Training Translation Style: A Q-Learning Approach to Enhancing Syntactic Diversity in AI-Generated English from Chinese Source Texts

This study investigates whether reinforcement learning can mitigate the stylistic flattening observed in AI-generated translations. By training a reward-aware T5-small model on a subset of the UN parallel corpus, we demonstrate that Q-learning can effectively enhance syntactic diversity in machine translations from Chinese to English. Our approach operationalizes three key stylistic features—parse tree depth, clause variety, and lexical distribution—as reward signals to guide the model toward more structurally diverse translations. Results show significant improvements in stylometric measures compared to standard GPT-4 outputs, with human evaluators rating reinforcement learning-optimized translations as more natural and stylistically varied in 73% of cases. This research contributes to translation studies by demonstrating that AI translation systems can be reoriented toward stylistic goals beyond mere fluency, potentially supporting human translators in maintaining stylistic richness across languages.