Snorkel weak supervision for relation extraction

This includes results in the results/summary.render.ipynb notebook conditioned on the following:

• All analysis and modeling pertains exclusively to the development corpus
• Labeling functions DO NOT include strongly supervised models
• The generative model is not trained with class balancing leading
• The generative model uses LFs with low coverage meaning a large portion of the weak labels are precisely .5
• Both the strong and weak supervision models were trained using 50% class balancing
• Metrics from both of the above are based on median scores across models in hyperparameter grid

Initial co-occurrence analysis

This tag maintains a commit with the initial results for analyzing co-occurrences of t cell type and protein mentions.