This repo contains the training, validation, and experimental code used in our submission to the SemEval 2019 Task 4: Hyperpartisan News Detection where our model came in 10th place internationally with an accuracy of 77.1% according to the leaderboard available here.
Team Clint Buchanan:
- Mehdi Drissi
- Pedro Sandoval Segura
- Vivaswat Ojha
Code for text preprocessing was provided to us by Professor Julie Medero as part of a course in Natural Language Processing. She was instrumental in helping us submit our model to the competition and write our workshop paper. This code can be found in extract_articles.py, extract_features.py, predict.py, preprocess.py.
We trained Bidirectional Encoder Representations from Transformers (BERT) models (Devlin et al., 2018) based on the implementation from pytorch-pretrained-BERT. Please refer to that repository for required dependencies.
To train our BERT model, you can use the train.sh bash script. You will need to download the articles here.