The emotional arcs of stories are dominated by six basic shapes | Hacker News Reader

The emotional arcs of stories are dominated by six basic shapes

arxiv.org

108 pointsehudla32 comments

Am I missing something? The result for the 6 main emotional arcs basically enumerates all the possible permutations of rise and fall of length 1, 2, and 3 which seems pretty obvious.

Length 1: rise, fall

Length 2: rise-fall, fall-rise

Length 3: rise-fall-rise, fall-rise-fall (must be interchanging because for example fall-fall-rise would probably just be considered fall-rise)

I agree... the figures were all just permutations of rise and fall. When all you have is one measure (average word happiness) I don't really see what else you can draw. I guess the main contribution is that many stories have relatively simple rise-fall curves, and don't often change their emotional state more than 3 times.

I thought the same, and as per fovc's comment, below, I think what their statistical analysis did was basically a coarse fourier transform.

In the Harry Potter example you can see that a higher frequency is very significant. But the exact frequency is probably rather arbitrary from book to book, whether there are ten peaks or five, say. So I'd imagine, over lots of texts, any particular higher frequencies is less significant, leaving the lowest modes to dominate, as you point out.

So overall, a rather uninspiring result, I felt.

Though if we're both missing something, it would be good to know!

A note some HNers might find tangentially interesting: in David Foster Wallace's Infinite Jest, Fourier transforms are a bedrock concept for the titular entertainment, which is considered such a perfect piece of media that it is lethally addictive. How nice it was to see this post about "shapes of narrative" be compared to Fourier transforms for a IJ nerd!

Very uninspiring indeed!

Let me know if you are interested in investigating this properly.

It actually sounds a lot like the old Vonnegut talk. https://www.youtube.com/watch?v=oP3c1h8v2ZQ

non video description https://www.washingtonpost.com/news/wonk/wp/2015/02/09/kurt-...

the talk is funny though, i saw him give a much longer version.

My first thought was that Vonnegut realized this decades ago

Actually, if all possible arcs are represented it is an even more interesting result, I think. Think about it, it goes against all the enumerations of story lines that have been offered.

I wonder how books with a lot of cuts between characters and stories look like.

Read this quickly, but my immediate thought when I saw the PCA result vectors was that they resembled the Fourier basis. Given a relatively slow-moving set of signals (which I think these are, given the 10,000 word sliding window), wouldn't you expect the PCA to give you exactly this? In other words, what is the null hypothesis here and is it rejected?

I think the idea is, of you compare short fiction to the Harry Potter graph, most stories have simple emotional arcs, rather than complicated back and forth. Though, if you limit the word length allowed... I wonder how they deal with "it was the best of times. it was the worst of times"...

You could mess up pretty easily at a local level with missed negations. For example, John Lennon's song "Imagine" invites people to envision a world without various things; if you assign negative sentiment to those things, you might infer that the song overall had very negative affect. Or religious prophecies about a future era in which people wouldn't have to suffer from various afflictions.

Jockers, in one of his posts, argues that it's less of a problem than you might think because it's a local problem that just adds noise to the data. Plus, his different sentiment analysis techniques seem to more-or-less agree.

On the other hand, he tells of showing a graph to his son, of a book his son had read, and being told the graph was completely wrong for part of the book---it turns out that part of the book was written from the antagonist's viewpoint and the "emotional valence" was apparently inverted.

See one of the follow-ups to http://www.matthewjockers.net/2015/02/02/syuzhet/.

The longer discussion is http://www.matthewjockers.net/2015/03/04/some-thoughts-on-an... (which gives examples of negation that sentiment analysis packages miss, but then some empirical evidence that analysis of sentiment in several literary works doesn't seem to be affected very much by this issue).

The antagonist's viewpoint issue appeared in footnote 1 of http://www.matthewjockers.net/2015/02/25/the-rest-of-the-sto... and it does seem like the sentiment analysis would clearly be backwards overall if this kind of thing continued for most of a book. (An example might be if an author depicted people enjoying themselves while committing horrible acts, and used more individual words related to the enjoyment than to the acts.)

Thanks for the references.

> people enjoying themselves while committing horrible acts

Now I'm kind of curious about sentiment analysis of something like Fight Club.

I thought the very same. I guess the most interesting result they obtain is the slight asymmetry in some of the low-frequency modes, which the Fourier basis does not exhibit.

Me too, interested in debunking this properly?

As expected, given the title, this[0] lecture by Vonnegut is mentioned in the article.

[0]: https://www.youtube.com/watch?v=oP3c1h8v2ZQ

Do you have any more of these? Since reading Dan Harmons story structure I have enjoyed American television a lot more.

http://channel101.wikia.com/wiki/Story_Structure_101:_Super_...

That linked series specifically and Dan Harmon's approach there is my favorite approach to these topics.

I wanted to also share https://en.wikipedia.org/wiki/Monomyth Joseph Campbell's very similar work that I think Dan gives credit to somewhere in that 101 series.

Any interest in: http://www.matthewjockers.net/2015/02/02/syuzhet/

This seemed interesting because I am working on visual story structure diagramming software (looking for beta users btw), but I wasn't very impressed with their results. Basically the only interesting conclusion I could see was "stories need emotional movement", e.g. there were no examples where things were happy or sad and continued at the same level throughout the events of the story.

I couldn't tell if this could even identify the stories shape from their sentiment graph -- e.g. if you fed it "Cinderella" could it identify the "Rags to Riches" plot that they identify?

It would be neat to see if you could extract and identify plot points -- deaths, fights, breakups, betrayals, (maybe with some of those newfangled neural network thinggies) and look for patterns there.

Still, it's encouraging to see people working on this. Maybe the results seem basic because this field is still unstudied. I expect software tools for storytelling to change a lot in the coming years with all the exciting new work in natural language processing and machine learning.

In the article they review several story plot taxonomies. Interesting that they do not include the one according to Borges that he describes in "Four cycles": (1) city sieged and defended (2) story of return (3) search (of treasure) (4) sacrifice of a god. Probably not the most complete taxonomy, but likely most dramatic one.

(5) story plots that belong to the Emperor

(6) those that tremble as if mad?

Last year, Matthew Jockers, an English professor at the University of Nebraska-Lincoln, had a series of blog posts on the subject[1]---I haven't heard of any actual publications from the work---that was pretty interesting.

I haven't read this paper yet (this post is falling off HN too fast!) but his technique was to use sentence-level sentiment analysis to get a time-varying signal, rub a Fourier transform against it, cut off all but the lowest frequencies, and returned it to the time domain to draw pretty pictures. He, too, came up with six basic arcs, I think, probably for the same reason that klue07 mentions but using statistics against a bunch of books.

There was a certain amount of press coverage at the time. The R package, syuzhet, is available on github[2]. Also, you can look at my notes on playing wiht syuzhet and R[3].

[1] Starting with http://www.matthewjockers.net/2015/02/02/syuzhet/

[2] https://github.com/mjockers/syuzhet

[3] http://maniagnosis.crsr.net/2015/08/exploring-syuzhet.html http://maniagnosis.crsr.net/2015/08/syuzhet-prodding-frequen...

I wonder if there is something like this applicable to pacing in video games? I wonder if video games could be adaptively paced based on user input?

Video games are already adaptively paced based on user input.

Valve's "Left 4 Dead" had a system like this called the "AI director" that was pretty well-received.

See the part about "Generate Dramatic Game Pacing" http://www.valvesoftware.com/publications/2009/ai_systems_of...

Interesting. Also reminds me of the Thomas Lennon & Robert Ben Garant screenwriting guide (excellent) which posits there is one - only one - blockbuster movie format, and it's most easily seen in Die Hard.

Introduce likeable character - get them stuck in trouble - get them unstuck.

This is awfully fun, I never saw Rom-coms the same after I noticed the trend of get together, break up, then get back together again (rise, fall, raise) in this paper.

Good stuff! Great to see people working on this.

Reply on news.ycombinator.com