Korean Traditional Document Translation Using Transformer In Bidirectional-CRF

초록

This paper proposes a solution to solve the Out Of Vocabulary(OOV) problem in a framework built with a transformer-based machine translation algorithm. The translation input is a traditional Korean document written in Chinese characters, and the output is a decoding of modern Korean paragraphs written in Korean alphabet. We used the word2vec algorithm to represent symbolic characters as numeric vectors and used them as input to the converter. Also, to solve the OOV problem, Bi-Directional LSTM + CRF has been used. To show the validity of the data set, the Annals of the Joseon Dynasty were presented as translations prepared by experts. Another source was collected at Kyungpook National University (Diary dataset), which is much smaller than the Annals of the Joseon Dynasty. According to the BLEU score, after learning the Annals of the Joseon Dynasty, fine-tune with data collected at Kyungpook National University showed a lower BLEU score than general machine translation in the results of applying CRF When learning only with the dataset collected at Kyungpook National University, it can be seen that a slightly high BLEU score was obtained.

키워드