Jejueo Datasets for Machine Translation and Speech Synthesis

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 2615–2621 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 2615 Jejueo Datasets for Machine Translation and Speech Synthesis Kyubyong Park, Yo Joong Choe, Jiyeon Ham Kakao Brain 20, Pangyoyeok-ro 241, Bundang-gu, Seongnam-si, Gyeonggi-do, Korea {kyubyong.park, yj.choe, jiyeon.ham}@kakaobrain.com Abstract Jejueo was classified as critically endangered by UNESCO in 2010. Although diverse efforts to revitalize it have been made, there have been few computational approaches. Motivated by this, we construct two new Jejueo datasets: Jejueo Interview Transcripts (JIT) and Jejueo Single Speaker Speech (JSS). The JIT dataset is a parallel corpus containing 170k+ Jejueo-Korean sentences, and the JSS dataset consists of 10k high-quality audio files recorded by a native Jejueo speaker and a transcript file. Subsequently, we build neural systems of machine translation and speech synthesis using them. All resources are publicly available via our GitHub repository. We hope that these datasets will attract interest of both language and machine learning communities. Keywords: Jejueo, Jeju language 1. Introduction Jejueo, or the Jeju language, is a minority language used on Jeju Island (O’Grady, 2015). It was classified as critically endangered by UNESCO in 2010.1 While there have been many academic efforts to preserve the language (Yang et al., 2017; Saltzman, 2017; Yang et al., 2018a; Yang et al., 2018b), data-driven approaches for Jejueo-related language tasks have been rare. Meanwhile, the natural language processing (NLP) com- munity has observed significant advances in both ma- chine translation (Sutskever et al., 2014; Cho et al., 2014; Vaswani et al., 2017) and speech synthesis (Oord et al., 2016; Wang et al., 2017; Shen et al., 2018), especially driven by deep learning in recent years. In particular, there has been growing attention towards low- resource scenarios (Zoph et al., 2016; Gu et al., 2018), which pose a unique challenge for existing deep learning methods. The challenge is unique not only in the sense that less data is available but also in the sense that low- resource languages often raise different syntactic, morpho- logical, and semantic challenges that are under-explored by systems optimized on major languages such as English, German, French, and Arabic (Bender, 2019). Tasks relevant to Jejueo often involve having to deal with such challenges. Motivated by the unique challenges that Jejueo presents as well as the lack of Jejueo resources available for computa- tional approaches, we develop a machine-readable Jejueo- Korean parallel corpus and a clean Jejueo single speaker speech dataset. Our contributions can be summarized as follows: • We present the Jejueo Interview Scripts (JIT) dataset, a Jejueo-Korean parallel corpus of more than 170k sen- tences. • We train neural machine translation models on the JIT dataset so that they can be the baselines for future stud- ies. 1http://www.unesco.org/languages-atlas/ en/atlasmap.html 제주어 구술자료집 PDF JIT Dataset for Machine Translation compile A B C D TXT TSV WAV JSS Script JSS Audios extract record for Speech Synthesis for Speech Synthesis Figure 1: Overview of dataset construction. The original pdf …

Content truncated to 500 words