Submission for the Fake News Challenge. The pipeline consists of text preprocessing, feature extraction, test-split generation, minority-class oversampling, followed by training and prediction using LightGBM.
This implementation was built and tested on Python 3.5. Dependencies can be installed via the following commands:
pip install -r requirements.txt
python -m spacy download en
LightGBM needs to be installed from source, by following the instructions here.
In addition, the FNC dataset is included as a submodule, and should be downloaded by running the following commands:
git submodule init
git submodule update
python fnc-1.py
During the first run, models and features will be generated for the first time, which may take several hours. Subsequent runs will use cached data, stored in the caches directory.