Improving the efficiency of contextual semantic matching of terms in fashion English translation based on the Dual Encoder mechanism

In the fashion e-commerce domain, accurately translating domain-specific content is essential to maintaining the integrity of the brand identity, cultural specificity, and customer connection. Contrasting general translation tasks, fashion translation is more challenging because it involves rapidly evolving terminology, stylistic nuances, and fine-grained descriptions of textures and styles that generic models often fail to capture. Traditional translation systems have struggled with properly interpreting context-reliant fashion terms, as they are generic language model systems with little to no domain knowledge. This paper proposes a new deep learning framework, using a domain-specific transformer model, Fashioner, in combination with a Dual Encoder architecture, to produce improved contextual semantic matching for translations of fashion English. Our model encodes the source and target sentences in an independent fashion using a feature fusion and interaction layer that approximates global semantics in addition to distinguishing stylistic elements. A lightweight classifier is then used to evaluate semantic alignment using the cosine similarity method. The model was evaluated on a fashion-specific bilingual dataset and achieved a high level of accuracy, 98.35%, a precision of 98.39%, a recall of 98.30%, and an F 1 -score of 98.34%. Results showed that the proposed approach outperformed baseline approaches such as DSSM, PICSO, and FashionSAP on benchmark datasets, serving as a comparative point of reference for all approaches. Overall, the proposed approach is an effective tool for cross-lingual fashion translation, as it produces a greater degree of accuracy and contextual sensitivity in matching fashion-specific sentences versus previous approaches, given the context of the fashion domain.