Disaster Tweet Classification

This is the repository for a mini-project where we classify whether a tweet is discussing a disastrous event (e.g flood) or not (see dataset for details). We use a number of pre-trained transformer models such as BERT and RoBERTa and perform an ablation study on how these models perform with various degrees of preprocessing. Additionally, we also studied the effect of data augmentation via back translation on the models and could generally observe an increase in performance:

For more details, see the report.

Installation & Running

To run the code:

1. Install Python 3.9+ and jupyter notebook.
2. Run all_models.ipynb which will install the remaining dependencies.