This notebook is a tool to help highlight the important parts of a piece of evidence in Policy Debate, hoping to make this frustrating process more efficient. It is currently a work in progress, although it can be used for very simple highlighting.
The notebook fine tunes the Longformer transformer model, which has an attention mechanism that scales with the longer debate cards. To fine tune this model, it uses data from DebateSum, learning what tokens to highlight based on the data from real debaters.
Open the main.ipynb in the Jupyter Notebook editor of your choice, and run all the cells. The notebook was created in Google Colab, but can be run locally if your computer is powerful enough.
Because the Longformer model and the inputs are so big, the batch_size is significantly smaller than normal to prevent memory errors. If you have access to a mighty computer, you can increase the batch_size to speed up the training process.
When trained, the model achieves a ~75% accuracy on the validation set. It highlights words and phrases relevant to the tag, but has trouble creating coherent sentences. These results can probably be improved with more training and a better way to represent the already highlighted words.