Computational narratology · a showing
The Shapes of Stories
Kurt Vonnegut said every story has a shape you can draw on graph paper — and that a society's stories are as worth mapping as its pots and spearheads. Decades later a computer read 1,327 books from Project Gutenberg's fiction collection and found he was mostly right: a handful of shapes cover most of them. Below are the real emotional arcs of 29 famous books, each computed from its own words. Pick one and watch its shape.
After the war Vonnegut studied anthropology at the University of Chicago, and the master's thesis he eventually submitted there argued that stories have shapes you can draw. The university rejected it. (The University's 2007 obituary has it rejected before he left Chicago in 1947; his biographer Charles Shields dates the story-shapes thesis to 1965, written while Vonnegut taught at Iowa.) He never quite got over how much fun the idea was.
“What has been my prettiest contribution to my culture? I would say it was a master's thesis in anthropology which was rejected by the University of Chicago a long time ago. It was rejected because it was so simple and looked like too much fun. One must not be too playful.” — Kurt Vonnegut, Palm Sunday (1981)
His method: two axes. A vertical one running from ill fortune at the bottom to good fortune at the top (he called it the G–I axis), and a horizontal one from the beginning of the story to its end. Then you draw the line. Here are three shapes he drew himself — these three names are genuinely his1:
That's the intuition. Now the question the Wasteland cares about: is it true? Can you actually measure the emotional shape of a real book — not by asking a critic, but by reading the words themselves?
The instrument
In 2016, five researchers, four at the University of Vermont and one at the University of Adelaide, did exactly that. They took 1,327 works of English fiction from Project Gutenberg and ran a sliding window across each one, scoring every word for happiness against a lexicon of 10,000+ words, each rated 50 times by people on Amazon's Mechanical Turk. The window's average traces an emotional arc — how the language rises and falls from first page to last.2
The tool below shows the result of doing the same thing to 29 books you know. Each was downloaded from Project Gutenberg, stripped of its front and back matter, and scored word-by-word with a revised version of the same labMT lexicon and the same “lens” that drops neutral words. Tap a title to draw its arc; tap more to compare. Hover or drag across the plot to read any point.
Loading the arcs…
The vertical position is real: grimmer books sit lower. Tom Sawyer and Heart of Darkness live near the floor; Little Women floats at the top. Everything sits above the lexicon's neutral line, because human language carries a measurable positivity bias — one of the same team's other findings.3 Flip “align shapes” to strip the height away and compare pure shape.
The six shapes
Across all 1,327 books, three different methods — a matrix decomposition (SVD), hierarchical clustering, and a self-organizing neural map — kept surfacing the same small family. The paper's exact and carefully hedged title: the emotional arcs of stories are dominated by six basic shapes. Not the only shapes — the dominant ones. They come as three arcs and their mirror images:
Tap a shape to load every book here that best fits it. The label under each book is its closest family and the correlation r — how tightly its real arc hugs that ideal. A low r is honest: some books (here, Crime and Punishment and The Wizard of Oz) match nothing cleanly, so we say so rather than force them.
The trap this instrument is built to reveal
Watch The Metamorphosis. Kafka's novella scores as a rise — its emotional line climbs toward the end. But the plot is the opposite: Gregor Samsa wakes as an insect and slowly dies. What lifts is the language — the final pages turn to the family's relief and the sister blooming into a young woman, and those words are warm. The arc measures the sentiment of the words, not the fate of the hero.
The paper's authors flag this themselves: an emotional arc “does not give us direct information about the plot.” Emma reads as a fall, Hamlet as a rise-fall-rise — surprises that dissolve the moment you remember what is actually being measured. That gap is the lesson, not a bug in it.
The check
Every arc above is recomputed from raw text by a script you can download and run (build_arcs.py: Python 3, standard library only). Two independent inputs, both fetched live: the labMT-en-v2 happiness lexicon from the study authors' own hedonometer.org, and the plain-text books from Project Gutenberg (public domain). Here is the whole method, and where it holds and where it doesn't.
What we reproduced — and what we couldn't
The paper's core move is a singular-value decomposition of the stacked, mean-centred arcs: the leading “modes” of that matrix are supposed to be the basic shapes. We ran the same decomposition on our 29 books (svd_check.py, which needs numpy and is in the site's repository only), centring each arc on its own mean as the paper's code does:
The honest verdict: mode 1, the overall rise-vs-fall axis, reproduces cleanly (r = 0.96 with a straight line) and is the single biggest way these stories' shapes differ, though only just (22% of the variance, against 19% for mode 2). But the tidy “mode 2 = one swing, mode 3 = two swings” ordering of the full study does not emerge at 29 books; the swing shapes scatter into weaker, noisier modes. That gap is exactly what the paper's 1,327 books buy you. We show a demonstration of the method on famous books, not a re-derivation of the six-shape result at scale — and we won't pretend otherwise.
The method, precisely
Lexicon. labMT (“language assessment by Mechanical Turk”): 10,222 English words, each rated 1–9 for happiness in 50 independent evaluations (Dodds et al., 2011). We use the authors' revised labMT-en-v2 list (10,187 words) from hedonometer.org. Reagan et al. used the earlier labMT — same family, slightly different vintage, which we note for honesty.
The lens. Following the hedonometer method, we drop every word whose score falls in the neutral band [4.0, 6.0] (Δh = 1). That removes function words and ambiguous ones, leaving 3,647 clearly-valenced words to score with. Without it, “the / of / and” drown the signal.
Window. Reagan et al. slid a fixed 10,000-word window (their books were 20k–100k words). To let short classics — Alice, the plays, The Metamorphosis — share the axis, we use a window of 10% of each book, capped at their 10,000 and floored at 2,500. For books over 100k words this is exactly the paper's window; for shorter works it shrinks so the arc spans ~90% of the book instead of a stub. Below ~10k words the signal is noisier — that's the price, and it's flagged here.
Normalising. Every arc is resampled to 100 points across 0–100% of the book, so a 30k-word novella and a 580k-word epic sit on one axis. Classification is the correlation of each arc with the six idealised shape templates; the best match (if |r| ≥ 0.20) is the book's family.
What this does NOT prove — the caveats, stated plainly
- Arc ≠ plot. We measure the happiness of the words, not the events. The Metamorphosis case above is the clearest proof.
- Dictionary sentiment is coarse. Word-level scoring ignores negation, irony, and context — “not happy” counts the happy. It performs worse than chance on single sentences; the large window is what rescues it, by averaging over thousands of words.
- “Dominated by,” not “only.” The six shapes are the most common emerging modes across an English-language Project Gutenberg corpus of mostly, but not all, fiction. That is not shown to be a law of all storytelling, though the authors do write of “the universality of these story types.”
- The “success” finding is confounded. The paper notes that Icarus, Oedipus and double “man in a hole” arcs are the most-downloaded — but download counts favour older, more canonical, classroom-assigned books, and the authors call downloads “only a rough proxy for success.” We don't repeat that ranking here as if it were causal.
- Our corpus is 29 hand-picked books, not a random sample — chosen to be recognisable and to span the shapes, not to be representative.
Sources — every claim's paper
- Reagan, Mitchell, Kiley, Danforth & Dodds, “The emotional arcs of stories are dominated by six basic shapes,” EPJ Data Science 5:31 (2016). doi:10.1140/epjds/s13688-016-0093-1 · arXiv:1606.07772
- Dodds, Harris, Kloumann, Bliss & Danforth, “Temporal Patterns of Happiness and Information in a Global Social Network: Hedonometrics and Twitter,” PLOS ONE 6(12):e26752 (2011) — the labMT lexicon. doi:10.1371/journal.pone.0026752 (CC-BY)
- Dodds et al., “Human language reveals a universal positivity bias,” PNAS 112(8):2389–2394 (2015). doi:10.1073/pnas.1411678112
- labMT-en-v2 scores: hedonometer.org. Texts: Project Gutenberg (public domain in the US).
- Vonnegut, Palm Sunday (1981) & A Man Without a Country (2005), “Here is a lesson in creative writing.” The famous “blackboard” lecture is where the drawn shapes appear.