~/projects/general-conference/rhetoric
How a point gets made
A paragraph can prove from scripture, cite an authority, bear personal witness, draw an analogy, pose a hypothetical, address the listener directly, or simply assert. This study tags every paragraph with its dominant move, draws the argument shape of a talk from opening to close, and compares the mix across eras, callings and speakers.
What stands out
- Plain assertion is the commonest move at 38% of paragraphs, then Scriptural proof (20%) and Direct address (15%). Analogy and hypothetical together are under 6%: the pulpit argues from texts and from experience, not from thought experiments.
- The shape: scripture and authority in the middle, witness and direct address at the close. Direct address rises from 11% of paragraphs in the first fifth to 19% in the last; personal witness from 9% to 11%.
- From the 1850s to the 2020s, scriptural proof went from 6% to 20% and personal witness from 6% to 11%.
- Apostles and Seventies differ most on plain assertion: apostles use it 5.1 points more than the other, once each decade's mean is removed.
- The tags are weakly supervised and the page says so. Against 50 paragraphs labelled by hand, the rules alone agree 74% of the time and the smoothed model 68%, on seven classes where chance is 14%.
Loading the study…
Model card
- Data
- Paragraphs of twenty words or more in talks of 300 words or more: 263,044 paragraphs.
- Weak labels
- Rules give every paragraph a provisional move, in priority order: scriptural proof (a scripture reference, a quotation the quotes study labelled scripture, three or more archaic forms, or a gospel-narrative cue); personal witness (two or more first-person narrative verbs with no citation cue); authority citation (a leader or author cue, a quotation of sixty characters or more, a bracketed source); hypothetical (imagine, suppose, what if…); analogy (is like, just as, parable…); direct address (brothers and sisters, I invite you, let us, a question put to you…); otherwise plain assertion.
- Model and validation
- A logistic regression over TF-IDF words and bigrams plus the cue flags is trained on the weak labels and its predictions replace them, which smooths rule gaps. A hundred paragraphs drawn at random were labelled by hand before the model ran; agreement with those labels is reported for the rules and for the model. Shapes are the share of each move in each fifth of a talk; the apostle–Seventy contrast is each group's share minus its decade's mean, since 1950.
- Known failure modes
- One move per paragraph is a simplification; a paragraph that quotes scripture and then bears witness is tagged by the rule that fires first. Hand agreement in the sixties on seven classes means a tag on any single paragraph is uncertain; the aggregates by era and calling are what the study claims. Analogy and hypothetical are under-detected, since many analogies use no cue word.
Methods and data for the corpus as a whole.