~/projects/byu-speeches/stories
The stories they tell
Doctrine is taught through stories: a pioneer handcart, a missionary knocking, a child's question, a hospital bed. This study pulls every story out of every speech using the paragraph-function classifier, embeds the stories with a sentence-embedding model, clusters them into types and the types into families, and asks what gets told, how that has changed by decade, and who tells which kind of story.
What stands out
- 13,167 stories in 1,931 of 1,971 speeches, a median of 123 words each. The largest family is stake, meeting, gospel, felt (26.24% of stories); the largest single type is university, students, campus (5.26%).
- A speech told 8.03 stories in the 1970s and 6.15 in the 2020s; the peak was the 1970s at 8.03.
- Rising: university, students, campus; school, university, student; married, love, marriage. The type that rose most (+5.8 points of share) opens, in one telling, “Often, during the hours that I have spent preparing this message, my thoughts have gone back to my own experiences at BYU.” (Dallin H. Oaks, 1979).
- Fading: old, looked, mother; prophet, revelation, book; water, feet. The type that faded most (-4.61 points) was told like this: “There was a man in his mid-sixties in the center.” (Ronald E. Terry, 2005).
- Presiding Bishopric tell the most stories per speech (9.34); BYU leadership the fewest (4.8).
- 24% of stories open a speech and 16% close one. The rest sit in the middle, where the shape study found the eras differ most.
Loading the study…
Model card
- Data
- Every run of consecutive paragraphs tagged story by the paragraph-function classifier (61% cross-validated accuracy on seven classes), thirty words or more, capped at 450 words, in speeches of 300 words or more: 13,167 stories. Names, years and citations are masked before embedding and term scoring, as in the other studies.
- Embedding and clustering
- all-MiniLM-L6-v2 sentence embeddings (384 dimensions, mean pooling, unit norm, first 256 tokens). K-means with 30 clusters on the embeddings gives the types; Ward linkage on the type centroids, cut into 7 groups, gives the families. Types are labelled by class-based TF-IDF: the terms most distinctive of the type's concatenated text against the other types.
- Trends and exemplars
- A type's share of each decade's stories (decades with thirty or more stories); rising and fading are the change in mean share between the earliest and latest third of decades. Exemplars are the stories nearest each centroid from different speakers, shown as their opening sentence with a link to the official text.
- Known failure modes
- The classifier's story tag is imperfect, so some 'stories' are extended quotations or vivid exhortation, and some stories told in a single short paragraph fall under the thirty-word floor. K-means types are a partition of a continuum, and a story that mixes two themes lands in one. Cluster labels are term lists, not titles; read the exemplars. The Journal of Discourses and OCR'd Conference Reports split paragraphs differently from modern transcripts, which lengthens or shortens their stories.
Methods and data for the corpus as a whole.