DUTNLP Machine Translation System for WMT21 Triangular Translation Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 331–335 November 10–11, 2021. ©2021 Association for Computational Linguistics 331 DUTNLP Machine Translation System for WMT21 Triangular Translation Task Huan Liu Junpeng Liu Kaiyu Huang∗ Degen Huang School of Computer Science, Dalian University of Technology {liuhuan4221,liujunpeng_nlp, kaiyuhuang}@mail.dlut.edu.cn {huangdg}@dlut.edu.cn Abstract This paper describes DUT-NLP Lab’s submis- sion to the WMT-21 triangular machine trans- lation shared task. The participants are not allowed to use other data and the translation direction of this task is Russian-to-Chinese. In this task, we use the Transformer as our baseline model, and integrate several tech- niques to enhance the performance of the baseline, including data filtering, data selec- tion, fine-tuning, and post-editing. Further, to make use of the English resources, such as Russian/English and Chinese/English parallel data, the relationship triangle is constructed by multilingual neural machine translation sys- tems. As a result, our submission achieves a BLEU score of 21.9 in Russian-to-Chinese. 1 Introduction The WMT2021 Shared Task on translating sen- tences from Russian into Chinese provides a challenging mixed-genre test for machine trans- lation systems and a triangular relationship for re- searchers to evaluate new techniques. The task focuses on translation between non-English lan- guages and optimally mixing direct and indirect parallel resources. In this task, the participants must use only the provided parallel training data and use of other data is not allowed. The provided data is shown in Table 1, including parallel data in three directions. Given the language pair (Russian- to-Chinese), the bulk of previous NMT work has pursued one of two strategies that are illustrated in Figure 1: Direct: Collect parallel Russian-to-Chinese data from the public resource, and train a Russian-to- Chinese translator. Pivot: Collect parallel Russian-to-English and English-to-Chinese data (usually larger than direct data), train two translators (Russian-to-English + ∗Corresponding author Direction # SENT Russian/Chinese 33,388,455 Russian/English 69,155,404 English/Chinese 28,528,290 Table 1: The provided training data in the constrained data track. ������� ������� ������� �������� ����� �������� ����� �������� ����� ����� ������ Figure 1: The illustration of two strategies for triangu- lar machine translation. English-to-Chinese), and make a cascade translator from Russian to Chinese. The DUTNLP submission to the constrained data track is based on the mainstream architecture Transformer (Vaswani et al., 2017). According to the scale of datasets, this shared task should be considered as the high-resource translation di- rection. We use the Transformer-big setting for better performance, on the contrary, a low-resource translation task often utilizes the Transformer-base due to the limited parallel training data. Moreover, to enhance the baseline model and investigate the usage of triangular direction data, we utilize two pipelines for combining English related resources: 1) incorporates English-to-Chinese translator into direct translation process to normalize the transla- tion of rare words. 2) adopts a multilingual training strategy to make use of English ↔ X parallel re- sources. Besides, some of the provided Russian ↔ Chinese parallel corpora are crawled from the 332 web, which has many noise issues. We filter the training bilingual corpora with several …

Content truncated to 500 words