This is the project of SNLP course. It is based on the fake news challenge FakeChallenge.org. We implement the fake news detection based on several machine learning algorithms and compare their results. It is mainly contained in the sourcefile stance_detection.py. We also fine tuned a RoBERTa model on our dataset. The sourcefile is stance_RoBERTa.py.
Note that if you want fine tune the RoBERTa model with GPU. You need to check the official installment documentation carefully in here.
For us, we fine tuned the RoBERTa model on a Linux machine with CUDA 10.1. We installed the pytorch with command:
pip install torch==1.5.0+cu101 torchvision==0.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
Other dependencies can be checked in the file requirements.txt
We use key_word_extraction.ipynb to extract the keywords features of the article bodies.
The datasets are in the folder fnc-1.
The report is in the folder report_files.
The results are also saved in the folder results.
Note that all the features are extracted already and saved in the folder features. If you delete this folder and run the program stance_detection.py, it will extract the features again and save them in that folder.