부산대학교 도서관

We study Japanese-English machine translation technology to support NHK's production of English news. NHK Japanese news tends to have long sentences, and English news tends to translate a Japanese sentence into multiple English sentences. Therefore, it is necessary to translate English with an appropriate number of sentences. In this paper, we propose a method to control the number of sentences in a neural machine translation (NMT). The proposed method trains the NMT model by adding a tag indicating the number of English sentences to the top of Japanese sentence. We evaluated our method with Japanese-English bilingual sentences of news, and we confirmed that the proposed method improved the rate of coincidence between the number of the output sentences and the number of the reference sentences from 60.4% to 96.6%.