Dynamically perturbed Ruppell Fox optimization with a context aware transformer for coherent document level English to Chinese translation

Document-level English–Chinese neural machine translation requires consistent handling of lexical choices, entities, anaphora, and discourse relations across sentence boundaries. Existing context-aware models partially improve cross-sentence consistency, but many rely on fixed context windows, manually selected configurations, or single-path document representations, limiting their adaptability to domains with different discourse structures and terminology densities. This study proposes a unified framework that integrates a Context-Aware Transformer (CAT) with dynamically perturbed Rüppell’s Fox Optimization (DPI-RFO) to jointly optimize document-context representation and model configuration. CAT separates sentence self-attention, document-context cross-attention, and discourse-aware attention and combines them through token-wise gated fusion. DPI-RFO searches mixed discrete–continuous architecture and training variables using elite-guided movement, cosine-decaying perturbation, and stagnation-triggered Lévy-flight escape under a joint quality–cost objective. Experiments were conducted across news, spoken, subtitle, legal, medical, and literary-style domains using five independent seeds, sentence- and document-level metrics, ablation and robustness analyses, statistical testing, professional human evaluation, and computational-efficiency assessment. On the News test set, the selected configuration achieved 50.8 BLEU, 49.5 document-level BLEU, and 67.8 Coherence-BLEU, exceeding the strongest context-aware baseline by 2.2 BLEU points. DPI-RFO reached the target fitness with approximately 30.5% fewer candidate evaluations than baseline RFO. Ablation results identified CAT as the primary source of translation-quality improvement, while paired bootstrap testing confirmed significant gains over all evaluated baselines. These findings demonstrate that coordinated discourse-aware representation learning and automated configuration search improve document-level English-Chinese translation quality and coherence. To clarify evidence priority, BLEU is the primary automatic translation-quality metric, Coherence-BLEU is the primary automatic discourse-level endpoint, blinded human judgments are the primary subjective evidence, and one-time search cost is reported separately from deployment-time inference cost.