Neural Machine Translation Strategies for Generating Honorific-style Korean

Abstract

Expression with honorifics is an important way of dressing up the language and showing politeness in Korean. For machine translation, generating honorifics is indispensable on the formal occasion when the target language is Korean. However, current Neural Machine Translation (NMT) models ignore generation of honorifics, which causes the limitation of the MT application on business occasion. In order to address the problem, this paper presents two strategies to improve Korean honorific generation ratio: 1) we introduce honorific fusion training (HFT) loss under the minimum risk training framework to guide the model to generate honorifics; 2) we introduce a data labeling (DL) method which tags the training corpus with distinctive labels without any modification to the model structure. Our experimental results show that the proposed two strategies can significantly improve the honorific generation ratio by 34.35% and 45.59%.