因應AI科技的發展，全球翻譯語言服務業者也將機器翻譯融入專業工作流程中，而且工作產值量不容忽視。在此趨勢下，在翻譯教學中訓練學生的機後編輯(Machine Translation Post-Editing, MTPE)能力變成不可或缺的一環。本論文主要採用質性分析法，以六篇2019年2月份紐約時報中文網國際縱覽網頁中最受歡迎文章為文本，分別於2019年和2021年利用Google神經機器翻譯系統(GNMT)，將這些文本之人工中譯文回譯成英文，並以英文原文為評價指標，然後應用適合各種語言配對且錯誤類型定義清楚的「多維度質量指標」(Multidimensional Quality Metrics, MQM)，分析與比較2019年與2021年之英文機器譯文文本的錯誤類型、質量及錯誤嚴重度(severity)之差異，最後提出機後編輯教學相關建議。研究結果只發現精確度、流暢度等二維度的錯誤類型。在「精確度」維度方面，主要錯誤集中在「誤譯」類型；整體而言，2021年版較2019版譯文的精確度錯誤類型數量明顯減少，在「漏譯」部分也稍有改善，「增譯」數量極少，而且兩版本沒有質量上之差別。在「流暢度」維度方面，本研究僅發現「連貫錯誤」和「文法錯誤」，而且兩版的機器英譯文在此兩方面的翻譯質量從2019到2021年沒有進展；錯誤數量也幾乎一樣，僅是文法錯誤較多些。此外，兩版英譯文的整體精確度錯誤嚴重度僅有小幅差異，然重大級及嚴重級的「誤譯」錯誤最多，而在流暢度錯誤嚴重度上可謂沒有差別。研究結果有助於了解GNMT在中譯英翻譯上的進展，也對機後編輯教學有啟示作用。

Translation services around the world have incorporated the use of machine translation (MT) in their professional workflow as a response to the development of artificial intelligence. Therefore, post-editing competency training has become indispensable in the translation instruction. The authors use qualitative analysis to track the changes in the quality of the MT output produced by Google Neural Machine Translation (GNMT) in the years of 2019 and 2021. The study selects six articles from the New York Times Chinese language websites in 2019 for analysis and comparison. GNMT is used to back-translate the manually translated Chinese versions into English, and the non-translated texts are used as an evaluation reference. The discrepancies in terms of quality and quantity of the machine translated texts between 2019 and 2021 are analyzed according to the MT error typology in Multidimensional Quality Metrics (MQM). As news texts are used in this study, only errors in accuracy and fluency are found. Errors in accuracy are mostly mistranslations. In the fluency category, only errors in “coherence” and “grammar” are found. Very similar mistakes in “coherence” and “grammar” in both years, but 2021 versions contain more errors in “grammar”. The severity of “accuracy” errors in both versions varies, with major and critical errors accounting for the most part. The severity of “fluency” errors in both years is about the same. The error types analyzed in these texts will help see the development of GNMT in the Chinese to English translation and serve in the teaching of post-editing skills.