Pre-editing Plus Neural Machine Translation for Subtitling: Effective Pre-editing Rules for Subtitling of TED Talks

摘要

In this study, the authors developed a set of pre-editing rules for TED Talk subtitling to translate Japanese source text into English. The simplified rules optimized for NMT (@TexTra® Minnano jido hon'yaku) were intended for use by a monolingual pre-editor of content to be disseminated in English. The rules were a) insert punctuation b) make implied subjects and objects explicit, and c) write proper nouns in English. The effectiveness of the rules was evaluated by human raters and BLEU score. Quality improvement was confirmed significant on human evaluation, although in some cases no changes or even degrade in quality were observed. However, one of the main concerns about the feasibility of this approach, the 21-character limit specified in the TED subtitling guidelines, was validated. The authors hold that pre-editing plus NMT is a promising approach to translating TED Talk subtitles.