Statistics #4: The Robot Revolt · 8:46pm Jan 10th, 2018
Well, it’s not quite statistics, but close enough.
You see, recently, I’ve been trying to train a neural network to write clop.
No, don’t run away, this post is actually quite safe for work, at least, it is the intention that it should be so. Mind you, I said nothing about it being safe for your sanity, and it is definitely not safe for the English language, but who cares about those two.
If you know me any, I expect you can easily imagine why I would find such a project interesting, but in case you don’t, here’s why: Some of the fiction observed around here is of such poor quality that I’m sure a computer could do better. Once that is conclusively demonstrated, maybe we could all go read something more interesting.
I’m a mad scientist. Gotta keep the rep up. Notice this does not require my projects to actually have the intended results, just that they work.
The neural network is a step up from my prior attempts to teach computers to write clop using Markov chains. For reference, here are some choice bits of what the Markov chain generator came up with, selected for comedic potential:
Luna listened to her lips. This will be intentional.
You grabbed Luna by the bed to lie like this for a moment of pleasure.
Water gushed out of reach.
Octavia sputtered, weeping.
Spitfire returned the nod. She gave Carrot his sandwich with a plate that was covered in various spots, and wiped away her worries. “At least I can do it…” She could have sworn that was the Wonderbolt on a mare, who could take care of him?
I sighed loudly and smiled wide as Shining moaned at this action, and the two deer breathed deeply, recovering from their carnal activities.
I scooted forward again, and took another bite of my dinner. Instead, he decided to say, “I know, dear,” giving a peck on her cheek and caress it, while the other wanted nothing more than to please Pinkie Pie, as there would be plenty of time for herself later.
Spitfire slowly circled Rainbow, examining her from all dangers, including her fidelity!
Two relatively thin plates of dark chitin made up the clubhouse’s walls, chilling her to the interior of the storm.
Milano ran to her neighbor’s house and frantically banged on the door below tore her out of the cooling, bubbly water and rest it on the floor, locked in an embrace in a puddle of water that began to swell in the colt’s ass. And then another, and another, and another, and another, and another, each spark of ecstasy multiplying exponentially.
A smile grew more and pulled out of the underbust corset.
Plenty more of this before we have all the ponies back then, but circumstances wouldn’t allow it.
Trixie came down to push herself back down, her tongue out, she suddenly pulled her against the sloppy look as she desperatly tried to put her on the counter checked the schedules.
It was enough to elicit a groan and rub your hooves feverishly rubbing a hoof on either side of the dark.
She shifted her back, or her words.
They all bathed in golden light slowly leaked into her panties again, letting down someone they clearly respected, and appeared to be done in a whimper as the warmth of her until only the common area, which let everypony know that was exactly like her sofa.
We kissed and licked Twilight’s hind leg over her floor.
You get the idea. Now, back then, I was dissatisfied with the results, and shelved the project. Tuning it to use longer chains gave more readable text, but also produced sentences that were too mismatched against each other, and it was obvious I need a more specific corpus selection to get anything workable. Also, Markov chains are hardly top of the line in text generation these days.
This time, I got a few hours to install TensorFlow, CUDA libraries, and give the idea a proper recurrent neural network treatment. An entirely different approach, a recurrent neural network doesn’t think in terms of words, and essentially has to figure English out from scratch. You know, like people do.
Unfortunately, I neglected to select a better corpus this time as well, so I’m still not happy with the results.
It’s somewhat a surprise that the neural network is extremely reticent to produce anything actually lewd from the very same corpus that was used to train the Markov chain generator. Markov chains would result in at least half of the sentences not fit for posting, but the RNN only turns up one in about twenty.
Now…
I’m not feeling well, and somehow this gives me the idea that nobody else deserves to either. So I’m going to quote examples of what the neural network came up with at you as well.
Now you can run away screaming. Mwa-ha-ha-ha-ha-ha-ha!
Ready?
Fluttershy was met, even fired verbal.
Able to lay silent a convulse that was up and fell out off underwear, but I touched her kiss. “Wait, we can know, right, Colgate’s and, trust it. You love yourself and I worked a little like she’d even gained, Twilight? I…” she called her tongue hard across the floor begpt today’s lungs. I smirked and picked up a gave over both list of Twilight as they rubbed her head out.
Withly caressed his wasted. She lay strong in the lips. Heavy, detrusping loving expression would be – lightfor was over all the kiss knew a buildain did not be giving. Twilight pushed inhers to her mouth for such a human bust at a moment liver uncomfortable between her lips was now. All on her, one another was being dippetually in my fault.
Wearing mines. Wobbing with the low grunt. However, Twist.
Snort when she adjusted its bared hoof.
Sparkle lifting up her body, she lolled into the chinky storpered girl of such nonetaining. Her mind wished that.
Without fell up stiffen. The pent and galloponies touching herself with licking in Ponyville.
Love smacker is seriously-founded quiver of the sentence too. The months inches and up into the unicorns.
Berry kneeled in the couch of pleasures and managed to certain the laid next wind flicking their labble fall. She’s like a fresh wave of coming to something like three wonderful girl ever heedsed forward, she was her initial color and jumped twisting the blunt.
Diamond Tiara. She expertly hovered almost up, I felt a teasing.
After all, and sets her lips against a love counterpart, the inhaling, and wiped her moans and hitions. I runs between her teeth to my lips and lead herself from wide walk in, aight and carelessly, leaving her hoof leaned down as she lifted towards the delied Lily was be next to when lonnesily hot both pony to let my tongue so good.
Sweetie closed his eyes and settled her head together in her lunchdincence; she missed close to her bridle in the cork.
For the area, I bet little events, you were lucky to notice he’s slid down on the floor that later.
Twilight felt a cry of boots slowly met doubled what the other was goodnight for not up and alone. Make her mouth widened, of course one of their flank and she had and used to be hunting her mind.
As Fluttershy’s reactionally drops over her thighs. Her face spurted up that the pleaphony moan on her. Macintosh pulled out of her hooves, and push as her eyes shifted the doorbeats sobbed a jew what determined. Sitting together. “But no processing, ya could work and switch!” she squeaked. Her eyes clenched to her face. “Every ya not happy is a mare’s up now, all thosely.”
Her eyes were broken up again. Derpined by her lips. When she came forward.
“Because I hear word comes as it would believe I’ve mit, let it, slowly,” she found that much incpit, moaned in surprise, so unfocused observant was. “Oh long, huh, I dunno!” Rubbing her tongue desperately twisted to met the dialously, at the shades and approval, “Hey was doing that she’s mented, so you’re honeyness, I think this was you being a second, just fixing deep into my address for me?”
Fluttershy never slowly talked to her body. “Don’t take it at your lips that–” I licked her resort.
No, I am entirely heartless and feel no shame whatsoever. For science.
For Science
Darn it, I'm at work and I'm just opposite my boss's office. I can't afford to laugh this much!
Seriously, hilarious stuff. RoboRosewater writes erotica.
4770874
The next paragraph, which doesn’t even mention Fluttershy.
It appears that to get good results, the training corpus needs to be selected by pairing, among other things. And it needs to be cleaned up manually, so that the RNN isn’t confused by occasional writers who don’t know what paragraphs are.
4770887
A character-based RNN is interesting precisely because it doesn’t have any preconceptions about the stream it attempts to imitate, so as long as the corpus can be reduced to a stream such as you describe, representing vectors as unique UTF-8 characters, it would be trivial, I wouldn’t even need to change the code. UTF-8 can handle enough symbols for any reasonable vocabulary for such a narrow purpose.
Unfortunately, converting the corpus into such vectors would be the really difficult part. In part, because people can’t spell.
4770917
Actually, it mostly makes up nonsense words because some of the stories in the corpus I used are mis-formatted, and have more linebreaks than they have paragraphs, including linebreaks in the middle of words. It imitates everything, including syntax – RNNs are actually very good at figuring out syntax, check the links in the post, which, among other things, demonstrate how RNNs imitate C code – but since the corpus is not syntactically uniform, it occasionally throws a loop.
Nothing a better corpus and some pre-normalization won’t eventually fix.
That said, feeding it through a spellchecker would probably help, if only because eye halve a spelling chequer…
Yay, other people doing mad science on ponyfiction! I've played around with Markov generators too (well, letter-frequency-based generators at least; I don't know if it completely fits the Markov definition). I haven't progressed to real machine learning yet. :)
I heard of someone who tried something similar. But their specific corpus wasn’t just clop, but one particular author’s body of fetish clop. And when they shared, someone else replied that the resulting word salad still turned him on, because it had just enough of the “right” words scattered through the nonsense to elicit some kind of Pavlovian response.
I bet you could post a botnik-style curated selection of these as a pretty awesome Finnegan's Wake style story.
4770995
That’s kind of the idea.
It sounds like you're trying to build a Great Automatic Grammarizer, good luck!
Also, the preferred term for your disorder is Malign Hypercognition
4771056
Only if I get caught.
Ah, good old AI-generated text. I've got an AI-generated title that would go well with your offerings: *Before the Rainbow, comes the perversion*.
I wonder what they use for the postmodernism generator.
(Then again, pomo is word salad anyway, so who would notice.)
4816029
Since it references the Sokal Affair, I expect it’s actually a Markov Chain generator, in late 90s they were all the rage.
You need Intel Inside <Dum, dumdumdumDUM!>
4816064
…Nah, not sexy enough.
4816079
Trust me, we are working on VERY sexy stuff.
4816153
…Like another side channel attack?
4816235
I can't really discuss a lot of it, outside of what has been published.