An English-Swahili parallel corpus and its use for neural machine translation in the news domain

An English–Swahili parallel corpus and its use for neural machine translation in the news domain Felipe S´anchez-Mart´ınez,‡ V´ıctor M. S´anchez-Cartagena,‡ Juan Antonio P´erez-Ortiz,‡ Mikel L. Forcada,‡ Miquel Espl`a-Gomis,‡ Andrew Secker,† Susie Coleman,† Julie Wall† ‡Dep. de Llenguatges i Sistemes Inform`atics, Universitat d’Alacant E-03690 Sant Vicent del Raspeig (Spain) {fsanchez,vmsanchez,japerez,mlf,mespla}@dlsi.ua.es †The British Broadcasting Corporation BBC Broadcasting House, Portland Place, London, W1A 1AA. (UK) {andrew.secker,susie.coleman,julie.wall}@bbc.co.uk Abstract This paper describes our approach to cre- ate a neural machine translation system to translate between English and Swahili (both directions) in the news domain, as well as the process we followed to crawl the neces- sary parallel corpora from the Internet. We report the results of a pilot human evalua- tion performed by the news media organisa- tions participating in the H2020 EU-funded project GoURMET. 1 Introduction Large news media organisations often work in a multilingual space in which they both publish their material in numerous languages and monitor the world’s media across video, audio, printed and on- line sources. As regards content creation, one way in which efficient use is made of journalistic endeav- our is the republication of news originally authored in one language into another; by using machine translation, and with the appropriate user interfaces, a journalist is able to take a news story or script, in the case of an audio or video report, and quickly obtain a preliminary translation that will be then manually post-edited to ensure it has the quality required to be presented to the audience. Concern- ing news gathering, expert monitors and journalists have to currently perform a lot of manual work to keep up with a growing amount of broadcast and social media streams of data; it is becoming im- perative to automate tasks, such as translation, in order to free monitors and journalists to perform more journalistic tasks that cannot be achieved with technology. c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. In order to cope with these requirements, promot- ing both the reach of the news published to under- served audiences and the world-wide broadcasting of local information, the H2020 EU-funded project GoURMET (Global Under-Resourced Media Trans- lation),1 aims at improving neural machine trans- lation (NMT) for under-resourced language pairs with special emphasis in the news domain. The two partner media organisations in the GoURMET project, the BBC in the UK and Deutsche Welle (DW) in Germany, publish news content in 40 and 30 different languages, respectively, and gather news in over 100 languages. In particular, both media partners gather news in and produce content in Swahili. According to Wikipedia, Swahili has between 2 and 15 million first-language speakers and 90 million second-language speakers. As one of the largest languages in Africa and the recognised lin- gua franca of the East African community, BBC and DW see Swahili as an important language in which to make content available. The NMT systems described and evaluated herein can be deployed to support them in …

Content truncated to 500 words