Implementations for the methods for detecting meaning changes (differences) with their representative instances. All source codes are available under src/. The details of the methods are available in the following paper.
Please cite the following paper if you use the above resources.
@inproceedings{nagata-et-al,
title = "Variance Matters: Detecting Semantic Differences without Corpus/Word Alignment",
author = "Ryo Nagata and Hiroya Takamura and Naoki Otani and Yoshifumi Kawasaki",
booktitle = "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
year = "2023"
}
In Subsection 3.1 in the paper above, max(Kt, Ks)/min(Kt, Ks) is described as a meaning change score. This is incorrect. Correctly, the reciprocal, min(Kt, Ks)/max(Kt, Ks) is used in the following experiment (the smaller the value, the larger the difference). Also, In 3.2, the mean of 1 - cos and 1 - min(Kt, Ks)/max(Kt, Ks) is used as a score for the Mean direction & coverage method. These are for the SemEval2020 semantic change detection experiment, and do not affect the implementations on this page.
Ryo Nagata (nagata-emnlp2023 [at] ml.hyogo-u.ac.jp).