Adaptive Transformer-Based Framework for Cross-Lingual Translation Similarity Detection with Bilingual Embedding Alignment
Jiao Jiao · Informatica · 2025
This study proposes a novel deep learning framework for bilingual translation similarity detection that addresses semantic gaps between structurally different languages through an Adaptive Transformer with dynamic masking as the core innovation. The framework features three key components: the adaptive transformer with dynamic content-based and structure-aware masking mechanisms that adjust attention weights based on cross-lingual semantic relevance, cross-lingual feature representation with supervised and unsupervised bilingual embedding alignment strategies, and a multi-dimensional similarity measurement framework incorporating semantic, syntactic, and pragmatic dimensions. Experiments on three language pairs (English-Chinese, English-German, and English-Urdu) demonstrate significant performance improvements, with the proposed method achieving an F1 score of 0.876 — a 7.2% relative improvement over the best baseline (0.817). Ablation studies confirm that adaptive masking and cross-lingual alignment are crucial for handling cultural adaptations and non-literal translations. This research has significant applications in machine translation quality assessment, cross-lingual information retrieval systems, and multilingual plagiarism detection.