ISTIC’s Triangular Machine Translation System for WMT2021

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 320–324 November 10–11, 2021. ©2021 Association for Computational Linguistics 320 ISTIC’s Triangular Machine Translation System for WMT’ 2021 Hangcheng Guo, Wenbin Liu, Yanqing He∗, Tian Lan, Hongjiao Xu, Zhenfeng Wu, You Pan Institute of Scientific and Technical Information of China, Beijing, China {guohc2020, liuwb2019, heyq, lantian, xuhj, wuzf, pany}@istic.ac.cn Abstract This paper describes the ISTIC’s submission to the Triangular Machine Translation Task of Russian-to-Chinese machine translation for WMT’ 2021. In order to fully utilize the provided corpora and promote the translation performance from Russian to Chinese, the pivot method is used in our system which pipelines the Russian-to-English translator and the English-to-Chinese translator to form a Russian-to-Chinese translator. Our system is based on the Transformer architecture and sev- eral effective strategies are adopted to improve the quality of translation, including corpus fil- tering, data pre-processing, system combina- tion, model averaging, model ensemble and reranking. 1 Introduction The Institute of Scientific and Technical Informa- tion of China (ISTIC) participated in the Triangular Machine Translation Task of Russian-to-Chinese in the Sixth Conference on Machine Translation1 (WMT’ 2021). This paper demonstrates the over- all framework of the ISTIC’s submission and its technical details. In this evaluation, we adopted the neural ma- chine translation architecture of Google Trans- former(Vaswani et al., 2017) as a part of our sys- tem. We use the three parallel corpora released by the evaluation organizer and adopted a two-stage method for data pre-processing. Several filtering methods of the corpus are explored to reduce the data noise and improve the data quality. As for model construction, we use the pivot method to get a Russian-to-Chinese translator by bridging the trained Russian-to-English translator and English- to-Chinese translator. Model averaging(Claeskens and Hjort, 2008), model ensemble(Lutellier et al., ∗Corresponding author: Yanqing He, heyq@istic.ac.cn. 1http://www.statmt.org/wmt21/ triangular-mt-task.html 2020) and reranking(Ng et al., 2019) strategies are adopted to generate the final output translation. We removed spaces between words and restored the target language translation results to the prescribed file format in data post-processing. In our experi- ment, the performance of the system under differ- ent settings was compared and further analyzed the experimental results. The structure of this paper is as follows: the sec- ond part introduces the technical architecture of our machine translation system; the third part describes the data pre-processing, parameter settings, experi- mental results, and related analysis; the fourth part gives the conclusion and future work. 2 System Overview The overall framework of the ISTIC’s triangular machine translation system is shown in Figure 1. 2.1 Single Transformer System Our baseline single system used in participated evaluation tasks is the Transformer based encoder- decoder architecture. Transformer is completely based on a self-attention mechanism. It can achieve algorithm parallelism, speed up model training, fur- ther alleviate long-distance dependence and im- prove translation quality(Zhang and Zong, 2020). The encoder and the decoder are formed by stack- ing N identical layer blocks, where N is set to 6. 2.2 Context-based Combination System As shown in Figure 2, …

Content truncated to 500 words