This repository contains the code and data used in the paper titled 'Cross-SEAN: A cross-stitch semi-supervised neural attention model for COVID-19 fake news detection'
link: https://www.sciencedirect.com/science/article/abs/pii/S1568494621003161
README is divided into 3 files.
- data.zip - Contains sample data (2000 Fake, 2000 Genuine and 5000 Unlabelled) according to the format expected by the pipeline.
- covid19.sh - Shell config file which ensures the end-to-end working.
- extract_features.py - Prepares the data by extracting various inputs from the tweet objects containing json files.
- preprocess.py - Prepares all the data according to the embedding vector inputs.
- model.py - Contains all the model component classes
- training.py - Runs the training epochs and evaluation steps
- main.py - Loads all the data and initialises training, followed by evaluation
- config.py - Contains all args which are parsed after running main.py, along with default values.
- flask_sean.py - Responsible for the chrome extension API (to run, replace your Tweepy credentials in the file)
- sentimentanalysis - A helper module to generate some tweet features in
preprocess.py. Uses a slightly modified version of williamscott701/sentimentanalysis to create the extended version of the data.
- 5GB GPU for training
- 1GB for inference
- Unzip data.zip from Sample CTF Dataset folder OR
- Add/Update data in data/covid19/ in json format, as fake.json, genuine.json, unlabelled.json. (sample data given in CTF are given here in json format)
- Install the requirements
pip install -r requirements.txt - Change the project and directory name in covid19.sh (optional)
- Run the pipeline (changing the paths appropriately, if any of the above steps are done differently)
sh ./covid19.sh
- To run live service, follow README_Chrome_SEAN.txt
- Separate readme are available for other aspects.
Please cite the paper if you are using the dataset or the model.