For those words with enough mentions, we can even look at the patterns for individual speakers, further demonstrating the way speakers adapt to newer meanings of words over their lifespan.

In other words, older adults are slightly slower than younger adults to adopt newer meanings, but only slightly. In general, adults adapt to new meanings quite quickly, as we can also see by looking at generation-wise plots.

Looking at all of the inferred senses, we report a small but significant effect of age in the vast majority of cases, with the majority of these being slightly positive. The average effect is around 0.13, meaning someone a generation older will sound just a few years out of date.

We justify this approximation, by noting that these interpretable models, with an explicit age effect, closely approximate the results of more expressive GAMMs that allow for arbitrary year by age interactions.

We then use GAMMs to model the probability of a speaker intending a particular sense of a term, as a function of both year and speaker age, where we explicitly model the effect of age as a linear offset in time. Here, for example, is the modeled probability of a particular sense of "articles".

After identifying words that have undergone apparent meaning change during this time, we use masked language models to infer sense clusters, and trace the rise and fall of different senses over time for just under 100 words.

This project, led by Gaurav Kamath, looks specifically at semantic change among adults. Because we needed a large scale dataset with broad coverage, including both timestamps and author age information, we turned to the Congressional Record, covering the years 1873-2010.

I am delighted to share our new #PNAS paper, with @grvkamath.bsky.social @msonderegger.bsky.social and @sivareddyg.bsky.social, on whether age matters for the adoption of new meanings. That is, as words change meaning, does the rate of adoption vary across generations? www.pnas.org/doi/epdf/10....