~/projects/byu-speeches/topics
What rose and what faded
A topic model does not know what a speech is about; it finds the thirty bundles of words that recur together and measures how much of each speech they explain. Stacked by decade, those bundles show which subjects rose, which quietly faded, and which never moved.
What stands out
- The rising topics are “didn’t · challenges · trials”, “covenant · temple · path”, “mission statement · aims · byu education”. “didn’t · challenges · trials” gained 0.205 points of share per century and peaked in the 2020s.
- The fading topics are “i’m · that’s · it’s”, “missionaries · missionary · stake”, “money · pay · income”. “i’m · that’s · it’s” lost 0.215 points per century after peaking in the 1970s.
- The largest topics are “didn’t · challenges · trials”, “i’m · that’s · it’s”, “charity · loved · christmas”, the main topic of 154, 141 and 128 speechs.
- Conference and BYU drifted apart. On one model fitted to both corpora, the distance between their topic mixes went from 0.2589 in the 1970s to 0.3628 in the 2020s; the joint page has the detail.
- A topic's words move even when its share does not. Open a topic below to see its vocabulary era by era; the model-assigned topic of every speech is now a filter in the database.
Loading the study…
Model card
- Data
- 1,972 speechs of 200 words or more with a known year, prose paragraphs only, masked as in the year study.
- Model
- TF-IDF over words and two-word phrases (in at least ten and at most half of the talks, 25,000 features, English stop words plus the corpus's own stock words such as brethren, gospel and amen removed), then non-negative matrix factorisation with thirty components (NNDSVD initialisation, 400 iterations). Each talk's topic weights are normalised to sum to one.
- Shares and slopes
- A decade's share of a topic is the mean normalised weight across its talks (decades with twenty or more). The slope is a least-squares fit of share on decade, expressed per century. A talk's model-assigned topic is its largest weight.
- Labels
- Each topic is labelled by its three top words; the full fifteen are on the topic panel. Era words are the top terms among the talks whose main topic it is, within forty-year eras.
- Known failure modes
- Thirty is a choice; a topic can mix two subjects or split one. Topics track vocabulary, so a change in house style (the switch from thee to you, the end of the Journal of Discourses reporters) can register as a change of subject. Early decades rest on few talks.
Methods and data for the corpus as a whole.