JoWUA

Abstract

The problem of semantic changes detection in natural language is one of the most important issues in Computational Linguistics and NLP, especially in the context of analysis of diachronic corpora at a large scale. Static embedding models fail to properly account for changes in the meaning of words over time and, therefore, exhibit poor interpretability and low sensitivity to changes in contexts. To solve these problems, it suggests an adaptive approach for tracing the semantic changes in Contemporary English based on large-scale digital corpora. It utilizes the framework combining Gaussian word embeddings, temporal alignment via Orthogonal Procrustes transformation, and incremental modeling of semantic drift to detect semantic changes over time. Moreover, the approach provides a probabilistic representation of words and, thus, allows uncertainty-aware semantic representation and better capability of modeling polysemic behavior. For evaluation, it uses a large-scale corpus consisting of around 12 million documents retrieved from such sources as Wikipedia, news archives, Common Crawl, and others within the timeframe of 2000–2025. Our results show that our adaptive model outperforms the baseline approaches, including static embedding models and Bayesian semantic change detection methods. The proposed approach obtains semantic change detection with an accuracy of around 92.6% as well as better precision, recall, F1 score, and AUC values, suggesting that it performs better than previous solutions in detecting lexical semantic change. In addition to this, the inclusion of temporal alignment and probabilistic embeddings leads to much better interpretability and stability when modeling semantic trajectories. Overall, the presented adaptive framework can be considered scalable and efficient for diachronic semantic analysis.