The development of globalization has led to increasingly stringent requirements for translation accuracy. This paper designs a cross-language English translation model based on deep reinforcement learning and translation quality assessment, and selects the Transformer architecture with multi-layer encoder-decoders. Through the reward and punishment mechanism of the intelligent NMT system, real-time probability calculations of contextual information are performed to select the most appropriate words at the semantic level as components of the target sentence. Combined with a supervised quality assessment module, the translated text is scored, and the next word selection is guided. Experiments show that after 934 iterations, the BLEU score stabilizes around 97.62%. The F1 score reaches 99.87% after 316 iterations, and the accuracy achieves a stable value of 90.74% after 815 iterations. In two-class cross-language English translation tasks, the model’s average BLEU scores were 90.67% and 93.08%.