<p>Although the MNMT model has shown some generalization ability in zero-shot translation, existing methods generally suffer from insufficient cross-lingual semantic representation alignment, language label confusion, and unstable translation quality due to the lack of identification of language pairs. To address these issues, this paper proposes a zero-shot translation enhancement algorithm based on dual semantic decoupling and explicit path regularization. This method first introduces a language-independent semantic subspace within the shared encoder. It decouples source-language semantics from language identity via joint constraints from an adversarial language discriminator and a semantic reconstruction loss. Secondly, during the training phase, a virtual zero-shot path X → Y is explicitly injected, where (X, Y) does not appear in the training data, and a bidirectional consistency regularization term is designed to enforce semantic equivalence across the X → Y and Y → X paths. At the same time, the introduction of the target language perceptual decoding initialization strategy enhances language purity. The effectiveness of the proposed method was validated on the Flores-101 and TED Talks benchmarks, and its translation performance was significantly better than that of the baseline model. The average Bilateral Evaluation Understudy (BLEU) was 27.8, the Language Purity Rate (LPR) was 96.3%, and the Code-switching Rate (CSR) was 2.1%, significantly improving the robustness and controllability of zero-shot translation in low-resource and cross-linguistic scenarios.</p>