This poster briefly summarizes the work done during my thesis:
- 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 Constraint AAAI competitions were obtained using a considerably simpler model.
- The project resulted in the publication of a paper in the International Seminar on Artificial Intelligence and Disinformation. https://sites.google.com/go.ugr.es/aidisinfo2022/registration?authuser=0&pli=1
- 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 |
- Comparison of clustering formed using tf-idf and word embeddings using the most commons clustering algorithms like KMeans, Gaussian Mixture Models and Agglomerative Clustering.
- Tuning of the hyperparameters of all models.
- Comparaison of the results using multiple clustering metrics (DBI, Silhoutte and Calinski).
- Bonus experiment using only the most relevant tf-idf words and partly solving the curse of dimensionality.
- Bonus experiment using word embeddings from Microsoft MiniLM-L12-H384.
- Final analysis using wordclouds and n-grams to identify the topics.
- Found insights about which algorithms and metrics work best for document clustering and why.
- I used cuML, Spark (PySpark) and sentence-transformers.
- Build a CNN (Convolutional Neural Network) to detect pneumonia with chest x-ray images.
- Fine-Tuned Resnet50 (trained with Imagenet) to obtain 94.54% accuracy.
- We apply ImageDataGenerator to balance the classes.
- We used TensorFlow and Scikit-Learn.
Still in progress...