~/projects/general-conference/audience
Who the talk is for
General conference has spoken to three rooms: the general sessions, the priesthood session of men and boys, and the women's sessions. Given only the words, can a model tell which room a talk was given in? And has the answer changed, which is to say, have the rooms grown more alike or less?
What stands out
- The room shows in the words. Without any form of address, the model names the audience of 93% of talks by speakers it never saw; priesthood-session talks are recognised 72% of the time and women's-session talks 84%. With greetings left in, 94%.
- The same speakers change with the room. Among the 206 speakers who addressed more than one kind of session, a model that is allowed to know each speaker's other talks still names the audience 92% of the time. It is the audience, not the voice.
- The priesthood session grew more distinct, not less. Separability (AUC) for priesthood-session talks went from 86 in the 1940s to 95 in the 2010s.
- Women's sessions moved slightly toward the general sessions. Their AUC went from 100 in the 1970s to 96 in the 2010s: still highly distinct, a little less so.
- The mistakes are mostly talks about the other room's subject. A general-session talk on the Aaronic Priesthood reads as a priesthood-session talk; a general-session talk on Relief Society reads as a women's-session talk. The explorer below lists them.
Loading the study…
Model card
- Data
- 8,878 talks of 300 words or more in a general, priesthood or women's session (welfare and leadership sessions excluded). Session kind comes from the session title; the women's sessions include the Relief Society and Young Women meetings held with conference. Since 2019 the rotating Saturday evening session is not tagged as a women's session, so the 2020s have no women's-session class.
- Features and model
- Masked TF-IDF over words and two-word phrases (names, years, magazines and citations removed, 30,000 features) and a multinomial logistic regression with balanced class weights. The forms-of-address model also removes brethren, brothers, sisters, tonight, this morning/afternoon/evening, dear, beloved, young men, young women and priesthood holders before vectorising.
- Validation
- Five folds grouped by speaker for the headline numbers. The within-speaker model restricts to speakers with talks in two or more session kinds and folds by talk, so a speaker's other talks are in the training set; it tests whether the audience, not the speaker, carries the signal. Decade AUCs are one-versus-rest on the out-of-fold probabilities.
- Known failure modes
- Priesthood talks about the priesthood and women's-session talks about Relief Society are what the model learns; a general-session talk on those subjects is scored as the other room. Speaker gender correlates with the women's session; the within-speaker model partly controls for it but women rarely spoke in general sessions before the 1980s.
Methods and data for the corpus as a whole.