Dynamic Alignment and Consistency‐Regularized Diachronic Semantic Embedding Learning for Corpus‐Based Knowledge Evolution in <scp>IIoT</scp> 5.0 Environments

ABSTRACT Industry 5.0 IIoT platforms increasingly depend on temporally evolving textual knowledge, yet semantic drift and representation non‐identifiability impede reliable cross‐time retrieval and monitoring. This work studies diachronic semantic embedding learning under practical constraints of sparse time slices, heterogeneous domains, and the need for geometrically comparable embedding spaces, with particular relevance to English diachronic corpora and evolving technical text. We propose DyAlignCL, a dynamic alignment and consistency‐regularized framework that jointly optimizes intra‐slice distributional learning, orthogonal Procrustes alignment, second‐order temporal smoothness, and structure‐preserving contrastive constraints. Specifically, the proposed framework integrates dynamic orthogonal alignment with temporal consistency regularization in a unified learning objective, enabling stable cross‐time semantic comparability while preserving genuine semantic evolution. Experiments on four public diachronic corpora show that DyAlignCL improves semantic shift detection accuracy while consistently reducing temporal alignment error and enhancing neighborhood stability against representative baselines. These results demonstrate that explicitly coupling alignment and temporal regularization provides robust and interpretable diachronic semantic embedding for evolving IIoT knowledge environments, including English knowledge evolution scenarios.