Description: This project aims to build a constrained generative model along the axes of politeness using exisitng politeness strategy models.

1. Install requirements.txt

pip install -r requirements.txt
pip install git+https://github.com/huggingface/transformers

2. Create a wandb account for recording model results and performance issues.

The main APIs to gather and preprocess tweets are contianed in "Experimentation/api/". A walk through of this process is started in "Experimentation" starting with "Scrapping_Cleaned_Unconstrained.ipynb". It uses snscrape to gather data and then processes the raw data for suitable use. Full_model_processing.ipynb completes the whole processing step post scrapping and cleaning. Once it is run, GPT2_model_new.ipynb is used to fine tune gpt-2 models for the constrained case and "GPT2_model_old.ipynb" is used to fine tune gpt-2 models that are unconstrained. The file "Generate_Human_Evaluation_Unconstrained.ipynb" generates csv files that can be used for human evaluation of output generation. The scoring for this output was done manually by human annotators located at https://docs.google.com/spreadsheets/d/1p1CCXQranUoIe5MO4VuEUfSCtboTBzyQIKcRcezktQ0/edit#gid=1528582561.

The full order to understand this project is in the "Experimentaion" folder: ** Scrapping_Cleaned_Unconstrained.ipynb --> Full_Model_Processing_Constrained --> GPT2_model_old --> GPT2_model_new --> Generate_Human_Evaluations_Unconstrained ** Baselines for politness are available through runing baselines_politeness which is the demo file from https://github.com/CornellNLP/ConvoKit/blob/master/examples/politeness-strategies/politeness_demo.ipynb to check that the API is functional for this instance (please note all credit for that notebook goes to https://convokit.cornell.edu/documentation/index.html).