Post-editing challenges in Chinese-to-English neural machine translation of movie subtitles

Abstract

Subtitle translation has been a longstanding factor hindering the overseas development of Chinese movies. The potential of using Neural Machine Translation (NMT) as an innovative solution has yet to be studied. This case study aims to integrate Google Neural Machine Translation (GNMT) into the Chinese-into-English subtitle translation of Chinese movies. The research identifies errors in GNMT-generated subtitles per Pedersen's FAR model and develops post-editing (PE) recommendations to address these errors. Firstly, the Chinese subtitles, human-translated subtitles, and GNMT-generated subtitles of a Chinese movie were collected. Then, the FAR model-based error analysis was conducted to explore the errors' features. Lastly, PE recommendations were proposed accordingly to modify these errors. Approximately a quarter of all subtitles contain errors, with functional equivalence errors the most prevalent (about half), followed by acceptability errors (about a third) and readability errors (14%). Regarding the severity of errors, standard errors rank first (42%), followed by serious errors (30%) and minor errors (28%). Subtitlers should focus on semantic, idiomaticity, grammar, and line length errors caused by incorrect translation of proper nouns, cultural-bound terms, incomplete sentences, etc. By exploring the features of errors and PE recommendations for GNMT of Chinese movie subtitles, this study offers a valuable framework for improving the quality of NMT-generated subtitles and presents a new solution to overcome the language barriers faced by the Chinese film industry.

Keywords

• Error analysis
• FAR model
• Machine translation (MT)
• Movie subtitles
• Neural machine translation (NMT)
• Post-editing (PE)