An English Translation Model with Transformer and Memory Augmentation

Abstract

Machine translation, as a core approach in natural language processing, plays a crucial role in promoting cross-lingual communication. This study proposes an English-to-target language translation model that integrates enhanced back-translation with translation memory, aiming to address the issues of low training efficiency, slow convergence, and weak domain adaptability in traditional machine translation models for English-to-multilingual tasks. The model accelerates translation training from English to the target language through enhanced back-translation and leverages translation memory to enhance the capture of specialized terminology and domain knowledge. Experimental results show that the proposed model reaches a loss function value of 0.060 at the final iteration, achieving an accuracy of 85.55% and an F1 score of 84.60%. In translation performance tests, the model achieves BLEU scores of 55.10, 51.47, 59.33, and 54.22 on four domain-specific datasets, all substantially higher than those of the comparative models. Case studies involving Russian, German, and French translations, along with heatmap visualization, further confirm its strengths in terminology accuracy and cross-lingual alignment. The innovation of this research lies in effectively integrating back-translation optimization with translation memory, significantly enhancing the efficiency, accuracy, and domain adaptability of machine translation. This provides a novel solution for high-performance and practical English-source language translation systems.