民汉稀缺资源神经机器翻译技术研究

The Study on Ethnic-to-Chinese Scare-Resource Neural Machine Translation

该文介绍了中国科学院自动化研究所参加第15届全国机器翻译大会(CCMT2019)翻译评测任务总体情况以及采用的技术细节.在评测中,中国科学院自动化研究所参加了3个翻译任务,分别是蒙汉日常用语机器翻译、藏汉政府文献机器翻译以及维汉新闻领域机器翻译;阐述了参评系统采用的模型框架、数据预处理方法以及译码策略;最后给出了不同设置下评测系统在测试数据集上的表现,并进行了对比和分析.

The overview and the technical details adopted by Institute of Automation Chinese Academy of Science(CASIA)to participate in the 15th China Conference on Machine Translation(CCMT 2019)evaluation tasks are described in the paper.In the conference,CASIA participates in three translation tasks,including Mongolian-Chinese daily language machine translation,Tibetan-Chinese government literature machine translation,and Uyghur-Chinese news machine translation.The content of the report describes the model framework,datasets pre-processing methods and decoding strategies.Lastly,the report gives the performance of the system on the evaluation dataset under different settings and conducts a comparative analysis.