~/projects/byu-speeches/audience
Who the speech is for
BYU's podium hosts four occasions: the devotional, the forum, commencement and the university conference. Given only the words, can a model tell which? The devotional and the forum are the interesting pair, since both are addressed to the same students on the same Tuesday.
What stands out
- The occasion shows in the words. Without any form of address, the model names the occasion of 93% of speeches by speakers it never saw: commencement 86% of the time, university conference 80%, devotional 96%.
- Forums are the hard case, and the rare one. Only 44 forum addresses have a transcript of 300 words or more, against 1,619 devotionals, and the model recognises 21% of them; most forums read as devotionals.
- Devotionals are marked by scripture and the sacred vocabulary. The words that most mark a devotional are lord, the lord, unto, the gospel, heavenly; a forum is marked by americans, revolution, century, religious, stevenson.
- The explorer lists the speeches that sound like another occasion. Most are devotionals by visiting scholars that read as forums, and forums by Church leaders that read as devotionals.
Loading the study…
Model card
- Data
- 1,971 speeches of 300 words or more tagged devotional, forum, commencement or university conference in the BYU Speeches archive.
- Features and model
- Masked TF-IDF over words and two-word phrases (30,000 features) and a multinomial logistic regression with balanced class weights. The forms-of-address model also removes graduates, commencement, congratulations, forum, devotional, faculty, staff, trustees and students before vectorising.
- Validation
- Five folds grouped by speaker. Decade AUCs are one-versus-rest on the out-of-fold probabilities, for decades with twenty or more speeches.
- Known failure modes
- Forums are too few to learn well, and the archive's tagging of early forums is uneven. Visiting speakers appear once, so the speaker-grouped design is strict for them.
Methods and data for the corpus as a whole.