Improving English Interpretation Style Transfer Capabilities by Self-Supervised Semantic Enhancement Strategies
L. Gu · Advanced Electromagnetics · 2026
To enhance stylistic adaptability in English interpretation systems, this paper addresses stylistic inauthenticity, which often appears as rigid or pragmatically inappropriate output when processing domain-specific terminology and technical standards. The objective is to achieve semantically faithful and style-controllable generation without explicit style annotations. A self-supervised semantic enhancement strategy is proposed. Six fine-grained interpretation styles are first defined: Formal-Diplomatic, Academic-Expository, Business-Negotiation, Media-Interview, Casual-Conversational, and Legal-Procedural. Semantically preserved style variants are generated using the T5-large model, and a dual-filtering mechanism is applied to construct high-quality augmented data. Multi-view contrastive learning is then performed using the SimCLR framework. DeBERTa-v3-base is used to extract style-invariant semantic representations. Semantic-style decoupling is achieved by integrating gradient reversal layers with learnable style prototypes. Finally, multi-task joint training optimizes semantic fidelity and stylistic control. Experimental results show that the model outperforms baselines in semantic fidelity, with BLEU-4 of 33.9, chrF++ of 59.8, and COMET-22 of 0.714. It achieves style classification accuracy of 82.6% and a mutual information score of 0.183, validating effective decoupling and high-fidelity multi-style interpretation generation.