Concept-aware contrastive representation learning for cross-translation semantic consistency modeling of the Analects

As a representative Chinese classical text, the Analects has been translated into English in many versions over a long historical period, and different translators show substantial variation in concept rendering and language style. To address the lack of a quantifiable and reproducible framework for comparing semantic consistency across multiple English translations of the Analects, this paper proposes a concept-aware semantic consistency modeling framework for cross-translation comparison. The proposed method first constructs a unified parallel corpus input based on sentence groups aligned by chapter and sentence IDs. It then introduces concept-term normalization and proper-name normalization to reduce surface lexical variation across translations. In addition, it combines a shared sentence encoder, supervised contrastive learning, and concept-aware constraints to learn sentence representations with cross-translation consistency. Based on the learned representations, the framework is evaluated through two primary tasks, namely cross-translation retrieval and low-consistency detection. Experimental results show that, within the five-translation dataset examined in this study, the proposed method achieves the strongest overall performance across the evaluated settings. The Full model achieves R@1 = 0.936 and F1 = 0.807. The supplementary association analysis further shows that higher model scores tend to be associated with translation pairs labeled as high consistency. These findings support the framework as a computational aid for cross-translation comparison and for identifying translation pairs that may warrant closer expert examination, rather than as a general-purpose method for automatic translation-quality assessment.