the idea being that words next to each other on the grid—up and down, left and right—sound similar to each other, and the model can generate a new word in between any two others.
there's still a ton of tweaking to be done on this but the main Poetic Vision of the project was "nonsense words in strange harmony stretching out as far as the eye can see" and I feel like I'm getting pretty close to that
for my oulipo friends, another thing I can do with this model: by zeroing out the possibility of certain letters during inference, I can automatically spell words phonetically without using particular letters. random examples:
emptier → imptiar
relies → rilis
checkers → chackirs
interpreter → inturpruttor
commenter → commnttor
ireene → irin
preference → proffrints
reiteration → ritiration
nineteen → ninitin
coherent → cohirant
breadbasket → brodbastach
repeal → rippill
groceries → grossiri's
trying to re-spell words without G (in these examples you can see how having some information about which phonemes are similar helps the model out)
gambling → banbblink
laggards → lakard's
pledging → pledjink
raging → rajink
sightseeing → citicine
ginseng → jinsen
gauges → bajiz
aggressors → acreasers
goings-ons → joinsins
signalling → cicdonalin
glowingly → klowinli
haggis → hakes
gigi → judjy
gutting → kuttin
gold-green → koldkrine
disengaged → disindajed
engineering → enjunerin
also: replacing the "nasal" feature with "voiced" & "stop," thus fulfilling my childhood dream of computationally modeling when Bilbo has a cold in The Hobbit and says "Thag you very buch":
macintosh's → backedtashes
nylons → dilods
misconduct → biscodduct
unshaken → adshaked
masonry → bacedry
emphasizing → ebfosizing
meeting → beatig
contradiction → codtrodictide
mimis → beeby's
masculine → basculad
insensitive → idsedsitive
drinkwine → drigquide
negatron → degatrad
modernized → bodderdized
turning up the temperature on my own name
0.10 allossen perish
0.23 allossen perish
0.35 alleson perish
0.47 alleson parish
0.60 allison perrish
0.72 alleson parrish
0.85 alasson parish
0.97 allodson perish
1.10 allison perish
1.23 allison paresh
1.35 allionsop periish
1.48 a pajrish
1.60 allyos'n parischf
1.73 ahlleoson parish
1.85 xaliason cparrzich
1.98 azlacenl pecresh-decxs
2.10 ahtellson patriosu
2.23 'lasaym yiarishc
2.35 aalizontsekq uiosirmyvr
2.48 pbiausilonfot pasurhstnu
@aparrish this is so good!
@aparrish I never thought https://www.goodreads.com/book/show/921445.The_Holes_in_Your_Nose could scale, this is fun
@aparrish Yup I definitely said all of these out loud 🤣
@aparrish Okean'sx is a prequel to Ocean's 11. 😊
@aparrish i didn't really pay attention to this thread earlier but it turns out to be a real winner. top masto-content, would chuckle again.
@brennen i know i've been on kind of a bender on this for a while with little else to say, so i totally understand people tuning out! but yeah i'm happy with the results i'm getting
@aparrish
throw in me crying uncontrollably from beginning to end and this toot is the movie Arrival (2016)
@skirts don't you mean crying uncontrollably from end to beginning? 🙃 (seriously though i cried through that movie too)
@aparrish
haha yeah, crying all at once for forever!
@aparrish
The latters look definitely like Indian divinities names
@aparrish I don't grasp this at all, but it's astounding! Sort of Douglas Adamsy in flavor I'd say. I'm into it.
turning up the softmax temperature on character sampling turns model into a pretty plausible pulp sci-fi name generator
silhouettes → plypousequeess
scribbling → skridblitic
minilab → mynillabbueh
myriam → nmirio
chains → okean'sx
binghamton → binmoicinme
circumstances → surkumagstmens'
palmist → paammist
accelerator → acquejalercater
precocious → xpricoclus
minimize → minymize
illiterate → aietaludret
fevers → fivxd-hask
jump → vgumppe
bounded → fraxptnud
shortest → shourjtls''t