Selective Segmentation Strategies for Chinese-to-English Translation of Long Patent Sentences Based on ParaPat

Long-sentence translation is a challenging task in the processing of Chinese-English patent texts. The traditional whole-sentence translation method (NOSPLIT) and the fully mechanical segmentation method (ALLSPLIT) are prone to information loss, contextual disruption, or under-translation. This study proposes the RLSEGMENT strategy, which uses reinforcement learning to train a policy network to selectively segment candidate sentence boundaries and adopts sentence-level BLEU gain as the reward to optimize translation quality. The experiment uses the ParaPat English-Chinese parallel patent corpus. The results show that the overall BLEU score of RLSEGMENT is 11.9833, which is higher than that of NOSPLIT (10.0127) and ALLSPLIT (9.9518). These results show that RLSEGMENT achieves better BLEU performance and uses fewer segments than ALLSPLIT. It also reduces suspected under-translation compared with NOSPLIT, suggesting that selective segmentation is more practical than uniformly splitting every candidate boundary for patent long-sentence translation. This study analyzes key experimental metrics, visualized results, and model behavior, providing a reference for research on long-sentence translation.