OPUS-MT – Building open translation services for the World

OPUS-MT – Building open translation services for the World J¨org Tiedemann Department of Digital Humanities HELDIG University of Helsinki Santhosh Thottingal Wikimedia Foundation 1 Introduction Equality among people requires, among other things, the ability to access information in the same way as others independent of the linguistic back- ground of the individual user. Achieving this goal becomes an even more important challenge in a globalized world with digital channels and infor- mation flows being the most decisive factor in our integration in modern societies. Language barriers can lead to severe disadvantages and discrimination not to mention conflicts caused by simple misun- derstandings based on broken communication. Lin- guistic discrimination leads to frustration, isolation and racism and the lack of technological language support may also cause what is known as the digital language death (Kornai, 2013). Machine translation (MT) has developed into a useful tool that diminishes and partially removes such language barriers. Modern MT engines enable people to communicate, to access information in foreign languages and to build efficient resources for new communities. The mission of OPUS-MT1 is to provide open translation services and tools that are free from commercial interests and restrictions. The idea is to make automatic translation accessible for anyone in a transparent and secure way without exploitation plans and hidden agendas compromis- ing privacy and placing marketing strategies. We also want to focus on the support of minority and low-resource languages with the aim to introduce a community effort for the benefit of all. OPUS-MT has successfully launched its first pi- lot system and currently collaborates with the Wiki- media foundation in the setup of translation ser- © 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1https://github.com/Helsinki-NLP/Opus-MT vices for the production of Wikipedia content in new languages based on more elaborated resources available in, e.g. English. Currently, the project provides over 1,000 pre-trained translation models that are free to download and use. OPUS-MT also contains open-source software for launching trans- lation services as web applications. The on-going effort focuses on the improvement of translation quality, language coverage and emphasizes specific test cases to study the applicability of the approach. More details about the implementation and current status of the project are given below. 2 OPUS-MT models The models that we train are based on state-of-the- art transformer-based neural machine translation (NMT). We apply Marian-NMT2 in our framework, a stable production-ready NMT toolbox with effi- cient training and decoding capabilities (Junczys- Dowmunt et al., 2018). Our models are trained on freely available parallel corpora collected in the large bitext repository OPUS3 (Tiedemann, 2012). The architecture is based on a standard transformer setup with 6 self-attentive layers in both, the en- coder and decoder network with 8 attention heads in each layer. The hyper-parameters follow the gen- eral recommendations given in the documentation of the software. All the details can be seen in the training procedures that we also release as open source in our GitHub repository.4 …

Content truncated to 500 words