More on automated sentiment-analysis, and Moby-Dick.
Paul Valéry tells the story: The painter Edgar Degas was backhanded-bragging to his friend Stéphane Mallarmé about the poems that he, Degas, had been trying to write. He knew they weren’t great, he said, “But I’ve got lots of ideas—too many ideas.” “But my dear Degas,” the poet replied, “poems are not made out of ideas. They’re made of words.”
Paintings, for that matter, are not made of pretty ballerinas or landscapes: they’re made of paint.
Which brings us to Syuzhet, Matthew Jockers’s new program that analyzes the words of a novel for their emotional value and graphs the sentimental shape of the book. Dan Piepenbring has explained it all here and here on the Daily, with links to the original postings and the various outcries, some of them in the comments, that have blown up around Jockers.
Many people apparently find Jockers’s research the latest assault of technocratic digitocracy on the citadel of deep humanistic feelings, but that’s not how I see it. What the graphs reveal about potboiler narrative structure versus high-literary arcs, for instance—Dan Brown’s higher average positivity than James Joyce’s, and his more regular cycle of highs and lows to force the reader through the book—is insightful, useful, and great.
In some ways, it’s hard for me to even see what the fuss is about. “It’s not that it’s wrong,” one commenter writes. “It’s just that it’s an extremely poor substitute for reading, enjoying, and discussing literature.” But who said anything about a substitute? Does this commenter not notice that the discussions of the graphs rest on having read the books and seeing how the graphs shed light on them? Another: “Okay, fuck this guy for comparing Dan Brown to James Joyce.” Well, how else can you say Joyce is better and Brown is worse? That’s what’s known as a comparison. Or do you think Joyce can’t take it?
Freak-outs aside, there are substantive rebuttals, too. What seems to be the most rigorous objection is from SUNY professor and fellow digital-humanities scholar Annie Swafford, who points out some failures in the algorithm. “I am extremely happy today” and “There is no happiness left in me,” for example, read as equally positive. And:
Longer sentences may be given greater positivity or negativity than their contents warrant, merely because they have greater number of positive or negative words. For instance, “I am extremely happy!” would have a lower positivity ranking than “Well, I’m not really happy; today, I spilled my delicious, glorious coffee on my favorite shirt and it will never be clean again.”
But let’s actually compare “Well, I’m not really happy; today, I spilled my delicious, glorious coffee on my favorite shirt and it will never be clean again” to “I’m sad.” The positivity or negativity might be the same, assuming there could be some kind of galvanometer or something attached to the emotional nodes of our brain to measure the “pure” “objective” “quantity” of positivity. But the first of those sentences is more emotional—maybe not more positive, but more expressive, more histrionic. Ranking it higher than “I’m sad” or even “I am very happy” makes a certain kind of sense.
“There is no happiness left in me” and “I am all sadness from now on” are the same seven words to a logician or a hypothetical emotiomometer, but not to a novelist or a reader. Everyone in advertising and political wordsmithing knows that people absorb the content of a statement much more than the valence: to say that something “is not horrific and apocalyptic” is a downer, despite the “not.” Or consider: “Gone for eternity is the delight that once filled my heart to overflowing—the sparkle of sun on the fresh morning dew of new experience, soft envelopments of a lover’s thighs, empyrean intellectual bliss, everything that used to give my life its alpenglow of hope and wonder—never again!” and “I’m depressed.” An algorithm that rates the first piece of writing off-the-charts positive is a more useful quantification of the words than one that would rate the emotional value of the two as the same.
Some years back, Orion Books produced a book called Moby-Dick in Half the Time, in a line of Compact Editions “sympathetically edited” to “retain all the elements of the originals: the plot, the characters, the social, historical and local backgrounds and the author’s language and style.” I have nothing against abridgments—I’ve abridged books myself—but I felt that what makes Melville Melville, in particular, is digression, texture, and weirdness. If you only have time to read half the book, which half the time is more worth spending? What elements of the original do we want to abridge for?
Moby-Dick in Half the Time seemed like it would lose something more essential than would Anna Karenina in Half the Time or Vanity Fair in Half the Time or Orion’s other offerings. I decided to find out. So I compiled every chapter, word, and punctuation mark that Orion’s abridger cut from Melville’s original Moby-Dick; or The Whale, and published the result, with its inevitable title, as a book of its own: a lost work by Herman Melville called ; or The Whale.
Half the Time keeps the plot arc of Ahab’s quest, of course, but ; or The Whale arguably turns out closer to the emotional ups and downs of Melville’s novel—and that tells us something about how Melville writes. His linguistic excess erupts at moments of emotional intensity; those moments of intensity, trimmed as excess from Half the Time, are what make up the other semibook. Chapter sixty-two, for example, consists of a single word, “hapless”—the only word Orion’s abridger cut from the chapter, trimming a 105-word sentence to 104, for some reason. That’s a pretty good sentiment analysis of Melville’s chapter as a whole. Reading ; or The Whale is a bit like watching a DVD skip ahead on fast forward, and it gets at something real about Melville’s masterpiece. About the emotion in the words.
So I would defend the automated approach to novelistic sentiment on different grounds than Piepenbring’s. I take plot as seriously as he does, as opposed to valorizing only the style or ineffable poetry of a novel; I also see Béla Tarr movies or early Nicholson Baker novels as having plots, too, just not eventful ones. Jockers’s program is called Syuzhet because of the Russian Formalist distinction between fabula, what happens in chronological order in a story, and syuzhet, the order of things in the telling (diverging from the fabula in flashbacks, for instance, or when information is withheld from the reader). It’s not easy to say how “plot” arises out of the interplay between the two. But having minimal fabula is not the same as having little or no plot.
In any case, fabula is not what Syuzhet is about. Piepenbring summarizes: “algorithms assign every word in a novel a positive or negative emotional value, and in compiling these values [Jockers is] able to graph the shifts in a story’s narrative. A lot of negative words mean something bad is happening, a lot of positive words mean something good is happening.” This may or may not be true, but novels are not made of things that happen, they are made of words. Again: “When we track ‘positive sentiment,’ we do mean, I think, that things are good for the protagonist or the narrator.” Not necessarily, but we do mean—tautologically—that things are good for the reader in the warm afternoon sunshine of the book’s positive language.
Great writers, along with everything else they are doing, stage a readerly experience and lead their readers through it from first word on first page to last. Mapping out what those paths might look like is as worthy a critical approach as any.