one more quick experiment: this is Lewis Carroll's "Jabberwocky" with each word "spelled" out using a visualization of the dimensions of the hidden state of a phoneme-to-grapheme model
(in particular, Pincelate's model, passed through PCA in order to reduce from 256 to 32 dimensions—I resized the length of the squiggle based on the number of phonemes in the word being spelled)
I think what I'm trying to do here are visualizations that appear "asemic" but that actually encode information about the text—the idea with this one is that you can actually see visual similarities between marks for words that sound similar—you can sorta see this in the end rhymes and in e.g. the "o frabjous day calloo callay" line (sixth from bottom)
@aparrish Sorry to be a bother: when I click on the "GloVe vector" link in the help dialog of the visualization, it just closes the dialog instead of opening the link.
Firefox 72.0.1 on Linux (Ubuntu 18.04)
Love this experiment. 👍 👏
@aparrish Are the phoneme embeddings sequential within each word-squiggle? I keep trying to find the rhyme pattern near the ends of the words, or something that looks like "wabe" near the end of "outgrabe" Regardless, I am super into this concept of hidden meanings, and I love how this totally looks like someone's messy handwriting.
@danima no, the spatial layout of the squiggles don't have anything to do with the order of the phonemes—the circular forms I used in the previous experiment are partially designed to avoid suggesting exactly that mapping. but I'm glad you like the piece!
@aparrish I really like this! I have the poem memorised, so I was able to 'read' through your visualisation and pick out some patterns/commonalities.
Hometown is adapted from Mastodon, a decentralized social network with no ads, no corporate surveillance, and ethical design.