유의어 한일 번역을 통해 알아본 ‘인공신경망 기계번역’의 의미 관계성 인지 능력 연구

Abstract

With the spread of "GNUMT (Google Neural Network Machine Translation)" programs, various machine translation errors based on previous "RBMT (Rule-Based Machine Translation)" or "Statistical Machine Translation (SMT)" methods have greatly diminished, leading to high satisfactions on the part of users. Unlike past methods, NMT (Neural Network Machine Translation) employs deep-learning methods. Thus, translation quality hinges on the amount of data in NMT. As translation quality will increase with the accumulation of data, artificial NMT is expected to continue improving at a fairly rapid pace. Compared to other language pairs, the Korean-Japanese pair often produces high quality translations since the two languages have quite similar syntactic structures and Chinese characters. Today, Artificial Intelligence identifies the context of the original text and draws appropriate expressions from accumulated corpus data. This study stemmed from the question of 'When translating synonyms, can MT (Machine Translation) recognize a situation and context and select appropriate expressions among synonyms?' The dictionary definition of synonyms is 'words with similar meaning', and even if the meanings are almost the same or similar, each word must be clearly distinguished according to the situation or context. When human translation is performed, professional translators are capable of selecting appropriate words suitable for the context and situation. Although translation quality has improved significantly due to the NMT, this paper aims to verify the ability of the machine to accurately recognize the situation and context in which the word is used. In this paper, by using a neural machine translation program, Naver's Papago, to translate six pairs of representative synonyms (Japanese) in both directions of Korean-Japanese and Japanese-Korean, we have confirmed that the neural machine translation system can recognize some degree of scenes and contexts through words in a sentence and lead to appropriate lexical selection.