Russian-English Bidirectional Machine Translation System

Proceedings of the 5th Conference on Machine Translation (WMT), pages 320–325 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 320 Russian-English Bidirectional Machine Translation System Ariel Xv 1 Beihang University, Beijing 100191, China zf1906251@buaa.edu.cn Wenhan Chao 2 Beihang University, Beijing 100191, China chaowenhan@buaa.edu.cn Abstract This review depicts our submission to the WMT20 shared news translation task. WMT is the conference to assess the level of machine translation capabilities of organizations in the word. We participated in one language pair and two language directions, from Russian to English and from English to Russian. We used official training data, 102 million parallel corpora and 10 million monolingual corpora. Our baseline systems are Transformer mod- els trained with the Sockeye sequence model- ing toolkit, supplemented by bi-text data filter- ing schemes, back-translations, reordering and other related processing methods. The BLEU value of our translation result from Russian to English is 35.7, ranking 5th, while from En- glish to Russian is 39.8, ranking 2th. 1 Introduction We participated in WMT20 shared news transla- tion task by building neural translation systems for one language pair and two language direc- tions, from English to Russian and from Russian to English. Our systems are based on the frame- work of the Transformer neural machine transla- tion model, using many techniques and approaches, including the use of BPE subword segmentation for open-vocabulary translation with a fixed vo- cabulary, large-scale back-translation,and model ensembling. Neural machine translation(Bahdanau et al., 2014) has emerged as the most promising machine translation approach in recent years, showing su- perior performance on public benchmarks. The proposed attention mechanism brought a new rev- olution in the neural machine translation in most cases, making the overall effect of translation much better than before. Then, the Transformer(Vaswani et al., 2017) that makes full use of the attention mechanism demonstrated outstanding performance and effectiveness. Up to now, most of work uses the structure of Transformer, and its superiority has been widely recognized. Since the beginning of machine translation re- search, the translation between Russian and En- glish has been extensively developed.As early as 1954, Georgetown University in the United States under the IBM company completed the English- Russian machine translation experiment with the IBM-701 computer, which opened the prelude of machine translation research. During the period, there are three core technologies, rule-based ma- chine translation, statistical machine translation and neural machine translation. However, as the application field of machine translation became more and more complex, the limitations of vari- ous technologies started to become obvious. Due to more application scenarios and higher require- ments for accuracy, model optimization problems appeared. The translation between Russian and English is extremely difficult because their linguistic features are distinguished and the lexical composition and grammatical structure of Russian are more com- plicated than those of English. In the early pe- riod,statistical machine translations were hoped to be implemented through phrase-based meth- ods(Marcu and Wong, 2002) and related techniques for language models and translation models. These methods have solved the Russian-English transla- tion …

Content truncated to 500 words