Rocket is Nowhere Near so Central as One Would Suppose

December 13, 2019 10:47 AM

Visualizing Gravity's Rainbow by Martin Paul Eve using TextPlot
posted by chavenet (16 comments total)

And it's a musical.
posted by Grangousier at 12:09 PM on December 13, 2019 [2 favorites]

I love this. I want to see the map for V., which always puts me in some kind of trance midway through where I develop weird memories of reading or experiencing things, but can never quite remember them clearly.
posted by spacewrench at 1:54 PM on December 13, 2019 [3 favorites]

okay so i don't tell many people this but really you can just read the byron the bulb part and skip the rest
posted by Reclusive Novelist Thomas Pynchon at 2:51 PM on December 13, 2019 [15 favorites]

Damn! I was sure it was Grigori the octopus.
posted by sjswitzer at 3:21 PM on December 13, 2019 [1 favorite]

Wanda, is that you?
posted by sjswitzer at 3:27 PM on December 13, 2019 [3 favorites]

> Wanda, is that you?

shhhhhh
posted by Reclusive Novelist Thomas Pynchon at 4:05 PM on December 13, 2019 [2 favorites]

I have a difficult time taking wordclouds seriously. That said, I would have thought "s-gerät" would figure prominently. I cannot find it, however.
posted by sensate at 4:29 PM on December 13, 2019 [3 favorites]

The dude's analysis is waaaaaaaaay more credit than I would give to a word cloud, even if you fancy it up with names like 'metric space embedding.' Especially when the dimensionality of the metric space is 2. This just looks like the kind of mess you get from projecting into woefully few dimensions...
posted by kaibutsu at 4:48 PM on December 13, 2019 [7 favorites]

Ask-ish question: Is there a word to describe the literary device of a book length novel that has various scenes taking place over wide spans of time & geography, with relatively thin connections between the scenes? I think Pynchon invented this concept with V, but I've seen it more often lately.
posted by ovvl at 5:46 PM on December 13, 2019 [1 favorite]

Kaibutsu: Can you say a bit more about the tools that would produce a higher-dimensional embedding, like you describe? And whether it’s possible to look at pairs of words together to examine more meaningful concepts (e.g. “black cat” vs. “black” and “cat”)?

This hews very close to a topic I’ve thought about quite a bit as an interested outsider, and I would love to learn more about the technical vocabulary in order to construct better search terms. :)
posted by Verg at 6:10 PM on December 13, 2019 [1 favorite]

To wit, my first AskMetafilter post was in search of a way to specify connections between pieces of text, albeit manually and not automatically as Text Map does.

This was in 2014, the same time that Text Map appears to have been in development. Since then, I’ve learned that the “embeddings” you see in the context of neural network and deep learning model visualizations are a good way to display highly-connected graphs. I’d love to more examples in the context of text analysis.
posted by Verg at 6:31 PM on December 13, 2019 [2 favorites]

High dimensional embeddings are easy! So long as you don't care about looking at them with your eyes!

As for methods, it depends a lot on what you want to do at the end of the day... For many algorithms, the dimensionality is just a number you choose, and (again, depending what you're trying to do) there's usually some sweet spot of 'enough for good performance.' And usually, two is waaay too few.

Word2Vec is probably something you'll find interesting (and probably already know about): it embeds a given word by deleting the word, and using the surrounding words. So, sentences like "The _____ cat purred on my lap" and "He had a pitch _____ sense of humor" would probably map to different places...

For a deeper look into general methods of 'putting related things close to one another' Fortunato's survey paper on community detection is a gold mine. Basically, for all of the Things, invent a similarity score between pairs of Things. Then take your pick of algorithms from the paper. (I'm partial to spectral methods, since they give some pretty intuitive indications of the 'right' number of dimensions for the data.) This sort of approach is less good for considering context, though, which is probably why things like Word2Vec took off so well.
posted by kaibutsu at 6:40 PM on December 13, 2019 [3 favorites]

okay so i don't tell many people this but really you can just read the byron the bulb part and skip the rest

Candy scene or GTFO.
posted by Nerd of the North at 10:04 PM on December 13, 2019 [3 favorites]

that scene is kinda gross though, if you think about it.

also there are other scenes that are definitely gross no matter how you think about it.

basically i put a lot of gross stuff in that book you don't have to read it all it's okay.
posted by Reclusive Novelist Thomas Pynchon at 12:25 AM on December 14, 2019 [1 favorite]

The classic technique for getting a "good-looking" 2D representation of points in a higher-dimensional space is t-SNE. UMAP is also recently fashionable. People often run these on word vectors and the results are good although tuning them is an art.
posted by vogon_poet at 3:44 PM on December 15, 2019 [2 favorites]

Thank you both for your answers!
posted by Verg at 9:43 PM on December 15, 2019