Word-domain feature-aware transformer learning for practical multi-domain neural machine translation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Multi-domain neural machine translation (NMT) faces challenges in simultaneously modeling domain-specific lexical patterns and preserving shared linguistic knowledge, particularly when polysemous words occur across heterogeneous domains. To address this problem, this study proposes a Word-Domain Feature-Aware learning framework (WDFS), a lightweight extension of the standard Transformer architecture for multi-domain translation. The framework introduces two complementary modules: a Domain Context Sensitive Mechanism (DCSM), which employs multi-scale convolution to capture contextual domain features at different granularities, and a Reinforced Domain Discriminative Mechanism (RDDM), which enhances weak domain-discriminative feature responses through a trainable activation mechanism. The resulting word-domain features are integrated into the multi-head attention and feed-forward network layers of the Transformer to guide domain-adaptive translation. Experiments are conducted on two multi-domain translation tasks, including English–Chinese translation on five domains of the UM-Corpus dataset and English–French translation on three professional domains of the OPUS dataset. Compared with the word-level WALDM baseline, WDFS improves the average BLEU score by 0.82 and 1.06 on the two tasks, respectively, while improving word-level domain discrimination accuracy by 10.07% and 18.06%. Additional evaluations using COMET and chrF++ reported in the Supplementary Materials show consistent performance trends. Comparisons with larger pretrained multilingual models further demonstrate that WDFS achieves competitive translation performance with substantially fewer parameters. Ablation and hyperparameter analyses confirm the complementary contributions of DCSM and RDDM. These results indicate that contextual word-domain feature learning can improve multi-domain translation quality while maintaining a practical balance between model performance and parameter complexity.