韓漢神經網絡機器翻譯質量考察 —以韓語同音異義動詞為中心

Abstract

The statistical-based translation before the neural network translation system was a method of translation in phrase units composed of words and words, but the neural network translation system is a method of interpreting and translating components in the context of the whole sentence. This system has the advantage of enabling contextual translation, thus translating more accurately into the output language. However, there are many cases where the current neural network translation system misinterprets the meaning of the homonym verb, which depends on the context, and produces a meaningless or ungrammatical sentence. Therefore, this study examined the ability of Papago, Google, and Baidu to determine the meaning of homonymous words in neural network machine translation, and analyzed the types of errors to analyze the data needed to enhance the functioning of machine translation systems. As a result of the analysis, Google Machine Translation still had many limitations in its contextual semantic ability, with significantly lower translation success rates compared to Papago Machine Translation and Baidu Machine Translation. Compared to Google, Papago and Baidu had a relatively high score in terms of semantics, but the translation success rate was affected by changes in the form of verbs or other polysemous words appearing in sentences, and the absence of words with Semantic Correspondence. In the future, Google will need to reinforce colloquial-oriented data and enhance semantic judgment ability of homonymous words, and it will also have to supplement and correct the limitations pointed out in the Baidu and Papago systems.