Repository of the Scientific Analysis and Web App of the Final Degree Project of Santiago González Silot.
The Fine-Tuning.ipynb file contains the analysis of the 2 datasets together with the different tests carried out in the work. The rest of the files correspond to the web app of the work.
- An analysis of the main datasets available for the detection of fake news has been carried out.
- Comparison between 7 BERT and RoBERTa models (4 for English and 3 for Spanish) with 4 different optimization and regularization techniques specialized for word embeddings. Giving a total of 28 different models tested.
- Results very close to the winners of the Iberlef and ConstraintAAAI competitions were obtained using a considerably simpler model.
- Implementation of a basic web interface for the use and access to the models. Currently it can be accessed through HuggingFace: https://huggingface.co/spaces/sgonzalezsilot/Fake-News-Twitter-Detection_from-my-Thesis
| Model | F1-Score | Place in the competition | Difference with the winner |
|---|---|---|---|
| English | 98.41 | 8 | 0.2 |
| Spanish | 73.77 | 5 | 2.89 |
