Notices by Allison Parrish (aparrish@mastodon.social), page 2

@ctrlaltchaos I definitely don't think there was malicious intent! just something I'm sensitive about lately.

the second being that the "our" means "all of our children, because children belong to all of us, right?" which feels kinda belittling to me as someone who belongs to a particular gender minority that often faces significant medical and bureaucratic barriers along the path to being a parent (if I even wanted to be a parent to begin with). I sorta prefer "bring *your* children to work" since the "your" there seems to have an implied "if you got 'em/if you want to"

@kragen that's the idea!

decoding the same underlying vectors from the VAE using the french spelling model, for some reason, sure, whatever

using the phonetic VAE to interpolate between US state names in a grid

(the phonetic vae was only trained on english words, and this is still using the smaller test model for french spelling, which helps to explain why some of these don't feel very similar to each other. but as a proof of concept I'm pretty happy with this)

predicting french & english spelling from the same random samples of phonetic nonsense space

en: koehn; fr: caine
en: dasle; fr: desel
en: kepled; fr: ceppéde
en: priphilds; fr: prépédes
en: shimacrats; fr: chane-cemeste
en: guenill; fr: gunille
en: phown; fr: phoune
en: ehrill; fr: hirele
en: dosinon; fr: déchénane
en: pracke; fr: praque
en: cargian; fr: cargiennes
en: haise; fr: hases
en: modrowns; fr: mardonne
en: atriane; fr: atraiener
en: capair; fr: capare
en: flassey's; fr: flacheses

the first model essentially "sounds out" the word like a French speaker would, and the second model writes down those sounds like an English speaker would.

(this output is from the test french model that I trained last night—expect better results when I train the model for real. only disappointment: the english spelling model doesn't know what to do with nasalized vowels 😐 I feel like most English speakers would write an "n" but of course there aren't any nasalized vowels in the English data)

this isn't perfect yet, but it's promising: seq2seq models translating from french orthography to phonetic features, then from phonetic features to english orthography:

foutaise → futes
touchées → tuchey
rapetissait → rapatese
aiguisant → eguiza
péchais → pechei
dandinant → dadina
grimperait → gripare
huer → yea
proposât → proposa
persistance → percistace
bâche → bash
trompèrent → trowpare
local → locale
dépouillant → depua
prouvés → pruvey
bridée → breday
connaîtront → konetro

@KitRedgrave this is an extremely good and important point

pronunciation key here: http://www.lexique.org/outils/Manuel_Lexique.htm#_Toc108519023

it works! (the model I mean, not necessarily the swapping of the rhotics). after 5 epochs w/a fairly small model (64 enc/dec dims), with the model's pronunciation and the "real" pronunciation in parens

camps → k@p (k@)
futaie → fytE (fytE)
métaphysique → metatitis (metafizik)
embuaient → @b8E (@b8E)
compilations → k§piRiRO (k§pilasj§)
coupure → kupyR (kupyR)
nettoyer → nEtRERe (netwaje)
métaboliser → mepalydize (metabolize)
centralise → t@tRiliz (s@tRaliz)
penchée → p@Se (p@Se)

I'm making a neural network model of french spelling and pronunciation right now and I replaced /ʀ/ with /r/ (or actually just "alveolar approximant"), let's see if anyone notices

@varx I guess what I'm wondering then is whether or not the output of GPT-2 is significantly more "plausible" than, say, a markov chain for this purpose. (this is a real question and also an empirical one I think!)

can we just admit that "let's use uvular rhotics" was one of the worst memes in history https://en.wikipedia.org/wiki/French_phonology#/media/File:Uvular_rhotics_in_Europe.png

waiting to see how long it takes for someone to suggest emacs as a solution to this problem

about 50% of the time that I press "play" in iTunes nothing happens (especially when I've had it open but idle for a while). which seems like a pretty fatal flaw for, you know, media-playing software