Predicting Propaganda using BERT

Medium article for the project

Guide to files:

• EDA.ipynb
  • Exploratory data analysis on training data.
  • Leverage spaCy, intervaltree, textstat, and Matplotlib for visualizations.
  • Features an approach for increasing training data by 160%!
• generate_data.ipynb
  • Create train, dev, and test .tsv files for classifiers.
  • Generate negative spans from news articles, increasing data by +160%.
• dummy_classifier.ipynb
  • Establish a baseline model for evaluating more sophisticated models.
  • Perform multi-class classification using training and dev data.
  • Evaluate performance using confusion matrix, classification report, and micro F1.
• logistic_regression.ipynb
  • Create a logistic regression model.
  • Perform grid search to adjust the model's hyperparameters.
  • Perform multi-class classification using training and dev data.
  • Evaluate performance using confusion matrix, classification report, and micro F1.
• bert_train_validate.ipynb
  • Fine-tune a pre-trained BERT model on the train data.
  • Perform hyperparameter tuning by evaluating different variations based on: model, epochs, learning rate, and batch size.
  • Evaluate average performance on validation sets using accuracy, precision, recall, and F1.
  • Save the best performing version for evalution on the dev set.
• bert_dev.ipynb
  • Run best performing version on dev data.
  • Evaluate performance using accuracy, precision, recall, and F1.
• bert_test.ipynb
  • Generate predictions for test data.
  • Results are to be submitted to Semeval 2020 task 11 (TC).

Photo by Toa Heftiba on Unsplash