Document-aligned Japanese-English Conversation Parallel Corpus

Proceedings of the 5th Conference on Machine Translation (WMT), pages 639–645 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 639 Document-aligned Japanese-English Conversation Parallel Corpus Mat¯ıss Rikters, Ryokan Ri, Tong Li and Toshiaki Nakazawa The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo, Japan {matiss, li0123, litong, nakazawa}@logos.t.u-tokyo.ac.jp Abstract Sentence-level (SL) machine translation (MT) has reached acceptable quality for many high- resourced languages, but not document-level (DL) MT, which is difficult to 1) train with lit- tle amount of DL data; and 2) evaluate, as the main methods and data sets focus on SL eval- uation. To address the first issue, we present a document-aligned Japanese-English conversa- tion corpus, including balanced, high-quality business conversation data for tuning and test- ing. As for the second issue, we manually identify the main areas where SL MT fails to produce adequate translations in lack of con- text. We then create an evaluation set where these phenomena are annotated to alleviate au- tomatic evaluation of DL systems. We train MT models using our corpus to demonstrate how using context leads to improvements. 1 Introduction The quality of machine translation (MT) for written text and monologue has vastly improved due to the increased amount of available parallel corpora and recent neural network technologies. However, there is much room for improvement in the context of dia- logue or conversation translation. One typical case is the translation from a pro-drop language to a non- pro-drop language where correct pronouns must be supplemented according to the context. The omis- sion of the pronouns occurs more frequently in spoken language than written language. Recently, context-aware MT models attract attention from many researchers (Tiedemann and Scherrer, 2017; Voita et al., 2019) to solve this kind of problem, however, there are almost no parallel conversation corpora with context information except the rather noisy Open Subtitles corpus (Tiedemann, 2016). A document and sentence-aligned conversation parallel corpus should be advantageous to push MT research in this field to the next stage. In this paper, we introduce a newly constructed document- aligned (DA) Japanese-English conversation cor- pus, which contains three sub-corpora: Busi- ness Scene Dialogue (BSD (Rikters et al., 2019)), Japanese translation of AMI Meeting Corpus (AMI (McCowan et al., 2005)) and Japanese translation of OntoNotes 5.0 (ON (Weischedel et al., 2011)). The corpus contains multi-person conversations in various situations: business scenes, meetings un- der specific themes, broadcast conversations and telephone conversations. We supplement the original BSD part with addi- tional data, increasing its size by almost three times. We also enrich the corpus with speaker information and other useful meta-data, and separate balanced versions of development and evaluation data sets. 2 Related Work There are many ready-to-use parallel corpora for training MT systems, but most of them are in writ- ten languages such as web crawl, patents (Goto et al., 2011), scientific papers (Nakazawa et al., 2016). Even though some parallel corpora are in spoken language, they are mostly monologues (Cet- tolo et al., 2012; Di Gangi et al., 2019) or contain a …

Content truncated to 500 words