Application of Minimax Optimization Mechanism in Chinese-English Machine Translation Quality Estimation

Machine Translation Quality Estimation (MTQE) is pivotal in bridging the gap between machine-generated translations and human translation quality, especially in real-time applications where post-editing is not feasible. Despite advancements with Neural Machine Translation (NMT), challenges such as mistranslation, omissions, and over-translation persist. Traditional MTQE models often suffer from incoherent optimization goals due to their dual phase architecture, limiting their effectiveness. In this study, we introduce a novel approach that integrates minimax optimization to unify the prediction and estimation phases under a common optimization goal. Utilizing the Ranger optimizer, our model comprises a generator based on the T5 (Text-to-Text Transfer Transformer) and a discriminator leveraging Convolutional Neural Networks (ConvNet). Additionally, we incorporate a data reliability screening module to ensure the discriminator is trained on high-quality data. Experiments conducted on the Chinese-English corpus demonstrate that the superiority of proposed approach. Furthermore, the generator maintains comparable BLEU scores to baseline models, confirming that the minimax optimization mechanism does not compromise its performance. Ablation studies highlight the optimal settings for the optimizer, pooling strategies, and learning rates, underscoring the importance of data reliability screening in achieving stable training outcomes. Our findings indicate that minimax optimization is a viable strategy for enhancing MTQE models, offering a path toward more accurate and reliable translation quality assessments.