Semantic change detection with Gaussian word embeddings

In this work, we present a novel approach to detecting semantic change over time by modeling words as Gaussian distributions instead of point embeddings. Unlike traditional word vectors, Gaussian embeddings capture uncertainty and allow for more nuanced representations of meaning, particularly useful when analyzing temporal language shifts.

We demonstrate that this probabilistic embedding approach improves semantic shift detection in diachronic corpora, outperforming baseline methods on standard benchmarks. Our experiments include a detailed analysis of word evolution in both English and Turkish, highlighting the effectiveness of Gaussian modeling in multilingual settings.

For implementation details and datasets used, please refer to the supplementary materials or contact the authors.

Recommended citation: Yüksel, A., Uğurlu, B., & Koç, A. (2021). "Semantic change detection with Gaussian word embeddings." *IEEE/ACM Transactions on Audio, Speech, and Language Processing*.
Download Paper