PROMT Systems for WMT 2020 Shared News Translation Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 248–253 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 248 Abstract This paper describes the PROMT submissions for the WMT 2020 Shared News Translation Task. This year we participated in four language pairs and six directions: English-Russian, Russian- English, English-German, German- English, Polish-English and Czech- English. All our submissions are MarianNMT-based neural systems. We use more data compared to last year and update our back-translations with better models from the previous year. We show competitive results in terms of BLEU in most directions. 1 Introduction This paper provides an overview of the PROMT submissions for the WMT 2020 Shared News Translation Task. This year we participate with neural MT systems for the third time. We participate in four language pairs and six directions. We describe our data preparation pipelines, models training setups and present the results on the newstest sets. The paper is organized as follows: Section 2 is a brief overview of the submitted systems. Section 3 describes the data preparation, preprocessing and statistics in detail. Section 4 provides a detailed description of the systems. In Section 5 we present and discuss the results. Section 6 concludes the paper. 2 Systems overview We submitted six systems based on the MarianNMT (Junczys-Dowmunt et al., 2018) toolkit: English-Russian, Russian-English, English-German, German-English, Polish-English and Czech-English. All systems are unconstrained (we use the allowed data, private data and publicly available unconstrained data like OpenSubtitles). The English-German and German-English systems have the same basic architecture. The English-Russian and Russian- English systems are slightly different as we use separate vocabularies. The Polish-English system was trained jointly in both directions. The Czech- English is a multilingual system trained to translate from Croatian, Serbian, Slovak and Czech to English. 3 Data We use all data provided by the WMT organizers, private in-house parallel data and other publicly available data, mainly from the OPUS website (Tiedemann, 2012). The human parallel data for the German-English system is exactly the same as for the English-German system, the two systems only have different synthetic back-translated data. This also applies to the English-Russian and Russian-English systems. We use the Tatoeba sets as our validation sets and the newstest2019 is our test set. The reason why we choose the Tatoeba corpus for validation is that we aim at building general-domain (and not just news-domain) models. Besides, the Tatoeba corpus is available for many language pairs beyond the scope of the WMT Translation Task. We only do fine-tuning for the Czech-English system. This will be described in detail in Section 3.4 below. 3.1 Data filtering There are several stages in our data filtering pipeline. The statistics for the final training data are shown in Table 1. Note that for the multilingual Czech|Croatian|Serbian|Slovak- English system the table provides statistics only for the Czech-English part. The size of the filtered versions of the Croatian, Serbian and Slovak parts PROMT Systems for WMT 2020 Shared News Translation Task Alexander Molchanov PROMT LLC 17E Uralskaya str. building …

Content truncated to 500 words