HSGAR: Hybrid Semantic-Guided Adversarial Reconstruction for semantic-preserving machine translation

Machine translation systems have achieved remarkable fluency and accuracy using Transformer-based architectures, yet they continue to struggle with preserving semantic consistency. To overcome these challenges, this paper proposes HSGAR, a novel Hybrid Semantic-Guided Adversarial Reconstruction framework designed to significantly advance semantic-preserving machine translation across multiple language pairs. HSGAR integrates three core innovations built upon a pre-trained mBART backbone: (1) the Semantic-Guided Attention Module (SGAM), which leverages sentence-level semantic embeddings from Sentence-BERT to dynamically guide the Transformer’s attention mechanism and overcome token-level limitations; (2) the Dynamic Adversarial Loss (DAL), an adaptive objective that dynamically prioritizes semantic fidelity over surface fluency during GAN-based training; and (3) Targeted Fine-Tuning on a curated dataset of 15,000 semantically challenging sentence pairs focusing on idioms, metaphors, and culturally nuanced expressions. Extensive experiments on WMT23 benchmarks and the custom semantic-complexity dataset demonstrate the effectiveness of HSGAR. On WMT23 DE-EN, HSGAR achieves a BLEU score of 35.3 and BERTScore of 0.89; on the custom dataset, it improves BLEU by 4.1 points to 19.3 and BERTScore to 0.91. Substantial gains are also observed in COMET, MEANT, and human evaluation of semantic consistency, particularly for idiomatic and metaphorical expressions. Ablation studies confirm the critical contribution of each proposed component.