Customizing Neural Machine Translation for Subtitling

Proceedings of the Fourth Conference on Machine Translation (WMT), Volume 1: Research Papers, pages 82–93 Florence, Italy, August 1-2, 2019. c⃝2019 Association for Computational Linguistics 82 Customizing Neural Machine Translation for Subtitling Evgeny Matusov∗, Patrick Wilken∗, Panayota Georgakopoulou∗ AppTek Aachen, Germany ematusov@apptek.com pwilken@apptek.com yota@athenaconsultancy.eu 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 subti- tling content and style and extended by a sim- ple, yet effective technique for utilizing inter- sentence context for short sentences such as dialog turns. The main contribution of the pa- per is a novel subtitle segmentation algorithm that predicts the end of a subtitle line given the previous word-level context using a recur- rent neural network learned from human seg- mentation decisions. This model is combined with subtitle length and duration constraints established in the subtitling industry. We con- ducted a thorough human evaluation with two post-editors (English-to-Spanish translation of a documentary and a sitcom). It showed a no- table productivity increase of up to 37% as compared to translating from scratch and sig- nificant reductions in human translation edit rate in comparison with the post-editing of the baseline non-adapted system without a learned segmentation model. 1 Introduction In recent years, significant progress was observed in neural machine translation (NMT), with its quality increasing dramatically as compared to the previous generation of statistical phrase-based MT systems. However, user acceptance in the subti- tling community has so far been rare. The rea- son for this, in our opinion, is that the state-of-the- art off-the-shelf NMT systems do not address the issues and challenges of the subtitling process in full. In this paper, we present a customized NMT system for subtitling, with focus on the entertain- ∗equal contribution ment domain. From the user perspective, we show how the quality of translation and subtitle seg- mentation can improve in such a way that sig- nificantly reduced post-editing is required. We believe that such customized systems would lead to greater user acceptance in the subtitling indus- try and would contribute to the wider adoption of NMT technology with the subsequent benefits the latter brings in terms of productivity gain and time efficiency in subtitling workflows. The paper is structured as follows. We start with the review of related research in Section 2. Sec- tion 3 describes the details of our baseline NMT system and how it compares to NMT systems from previous research. Section 4 presents the details of the changes to the MT system that were nec- essary to boost its performance on the subtitling tasks for entertainment domain, with a focus on Latin American Spanish as the target language. In Section 5, we present a novel algorithm for automatic subtitle segmentation that is combined with rule-based constraints which are necessary for correct subtitle representation on the screen. Finally, Section 6 describes the automatic and hu- man evaluation of the proposed system, including post-editing experiments and feedback from pro- …

Content truncated to 500 words