~/projects/general-conference/quotes
Who gets quoted
A talk borrows its authority from what it quotes. This study finds every quotation in every talk and classifies its source: scripture, a leader of the Church, a secular author, or nobody named. The share of each over time is a measure of how outward-looking the pulpit has been.
What stands out
- Scripture is 47% of all quotations, Church leaders 37%, secular authors 6%. The rest, 10%, carry no attribution the method can resolve.
- Secular quotation peaked in the 1930s at 11% and is 2% in the 2020s; scripture went from 26% of quotations in the 1850s to 51%.
- Joseph Smith is the most-quoted leader, Benjamin Franklin the most-quoted secular author. 874 quotations name Joseph Smith; 11 name Benjamin Franklin, with George Washington, Thomas Carlyle, Edgar Hoover next.
- Talks carry 11.12 quotations each in the 2020s, against 8.56 in the 1850s.
- Patriarch speakers quote scripture most (65% of their quotations); Other speakers quote secular authors most (10%).
- The classifier is honest about its limits. On the rule-labelled quotations it is 87% accurate out of fold, strong on scripture, weakest on secular authors, so the secular share is a floor rather than a ceiling.
Loading the study…
Model card
- Data
- 96,722 spans inside double quotation marks of six words or more, in 10,686 talks of 100 words or more. The sentence before the opening mark and 160 characters after the closing mark form the context.
- Rules
- Scripture: at least half the quotation's six-word shingles occur in the standard works (every verse, with runs across verse boundaries). Leader: the context names a leader in the leadership database (a Church title and a name, resolved to the most senior bearer of a surname called by that year; first and last name without the initial also resolves), or the quotation matches an earlier talk in the corpus, or the context cites conference, the Twelve or a Church magazine. Secular: the context names another person next to an attribution verb. 60,926 quotations were settled this way.
- Classifier
- A logistic regression (balanced classes) over TF-IDF of the context and the quotation, trained on the rule-labelled quotations, labels the rest; a prediction under 60% confidence is marked unattributed. Out-of-fold accuracy on the rule-labelled set, grouped by talk, is reported above and per class in the data file.
- Known failure modes
- Dialogue in a story is quoted too, and reads as unattributed or secular. A leader quoted without a title and without a middle initial can be missed. The Journal of Discourses and the OCR'd Conference Report use quotation marks loosely. A paraphrased scripture is not matched.
Methods and data for the corpus as a whole.