Q's discourse
What the 4,735 non-empty drops say (themes, vocabulary, ritual formulas, style), and how the way they say it changes.
Themes
Non-negative matrix factorisation (NMF) on TF-IDF weights of words and bigrams, 12 themes. Themes are not hand-labelled, so their most characteristic terms are shown. Share = average weight of the theme in the corpus.
| # | Characteristic terms | Share | Typical drops (#) |
|---|---|---|---|
| 1 | think, potus, did, public, control, important, does, time, fbi, thinking | 43.4% | 2937, 1318, 236 |
| 2 | people, stupid, people stupid, sick, people sick, truth, power, evil, enemy, enemy people | 10.3% | 3200, 4323, 3609 |
| 3 | news, fake, fake news, unlocks, news unlocks, news media, media, coming, narrative, attacks | 9.2% | 2002, 268, 2950 |
| 4 | god, bless, god bless, thank, patriots, country, fight, love, good, stand | 8.5% | 1526, 323, 1171 |
| 5 | know, past, future, proves, proves past, future proves, soon, choice, choice know, knowing know | 5.1% | 3162, 1562, 2046 |
| 6 | coincidence, post, amazing, msm, error, ahead, day, guard, enjoy, drop | 4.8% | 800, 4561, 2457 |
| 7 | coincidences, believe coincidences, believe, think, mathematically impossible, mathematically, impossible, panic, light, enjoy | 3.9% | 3183, 901, 3044 |
| 8 | confirmed, trip, hunting, happy, happy hunting, good, panic, graphic, correct, marker | 3.8% | 3795, 3500, 2944 |
| 9 | patriot, vip, vip patriot, enjoy, enjoy vip, eyes, eyes vip, vip patriots, thank patriot, thank | 3.8% | 3273, 3274, 3272 |
| 10 | wwg1wga, united, win wwg1wga, stand, stand wwg1wga, win, patriot, awakening wwg1wga, proud, digital | 3.3% | 4113, 4071, 4070 |
| 11 | remembering, worth remembering, worth, note, soon, following, infiltration, planned, fight, border | 2.1% | 3699, 3711, 3813 |
| 12 | test, sec test, device, sec, trip, possible, failed, infiltrated, activate, posts | 1.9% | 401, 406, 2907 |
Theme weights, month by month
Average weight of each theme in the month's drops. Theme named by its first three terms. 2022 (13 drops) is not shown.
Distinctive words of each year
Log-odds ratio with informative Dirichlet prior (Monroe, Colaresi & Quinn, 2008) picks out words over-represented in a year relative to the others (z-score).
| 2017 | relevant, bank, central, crumbs, huma, map, controls, military, does, soros, alice, picture |
|---|---|
| 2018 | huber, sessions, trust, know, boom, doj, think, comms, coincidence, fisa, panic, eyes |
| 2019 | barr, evidence, news, narrative, transparency, fake, illegal, facts, blockade, mueller, attempt, knowingly |
| 2020 | biden, wwg1wga, democratic, child, political, humanity, warfare, election, sex, control, party, media |
Ritual formulas
Monthly occurrences of the most frequent glossary formulas.
Style over time
Monthly share of drops containing a question, and of drops containing a link (months with at least 5 drops, up to December 2020).
Exploratory stylometry: one author or several?
In 2022, two independent teams (OrphAnalytics; Florian Cafiero and Jean-Baptiste Camps) concluded through stylometry that the early drops were written by Paul Furber and later ones by Ron Watkins, with a switch around late 2017 – early 2018. A simple, transparent version of this kind of analysis is reproduced here, without reference texts from candidate authors.
The sharpest split falls on Nov 25, 2017. A second split, in the later part, appears around Jul 29, 2019. This date coincides with the first drop on 8ch /pol (Nov 25, 2017), so a change of author and a change of platform cannot be told apart here.
The five best split dates (between/within ratio) — differences this small show the exact date is fragile: Nov 25, 2017 (1.127), Dec 9, 2017 (1.125), Dec 19, 2017 (1.118), Nov 20, 2017 (1.117), Nov 11, 2017 (1.115)
Stylistic position of each chunk
Negative values are closer to chunks before the split, positive values closer to those after. The vertical rule marks the split.
Method
Consecutive drops are grouped into 89 chunks of about 1,500 words, and each chunk is described by the frequencies of the 150 most frequent words (mostly function words). Burrows' Delta measures the distance between chunks, and we look for the split date that best separates them (between-group / within-group distance ratio).
The limits come from very short texts, shifting topics, quotes and link lists. A change in style may also reflect a change of platform, topic or format. This analysis does not identify any author.