~/projects/byu-speeches/shape
The shape of a speech
Does a speech have a standard shape, an opening, a middle and a close that most speakers follow? Using the paragraph functions of the previous study in sequence, this one draws the average speech by era, from first paragraph to last, and measures how much individual speechs depart from it, which is to say when the shape hardened.
What stands out
- The average speech opens with other and story and closes with testimony and exhortation. The first tenth is where greetings and stories sit; the last tenth is where the testimony formula and the closing charge sit, in every era.
- The shape hardened. The mean distance of a speech's profile from its decade's average went from 0.8878 in the 1970s to 0.8604 in the 2020s; the tightest decade is the 1980s (0.8308) and the loosest the 2010s (0.9242).
- The middle is where eras differ. Compare the panels: the share of story and of doctrine in the middle tenths moves between eras while the ends stay fixed.
- Any speech can be read as a strip. Search below to see the sequence of moves a speech makes; the tags come from the paragraph-function classifier and carry its error.
Loading the study…
Model card
- Data
- 1,971 speechs of 300 words or more, each paragraph of twelve words or more tagged by the paragraph-function classifier (61% cross-validated accuracy on seven classes).
- Shape
- Each speech's paragraphs are assigned to ten equal position bins; the share of each function per bin is averaged across speechs, overall and within three eras.
- Hardening
- Each speech with eight or more paragraphs gets a five-bin by seven-function profile; the spread for a decade is the mean Euclidean distance from each profile to the decade's mean profile. Lower is more uniform.
- Known failure modes
- Classifier error adds noise to every profile but affects decades alike, so the comparison across decades is fairer than any one profile. Short speechs have coarse bins. The Journal of Discourses and the OCR'd Conference Report split paragraphs differently from the modern transcripts.
Methods and data for the corpus as a whole.