~/projects/general-conference/translation
Lost in translation
Most of the Church hears conference in a language other than English, through official translations. This study takes every fourth talk from the first year each language is available on the Church's site, aligns the translation to the English paragraph by paragraph, and measures how close they stay in a multilingual sentence-embedding space, by language, over time, and talk by talk.
What stands out
- Spanish stays closest to the English (mean similarity 0.883); Korean drifts furthest (0.794). The order follows how well the embedding model knows each language as much as how the translators work, which is why the model card matters here.
- Samoan runs to 1.431 words per English word; Portuguese to 0.981; Japanese to 2.651 characters and Korean to 3.049 characters. Length ratios are the model-free measure and the one that holds for Tongan and Samoan too.
- Spanish went from 0.854 in 1990–1994 to 0.892 in 2025–2026. Translations have moved closer to the English, as the talks themselves moved from prepared text read aloud to text published in advance in every language.
- Fidelity is flat across a talk. By fifths of the English, Spanish pairs score 0.884, 0.878, 0.882, 0.879, 0.873: openings, middles and closes translate alike, within a hundredth.
- The Spanish that drifts most is David B. Haight's “This Work Is True” (1996), at 0.671. Audit and statistical reports sit near the top of every language's list, since numbers and proper names embed poorly; the rest are free translations, or transcripts that differ from the text the translators were given.
- What survives is doctrine; what drifts is quotation. Paragraphs with christ, church, satan, god, resurrection, father come through most faithfully; those with savior, tithing, peace peace, zion, right heart, wide least, the vocabulary of hymns and the King James Bible, which each language renders from its own hymnbook and scripture.
Loading the study…
Model card
- Data
- Every fourth conference talk from 1971 by sorted id was requested from the Church's content API in Spanish, Portuguese, French, German, Japanese, Korean, Tongan and Samoan; 623 talks have at least one translation, Spanish from 1990. Translated texts stay off the repo; only per-talk scores are published.
- Alignment
- English and translated paragraphs of six words or more are embedded with paraphrase-multilingual-MiniLM-L12-v2 (384 dimensions, mean pooling, unit norm). Paragraphs are aligned monotonically by dynamic programming on the cosine matrix with a gap cost of 0.25, so a merged or dropped paragraph costs a gap rather than a mismatch. The talk's score is the mean cosine over aligned pairs; unaligned is the share of English paragraphs left without a partner.
- Word model
- A ridge regression (alpha 5) of pair similarity on the English paragraph's TF-IDF words and bigrams, pooled over the six embedded languages, gives the words whose paragraphs translate least and most faithfully.
- Known failure modes
- The embedding model's quality differs by language, so cross-language comparisons partly measure the model; Tongan and Samoan are not among its languages and get structural measures only. Scripture quoted from each language's own Bible edition scores low without being a translation error. Older transcripts differ from delivered text; translators worked from the prepared text. One talk in four is a sample, not the whole.
Methods and data for the corpus as a whole.