and inexpensively, is expected to become widespread, there are various technical terms and new words that represent the latest technology in the translation of scientific and technological information, so it is desired to improve the accuracy of machine translation. ..
Against this background, from 2013, Kyoto University Graduate School of Informatics, Science and Technology Promotion Organization Information Planning Department, and China Science and Technology Information Research Institute have collaborated to create high-precision neural machines for science and technology papers. We have been working on the "Japan-China / China-Japan Machine Translation Practical Application Project" to promote the practical application of translation systems.
In the project, we have prepared more than 400 million Chinese / Japanese bilingual corpora (= teacher data used for machine translation training) based on science and technology papers.These were learned by the neural machine translation engine developed this time, and mistranslations and omissions were reduced by preparing a bilingual dictionary.As a result, 97% of the evaluation criteria set in the development "contains most important information" and "easily understandable without excess or deficiency of information" achieved translation accuracy of about 6%.In addition, in the evaluation at the international machine translation workshop "WAT2016" for Asian languages held in December 12, it showed the highest accuracy in the machine translation task of science and technology information.
It is expected that the Japan-China / China-Japan machine translation system will be able to provide highly accurate translations even for cutting-edge technical information through the development of bilingual corpora and bilingual dictionaries.
* Neural networks are computational models devised to imitate the mechanism of information transmission in human brain cells.Computers are used for "deep learning" to understand complicated things, and in recent years, they are used not only in machine translation but also in a great many fields such as advanced image recognition, Go, and shogi.
