The University of Helsinki submissions to the IWSLT 2018 low-resource translation task

The University of Helsinki submissions to the IWSLT 2018 low-resource translation task Yves Scherrer Department of Digital Humanities University of Helsinki, Finland yves.scherrer@helsinki.fi Abstract This paper presents the University of Helsinki submissions to the Basque–English low-resource translation task. Our primary system is a standard bilingual Transformer system, trained on the available parallel data and various types of syn- thetic data. We describe the creation of the synthetic datasets, some of which use a pivoting approach, in detail. One of our contrastive submissions is a multilingual model trained on comparable data, but without the synthesized parts. Our bilingual model with synthetic data performed best, obtain- ing 25.25 BLEU on the test data. 1. Introduction The University of Helsinki has participated in the IWSLT low-resource translation task on Basque-to-English transla- tion with one primary and two contrastive systems. Our ex- periments mainly focused on creating synthetic training data for classical supervised neural machine translation models. In particular, we show that a bilingual system trained on partly synthetic data performs better than a multilingual sys- tem that includes the same data in their original, non syn- thetic form. Our best submitted system obtained a BLEU score of 25.25. Section 2 describes the available Basque–English parallel datasets at the basis of our systems, as well as a baseline sys- tem trained on these parallel datasets alone. In Section 3, we present additional datasets that contain either Basque or En- glish text, but not both. We discuss several strategies for syn- thetically creating parallel Basque–English datasets out of these sources, and show the impact of these synthetic datasets on translation quality. In Section 4, we present a contrastive system that uses the additional datasets in their original state, without the synthesized parts. Section 5 summarizes our sub- missions and details the post-processing steps carried out at prediction time. 2. Parallel Basque–English data The IWSLT organizers released an in-domain data set for Basque-to-English translation containing 64 TED talks for training and 10 TED talks for development [1]. Another 10 TED talks have been held out for testing. The only allowed out-of-domain data source containing parallel Basque–English datasets is OPUS [2, 3]: it contains computer program localization files (repositories GNOME, KDE4 and Ubuntu), crowd-sourced translations (Tatoeba) and film subtitles (OpenSubtitles2018). We only selected OpenSubtitles2018 as the largest and most domain-similar dataset for our experiments. Table 1 summarizes the avail- able parallel data.1 Source Talks Lines EU tokens EN tokens TED train 64 5623 97k 128k OST — 806k 4.8M 6.5M TED dev 10 1140 20k 27k Table 1: Basque–English parallel data. 2.1. Baseline system We trained a baseline system using only the parallel data mentioned in the previous section. Data were tokenized and truecased using the Moses scripts [4]; no effort was spent on adapting the tokenization tools to Basque. Following the good results on various typologically diverse language pairs, we used the Transformer model setup [5] as implemented in Marian-NMT [6] (see Appendix). We used an initial set- ting of 20 000 BPE units [7] shared …

Content truncated to 500 words