A High-Quality Multilingual Dataset for Structured Documentation Translation

Proceedings of the Fourth Conference on Machine Translation (WMT), Volume 1: Research Papers, pages 116–127 Florence, Italy, August 1-2, 2019. c⃝2019 Association for Computational Linguistics 116 A High-Quality Multilingual Dataset for Structured Documentation Translation Kazuma Hashimoto Raffaella Buschiazzo James Bradbury∗ Teresa Marshall Richard Socher Caiming Xiong Salesforce {k.hashimoto,rbuschiazzo,james.bradbury, teresa.marshall,rsocher,cxiong}@salesforce.com Abstract This paper presents a high-quality multilingual dataset for the documentation domain to ad- vance research on localization of structured text. Unlike widely-used datasets for transla- tion of plain text, we collect XML-structured parallel text segments from the online doc- umentation for an enterprise software plat- form. These Web pages have been profession- ally translated from English into 16 languages and maintained by domain experts, and around 100,000 text segments are available for each language pair. We build and evaluate transla- tion models for seven target languages from English, with several different copy mecha- nisms and an XML-constrained beam search. We also experiment with a non-English pair to show that our dataset has the potential to explicitly enable 17 × 16 translation settings. Our experiments show that learning to trans- late with the XML tags improves translation accuracy, and the beam search accurately gen- erates XML structures. We also discuss trade- offs of using the copy mechanisms by focusing on translation of numerical words and named entities. We further provide a detailed human analysis of gaps between the model output and human translations for real-world applica- tions, including suitability for post-editing. 1 Introduction Machine translation is a fundamental research area in the field of natural language processing (NLP). To build a machine learning-based trans- lation system, we usually need a large amount of bilingually-aligned text segments. Exam- ples of widely-used datasets are those included in WMT (Bojar et al., 2018) and LDC1, while new evaluation datasets are being actively cre- ated (Michel and Neubig, 2018; Bawden et al., ∗Now at Google Brain. 1https://www.ldc.upenn.edu/ - Example (a) English: You can use this report on your Community Management Home dashboard or in <ph>Community Workspaces</ph> under <menucascade><uicontrol>Dashboards</uicontrol><uicontrol>Home </uicontrol></menucascade>. Japanese: このレポートは、[コミュニティ管理] のホームのダッシュボード、または <ph>コミュニティワークスペース </ph>の <menucascade><uicontrol>[ダッシュボード]</uicontrol> <uicontrol>[ホーム]</uicontrol></menucascade> で使用できます。 - Example (b) English: Results with <b>both</b><i>beach</i> and <i>house</i> in the searchable fields of the record. Japanese: レコードの検索可能な項目に <i>beach</i> と <i>house</i> の <b>両方</b>が含まれている結果。 - Example (c) English: You can only predefine this field to an email address. You can predefine it using either T (used to define email addresses) or To Recipients (used to define contact, lead, and user IDs). Japanese: この項目はメールアドレスに対してのみ事前に定義できます。 この項目は [宛先] (メールアドレスを定義するために使用 ) または [宛 先受信者] (取引先責任者、リード、ユーザ ID を定義するために使用 ) のいずれかを使用して事前に定義できます。 Figure 1: English-Japanese examples in our dataset. 2018; M¨uller et al., 2018). These existing datasets have mainly focused on translating plain text. On the other hand, text data, especially on the Web, is not always stored as plain text, but of- ten wrapped with markup languages to incorpo- rate document structure and metadata such as for- matting information. Many companies and soft- ware platforms provide online help as Web doc- uments, often translated into different languages to deliver useful information to people in different …

Content truncated to 500 words