机翻字幕质量评估的模型构建研究——以“人人译视界”和“讯飞听见字幕”为例

The Proposition of Machine-generated Subtitling Quality Assessment Model——Case Studies of“YSJ”and“Iflyrec”

AI时代神经网络翻译技术的飞速发展催生了一批视频译制平台。以“人人译视界”和“讯飞听见字幕”为研究对象,以动画作品《中华勤学故事》机翻字幕为研究素材,探究国内视频译制平台机翻字幕质量。基于FAR语际字幕评估模型提出了机翻字幕评估模型(FAR 2.0),从功能对等、接受程度、阅读体验三方面对机翻字幕的语义错误、术语翻译错误、译文冗余等问题进行了深入分析。根据研究结果,提出增强语音数据标注积累、加大机器翻译语料训练及强化翻译引擎的意群切分等三点优化建议。

With the rapid development of neural machine translation technology in the AI era,a number of video translation platforms have emerged.In order to evaluate the quality of machine-generated subtitling,this paper takes“YSJ”and“Iflyrec”as the research objects,and the machine-generated subtitles of the cartoon“Chinese Stories of Study”as the material when doing the assessing.This study tries to propose a model for assessing machine-generated subtitle(FAR 2.0)based on the FAR model by Jan Pedersen,and analyses semantic error,term error,and redundancy phenomenon from three parameters:functional equivalence,acceptability and readability.This paper gives the following three suggestions for optimization:firstly,enhance the accumulation of speech data annotation;secondly,increase data training on the various corpus;thirdly,enhance the segmentation and combination ability of the translation engine.