基于强化学习和机器翻译质量评估的中朝机器翻译研究

针对目前机器翻译模型存在的曝光偏差和译文多样性差的问题,提出一种基于强化学习和机器翻译质量评估的中朝神经机器翻译模型QR-Transformer.首先,在句子级别引入评价机制来指导模型预测不完全收敛于参考译文;其次,采用强化学习方法作为指导策略,实现模型在句子级别优化目标序列;最后,在训练过程中融入单语语料并进行多粒度数据预处理以缓解数据稀疏问题.实验表明,QR-Transformer有效提升了中朝神经机器翻译性能,与Transformer相比,中—朝语向BLEU值提升了5.39,QE分数降低了5.16,朝—中语向BLEU值提升了2.73,QE分数下降了2.82.

Research on Chinese-Korean machine translation based on reinforcement learning and machine translation quality estimation

This paper proposed a Chinese-Korean neural machine translation model called QE-Transformer based on reinforcement learning and machine translation quality estimation to solve the problems about exposure bias and diversity in current machine translation. First of all, this model introduced an evaluation mechanism into the sentence level to guide the model when the prediction couldn''t converge completely on the ground truth. And furthermore, it used reinforcement learning as a guiding strategy to optimize the sequence of objectives at the sentence level. Finally, by using monolingual data and different granularity data preprocessing in the training process, it alleviated the problem of data sparsity. Experimental results show that the QE-Transformer can improve effectively the Chinese-Korean machine translation performance compared with Transformer, the BLEU value increases by 5.39 and the QE score decreases by 5.16 in the direction of the Chinese to Korean machine translation, the BLEU value increases by 2.73 and the QE score decreases by 2.82 in the direction of the Korean to Chinese machine translation.