微信神经机器翻译系统用于WMT20
计算与语言
2020-11-22 v2 人工智能
机器学习
摘要
我们参与了WMT 2020中英新闻翻译共享任务。我们的系统基于Transformer (Vaswani et al., 2017a)及其有效变体与DTMT (Meng and Zhang, 2019)架构。在我们的实验中,我们采用了数据筛选、若干合成数据生成方法(即回译、知识蒸馏和迭代式域内知识迁移)、先进微调方法以及基于self-bleu的模型集成。我们的受限中英系统取得了36.9的区分大小写BLEU分数,为所有提交系统中最高。
引用
@article{arxiv.2010.00247,
title = {WeChat Neural Machine Translation Systems for WMT20},
author = {Fandong Meng and Jianhao Yan and Yijin Liu and Yuan Gao and Xianfeng Zeng and Qinsong Zeng and Peng Li and Ming Chen and Jie Zhou and Sifan Liu and Hao Zhou},
journal= {arXiv preprint arXiv:2010.00247},
year = {2020}
}
备注
Accepted at WMT 2020. Our Chinese to English system achieved the highest case-sensitive BLEU score among all submissions