Our overarching objective is to build systems that investigate biases in social media recommendation algorithms especially in politics. To this end, we are interested in measuring political partisanship, toxicity, and hate speech on online social media platforms. Our goals are to: (1) build reliable data collection infrastructure using platform APIs or web scraping tools for Facebook, Reddit, and YouTube; (2) build machine learning models to classify political ideology and other attributes of social media posts.

We will be using NLP techniques with deep learning including fine tuning transformers to train models that specialize in classifying text based on toxicity. Our current focus at the moment is Bidirectional Encoder Representations from Transformers (BERT). BERT is a neural network transformer model that is pre trained to be able to understand and process language. Our goal is to use our own data to apply transfer learning on a BERT model and make it specialize in detecting toxicity and hate speech on online social media platforms. One of our options is to use the Hugging Face Library which provides pretrained models (e.g. BERT) on a large corpora of text. We can then fine-tune the model for our downstream task. Our team will need to run experiments and find the parameters that work best for our purposes. We will be using web scraping techniques and libraries like Selenium to gather our own data from political websites.

Install the requirements by running the following commands:

git clone  https://github.com/rezhv/UCD_Noyce.git
pip install -r UCD_Noyce/requirements.txt

You can run the train the classifier through the command line. First, Clone the repository:

git clone  https://github.com/rezhv/UCD_Noyce.git

Here is a list of possible arguments to pass to the trainer:

python3 ./UCD_Noyce/Noyce/train_classifier.py -h 
usage: train_classifier.py [-h] [-e EPOCHS] [-b BATCH_SIZE] [-l LOGGING_STEPS] 
[-lr LEARNING_RATE] [-v] [-ds DATASET] [-tl TOKENIZATIONLENGTH] [-m MODEL] [-s]

optional arguments:
  -h, --help show this help message and exit
  -e EPOCHS, --epochs EPOCHS Number of Training Epochs
  -b BATCH_SIZE, --batch_size BATCH_SIZE Batch Size
  -l LOGGING_STEPS, --logging_steps LOGGING_STEPS  Number of Steps to Run Evaluation
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE Value of Learning Rate
  -v, --verbose         
  -ds DATASET, --dataset DATASET Path/Name of the Dataset to Train On
  -tl TOKENIZATIONLENGTH, --tokenizationlength TOKENIZATIONLENGTH Tokenization Max Length
  -m MODEL, --model MODEL  Name of the model: bert, xlmr, roberta
  -s, --scheduler Use Learning Rate Scheduler

Run the training script:

python3 ./UCD_Noyce/Noyce/train_classifier.py