The Impact of Transformer Model on Englishchinese Translation Accuracy: a Comparative Study

In response to the problems of insufficient modeling of long-distance dependencies and poor adaptability to word order differences in traditional English-Chinese translation models, this paper constructs an encoder-decoder architecture based on Transformer, uses the pre-trained language model parameters to initialize the weights of the Transformer encoder, and uses the rich language features learned by the model to accelerate the training process and improve translation quality. At the same time, small adversarial perturbations are added to the input to improve model performance. The Transformer model outperforms GNMT (Google Neural Machine Translation) in BLEU-4 scores and shows higher accuracy in translating long sentences. The average word order error rate of the Transformer model is about 0.22 %, which is much lower than the GNMT model. This study provides new perspectives and methods for translation practice and research, and promotes the further development of machine translation technology.