/Fake_Abstract_Generator

Generate a Fake Abstract Based on CVPR Entries from the last 5 years

Primary LanguageJupyter Notebook

README

Author: Elisa Warner Date: Dec 19, 2019

Fake Abstract Generator

The concept of this pet project is to download all abstracts from the 2015-2019 CVPR conferences and create a fake abstract based on the current abstracts. The generator uses a simple n-th order Markov Chain to generate results. Users can input the target word count and order of the model.

Potential Use: As filler text in a LaTex document

How to Use:

  1. Use the Jupyter Notebook fake_abstract.ipynb or run the fake_abstract.py code.
  2. You will be guided to a localhost port (e.g. output Running on http://127.0.0.1:5000/ (Press CTRL+C to quit). The port ID is http://127.0.0.1:5000/).
  3. Copy and paste the port ID into your browser address bar. There will appear an html page that asks you to input the order N of the model and a word count. If you don't know, input 1. Try playing around with values from 1 to 7 to compare results.
  4. Click submit to see results. Results will show the order and word count of the model, which will be within the range of 80% of the target word count to the nearest possible "natural" sentence end.
  5. Troubleshooting: If you encounter any issues with loading the Abstract the page, just refresh and try again.

Images

Landing Page: Main Page Abstract Page: Abstract

Requirements

Requires python 3.6 with Anaconda to run the Jupyter Notebook, and just python 3.6 to run the .py code. Install all packages in requirements.txt (Type pip install -r requirements.txt into your console if you have pip)
Tested on a Mac only but should work in Windows

Final Notes

  • Note that the generator is not that great at the moment. Since I'm just using a simple N-th order Markov Model with about 979 abstracts from CVPR (and I didn't adjust commas and periods yet), the model is still rough. However, this is just a fun little project for me to have fun with Markov models
  • Please don't use this incessantly. I need to introduce a delay into the requests still, so if you overuse the app you may (no promises) get kicked off the server of the CVPR website.
  • Included is a cached version of the website, so you don't have to hit the server (cache_file.json).
  • I also expanded caching to include the model for different orders of the model (model.pkl). If you delete this, the model will just have to be recreated on its own, which will take a little time.