Russian Jokes Generator

Description

This project is part of the compulsory course Programming for Linguists in National Research University Higher School of Economics. It is a joke generator, that is a special case of text generation.

Input: a collection of jokes.  
Output: generated text based on the input collection.

The program is trained on two models, which are based on the same architecture — RuGPT3. Before starting generation it is possible to select a model, a topic of jokes (if it is a ‘Наша’ model) and the length of the generated text. Besides, you need to enter a prompt for a joke and the program will continue the text.

Creators

Team-lead / NLP-engineer: Polina Savinova
Logic and Interface developer: Egor Zharikov
QA / Tech Writer: Vita Zaytseva
Product Manager: Evgenia Pareshina

---

Installing

To install the program you need to open Git on your computer and write the command:

git clone https://github.com/posavinova/russian_jokes_generator.git

---

Focal points

Contents

The generator consists of the following folders:

1. models_config

   Two models are stored here. Both are based on the same architecture — RuGPT3. The difference between them is fine-tuning on different datasets. Our model was fine-tuned on a collection of texts from humornet.ru. As for the other model, we are giving credits to Neural Shit.

2. spider

   Here you can find scraper for collecting dataset from the site humornet.ru

3. src

   It is the folder where the GUI, the model request and the saving the history of jokes are stored.

4. tests

   Here you can find unit tests on models.

Technical solutions

Module	Description	Component
pathlib	module for working with file paths
pylint	module for code analyser
gdown	module for downloading big files from Google Drive	models
torch	module for machine learning	models
transformers	module for downloading and training models	models
BeautifulSoup4	module for finding information on web pages	scraper
pandas	module for working with tabular data	scraper
requests	module for downloading web pages	scraper
tqdm	module for tracking progress	scraper

---

Running

Program launch

Before starting check that you have installed all the required modules. If not, open the console and write the command:

pip install -r requirements.txt

Now you are ready to start!
To run the generator you need to open the file main.py in the folder src and run the method:

if __name__ == '__main__':
    app = Application()
    app.run

Interface

When you start the generator, the Joke Generator front-end application will open. At the bottom you can see 2 drop-down lists: model and length. You can choose "Чужая" or "Наша" model. If you select "Наша" model, you can select a tag or, in other words, a topic of the future generated joke. There are 10 themes to choose from:

1. Food
2. Politics
3. Cats
4. Vulgar
5. Job
6. Computers
7. Children
8. Stirlitz
9. Students
10. Neighbours

Besides, you can choose the length of the generated joke from 30 to 100 characters.

Above the drop-down lists there is a field where you can write the prompt for a joke. To the right of this field there is a button "Отправить". After clicking on it or pressing Enter, in the window above you can see your prompt and after a while the generated anecdote itself.

Running

To get the generated joke, you need to enter a prompt for a joke in the field, select all categories and click on "Отправить" or press Enter. Then your prompt and the generated joke will appear.

Saving

Our program has an automatic function of saving jokes locally. When you run the generator for the first time, the file jokes_history.txt is created, where you can see all the generated jokes.