DA-MK: Dynamic Attention and Multimodal Knowledge Fusion for Error Tracing and Proofreading Optimization in Neural Machine Translation

Abstract

With the widespread application of neural translation systems, challenges such as difficult error tracing and low proofreading efficiency remain significant. To address this, a DA-MK model is proposed, integrating dynamic attention weighting with multimodal knowledge fusion, enabling fine-grained error localization and correction. The model incorporates explicit grammar and syntax metrics, including syntactic dependency parsing accuracy and grammar correction rate, ensuring robust linguistic consistency. Experiments on the WMT 2014 English-German and IWSLT 2016 German-English datasets benchmark DA-MK against advanced models such as ErrorFocus, KG-Translate, and BERT-Fix. Results demonstrate an error location accuracy of 90.3%, error type classification accuracy of 87.2%, proofreading suggestion adoption rate of 79.8%, BLEU score improvement of +13.0, syntactic parsing accuracy of 88.6%, and grammar correction rate of 85.1%. These findings confirm DA-MK’s superior capability in enhancing translation reliability, grammatical integrity, and proofreading efficiency. The study contributes a technically grounded pathway for optimizing neural translation systems with strong theoretical and practical significance.