Abstract:
Comparisons between machine-generated renderings and official subtitles are accumulating, yet it remains unclear where the two diverge in documentary subtitling, a genre in which the image carries part of the meaning. This review examines 25 items bearing on that question: 17 Chinese journal articles, six monographs, one dissertation, and one web source, 12 of them published after 2020, retrieved from CNKI and through reference-list tracing. Four lines of research are distinguished: comparisons of machine and human translation, the constraints of subtitling, documentary subtitle research, and translation quality assessment. Three findings emerge. First, the studies reviewed broadly agree that machine output is no longer inferior to human translation in semantic fidelity, and that its weaknesses concentrate in fluency, terminological consistency, and contextual cohesion. Second, most of these studies argue from illustrative examples rather than item-level counts and explicit criteria, so their conclusions are hard to verify. Third, research on subtitles has described the constraints of the medium in detail, and international work has turned them into computable parameters such as line limits and characters per second. A gap follows from the three. Existing comparisons treat the official translation as the sole standard, without separating genuine mistranslation from legitimate variation. The coordination of subtitles with the image, and the consistency of proper-name renderings across a whole film, has rarely been examined within the scope of this search. The review therefore proposes a three-way alignment that takes the source text and the image as joint reference, and a three-level scheme of judgement — word-and-sentence, contextual, and whole-film — counted level by level. Both are illustrated with three examples from Season 2 of the documentary The Potala Palace, in order to delimit what current large language models can and cannot achieve here.