As the significance of English translation training in multilingual higher education grows, the introduction of well-developed AI technologies can offer a reasonably good opportunity of improving the outcomes on the side of learners. The paper presents a novel AI-advised model, which permits enhancing the learning of English translation within the frames of college education, which refers to the specially created Bidirectional Error Prediction and Correction Network (BEPC-Net). BEPC-Net is simply a trained education-based model which employs two attention paths and the contextual embedding layers to identify and correct errors in translations completed at all levels of grammar, lexical and semantic levels. An experimental setting has been applied to the system in a college translation course where students were supposed to work with the tool and get instructions of the instructors. Evaluation measures, i.e. precision, recall, accuracy of corrections revealed that BEPC-Net exhibited superior performance compared to the traditional grammar-checking software in the case of idiomatic phrases and domain-specific vocabulary. There was a significant improvement in the performance, regarding translation fluency and structural diversity, which can be explained by the survey feedback data that demonstrates the improved confidence and engagement. The system that is also suggested in this paper does not only automate the detailed feedback but also maintains formative assessment, which reveals the error pattern of the learner in due time. The strength and applicability of BEPC-Net in actual classroom contexts is further justified by an independent assessment with the help of baseline tools and key performance measures.