Autoregressive translation model design integrating syntactic encoding computation and reinforcement learning: a framework for enhancing translation optimization
Huo Li, Liming Cui, Kemal Polat · PeerJ Computer Science · 2025
This study proposes an autoregressive translation model that integrates syntactic encoding computation and reinforcement learning. The model enhances positional encoding by leveraging the strengths of linear transformer and adaptive Fourier transform (AFT), thereby achieving a self-attention mechanism with O(nlogn) computational complexity. To improve translation accuracy and grammatical correctness, the proposed approach incorporates grammatical information from input sentences into the encoder and introduces a component attention module (CAM). This syntactic-aware mechanism significantly improves the model’s capacity to capture hierarchical grammatical structures, yielding a 12.7% relative improvement in translation accuracy on complex syntactic constructions. Addressing the issue of reduced translation quality in noisy input texts, the study employs a gradient-based attack method within reinforcement learning to facilitate adversarial training. Evaluated on the WMT14 En-Fr and WMT17 En-De datasets, our model is compared against several baselines using bilingual evaluation understudy (BLEU) and TwoBLEU scores as evaluation metrics. Experimental results on the WMT14 En-Fr and WMT17 En-De datasets demonstrate the model’s superior performance, with BLEU scores and both twoBLEU scores (bbs) values of 27.31, 8.9, and 20.3, 6.7, respectively. Compared to existing translation models, the proposed model achieves a BLEU score improvement of 4.7% and has a better balance between translation quality and the text generation rate of the Transformer. In conclusion, the autoregressive translation model integrating syntactic encoding computation and reinforcement learning demonstrates significant improvements in optimizing the translation framework and enhancing translation accuracy and efficiency. This research not only introduces innovative methodologies for advancing machine translation technologies but also provides robust support for optimizing language education and translation training programs.