Traditional machine translation systems mainly focus on linguistic and semantic consistency but struggle to handle the ambiguity and context sensitivity inherent in cultural content. To address this, we propose a retrieval-augmented two-stage translation framework R2-MT to improve the accuracy and cultural relevance of translating culture-specific items. In the first stage, knowledge base retrieval is performed based on label similarity and description semantics to disambiguate culture-specific items and generate culturally adapted translations. In the second stage, generated translations as hard constraints are injected into the translation prompt to ensure semantic precision and cultural conformity. To support this process, we constructed a multilingual culture-specific item knowledge base covering 15 language variants and approximately 1.9 million entities, providing a foundation for entity retrieval and disambiguation. Additionally, we designed an instruction fine-tuning dataset covering 10 languages and containing 160000 samples, and jointly trained a large language model on two tasks: entity translation generation and constrained name translation, thereby enhancing the model’s ability to translate culture-specific items. Experimental results demonstrate that R2-MT significantly outperforms strong baselines on multiple cross-cultural translation benchmarks in terms of BLEU, COMET, and cultural relevance metrics. Ablation studies and human evaluations further confirm the effectiveness and complementary advantages of the two-stage mechanism in improving the translation accuracy of culture-specific items and maintaining overall translation quality.