How to train a neural network to generate phrases?

Question

Neural networks

I want to train the neural network to generate phrases similar to those that I feed it. For example: I have a dataset of 1000 VK posts and based on them, the neural network should generate a new post. In which direction to dig? What technologies are used?

Answer 1

There is already an application " VK Post Generator " by Denis Olshin .
The search will find a bunch of tutorials and articles about LSTM. For example .

Answer 2

semantic model.

Answer 3

Obviously, the neural network should not work with words, but with the semantic load of words and phrases, otherwise the result will differ little from a random combination of words.
Therefore, each word must be marked with a semantic load. For example, "joy", "anger", "good mood", "cry-cry" and the like. Those. make every word meaningful. Then give meaning to word combinations and short phrases. The neural network will need to feed exactly the tags from the training set of phrases (1000 may not be enough). The neural network will learn to determine the "meaning and emotions" of the input phrase. After that, it will be possible to try to reverse the process: you set a set of tags for input, and the neural network produces a set of words with such a meaning. It will be clumsy, because you need to decline the words in accordance with the rules of the Russian language, you can fasten the MS-word for post-processing of the text ...

Answer 4

https://habr.com/ru/news/t/504588/
https://habr.com/ru/company/itsumma/blog/511690/
https://habr.com/ru/post/440564/