Neural Machine Translation Error Correction and Optimization System Driven by Machine Learning and Hybrid Algorithms

Neural machine translation (NMT) models are often confronted with mistranslated terminology and ungrammatical sentences due to domain adaptability issues when applying the NMT system to specialized texts, which severely hinders the popular application of the NMT system in specialized fields. Therefore, this paper proposes a novel hybriddriven NMT error correction and optimization system. The system architecture includes rule-based and knowledge-graph preprocessing, LightGBM-based intelligent router to dynamically select error-correcting path, and multi-model collaborative module (using BERT and T5) to deeply correct errors and vote. Experimental evaluation results of specialized financial texts clearly show the superiority of the system. The system obtains the BLEU score of 76.7 and key term accuracy of 93.2%, which is much better than all the baseline models. The intelligent routing module can keep high correction accuracy (89.8%) and reduce the computational resource consumption about 45% compared with the full-path execution strategy. In total, the system can effectively improve the translation quality and resource consumption in specialized fields.