The analysis of thousands of time series over two centuries in different languages shows that the frequency of word usage presents oscillations. We interpret these oscillations as cycles of growth and saturation, similar to fashion cycles, which can be modeled using a logistic equation with distributed delay. Fitting this model indicates that the system is tuned to a Hopf bifurcation, allowing words to be described, in terms of their usage, as phase oscillators. We propose that the statistical organization of words into semantic communities can be explained by an adaptive network of phase oscillators governed by homophily and homeostasis, two forces ubiquitous in nature. This model generates a distribution of communities and effective oscillatory frequencies that align with those observed experimentally.