Multilingual Parallel Corpus for Global Communication Plan

Multilingual Parallel Corpus for Global Communication Plan Kenji Imamura and Eiichiro Sumita National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0289, Japan {kenji.imamura,eiichiro.sumita}@nict.go.jp Abstract In this paper, we introduce the Global Communication Plan (GCP) Corpus, a multilingual parallel corpus being developed as part of the GCP. The GCP Corpus is intended to be develop speech translation systems; thus, it primarily consists of pseudo-dialogues between foreign visitors and local Japanese people. The GCP Corpus is sentence-aligned and covers four domains and ten languages, including many Asian languages. In this paper, we summarize the GCP and the current status of the GCP Corpus. Then, we describe some of the corpus’ basic characteristics from the perspective of multilingual machine translation and compare direct, pivot, and zero-shot translation techniques. Keywords: Multilingual Parallel Corpus, Global Communication Plan, Asian Languages, Pivot/Zero-shot Translation 1. Introduction In 2014, the Ministry of Internal Affairs and Communi- cations in Japan implemented the Global Communication Plan (GCP; c.f., Sec. 2.), whose mission is to eliminate global “language barriers.” An important aspect of this plan is the public demonstration of multilingual speech trans- lation technologies. With this in mind, we are improving the usability of the multilingual speech translation system by improving translation quality and developing user inter- faces that enable it to be used in various places, such as hospitals, malls/stores, and tourist spots. Multilingual translation corpora are required to develop multilingual speech translation systems, and the GCP in- cludes a corpus development program for machine transla- tion (MT). In this paper, we summarize the GCP and the current status of the GCP Corpus, which is a multilingual parallel corpus used for public demonstration 1. This corpus primarily con- sists of pseudo-dialogues between foreign visitors and local Japanese people. (It also includes some isolated utterances, such as phrases frequently associated with travelling.) The utterances in the dialogues are translated into ten languages (including Japanese) targeted by the GCP. Therefore, the corpus is sentence-aligned. The target domains are medical care, disaster prevention, shopping, and tourism. Although the GCP Corpus is being developed for use in speech translation systems, it could also be used in various other research fields because it has the following character- istics. 1. It is a multilingual sentence-aligned corpus that cov- ers ten languages, including Asian languages. There- fore, this allows 90 different MT systems to be con- structed. Furthermore, it can also be applied in com- parative studies of pivot translation (Utiyama and Isa- hara, 2007; Cohn and Lapata, 2007) and zero-shot translation (Johnson et al., 2016). 1We are also developing speech corpora for speech recognition and synthesis tasks; however, we only describe the parallel corpus in this paper. 2. It covers four domains, namely, medical care, disas- ter prevention, shopping, and tourism. It can be ap- plied to domain adaptation studies (e.g., (Imamura and Sumita, 2016)). 3. It consists of pseudo-dialogues. Therefore, it can also be applied to discourse studies that consider long- distance contexts. Note that such contexts are simpler than those …

Content truncated to 500 words