Business English Translation Quality Assessment Using Multimodal BERT

Business translation quality assessment faces numerous challenges, particularly for professional technical documents in engineering and industrial applications, where accurate terminology, numerical information, images, and structured tables are essential for reliable information transmission. Such multimodal consistency is also critical for documentation management and knowledge exchange in advanced electromagnetic systems and communication engineering. To address these issues, this paper proposes a multimodal BERT framework for no-reference translation quality assessment. The framework jointly integrates three modalities: text (source document and translation), images, and structured tables. Textual semantics are encoded using BERT, visual features are extracted by ResNet50, and tabular information is represented through TUTA. A multi-head cross-modal attention mechanism is then employed for lexical and cell alignment, while gated weighted pooling adaptively fuses heterogeneous modal representations to regress a unified translation quality score. To validate the proposed approach, a multimodal business dataset is constructed from publicly available reports issued by State Grid Corporation of China and Sinopec during 2023–2024, comprising approximately 12,000 documents, 48,000 sentence pairs, 20,000 image fragments, and 15,000 structured tables, with gold-standard annotations generated using an extended MQM evaluation scheme. Experimental results demonstrate that the proposed method achieves an RMSE of 0.06 for sentence-level quality regression, with Spearman and Pearson correlations of 0.87 and 0.88, respectively, relative to human assessment. In simulated business evaluation scenarios, the framework reduces economic risk by 38.64% and lowers the customer complaint rate to 3.21 %. These results demonstrate that multimodal BERT provides an effective and verifiable solution for translation quality assessment while offering valuable support for multilingual technical documentation, electromagnetic engineering communication, and intelligent information management in complex industrial environments.