Week 5 · From language to numbers
Wikipedia copies the paperwork, not the plot
KRAK! This week the Marvel network gets its text: 303 plain-text Wikipedia articles, one per character, the same 303 nodes we have been drawing since week 1. Superhero plots are famously shared (team-ups, crossovers, the same event told from six sides), so we expected the pages to be full of recycled plot summary. We asked: when two Marvel pages say exactly the same thing, word for word, what is it they share?
It's mostly not the story. The copied text is the real-world paperwork: who created the character and why, which series they launched in, and which trade paperbacks to buy. The fictional biographies, over a third of all the prose, are almost never copied. And every copied paragraph we found is between two pages that are already linked (one links to the other).
What we did
- Prose only. Each page is cut at its first References / External links / See also / Notes heading, so citation lists can't count as copying. What's left is lower-cased and split into word tokens (730,054 in total).
- Shingles. Every run of 8 consecutive tokens is an 8-gram. We index which pages each 8-gram occurs on. An 8-gram on more than 5 pages is boilerplate, not copying ("is a fictional character appearing in American comic books published by Marvel Comics" is on 282 pages), so we ignore those.
- Passages. For each pair of pages, overlapping shared 8-grams are merged into passages, and we keep passages of 12 tokens or more. Two pages are "copying" if they share at least 30 tokens that way. We call a pair a paragraph copy if its longest passage is 40+ tokens, otherwise a stock phrase.
- Sections. Each token is tagged with the top-level heading it sits under (Publication history, Fictional character biography, Other versions, …), so we can ask where on the page the copied text lives.
1 · The copying is in the paperwork
All the copied text (8,586 tokens, about 1.2% of the corpus), split by the section it sits in, next to how all prose is split. If copying were spread evenly over the page, the two bars would match.
Share of tokens per top-level section, for the whole corpus (grey) and for the copied passages only (pink). The 10 sections that hold the most copied text are shown. Variant headings are merged ("Conception and creation" → Creation, "Alternate versions" → Other versions).
Fictional character biography is the biggest section of a Marvel page, 37% of all prose, but only 5.5% of what gets copied. The pages share crossovers, but each one retells them in its own words. Publication history, Creation and Collected editions go the other way: 20% of the prose, half of the copied text. Collected editions alone is 0.6% of the prose and 7% of the copies. Fairness check: the copying pages might just have unusually long publication histories. They don't: on those 60 pages the section mix is almost the same as the corpus (publication history 19%, biography 30%).
2 · Who copies whom
Now draw it as a network: one node per page, one edge per copying pair. Yellow edges are paragraph copies, grey edges stock phrases, dashed edges pairs with no link between the pages. Click any edge to read the passage the two pages share; click a page to read all of them.
The copying graph is 21 small islands, none bigger than seven pages. The 14 paragraph copies come in four kinds, and we read every one:
- One mantle, several pages. Spider-Woman is an overview page for the name. Its Other versions section shares 150–170-token alternate-universe stories word for word with Spider-Girl's and Mayday Parker's pages, and its publication history shares a long Stan Lee quote with Jessica Drew's page: "…I suddenly realized that some other company may quickly put out a book like that and claim they have the right to use the name… I wanted to protect the name…" The same pattern shows up for Ghost Rider / Johnny Blaze, Union Jack / Joseph Chapman, and Bucky / Rikki Barnes.
- Same team, same launch. Ch'od and Raza Longknife both open with the same story about Dave Cockrum pitching the Starjammers and being told Marvel's try-out titles were "booked for two years solid". Blue Diamond and Jack Frost share the Liberty Legion's history. On these small pages that is 14–25% of the prose.
- One reading list, three pages. Spitfire and both Union Jack pages carry the same Collected editions list, trade paperbacks and ISBNs included (a 140-token stretch).
- A shared behind-the-scenes story. Cyclops and Jean Grey share 607 tokens in 19 passages, including a 104-token account of the X-Factor relaunch: Jean's resurrection, Cyclops leaving his wife and child, Claremont's "dismay", Bob Layton's Ghostbusters inspiration. Eddie Brock and Venom share the origin of the alien costume: John Byrne's idea of a self-repairing costume, passed to Roger Stern, then to Tom DeFalco and Ron Frenz, who made it a living alien. In both pairs the history belongs equally to both characters, so both pages tell it, in the same words.
3 · The small pages are the most borrowed
Per page: what share of its prose also appears, word for word, on another page?
| Page | prose shared | tokens | in-degree |
|---|---|---|---|
| loading… | |||
The top of the list is minor characters: Ch'od (in-degree 4), Blue Diamond (1), Raza Longknife (2), Jack Frost (1). For each, the team or mantle they belong to is most of what there is to say, so a quarter of the page is the team's story. Among the 60 copying pages, the share borrowed goes down as in-degree goes up (r = −0.31 with log in-degree). The famous pages copy too (13% of Cyclops' page is shared, almost all of it with Jean Grey), but it's a small part of a much longer page.
What surprised us
We expected plot, and we got bibliography. With hindsight it makes sense. A plot summary is written from one character's point of view, so the same battle gets retold from each side. A publication history is a fact about the comic, not the character, so when two characters debuted in the same series there is one true paragraph and Wikipedia uses it twice. The one place plot does get copied is an overview page quoting a relative's story (Spider-Woman's Other versions, the Ghost Rider hosts).
Is it robust?
Our three thresholds are choices, so we reran with the boilerplate cut at 3, 5 and 10 pages and the minimum passage at 8, 12 and 20 tokens:
| boilerplate if on > pages | min passage | copy pairs | linked | paragraph copies (linked) | publication + creation | biography |
|---|---|---|---|---|---|---|
| loading… | ||||||
The number of copying pairs swings a lot (21 to 236). Loose settings let in many more short stock phrases of the kind described above. The main result doesn't move: the same 14 paragraph copies, all linked, in every setting, and biography never above 8% of copied text. The row we report is highlighted.
Limitations (friendly criticism welcome)
- Verbatim only. 8-grams find copy-paste, not paraphrase. The biographies may well share content that was reworded, and this method can't see it. "The plot isn't copied" really means "the plot isn't copied word for word". Next week's TF-IDF and cosine similarity are the natural tools for the paraphrased version.
- No direction. A single snapshot can't say which page had the paragraph first. The edit histories could, and we didn't check them.
- Section tags are approximate. Plain text loses heading levels, so we tag tokens with the last heading from a list of known top-level ones. On Ghost Rider, the per-host histories sit under Powers and abilities and get tagged that way.
Reproduce: cd analysis && python week5_copying.py (about a minute; writes
assets/week5_copying.json). Data: the course's frozen Marvel snapshot
(marvel_pages.zip + week-1 network, 2026-08-26).