V&A: run load, NER, NEL; train NEL

[Kalyan Dutia]

• check data quality (joins etc)
• run load
• fix NEL for large datasets
• label NEL training data
• test model
• (after #342) add disambiguating descriptions to training data
• rewrite load function so more memory efficient (doesn't rely on full dataframe) -> ended up using pd.DataFrame(series.tolist()) instead of series.apply(pd.Series), and using categorical instead of string types for descriptions
• run with trained model

[Kalyan Dutia]

load with a threshold of 0.8