To provide information to foreigners promptly and efficiently, we are conducting research on machine translation for texts of news reports, broadcast programs and newspaper articles and on news scripts with reading assistance information to help understand Japanese news.

■Machine translation of news and program content

  Broadcasters translate Japanese into foreign language to provide information for non-native speakers. For the speedy and efficient production of foreign language content, we are researching machine translation⁽¹⁾. The mainstream of recent machine translation technologies is a method that collects a huge amount of parallel translation data and trains a translation model using neural networks. In FY 2018, we created 500,000 pairs of Japanese and English news sentences by translating Japanese news manuscripts manually. Using this data, we prototyped a Japanese-English machine translation system. In addition, we developed an interface that allows easy manual correction of the machine translation results by mapping Japanese and English words and assigning the degree of translation reliability for English words. This research was conducted in cooperation with the National Institute of Information and Communications Technology (NICT).
  Furthermore, we are researching English-Spanish machine translation with the aim of supporting the production of Spanish closed captions for video-on-demand (VOD) service programs on NHK WORLD. In FY 2018, we newly prepared 40,000 pairs of closed caption sentences from VOD programs and 10,000 pairs of sentences from scripts of NHK WORLD RADIO JAPAN as parallel data of English and Spanish.

■Machine translation of newspaper articles

  With the aim of facilitating communication between non-Japanese and Japanese in business scenes, we are researching machine translation technologies for conversations and small talk in meetings and social occasions and newspaper articles in cooperation with five external institutions. NHK is in charge of a newspaper article translation technology. In FY 2018, we conducted Japanese-English machine translation experiments using parallel data generated by automatically mapping Japanese sentences and English sentences of Jiji Press. The experiments achieved better results both in objective evaluation and subjective evaluation than existing translation systems. We also identified research subjects that we should work on in the future, such as support for translation of new words, under-translation, mistranslation, over-translation, and translation with context. This research was supported by NICT as part of a project titled "R&D of Deep Learning Technology for Advanced Multilingual Speech Translation."

■News scripts with reading assistance information

  In FY 2018, we produced 64 articles of news scripts with reading assistance information for experimental purpose and investigated the effect of such news scripts for non-native speakers in Japan. The results demonstrated that reading assistance information helps improve their understanding of news. This research was conducted in cooperation with Utsunomiya University.

 

[References]
(1)	I. Goto and H. Tanaka: "Detecting Untranslated Content for Neural Machine Translation," Journal of Natural Language Processing, Vol. 25, No. 5, pp.577-598 (2018) (in Japanese)
(2)	I. Goto, H. Mino, T. Kumano and H. Tanaka: "Proposal of News with Reading Assistance," ITE Journal, Vol.72 No.8, p. J113-J115 (2018) (in Japanese)