面向机器的NMT英语翻译系统研究

Research on machine oriented NMT English translation system

为克服传统NMT系统普适性差的缺陷,在面向人类用户的NMT模型和面向机器的NMT模型基础上,提出了一个面向机器的NMT多任务系统。系统根据奖励标准化策略衡量每个样本在训练过程中的重要性;同时为保证对候选翻译样本进行抽样时的平衡性,采用了Dropout算法提高模型灵敏性,从而避免陷入局部最优。通过仿真分析得出,所提奖励标准化机制训练模型的收敛速度更快,性能更优,扩展NMT系统准确率达到84.5%,明显高于单任务系统和通用NMT系统。

In order to overcome the disadvantage of poor universality of the traditional NMT system,this paper proposes a machine oriented NMT multi-task system based on the study of human user oriented NMT model and machine oriented NMT model.The system measures the importance of each sample in the training process according to the reward standardization strategy.At the same time,in order to ensure the balance of the candidate translation samples,Dropout algorithm is used to improve the sensitivity of the model,so as to avoid falling into the local optimum.Through simulation analysis,the results show that the proposed reward standardization mechanism has faster convergence speed and better performance.Meanwhile,the accuracy of the proposed extended NMT system is 84.5%,which is significantly higher than the single task system and the general NMT system.