Customizing Neural Machine Translation for Subtitling

Abstract

In this work, we customized a neural machine translation system for translation of subtitles in the domain of entertainment. The neural translation model was adapted to the subtitling content and style and extended by a simple, yet effective technique for utilizing intersentence context for short sentences such as dialog turns. The main contribution of the paper is a novel subtitle segmentation algorithm that predicts the end of a subtitle line given the previous word-level context using a recurrent neural network learned from human segmentation decisions. This model is combined with subtitle length and duration constraints established in the subtitling industry. We conducted a thorough human evaluation with two post-editors (English-to-Spanish translation of a documentary and a sitcom). It showed a notable productivity increase of up to 37% as compared to translating from scratch and significant reductions in human translation edit rate in comparison with the post-editing of the baseline non-adapted system without a learned segmentation model.

Is 42 the Answer to Everything in Subtitling-oriented Speech Translation?

“…Recently, after the advent of the neural machine translation paradigm, Matusov et al (2019) presented an NMT system customised to subtitling. The main contribution of the paper is a segmenter module trained on human segmentation decisions, which splits the resulting translation into subtitles.…”

“…These source language subtitles (also called captions) are already compressed and segmented to respect the subtitling constraints of length, reading speed and proper segmentation (Cintas and Remael, 2007;Karakanta et al, 2019). In this way, the work of an NMT system is already simplified, since it only needs to translate matching the length of the source text (Matusov et al, 2019;Lakew et al, 2019). However, the essence of a good subtitle goes beyond matching a predetermined length (as, for instance, 42 characters per line in the case of TED talks).…”

Towards Automatic Subtitling: Assessing the Quality of Old and New Resources

“…This has been shown in Karakanta, Negri, and Turchi (2020a), where training on the MuST-Cinema data, annotated with the special symbols corresponding to subtitle and line breaks, achieved high conformity to the subtitling constraints, without any modification in the NMT architecture. Similarly, the segmentation module presented in (Matusov, Wilken, and Georgakopoulou 2019), which is trained on the monolingual OpenSubtitles data (which contain breaks as metadata), led to significant reductions in post-editing effort.…”

Proceedings of the 17th International Conference on Spoken Language Translation

“…One way to automatise this labour-intensive process, especially in settings where several target languages are involved, is creating a subtitle template of the source language (Georgakopoulou, 2019). A subtitle template is an enriched transcript of the source language speech where the text is already compressed, timed and segmented into proper subtitles.…”

“…content, still rely heavily on human effort. In a typical multilingual subtitling workflow, a subtitler first creates a subtitle template (Georgakopoulou, 2019) by transcribing the source language audio, timing and adapting the text to create proper subtitles in the source language. These source language subtitles (also called captions) are already compressed and segmented to respect the subtitling constraints of length, reading speed and proper segmentation (Cintas and Remael, 2007;Karakanta et al, 2019).…”