Research On Korean-chinese Neural Machine Translation Method Based On Reinforcement Learning And Quality Estimation

Machine translation is a core research direction in the field of natural language processing.With the continuous development of deep learning technology,neural machine translation technology has made a breakthrough driven by the combination of large-scale parallel corpus,end-to-end large models and sufficient computational resources.Despite the excellent performance of neural machine translation models on standard datasets,there are still some pressing problems in the field of Korean-Chinese neural machine translation.This thesis investigates a Korean-Chinese neural machine translation approach based on reinforcement learning and quality estimation for three problems of Korean low resources,exposure bias,and poor translation diversity in Korean-Chinese translation tasks.First of all,a quality estimation model based on cross-language pre-training model is proposed to address the shortage of Korean language resources.Based on the concept of attention,this model uses a sentence embedding method that integrates cross-linguistic information.The model pays attention to cross-language information from two aspects: linguistic attention and token attention.This method solves the difficult problem of Korean few-shot quality estimation and enhances the representation ability of cross-language sentence coding.Secondly,in view of exposure bias,reinforcement learning is used to guide translation model training so that the model can avoid using the teacher forcing training strategy.In the training process,the neural machine translation model,as an agent of reinforcement learning,interacts with the environment continuously to receive rewards and decides the best translation target sentence.Finally,to solve the problem of poor translation diversity in machine translation,a quality estimation model is introduced into translation tasks.At each time step of the neural machine translation model decoding,the source sentence and the generated translation snippet were evaluated using the quality estimation model,and the reinforcement learning reward function was set as the combination of BLEU value and QE evaluation score.This method makes the model converge rapidly and increases translation diversity estimation.The Pearson correlation coefficient of the translation quality estimation model proposed in this paper has been improved by 0.226,0.156 and 0.034,respectively,compared with the mainstream models in the quality estimation task field such as Qu Est++,Bilingual Expert and Trans Quest.Spearman correlation coefficient increased by 0.123,0.038 and 0.026,respectively.The QR-Transformer model proposed in this paper effectively improves the performance of the Korean-Chinese neural machine translation.Compared with Transformer,the BLEU value of the Chinese and Korean languages increases by 5.39,the QE score decreases by 5.16,the BLEU value of the KoreanChinese language increases by 2.73,and the QE score decreases by 2.82.In terms of translation diversity,the P-BLEU value of QR-Transformer decreased by 14.47 at most compared with the classical beam-Search model.The experimental results show that the proposed translation quality estimation model and the Korean-Chinese neural machine translation model can effectively improve the performance of corresponding tasks,and significantly improve the Korean low resources,exposure bias and translation diversity.