The site in three walkthroughs

Muted, no sound track; captions are shown in the video and listed beside it. MP4 of Compare two speakers · WebVTT captions for Compare two speakers

Steps

1. 0:00Play from step 1: Distinctive words compares the vocabulary of any two groups of campaign documents.
2. 0:03Play from step 2: Group A: choose a candidate, Hillary Clinton, and the 2016 cycle.
3. 0:06Play from step 3: Group B: Bernie Sanders, in the same cycle.
4. 0:10Play from step 4: Each side's document count and indexed words, and how many words pass |z| = 1.96.
5. 0:14Play from step 5: Every word gets a Fightin' Words z-score; the funnel plot shows all of them at once.
6. 0:18Play from step 6: The ranked lists: a z-score bar and uses per 10,000 words in each group.
7. 0:22Play from step 7: Select a word to see the documents in each group that use it most.
8. 0:27Play from step 8: Every example links back to the full text on The American Presidency Project.
9. 0:30Play from step 9: Resampling documents 200 times shows how stable each top word is.

Muted, no sound track; captions are shown in the video and listed beside it. MP4 of Sixty years of debates · WebVTT captions for Sixty years of debates

Steps

1. 0:00Play from step 1: 179 debate transcripts from 1960 to 2024, split into 44,255 speaking turns.
2. 0:03Play from step 2: One dot per debate: the share of words spoken by moderators and panellists.
3. 0:07Play from step 3: Switch the measure to words per candidate turn: turns have shortened since the 1960s.
4. 0:13Play from step 4: Select a dot to open a debate: Kennedy and Nixon in Chicago, 26 September 1960.
5. 0:16Play from step 5: The debate turn by turn: each block is one turn, as wide as it is long.
6. 0:20Play from step 6: Share of words, a stand-in for talk time, and the length of every turn by speaker.
7. 0:25Play from step 7: Back to the overview to jump 64 years ahead.
8. 0:30Play from step 8: Harris and Trump in Philadelphia, 10 September 2024.
9. 0:39Play from step 9: Four times as many candidate turns, averaging 123 words where 1960 averaged 328.

Muted, no sound track; captions are shown in the video and listed beside it. MP4 of Classical vs LLM topic labels · WebVTT captions for Classical vs LLM topic labels

Steps

1. 0:00Play from step 1: Topic labels: a keyword dictionary and a language model label the same gold-labelled excerpts.
2. 0:03Play from step 2: The keyword baseline runs for everyone: agreement and Cohen's kappa with 95% intervals.
3. 0:06Play from step 3: The model is optional and runs on the visitor's own key. Open AI settings.
4. 0:11Play from step 4: Paste a key (here a placeholder, not a real key). It stays in this browser tab.
5. 0:18Play from step 5: Save. Requests go straight from the browser to the provider, never to this site.
6. 0:22Play from step 6: A seeded sample of 40 excerpts, sent in four requests. Mocked AI response for illustration.
7. 0:31Play from step 7: Both labellers scored on the same excerpts: intervals, paired difference and McNemar's test.
8. 0:36Play from step 8: Every label is marked AI-generated. Correct one, then accept the run.
9. 0:42Play from step 9: Every call is in the AI audit log with prompt, answer, latency, tokens and decision.