Add script to clean training data

[danielwatson6]

Tokenization is necessary, but preprocessing should be kept minimal for simplicity and model robustness.

Documentation on script should be added to the README.