The ADAPT System Description for the IWSLT 2018 Basque to English Translation Task

The ADAPT System Description for the IWSLT 2018 Basque to English Translation Task Alberto Poncelas, Andy Way ADAPT Centre, School of Computing, Dublin City University, Dublin, Ireland {firstname.lastname}@adaptcentre.ie Kepa Sarasola Ixa Group (UPV/EHU), Faculty of Informatics University of the Basque Country kepa.sarasola@ehu.eus Abstract In this paper we present the ADAPT system built for the Basque to English Low Resource MT Evaluation Campaign. Basque is a low-resourced, morphologically-rich language. This poses a challenge for Neural Machine Translation mod- els which usually achieve better performance when trained with large sets of data. Accordingly, we used synthetic data to improve the trans- lation quality produced by a model built using only authentic data. Our proposal uses back-translated data to: (a) create new sentences, so the system can be trained with more data; and (b) translate sentences that are close to the test set, so the model can be fine-tuned to the document to be translated. 1. Introduction We participated in the Basque to English Low Resource MT Evaluation Campaign as part of the International Workshop on Spoken Language Translation (IWSLT) 2018. In this task, we aimed to build an MT model to translate subtitles of TED (Technology, Entertainment, Design) talks from Basque into English. Basque (or Euskera), which is mainly spoken in the Basque Country in Northern Spain and Southern France, is considered an isolated language. Linguistically, it is an agglutinative language, and morphologically more complex than English. Furthermore, Basque is a low resource lan- guage. Due to these characteristics, creating an MT system that deals with Basque is a challenging task. As the MT Evaluation Campaign consists of translating subtitles from TED talks, we built our MT engines mainly from available subtitles. TED Talks1 is an event where ex- perts in different fields, such as education, business, science, etc. give a talk of up to 18 minutes to disseminate their ideas. The use of subtitles as training data is potentially prob- lematic as they may not be literal translation, causing the 1https://www.ted.com original and translated sentences not to be truly parallel. This is because subtitles are subjected to a great deal of adapta- tion. Localization strategies (adapting the text to suit con- sumers of a particular locale or culture), combined with the requirement to meet time constraints (where sentences in the source and target languages which have different length are supposed to appear on the screen within the same time frame), results in sentences which are comparable but not necessarily parallel [1]. Although the adaptation does not hinder human compre- hension of the intended message, when these sentences are used as training data for an MT model, the translation inac- curacies become obstacles for the system to correctly learn to translate. The system presented in this paper aims to overcome the aforementioned problems. First, the creation of synthetic data has two purposes: (i) it provides a new set of parallel sentences that mitigates the problem of Basque being a low resourced language; and (ii), artificially-created sentences tend to be more literal than …

Content truncated to 500 words