Structural Optimization of Generative AI Intelligent Machine Translation System Based on Target Search Algorithm

This article focuses on the research of iterative NAT and complete NAT models. By analyzing their development status, two optimization models are proposed: the iterative NAT model based on N-gram mask optimization and the iterative NAT model CMLM. Their shortcomings are analyzed, and two training strategies for optimizing the model are proposed: N-gram mask and consistency learning regularization. Through extensive experimental testing on commonly used public datasets in NMT, the effectiveness of the proposed method has been demonstrated: the N-gram masking strategy can help the model capture coarse-grained information in the target language, which to some extent alleviates the drawbacks of the assumption of a complete NAT model. Consistency learning regularization optimizes the instability caused by operations during model training, thereby improving the performance of the model. This model not only has the advantage of acceleration, but also narrows the gap between its translation quality and the AT model.