Artificial Intelligence Identifies the Six Main Arcs in Storytelling: Welcome to the Brave New World of Literary Criticism

Is the sin­gu­lar­i­ty upon us? AI seems poised to replace every­one, even artists whose work can seem like an invi­o­lably human indus­try. Or maybe not. Nick Cave’s poignant answer to a fan ques­tion might per­suade you a machine will nev­er write a great song, though it might mas­ter all the moves to write a good one. An AI-writ­ten nov­el did almost win a Japan­ese lit­er­ary award. A suit­ably impres­sive feat, even if much of the author­ship should be attrib­uted to the program’s human design­ers.

But what about lit­er­ary crit­i­cism? Is this an art that a machine can do con­vinc­ing­ly? The answer may depend on whether you con­sid­er it an art at all. For those who do, no arti­fi­cial intel­li­gence will ever prop­er­ly devel­op the the­o­ry of mind need­ed for sub­tle, even mov­ing, inter­pre­ta­tions. On the oth­er hand, one group of researchers has suc­ceed­ed in using “sophis­ti­cat­ed com­put­ing pow­er, nat­ur­al lan­guage pro­cess­ing, and reams of dig­i­tized text,” writes Atlantic edi­tor Adri­enne LaFrance, “to map the nar­ra­tive pat­terns in a huge cor­pus of lit­er­a­ture.” The name of their lit­er­ary crit­i­cism machine? The Hedo­nome­ter.

We can treat this as an exer­cise in com­pil­ing data, but it’s arguable that the results are on par with work from the com­par­a­tive mythol­o­gy school of James Fra­zier and Joseph Camp­bell. A more imme­di­ate com­par­i­son might be to the very deft, if not par­tic­u­lar­ly sub­tle, Kurt Von­negut, who—before he wrote nov­els like Slaugh­ter­house Five and Cat’s Cra­dle—sub­mit­ted a master’s the­sis in anthro­pol­o­gy to the Uni­ver­si­ty of Chica­go. His project did the same thing as the machine, 35 years ear­li­er, though he may not have had the where­with­al to read “1,737 Eng­lish-lan­guage works of fic­tion between 10,000 and 200,000 words long” while strug­gling to fin­ish his grad­u­ate pro­gram. (His the­sis, by the way, was reject­ed.)

Those num­bers describe the dataset from Project Guten­berg fed into the The Hedo­nome­ter by the com­put­er sci­en­tists at the Uni­ver­si­ty of Ver­mont and the Uni­ver­si­ty of Ade­laide. After the com­put­er fin­ished “read­ing,” it then plot­ted “the emo­tion­al tra­jec­to­ry” of all of the sto­ries using a “sen­ti­ment analy­sis to gen­er­ate an emo­tion­al arc for each work.” What it found were six broad cat­e­gories of sto­ry, list­ed below:

1. Rags to Rich­es (rise)
2. Rich­es to Rags (fall)
3. Man in a Hole (fall then rise)
4. Icarus (rise then fall)
5. Cin­derel­la (rise then fall then rise)
6. Oedi­pus (fall then rise then fall)

How does this endeav­or com­pare with Vonnegut’s project? (See him present the the­o­ry below.) The nov­el­ist used more or less the same method­ol­o­gy, in human form, to come up with eight uni­ver­sal sto­ry arcs or “shapes of sto­ries.” Von­negut him­self left out the Rags to Rich­es cat­e­go­ry; he called it an anom­aly, though he did have a head­ing for the same ris­ing-only sto­ry arc—the Cre­ation Story—which he deemed an uncom­mon shape for West­ern fic­tion. He did include the Cin­derel­la arc, and was pleased by his dis­cov­ery that its shape mir­rored the New Tes­ta­ment arc, which he also includ­ed in his schema, an act the AI sure­ly would have judged redun­dant.

Con­tra Von­negut, the AI found that one-fifth of all the works it ana­lyzed were Rags-to-Rich­es sto­ries. It deter­mined that this arc was far less pop­u­lar with read­ers than “Oedi­pus,” “Man in a Hole,” and “Cin­derel­la.” Its analy­sis does get much more gran­u­lar, and to allay our sus­pi­cions, the researchers promise they did not con­trol the out­come of the exper­i­ment. “We’re not impos­ing a set of shapes,” says lead author Andy Rea­gan, Ph.D. can­di­date in math­e­mat­ics at the Uni­ver­si­ty of Ver­mont. “Rather: the math and machine learn­ing have iden­ti­fied them.”

But the authors do pro­vide a lot of their own inter­pre­ta­tion of the data, from choos­ing rep­re­sen­ta­tive texts—like Har­ry Pot­ter and the Death­ly Hal­lows—to illus­trate “nest­ed and com­pli­cat­ed” plot arcs, to pro­vid­ing the guid­ing assump­tions of the exer­cise. One of those assump­tions, unsur­pris­ing­ly giv­en the authors’ fields of inter­est, is that math and lan­guage are inter­change­able. “Sto­ries are encod­ed in art, lan­guage, and even in the math­e­mat­ics of physics,” they write in the intro­duc­tion to their paper, pub­lished on Arxiv.org.

“We use equa­tions,” they go on, “to rep­re­sent both sim­ple and com­pli­cat­ed func­tions that describe our obser­va­tions of the real world.” If we accept the premise that sen­tences and inte­gers and lines of code are telling the same sto­ries, then maybe there isn’t as much dif­fer­ence between humans and machines as we would like to think.

via The Atlantic

Relat­ed Con­tent:

Nick Cave Answers the Hot­ly Debat­ed Ques­tion: Will Arti­fi­cial Intel­li­gence Ever Be Able to Write a Great Song?

Kurt Von­negut Dia­grams the Shape of All Sto­ries in a Master’s The­sis Reject­ed by U. Chica­go

Kurt Von­negut Maps Out the Uni­ver­sal Shapes of Our Favorite Sto­ries

Josh Jones is a writer and musi­cian based in Durham, NC. Fol­low him at @jdmagness