Reinforcement Learning-Based Dynamic Model for Enhancing Bilingual Expression in English Translation Teaching

Traditional evaluation in English translation training often misses learners' expressive ability, and existing incentives lack fine granularity. This paper proposes a dynamic teaching framework based on proximal policy optimization. Multi-round learner submissions form temporal state sequences from syntactic and semantic features. A composite reward combines BERTScore-based semantic alignment with n-gram diversity to guide policy learning, while step-size constraints ensure stable proximal policy optimization convergence. The action space spans vocabulary replacement, word-order adjustment, and sentence reconstruction. The optimized policy generates proficiency-aware, personalized feedback that fosters expressive development. Experiments show gains in accuracy and expressive richness (Word Error Rate 0.9%, Semantic Mismatch Rate 2.2%, Textual Lexical Diversity 0.782–0.813, Syntactic Variation Degree 2.41–2.56), surpassing conventional grading and incentive schemes. This paper offers a practical pathway to embed reinforcement learning in translation pedagogy and refine incentive mechanisms at fine granularity.