Papago’s Submissions to the WMT21 Triangular Translation Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 341–346 November 10–11, 2021. ©2021 Association for Computational Linguistics 341 Papago’s Submissions to the WMT21 triangular Translation Task Jeonghyeok Park∗ Papago, Naver Corp. 117033990011@sjtu.edu.cn Hyunjoong Kim Papago, Naver Corp. soy.lovit@navercorp.com Hyunchang Cho Papago, Naver Corp. hyunchang.cho@navercorp.com Abstract This paper describes Naver Papago’s submis- sion to the WMT21 shared triangular MT task to enhance the non-English MT system with tri-language parallel data. The provided paral- lel data are Russian-Chinese (direct), Russian- English (indirect), and English-Chinese (indi- rect) data. This task aims to improve the qual- ity of the Russian-to-Chinese MT system by exploiting the direct and indirect parallel re- sources. The direct parallel data is noisy data crawled from the web. To alleviate the issue, we conduct extensive experiments to find ef- fective data filtering methods. With the empir- ical knowledge that the performance of bilin- gual MT is better than multi-lingual MT and related experiment results, we approach this task as bilingual MT, where the two indirect data are transformed to direct data. In addition, we use the Transformer, a robust translation model, as our baseline and integrate several techniques, averaging checkpoints, model en- semble, and re-ranking. Our final system pro- vides a 12.7 BLEU points improvement over a baseline system on the WMT21 triangular MT development set. In the official evaluation of the test set, ours is ranked 2nd in terms of BLEU scores. 1 Introduction We participate in the WMT21 triangular machine translation task, using the direct and indirect par- allel data to improve Russian-to-Chinese machine translation. The provided data consists of one noisy web corpus (Russian-Chinese, direct translation) and two combined bitexts from several public re- sources (English-Chinese/Russian, indirect). Such cases frequently occur in both actual translation services and research. In particular, this task is crucial in scenarios where we need to improve the performance of non-English translations or low- resource languages with high-resource parallel data. ∗Work done during internship at Naver Corp. Previous works deal with the triangular MT using several methods such as pivot-translation (Cheng et al., 2017), transfer learning (Kim et al., 2019), pre-trained multi-lingual MT (Liu et al., 2020; Tang et al., 2020) and so on. In this paper, we explore existing novel tech- niques to integrate them for the triangular MT tasks. The original direct parallel degrades the transla- tion quality of the model due to the noisy parts containing not well-aligned sentence pairs, erro- neous characters, or the wrong language ID. To discard the noise parts of the noisy web corpus, we filter out data with sequence length, length ra- tio, language ID, de-duplication, and the sentence similarity computed with pre-trained multi-lingual language model (LaBSE, Feng et al., 2020). In pre- liminary experiments, we approached this task in three main ways: bilingual MT, multi-lingual MT, and fine-tuning the pre-trained multi-lingual trans- lation model (i.e., mBART). As shown in Section 4, we found that the bilingual MT outperforms the others. Thus, to augment the Russian-to-Chinese corpus, we conduct two types of data …

Content truncated to 500 words