~/projects/general-conference/citations
Which verse comes next
A speaker builds toward a scripture, and the words around it give it away. This study takes every paragraph since 1971 that cites one of the four hundred most-cited verses, hides the citation and any quoted scripture, and asks a model to name the verse from the surrounding words. Where it is right, the verse has a vocabulary; where it is confidently wrong, a speaker reached for something unexpected.
What stands out
- The right verse is the model's first guess 20% of the time, and in its top five 36%. Chance is 0.25%: a verse has a vocabulary that precedes it.
- Moses 1:39 is the most predictable verse (91% first guesses from 115 paragraphs; its signal words are work, purpose, immortality, his work); 2 Nephi 31:20 the least among often-cited verses (18%).
- Doctrine and Covenants is the most predictable book at 31% top-five, and prediction was harder in the 2020s (32%) than in the 1970s (39%).
- The surprises are the point. In 2002 a speaker cited Doctrine and Covenants 20:37 where the model was 88% sure of Doctrine and Covenants 20:77; the list below has forty such.
- The browser demo is a lighter model, nearest-centroid in a 160-dimensional space, with top-five accuracy 38% on the same folds; it loads once and runs locally.
Loading the study…
Model card
- Data
- 17,355 paragraphs of churchofjesuschrist.org talks (1971 on) contain a linked scripture; 4,881 of them cite exactly one of the 399 most-cited verses and have at least 25 words after the citation text and any quoted scripture in quotation marks are removed. Names, years, magazines and citations are masked.
- Model
- TF-IDF over words and two-word phrases (30,000 features) and a multinomial logistic regression with 400 classes, five folds grouped by talk. The full model's coefficients give each verse's signal words.
- Demo
- The same paragraphs in an 8,000-word TF-IDF space reduced to 160 dimensions by truncated SVD; each verse is a unit centroid and a paragraph is scored by cosine similarity. Accuracy is reported on the same grouped folds, recomputing centroids within each fold.
- Known failure modes
- A verse quoted without a link is invisible, and a paragraph citing several verses is excluded, so the training set favours single, linked citations. Verses that share a theme (Moroni 10:4 and 10:5, Alma 7:11 and 7:12) are confused with each other and the top-one score punishes that. The demo is a different, weaker model than the headline one.
Methods and data for the corpus as a whole.