Skip to content
james_tavita

~/projects/byu-speeches/meaning

How words changed meaning

A word is known by the company it keeps. Embeddings trained separately on each era of the corpus and then aligned let the same word be compared across eras: the words nearest to it in the 1850s against the words nearest to it today. The distance between the two is how far its meaning moved.

Unsupervised3 eras1,841 words measuredWord lookup ~/byu-speeches

What stands out

  • “press” moved most. In 1950–1979 it kept company with freedom, media, enterprise, economic; in 2000–2026 with steadfastness, feasting, forward, move. Drift 0.695.
  • The next biggest moves are “unknown”, “normal”, “balance”, “relative”. Most are words whose referent changed: an institution reorganised, a practice renamed, a technology arrived.
  • On the watch list, “disciple” moved most (drift 0.5766), then “ministering” (0.542), “gathering” (0.4398), “gentiles” (0.4326).
  • “years” barely moved (drift 0.106): the core religious vocabulary is the most stable part of the language.
  • The eras are 1950–1979, 1980–1999, 2000–2026, with 1.1M, 2.5M, 3.8M words of text respectively.
Loading the study…

Model card

Data
Prose paragraphs of every speech of 100 words or more, split into sentences, lowercased, with names, years, magazines and citations masked as in the year study. Eras: 1950–1979 (250 speechs); 1980–1999 (588 speechs); 2000–2026 (1,133 speechs).
Embeddings
Skip-gram word2vec per era: 100 dimensions, window 6, ten negative samples, eight epochs, words with fifteen or more uses. Each era's space is aligned to the latest era by orthogonal Procrustes on the 5,000 most frequent shared words (Hamilton, Leskovec and Jurafsky, 2016).
Drift
Cosine distance between a word's aligned vectors in the first and last era in which it has fifteen or more uses; step drift between consecutive eras. Ranked among words with forty or more uses in every era. Neighbours are the ten nearest words with twenty or more uses in that era.
What is published
Only the words that moved most and least and the watch list, with their neighbours per era; the embeddings themselves stay offline. A word not on the page did not meet the frequency threshold in every era.
Known failure modes
Alignment is imperfect, so a small drift is noise. A word used in a fixed phrase (a hymn line, a scripture) looks stable whatever it means. Transcription changes (the Journal of Discourses reporters, OCR) shift neighbourhoods. Masking removes names, so a word that moved because of whom it was used about will show the move indirectly.

Methods and data for the corpus as a whole.