How Should Markup Tags Be Translated?

Proceedings of the 5th Conference on Machine Translation (WMT), pages 1160–1173 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 1160 How Should Markup Tags Be Translated? Greg Hanneman and Georgiana Dinu Amazon AI {ghannema, gddinu}@amazon.com Abstract The ability of machine translation (MT) mod- els to correctly place markup is crucial to gen- erating high-quality translations of formatted input. This paper compares two commonly used methods of representing markup tags and tests the ability of MT models to learn tag placement via training data augmentation. We study the interactions of tag representation, data augmentation size, tag complexity, and language pair to show the drawbacks and ben- efits of each method. We construct and release new test sets containing tagged data for three language pairs of varying difficulty. 1 Introduction The quality of machine translation (MT) has dras- tically improved in recent years, making MT tech- nology more widely used than ever before in ap- plications ranging from financial services (Nun- ziatini, 2019) to fashion, social media, and other user-generated content (Kosmaczewska and Train, 2019; Birch et al., 2019; Michel and Neubig, 2018), among other tasks. A large amount of content requiring translation is not isolated plain text. Rather, it originates in the context of structured documents, using document format specifications such as HTML, Microsoft Word, PDF, etc. Currently, the translation industry addresses the translation of structured documents by dividing the task between a translation management system (TMS) and an underlying MT system (e.g. Fed- erico et al. (2014)). Figure 1 shows a schematic of the process. The TMS parses, manipulates, and validates the higher-level document structure. It is responsible for finding the translatable portions of the input document, performing sentence segmen- tation on the content, sending it to an underlying MT system for translation, and placing the result back into the document structure. For example, the TMS may pass over material contained in HTML <script>...</script> tags, while sending to MT the contents of <p>...</p> tags and the string values of <img alt="..."/> attributes. Properly transferring formatting tags within the translatable content (bold, italic, hyperlink, super- script, etc.) remains the responsibility of the MT process rather than the TMS. The correct preserva- tion and transfer of inline markup from source to target thus forms a crucial component of the overall quality of the MT system. An accurate placement within the segment of inline markup provides a more readable and usable document in its raw MT form, in addition to saving human time and effort if post-editing is used. Despite the key role of the MT system in pro- cessing structured documents — both in terms of TMS expectations and in lowering translation costs — the setup of translation in the context of struc- ture is rarely addressed in the standard MT eval- uation benchmarks. They typically focus on the pure-content, string-based tasks (Joanis et al., 2013; M¨uller, 2017). In this paper, we study the question of how in- line markup should be represented in and processed by the MT system in …

Content truncated to 500 words