Homemade bot writes erotic stories
Erotic literature is stereotyped: the same exciting words, epithets and, of course, the same actions. These are the ideal conditions for creativity for the bot!
Journalist Daniela Hernandez (pictured) decided to test this theory .
A familiar programmer told her that at least 750,000 words were needed to train a neural network, so it took some time to search for texts. She downloaded works from several specialized electronic libraries, where erotica is published under the Creative Commons license, and also asked the authors to share their novels for the experiment.
For training "erotic bot" took a free program char-rnn- a multilevel recurrent neural network that learns to predict the next letters in a sequence. On samples of erotica like “Immortal Simpsons” from Daniela’s library, the neural network had to learn how to compose the corresponding words and sentences.
The girl had no programming experience, but he was not needed. On the Internet you can find understandable instructions on how to train char-rnn. By the way, the German developer Samim Winiger gave a lecture on TED on this topic. He had previously created a neural network that can recognize porn on video. That is, Winiger is an experienced specialist in, say, neural networks.
Then Daniela registered in the Terminal cloud service, uploaded the collected texts there and started the process.
A virtual machine was not too expensive: $ 0.006 per hour. After several days of training, Erotibot was ready to compose his own works.
For the bot to work, you need to enter the first phrases, and then he himself. For example, here are some samples in the text and audio option (speech synthesizer from Acapela Box ).
I must admit, the bot’s vocabulary is strange, sometimes it even uses non-existent words. But this is just an amateur craft on a small vocabulary base. It’s another matter if professionals create such programs ...
Journalist Daniela Hernandez (pictured) decided to test this theory .
A familiar programmer told her that at least 750,000 words were needed to train a neural network, so it took some time to search for texts. She downloaded works from several specialized electronic libraries, where erotica is published under the Creative Commons license, and also asked the authors to share their novels for the experiment.
For training "erotic bot" took a free program char-rnn- a multilevel recurrent neural network that learns to predict the next letters in a sequence. On samples of erotica like “Immortal Simpsons” from Daniela’s library, the neural network had to learn how to compose the corresponding words and sentences.
The girl had no programming experience, but he was not needed. On the Internet you can find understandable instructions on how to train char-rnn. By the way, the German developer Samim Winiger gave a lecture on TED on this topic. He had previously created a neural network that can recognize porn on video. That is, Winiger is an experienced specialist in, say, neural networks.
Then Daniela registered in the Terminal cloud service, uploaded the collected texts there and started the process.
char-rnn
in a virtual machine. A virtual machine was not too expensive: $ 0.006 per hour. After several days of training, Erotibot was ready to compose his own works.
For the bot to work, you need to enter the first phrases, and then he himself. For example, here are some samples in the text and audio option (speech synthesizer from Acapela Box ).
I must admit, the bot’s vocabulary is strange, sometimes it even uses non-existent words. But this is just an amateur craft on a small vocabulary base. It’s another matter if professionals create such programs ...