A Study on Korean Language Honorific Translation using Transformers

초록

Due to anonymous Internet use and the widespread use of online communities, impolite expressions increase in the Korean language. To alleviate the side effects, we propose a Korean language honorific translation, impolite-polite translation in the same language, using Transformers. However, there are few studies on the conversion of Korean honorifics through deep learning. Especially, in our study, we newly constructed an impolite-polite dataset which amounts to about 20,000 datasets by combining a selected data from DC Inside bulletin with the data from AI-HUB. Moreover, we explore the optimal tokenization methods and the optimal numbers of encoder and decoder layers in Transformers. Through experiments, we achieved a high BLEU score - 66.71 and verified that the BLEU metric is highly correlated with human evaluation.

키워드