Evaluating the timing and magnitude of semantic change in diachronic word embedding models

Abstract

Recent studies have suggested that diachronic word embedding models are able to track the direction of changes in public perception. Building on these works, we evaluate the ability of diachronic word embedding models to accurately capture such changes both qualitatively and quantitatively, such as their timing and magnitudes. Using a longitudinal dataset on public perception of brands, we found that evolution of word meaning as captured by diachronic word embedding models, trained on New York Times articles, reflected the timing and magnitudes of general consumer awareness of companies. In contrast, this was not the case for other readily available characteristics, such as stock market prices. This comparison is enabled by a new feature extraction method which summarizes the semantic changes encoded in diachronic word embeddings.

Tasks

Reproductions