GoURMET – Machine Translation for Low-Resourced Languages

GoURMET – Machine Translation for Low-Resourced Languages Peggy van der Kreeft Deutsche Welle, Bonn, Germany peggy.van-der-kreeft@dw.com Sevi Sariisik BBC, London, UK sevi.sariisik@bbc.co.uk Wilker Aziz Univ. of Amsterdam, Netherlands w.ferreiraaziz@uva.nl Alexandra Birch Univ. of Edinburgh, UK A.Birch@ed.ac.uk Felipe Sánchez-Martínez Univ. of Alicante, Spain fsanchez@dlsi.ua.es Abstract The GoURMET project, funded by the EU H2020 research and innovation action (under grant agreement 825299), develops models for machine translation, in partic- ular for low-resourced languages. Data, models and software releases as well as the GoURMET Translate Tool are made available as open source. 1 The Project GoURMET (Global Under-Resourced Media Translation) started in January 2019 and runs until 30 June 2022. The consortium consists of five partners: The University of Edinburgh (coordinator), University of Alicante, University of Amsterdam, and user partners BBC and Deutsche Welle (DW). 1 The aim is to significantly improve the robustness and applicability of neural machine translation (NMT) for low-resourced language pairs and domains. This is in particular important because machine translation (MT) is increasingly used as a technology for supporting communication in a globalized world. The two international broadcasters participating in GoURMET are faced with the need to use MT to support their editorial work, especially for languages for which such tools are currently hard to find or lack quality. © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CCBY-ND. The main objectives of the project are: • to advance deep-learning for natural language applications • to arrive at high-quality MT for low- resourced languages and diverse language pairs and domains • to develop tools for media analysts and journalists in the form of a sustainable and maintainable platform and services. The work is built around three use cases. The first is global content creation, where we use MT in multilingual content production, with editorial control. The second use case is media monitoring for low-resourced and especially strategically im- portant languages. The third use case focuses on a specific topic, and the health sector, in particular COVID, was selected for this purpose, fitting the news requirements over the past two years. The objective in the last use case is to apply transfer learning between topical domains. 2 Languages Covered MT models were selected for sixteen low- resourced languages, jointly by user and technology partners and developed in different phases of the project. These languages are: Amharic, Bulgarian, Burmese, Gujarati, Hausa, Igbo, Kyrgyz, Macedonian, Pashto, Serbian, Swahili, Tamil, Tigrinya. Turkish, Urdu, and Yoruba – all of them from and into English. Pashto was selected as a “surprise language” and 1 https://gourmet-project.eu/ developed as a special case upon request in a period of two months. Different factors were taken into account for the selection process, including strategic importance for the news partners, proximity of languages, re- search interest and complexity for the develop- ment of the models. 3 Research and Development of the Models Each selected language was assigned to one technology partner, who developed the model. Different methods were used among …

Content truncated to 500 words