《预训练时代的机器翻译.pdf》由会员分享,可在线阅读,更多相关《预训练时代的机器翻译.pdf(57页珍藏版)》请在三个皮匠报告上搜索。
1、Pre-training Methods for Neural Machine Translation 1 2 Language Distributions More than 5000 different languages in the world 3 Machine Translation has increased international trade by over 10% Equality to make the world smaller than 26% 4 Global Footprints of Bytedance 5 Machine Translation: Condi
2、tional Sequence Generation 你 好 吗 ? Encoder Layer Howare Decoder Layer Linear +Softmax Linear +Softmax Linear +Softmax How are you Encoder Beam Search Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder Layer Encoder La
3、yer Decoder Layer Decoder Layer Decoder Layer Decoder Layer Decoder Layer Decoder Layer Decoder Layer Decoder Layer Decoder (|) = =1 (|Fr En-De En-Zh Derived Models Data scarcity for low/zero resource languages. 22 Why Training Multilingual MT Jointly? Arivazhagan et al. 2019 23 Why Training Multilingual MT Jointly? 1 year1 year 1 year3 months Data scarcity for low/zero resource languages. Transfe