>>10913699> train a NN to generate more articulate kinds of text structures like visual novels and 4chan threadsMaking a language model trained on thread data wouldn't be very difficult, then you put this language model as a encoder. Then you analyze each post in the thread you want to generate a new post in, I'd use some sort of attention based system depending on the amount of replies each post gets to determine what posts gets highest significance in the final context vector. When encoder has provided a thread-context vector, simply let decoder generate text somehow.
The problem is that there's not really any dataset. Maybe some sort of language model which simply gets a context vector fed into it? Then possibly some sort of cost function depending the amount of replies the post the AI gets.
The biggest problem is hardware. To get a well functioning text-generator you need at least 150M+ parameters, and then a batch size large enough to enable training. That shit's expensive.
