神经机器翻译的系统融合方法

(苏州大学计算机科学与技术学院,江苏 苏州 215006)

神经机器翻译; 循环神经网络; 系统融合; 注意力机制; 门机制

System combination method for neural machine translation
TAN Min,YIN Mingming,DUAN Xiangyu*

(School of Computer Science and Technology,Soochow University,Suzhou 215006,China)

neural machine translation; recurrent neural network; system combination; attention mechanism; gate mechanism

DOI: 10.6043/j.issn.0438-0479.201903015

备注

为了提高机器翻译模型的泛化能力,基于神经机器翻译系统,将系统融合技术应用于模型训练过程.在神经机器翻译系统的基本结构——编码器-解码器结构的基础上,提出5种融合方法(平均融合、权重融合、拼接融合、门机制融合和注意力机制融合)分别应用于多个编码器-一个解码器的融合、多个编码器-多个解码器的融合和一个编码器-多个解码器的融合.在中英翻译任务上进行实验,相对于基准系统,系统融合方法改进的机器翻译模型的机器双语互译评估(BLEU)值最终提升了0.59~3.01个百分点.实验结果表明,系统融合能有效地提升译文质量.

To improve the generalization ability of the machine translation model, we have applied the system combination technology to the model training process based on the neural machine translation system.According to the encoder-decoder structure of the neural machine translation system, five combination methods are proposed,including average combination,weight combination, concatenation combination, gate mechanism combination and attention mechanism combination.They are applied to the combination of multiple encoders and one decoder,the combination of multiple encoders and multiple decoders and the combination of one encoder and multiple decoders,respectively.Subsequently,the method is applied to the Chinese-English translation task.Compared with the benchmark system,the BLEU value of the machine translation model improved by the system combination method is finally increased by 0.59-3.01 percentage point.Experimental results show that the system combination can effectively improve the translation quality.