Automatic Detection of Word Substitutions within a Language over Periods of Time

Abstract

As natural languages change over time, automatically identifying this change is an area of interest. In this work, we develop a system that detects phrase substitution, new appearing phrases, and extinguishing phrases within a language. We also measure the overall language change. We achieve that using unsupervised learning, i.e., without time-based parallel corpora. Our proposed model employs an adversarial training procedure to learn how to align between time-based word embeddings spaces and time-independent global word embeddings space, within the same language. We also propose a simple and effective dictionary-based validation method. We apply our approach to the Turkish language. As a dataset, we use the Turkish parliamentary minutes from 1920 to 2014. We accomplish promising results; %54.76 accuracy measured by our validation method on words with at least %0.01 probability.