華語武術奇幻遊戲的神經機器翻譯比較

Abstract

本文以華語武術奇幻遊戲為例，比較兩種廣受歡迎的神經機器翻譯（neural machine translation，NMT）工具用於美式英文翻譯的品質。NMT 系統運用深度學習演算法，能夠產出愈加準確且自然的譯文，近年來機器翻譯（machine translation，MT）的品質已大幅提升。然而，礙於互動式多媒體軟體的創造力與複雜性，MT 工具要產出優質的電玩遊戲翻譯絕非易事。有鑒於以華語電玩遊戲為主題的研究尚且不足，加上各產業與學界對 NMT 的興趣與日俱增，本文以自訂的動態品質架構（dynamic quality framework，DQF）暨多維品質指標（multidimensional quality metrics，MQM）錯誤類型架構為基礎，旨在透過質性研究探討當今最先進的 MT 工具之優劣。本研究針對華語武術奇幻體裁及角色扮演遊戲的主要文字類型，選定兩款具有代表性的電玩遊戲：《古劍奇譚三：夢付千秋星垂野》與《俠客風雲傳》作為原文文本，並檢視 DeepL翻譯（DeepL Translator）和 Google 翻譯（Google Translate）這兩套當代 NMT 系統應用於翻譯時的特定錯誤與模式。

This paper compares the translation quality of popular neural machine translation (NMT) tools used in the translation of Chinese martial arts fantasy games into American English. Though there have been significant leaps in machine translation (MT) quality in recent years thanks to NMT systems that utilize deep learning algorithms to generate increasingly accurate and natural-sounding output, producing adequate video game translations still poses a great challenge for MT tools due to the creativity and complexity of interactive multimedia software. In light of the understudied field of Chinese-language video games and rising interest from both industry and academia in neural machine translation, this paper aims to explore the strengths and weaknesses of current state-of-the-art MT tools through qualitative analysis based on a modified DQF-MQM error typology framework. It examines specific translation errors and patterns of these tools when they are applied to Chinese martial arts fantasy genres and major text types of role-playing games. DeepL Translator and Google Translate were used to represent current NMT systems and source text was borrowed from two games representing Chinese martial arts fantasy video games: Gujian 3 and Tale of Wuxia.