The AFRL WMT20 News Translation Systems

Proceedings of the 5th Conference on Machine Translation (WMT), pages 207–212 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 207 The AFRL WMT20 News­Translation Systems Jeremy Gwinnup and Timothy Anderson Air Force Research Laboratory {jeremy.gwinnup.1, timothy.anderson.20}@us.af.mil Abstract This report summarizes the Air Force Research Laboratory (AFRL) machine translation (MT) systems submitted to the news­translation task as part of the 2020 Conference on Machine Translation (WMT20) evaluation campaign. This year we largely repurpose strategies from previous years’ efforts with larger datasets and also train models with precomputed word alignments under various settings in an effort to improve translation quality. 1 Introduction As part of the 2020 Conference on Machine Trans­ lation (wmt, 2020) news­translation shared task, the AFRL human language technology team par­ ticipated in the Russian–English portion of the competition. We largely employed our strategies from last year including language­based filtering of training corpora with fastText (Joulin et al., 2016b,a), employing transformer­based (Vaswani et al., 2017) translation models and once again ut­ litizing system combination to fuse outputs from OpenNMT (Klein et al., 2018), Marian (Junczys­ Dowmunt et al., 2018) and Moses (Koehn et al., 2007) systems. We also examine the effects of training Marian models with externally generated word alignments as described in (Alkhouli et al., 2018). 2 Data processing For purposes of training our systems, we use the following parallel corpora: Commoncrawl (Smith et al., 2013), Yandex1, UN v1.0 (Ziemski et al., 2016), Paracrawl2(Esplà et al., 2019), Wikimatrix (Schwenk et al., 2019), and backtranslated data from our WMT17 system (Gwinnup et al., 2017) as well as Edinburgh’s WMT17 system (Sennrich 1https://translate.yandex.ru/corpus?lang=en 2Version 1 Russian–English parallel data et al., 2017) yielding a raw corpus of over 76.3 mil­ lion lines. We prepare training corpora in a similar manner described in (Gwinnup et al., 2018), however this year, we utilize SentencePiece (Kudo and Richard­ son, 2018) with a 46k­entry vocabulary3 for pro­ cessing subword units instead of byte­pair encod­ ing (BPE) (Sennrich et al., 2016). 2.1 Language­ID based data filtering As with last year’s efforts, we again employ fast­ Text (Joulin et al., 2016b,a) to filter the various par­ allel corpora with a utility examining the source and target sentence pairs, discarding pairs where either (or both) sentence in the pair falls below a threshold score of 0.8. We wished to explore dif­ ferent threshold values, but our team did not have access to the majority of our computational assets due to the COVID­19 pandemic, limiting the band­ width available for experiments. We show the results of language­ID based filter­ ing in Table 1. On average, 76.79% of the original training data is retained, with our WMT17 back­ translated data retaining the largest percentage of lines at 93.22% ­ this is interesting since that data originated as English and was translated to Rus­ sian with a very shallow Amun (Hoang et al., 2018) model. Again, Paracrawl yielded the least percent­ age of retained lines at 42.90%, but is understand­ able due to the “raw” nature of this particular re­ …

Content truncated to 500 words