ran across this very good overview of text generation techniques with neural networks https://blog.usejournal.com/generating-natural-language-text-with-neural-networks-e983bb48caad though it's notable mostly for this very odd/hilarious illustration of "meaning space"
i hope someone has written an extensive fanfic that intrafictionally explains the terrible kerning in destiny 2 https://mastodon.social/media/cs0GjOCNGCtmmwsoivI
okay this is going to mean absolutely nothing to anyone but it's exciting to me, so. I'm using a seq2seq model with CMUdict data to predict phoneme features from the way words are spelled. each phoneme in the word is associated with a list of features (see https://github.com/aparrish/phonetic-similarity-vectors/blob/master/featurephone.py) and the network is trying to "translate" one-hot encoded letters to those features. after just a few epochs it's sounding out "fediverse" (not in the training set) as something like "fevezunz," which isn't bad?
now this... this is an LSTM conditioned on *both* the GloVe 50d vector *and* the phonetic vector (as separate inputs to the model) for ~100k words, the idea being that you can keep the sound of the word constant while changing the meaning (or vice versa), or invent plausible soundalike neologisms. here I'm resampling arrays of phonetic and semantic representations of words in the same line into progressively smaller arrays and predicting from both for each word respectively
typography, mouths, selfie(?), eye contact (briefly) Show more
playing with clmtrackr and opentype.js, letting those counters sing
anyway here's a neural network attempting to spell words based on their phonetic vectors (using the vector to condition an LSTM). caveats: I stopped the training after uh, 20k samples because I have to go home now, so it's only trained on a few hundred words or so? also I should have stripped the alternate pronunciation notation from CMU dict and used start/end tokens. still, a promising start!
i have a new piece in queer.archive.work v1.1, being sold by editor Paul Soulellis at the fair https://queer.archive.work/ and i'm so happy with how it turned out (black riso on dark rich folded paper)
great moments in the history of animated text (from Feuillade's _Les Vampires_, 1915 https://archive.org/details/lesVampires1915Episode3-theRedCodebook)
poet, programmer, game designer, computational creativity researcher. assistant arts professor at NYU ITP. she/her.
Follow friends and discover new ones. Publish anything you want: links, pictures, text, video. This server is run by the main developers of the Mastodon project. Everyone is welcome as long as you follow our code of conduct!